kernel/bpf/hashtab.c

   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/rcupdate_wait.h>
  11 #include <linux/random.h>
  12 #include <uapi/linux/btf.h>
  13 #include <linux/rcupdate_trace.h>
  14 #include <linux/btf_ids.h>
  15 #include "percpu_freelist.h"
  16 #include "bpf_lru_list.h"
  17 #include "map_in_map.h"
  18 #include <linux/bpf_mem_alloc.h>
  19
  20 #define HTAB_CREATE_FLAG_MASK                                           \
  21         (BPF_F_NO_PREALLOC | BPF_F_NO_COMMON_LRU | BPF_F_NUMA_NODE |    \
  22          BPF_F_ACCESS_MASK | BPF_F_ZERO_SEED)
  23
  24 #define BATCH_OPS(_name)                        \
  25         .map_lookup_batch =                     \
  26         _name##_map_lookup_batch,               \
  27         .map_lookup_and_delete_batch =          \
  28         _name##_map_lookup_and_delete_batch,    \
  29         .map_update_batch =                     \
  30         generic_map_update_batch,               \
  31         .map_delete_batch =                     \
  32         generic_map_delete_batch
  33
  34 /*
  35  * The bucket lock has two protection scopes:
  36  *
  37  * 1) Serializing concurrent operations from BPF programs on different
  38  *    CPUs
  39  *
  40  * 2) Serializing concurrent operations from BPF programs and sys_bpf()
  41  *
  42  * BPF programs can execute in any context including perf, kprobes and
  43  * tracing. As there are almost no limits where perf, kprobes and tracing
  44  * can be invoked from the lock operations need to be protected against
  45  * deadlocks. Deadlocks can be caused by recursion and by an invocation in
  46  * the lock held section when functions which acquire this lock are invoked
  47  * from sys_bpf(). BPF recursion is prevented by incrementing the per CPU
  48  * variable bpf_prog_active, which prevents BPF programs attached to perf
  49  * events, kprobes and tracing to be invoked before the prior invocation
  50  * from one of these contexts completed. sys_bpf() uses the same mechanism
  51  * by pinning the task to the current CPU and incrementing the recursion
  52  * protection across the map operation.
  53  *
  54  * This has subtle implications on PREEMPT_RT. PREEMPT_RT forbids certain
  55  * operations like memory allocations (even with GFP_ATOMIC) from atomic
  56  * contexts. This is required because even with GFP_ATOMIC the memory
  57  * allocator calls into code paths which acquire locks with long held lock
  58  * sections. To ensure the deterministic behaviour these locks are regular
  59  * spinlocks, which are converted to 'sleepable' spinlocks on RT. The only
  60  * true atomic contexts on an RT kernel are the low level hardware
  61  * handling, scheduling, low level interrupt handling, NMIs etc. None of
  62  * these contexts should ever do memory allocations.
  63  *
  64  * As regular device interrupt handlers and soft interrupts are forced into
  65  * thread context, the existing code which does
  66  *   spin_lock*(); alloc(GFP_ATOMIC); spin_unlock*();
  67  * just works.
  68  *
  69  * In theory the BPF locks could be converted to regular spinlocks as well,
  70  * but the bucket locks and percpu_freelist locks can be taken from
  71  * arbitrary contexts (perf, kprobes, tracepoints) which are required to be
  72  * atomic contexts even on RT. Before the introduction of bpf_mem_alloc,
  73  * it is only safe to use raw spinlock for preallocated hash map on a RT kernel,
  74  * because there is no memory allocation within the lock held sections. However
  75  * after hash map was fully converted to use bpf_mem_alloc, there will be
  76  * non-synchronous memory allocation for non-preallocated hash map, so it is
  77  * safe to always use raw spinlock for bucket lock.
  78  */
  79 struct bucket {
  80         struct hlist_nulls_head head;
  81         raw_spinlock_t raw_lock;
  82 };
  83
  84 #define HASHTAB_MAP_LOCK_COUNT 8
  85 #define HASHTAB_MAP_LOCK_MASK (HASHTAB_MAP_LOCK_COUNT - 1)
  86
  87 struct bpf_htab {
  88         struct bpf_map map;
  89         struct bpf_mem_alloc ma;
  90         struct bpf_mem_alloc pcpu_ma;
  91         struct bucket *buckets;
  92         void *elems;
  93         union {
  94                 struct pcpu_freelist freelist;
  95                 struct bpf_lru lru;
  96         };
  97         struct htab_elem *__percpu *extra_elems;
  98         /* number of elements in non-preallocated hashtable are kept
  99          * in either pcount or count
 100          */
 101         struct percpu_counter pcount;
 102         atomic_t count;
 103         bool use_percpu_counter;
 104         u32 n_buckets;  /* number of hash buckets */
 105         u32 elem_size;  /* size of each element in bytes */
 106         u32 hashrnd;
 107         struct lock_class_key lockdep_key;
 108         int __percpu *map_locked[HASHTAB_MAP_LOCK_COUNT];
 109 };
 110
 111 /* each htab element is struct htab_elem + key + value */
 112 struct htab_elem {
 113         union {
 114                 struct hlist_nulls_node hash_node;
 115                 struct {
 116                         void *padding;
 117                         union {
 118                                 struct pcpu_freelist_node fnode;
 119                                 struct htab_elem *batch_flink;
 120                         };
 121                 };
 122         };
 123         union {
 124                 /* pointer to per-cpu pointer */
 125                 void *ptr_to_pptr;
 126                 struct bpf_lru_node lru_node;
 127         };
 128         u32 hash;
 129         char key[] __aligned(8);
 130 };
 131
 132 static inline bool htab_is_prealloc(const struct bpf_htab *htab)
 133 {
 134         return !(htab->map.map_flags & BPF_F_NO_PREALLOC);
 135 }
 136
 137 static void htab_init_buckets(struct bpf_htab *htab)
 138 {
 139         unsigned int i;
 140
 141         for (i = 0; i < htab->n_buckets; i++) {
 142                 INIT_HLIST_NULLS_HEAD(&htab->buckets[i].head, i);
 143                 raw_spin_lock_init(&htab->buckets[i].raw_lock);
 144                 lockdep_set_class(&htab->buckets[i].raw_lock,
 145                                           &htab->lockdep_key);
 146                 cond_resched();
 147         }
 148 }
 149
 150 static inline int htab_lock_bucket(const struct bpf_htab *htab,
 151                                    struct bucket *b, u32 hash,
 152                                    unsigned long *pflags)
 153 {
 154         unsigned long flags;
 155
 156         hash = hash & min_t(u32, HASHTAB_MAP_LOCK_MASK, htab->n_buckets - 1);
 157
 158         preempt_disable();
 159         local_irq_save(flags);
 160         if (unlikely(__this_cpu_inc_return(*(htab->map_locked[hash])) != 1)) {
 161                 __this_cpu_dec(*(htab->map_locked[hash]));
 162                 local_irq_restore(flags);
 163                 preempt_enable();
 164                 return -EBUSY;
 165         }
 166
 167         raw_spin_lock(&b->raw_lock);
 168         *pflags = flags;
 169
 170         return 0;
 171 }
 172
 173 static inline void htab_unlock_bucket(const struct bpf_htab *htab,
 174                                       struct bucket *b, u32 hash,
 175                                       unsigned long flags)
 176 {
 177         hash = hash & min_t(u32, HASHTAB_MAP_LOCK_MASK, htab->n_buckets - 1);
 178         raw_spin_unlock(&b->raw_lock);
 179         __this_cpu_dec(*(htab->map_locked[hash]));
 180         local_irq_restore(flags);
 181         preempt_enable();
 182 }
 183
 184 static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node);
 185
 186 static bool htab_is_lru(const struct bpf_htab *htab)
 187 {
 188         return htab->map.map_type == BPF_MAP_TYPE_LRU_HASH ||
 189                 htab->map.map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH;
 190 }
 191
 192 static bool htab_is_percpu(const struct bpf_htab *htab)
 193 {
 194         return htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH ||
 195                 htab->map.map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH;
 196 }
 197
 198 static inline void htab_elem_set_ptr(struct htab_elem *l, u32 key_size,
 199                                      void __percpu *pptr)
 200 {
 201         *(void __percpu **)(l->key + key_size) = pptr;
 202 }
 203
 204 static inline void __percpu *htab_elem_get_ptr(struct htab_elem *l, u32 key_size)
 205 {
 206         return *(void __percpu **)(l->key + key_size);
 207 }
 208
 209 static void *fd_htab_map_get_ptr(const struct bpf_map *map, struct htab_elem *l)
 210 {
 211         return *(void **)(l->key + roundup(map->key_size, 8));
 212 }
 213
 214 static struct htab_elem *get_htab_elem(struct bpf_htab *htab, int i)
 215 {
 216         return (struct htab_elem *) (htab->elems + i * (u64)htab->elem_size);
 217 }
 218
 219 static bool htab_has_extra_elems(struct bpf_htab *htab)
 220 {
 221         return !htab_is_percpu(htab) && !htab_is_lru(htab);
 222 }
 223
 224 static void htab_free_prealloced_timers_and_wq(struct bpf_htab *htab)
 225 {
 226         u32 num_entries = htab->map.max_entries;
 227         int i;
 228
 229         if (htab_has_extra_elems(htab))
 230                 num_entries += num_possible_cpus();
 231
 232         for (i = 0; i < num_entries; i++) {
 233                 struct htab_elem *elem;
 234
 235                 elem = get_htab_elem(htab, i);
 236                 if (btf_record_has_field(htab->map.record, BPF_TIMER))
 237                         bpf_obj_free_timer(htab->map.record,
 238                                            elem->key + round_up(htab->map.key_size, 8));
 239                 if (btf_record_has_field(htab->map.record, BPF_WORKQUEUE))
 240                         bpf_obj_free_workqueue(htab->map.record,
 241                                                elem->key + round_up(htab->map.key_size, 8));
 242                 cond_resched();
 243         }
 244 }
 245
 246 static void htab_free_prealloced_fields(struct bpf_htab *htab)
 247 {
 248         u32 num_entries = htab->map.max_entries;
 249         int i;
 250
 251         if (IS_ERR_OR_NULL(htab->map.record))
 252                 return;
 253         if (htab_has_extra_elems(htab))
 254                 num_entries += num_possible_cpus();
 255         for (i = 0; i < num_entries; i++) {
 256                 struct htab_elem *elem;
 257
 258                 elem = get_htab_elem(htab, i);
 259                 if (htab_is_percpu(htab)) {
 260                         void __percpu *pptr = htab_elem_get_ptr(elem, htab->map.key_size);
 261                         int cpu;
 262
 263                         for_each_possible_cpu(cpu) {
 264                                 bpf_obj_free_fields(htab->map.record, per_cpu_ptr(pptr, cpu));
 265                                 cond_resched();
 266                         }
 267                 } else {
 268                         bpf_obj_free_fields(htab->map.record, elem->key + round_up(htab->map.key_size, 8));
 269                         cond_resched();
 270                 }
 271                 cond_resched();
 272         }
 273 }
 274
 275 static void htab_free_elems(struct bpf_htab *htab)
 276 {
 277         int i;
 278
 279         if (!htab_is_percpu(htab))
 280                 goto free_elems;
 281
 282         for (i = 0; i < htab->map.max_entries; i++) {
 283                 void __percpu *pptr;
 284
 285                 pptr = htab_elem_get_ptr(get_htab_elem(htab, i),
 286                                          htab->map.key_size);
 287                 free_percpu(pptr);
 288                 cond_resched();
 289         }
 290 free_elems:
 291         bpf_map_area_free(htab->elems);
 292 }
 293
 294 /* The LRU list has a lock (lru_lock). Each htab bucket has a lock
 295  * (bucket_lock). If both locks need to be acquired together, the lock
 296  * order is always lru_lock -> bucket_lock and this only happens in
 297  * bpf_lru_list.c logic. For example, certain code path of
 298  * bpf_lru_pop_free(), which is called by function prealloc_lru_pop(),
 299  * will acquire lru_lock first followed by acquiring bucket_lock.
 300  *
 301  * In hashtab.c, to avoid deadlock, lock acquisition of
 302  * bucket_lock followed by lru_lock is not allowed. In such cases,
 303  * bucket_lock needs to be released first before acquiring lru_lock.
 304  */
 305 static struct htab_elem *prealloc_lru_pop(struct bpf_htab *htab, void *key,
 306                                           u32 hash)
 307 {
 308         struct bpf_lru_node *node = bpf_lru_pop_free(&htab->lru, hash);
 309         struct htab_elem *l;
 310
 311         if (node) {
 312                 bpf_map_inc_elem_count(&htab->map);
 313                 l = container_of(node, struct htab_elem, lru_node);
 314                 memcpy(l->key, key, htab->map.key_size);
 315                 return l;
 316         }
 317
 318         return NULL;
 319 }
 320
 321 static int prealloc_init(struct bpf_htab *htab)
 322 {
 323         u32 num_entries = htab->map.max_entries;
 324         int err = -ENOMEM, i;
 325
 326         if (htab_has_extra_elems(htab))
 327                 num_entries += num_possible_cpus();
 328
 329         htab->elems = bpf_map_area_alloc((u64)htab->elem_size * num_entries,
 330                                          htab->map.numa_node);
 331         if (!htab->elems)
 332                 return -ENOMEM;
 333
 334         if (!htab_is_percpu(htab))
 335                 goto skip_percpu_elems;
 336
 337         for (i = 0; i < num_entries; i++) {
 338                 u32 size = round_up(htab->map.value_size, 8);
 339                 void __percpu *pptr;
 340
 341                 pptr = bpf_map_alloc_percpu(&htab->map, size, 8,
 342                                             GFP_USER | __GFP_NOWARN);
 343                 if (!pptr)
 344                         goto free_elems;
 345                 htab_elem_set_ptr(get_htab_elem(htab, i), htab->map.key_size,
 346                                   pptr);
 347                 cond_resched();
 348         }
 349
 350 skip_percpu_elems:
 351         if (htab_is_lru(htab))
 352                 err = bpf_lru_init(&htab->lru,
 353                                    htab->map.map_flags & BPF_F_NO_COMMON_LRU,
 354                                    offsetof(struct htab_elem, hash) -
 355                                    offsetof(struct htab_elem, lru_node),
 356                                    htab_lru_map_delete_node,
 357                                    htab);
 358         else
 359                 err = pcpu_freelist_init(&htab->freelist);
 360
 361         if (err)
 362                 goto free_elems;
 363
 364         if (htab_is_lru(htab))
 365                 bpf_lru_populate(&htab->lru, htab->elems,
 366                                  offsetof(struct htab_elem, lru_node),
 367                                  htab->elem_size, num_entries);
 368         else
 369                 pcpu_freelist_populate(&htab->freelist,
 370                                        htab->elems + offsetof(struct htab_elem, fnode),
 371                                        htab->elem_size, num_entries);
 372
 373         return 0;
 374
 375 free_elems:
 376         htab_free_elems(htab);
 377         return err;
 378 }
 379
 380 static void prealloc_destroy(struct bpf_htab *htab)
 381 {
 382         htab_free_elems(htab);
 383
 384         if (htab_is_lru(htab))
 385                 bpf_lru_destroy(&htab->lru);
 386         else
 387                 pcpu_freelist_destroy(&htab->freelist);
 388 }
 389
 390 static int alloc_extra_elems(struct bpf_htab *htab)
 391 {
 392         struct htab_elem *__percpu *pptr, *l_new;
 393         struct pcpu_freelist_node *l;
 394         int cpu;
 395
 396         pptr = bpf_map_alloc_percpu(&htab->map, sizeof(struct htab_elem *), 8,
 397                                     GFP_USER | __GFP_NOWARN);
 398         if (!pptr)
 399                 return -ENOMEM;
 400
 401         for_each_possible_cpu(cpu) {
 402                 l = pcpu_freelist_pop(&htab->freelist);
 403                 /* pop will succeed, since prealloc_init()
 404                  * preallocated extra num_possible_cpus elements
 405                  */
 406                 l_new = container_of(l, struct htab_elem, fnode);
 407                 *per_cpu_ptr(pptr, cpu) = l_new;
 408         }
 409         htab->extra_elems = pptr;
 410         return 0;
 411 }
 412
 413 /* Called from syscall */
 414 static int htab_map_alloc_check(union bpf_attr *attr)
 415 {
 416         bool percpu = (attr->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
 417                        attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
 418         bool lru = (attr->map_type == BPF_MAP_TYPE_LRU_HASH ||
 419                     attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
 420         /* percpu_lru means each cpu has its own LRU list.
 421          * it is different from BPF_MAP_TYPE_PERCPU_HASH where
 422          * the map's value itself is percpu.  percpu_lru has
 423          * nothing to do with the map's value.
 424          */
 425         bool percpu_lru = (attr->map_flags & BPF_F_NO_COMMON_LRU);
 426         bool prealloc = !(attr->map_flags & BPF_F_NO_PREALLOC);
 427         bool zero_seed = (attr->map_flags & BPF_F_ZERO_SEED);
 428         int numa_node = bpf_map_attr_numa_node(attr);
 429
 430         BUILD_BUG_ON(offsetof(struct htab_elem, fnode.next) !=
 431                      offsetof(struct htab_elem, hash_node.pprev));
 432
 433         if (zero_seed && !capable(CAP_SYS_ADMIN))
 434                 /* Guard against local DoS, and discourage production use. */
 435                 return -EPERM;
 436
 437         if (attr->map_flags & ~HTAB_CREATE_FLAG_MASK ||
 438             !bpf_map_flags_access_ok(attr->map_flags))
 439                 return -EINVAL;
 440
 441         if (!lru && percpu_lru)
 442                 return -EINVAL;
 443
 444         if (lru && !prealloc)
 445                 return -ENOTSUPP;
 446
 447         if (numa_node != NUMA_NO_NODE && (percpu || percpu_lru))
 448                 return -EINVAL;
 449
 450         /* check sanity of attributes.
 451          * value_size == 0 may be allowed in the future to use map as a set
 452          */
 453         if (attr->max_entries == 0 || attr->key_size == 0 ||
 454             attr->value_size == 0)
 455                 return -EINVAL;
 456
 457         if ((u64)attr->key_size + attr->value_size >= KMALLOC_MAX_SIZE -
 458            sizeof(struct htab_elem))
 459                 /* if key_size + value_size is bigger, the user space won't be
 460                  * able to access the elements via bpf syscall. This check
 461                  * also makes sure that the elem_size doesn't overflow and it's
 462                  * kmalloc-able later in htab_map_update_elem()
 463                  */
 464                 return -E2BIG;
 465         /* percpu map value size is bound by PCPU_MIN_UNIT_SIZE */
 466         if (percpu && round_up(attr->value_size, 8) > PCPU_MIN_UNIT_SIZE)
 467                 return -E2BIG;
 468
 469         return 0;
 470 }
 471
 472 static struct bpf_map *htab_map_alloc(union bpf_attr *attr)
 473 {
 474         bool percpu = (attr->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
 475                        attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
 476         bool lru = (attr->map_type == BPF_MAP_TYPE_LRU_HASH ||
 477                     attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
 478         /* percpu_lru means each cpu has its own LRU list.
 479          * it is different from BPF_MAP_TYPE_PERCPU_HASH where
 480          * the map's value itself is percpu.  percpu_lru has
 481          * nothing to do with the map's value.
 482          */
 483         bool percpu_lru = (attr->map_flags & BPF_F_NO_COMMON_LRU);
 484         bool prealloc = !(attr->map_flags & BPF_F_NO_PREALLOC);
 485         struct bpf_htab *htab;
 486         int err, i;
 487
 488         htab = bpf_map_area_alloc(sizeof(*htab), NUMA_NO_NODE);
 489         if (!htab)
 490                 return ERR_PTR(-ENOMEM);
 491
 492         lockdep_register_key(&htab->lockdep_key);
 493
 494         bpf_map_init_from_attr(&htab->map, attr);
 495
 496         if (percpu_lru) {
 497                 /* ensure each CPU's lru list has >=1 elements.
 498                  * since we are at it, make each lru list has the same
 499                  * number of elements.
 500                  */
 501                 htab->map.max_entries = roundup(attr->max_entries,
 502                                                 num_possible_cpus());
 503                 if (htab->map.max_entries < attr->max_entries)
 504                         htab->map.max_entries = rounddown(attr->max_entries,
 505                                                           num_possible_cpus());
 506         }
 507
 508         /* hash table size must be power of 2; roundup_pow_of_two() can overflow
 509          * into UB on 32-bit arches, so check that first
 510          */
 511         err = -E2BIG;
 512         if (htab->map.max_entries > 1UL << 31)
 513                 goto free_htab;
 514
 515         htab->n_buckets = roundup_pow_of_two(htab->map.max_entries);
 516
 517         htab->elem_size = sizeof(struct htab_elem) +
 518                           round_up(htab->map.key_size, 8);
 519         if (percpu)
 520                 htab->elem_size += sizeof(void *);
 521         else
 522                 htab->elem_size += round_up(htab->map.value_size, 8);
 523
 524         /* check for u32 overflow */
 525         if (htab->n_buckets > U32_MAX / sizeof(struct bucket))
 526                 goto free_htab;
 527
 528         err = bpf_map_init_elem_count(&htab->map);
 529         if (err)
 530                 goto free_htab;
 531
 532         err = -ENOMEM;
 533         htab->buckets = bpf_map_area_alloc(htab->n_buckets *
 534                                            sizeof(struct bucket),
 535                                            htab->map.numa_node);
 536         if (!htab->buckets)
 537                 goto free_elem_count;
 538
 539         for (i = 0; i < HASHTAB_MAP_LOCK_COUNT; i++) {
 540                 htab->map_locked[i] = bpf_map_alloc_percpu(&htab->map,
 541                                                            sizeof(int),
 542                                                            sizeof(int),
 543                                                            GFP_USER);
 544                 if (!htab->map_locked[i])
 545                         goto free_map_locked;
 546         }
 547
 548         if (htab->map.map_flags & BPF_F_ZERO_SEED)
 549                 htab->hashrnd = 0;
 550         else
 551                 htab->hashrnd = get_random_u32();
 552
 553         htab_init_buckets(htab);
 554
 555 /* compute_batch_value() computes batch value as num_online_cpus() * 2
 556  * and __percpu_counter_compare() needs
 557  * htab->max_entries - cur_number_of_elems to be more than batch * num_online_cpus()
 558  * for percpu_counter to be faster than atomic_t. In practice the average bpf
 559  * hash map size is 10k, which means that a system with 64 cpus will fill
 560  * hashmap to 20% of 10k before percpu_counter becomes ineffective. Therefore
 561  * define our own batch count as 32 then 10k hash map can be filled up to 80%:
 562  * 10k - 8k > 32 _batch_ * 64 _cpus_
 563  * and __percpu_counter_compare() will still be fast. At that point hash map
 564  * collisions will dominate its performance anyway. Assume that hash map filled
 565  * to 50+% isn't going to be O(1) and use the following formula to choose
 566  * between percpu_counter and atomic_t.
 567  */
 568 #define PERCPU_COUNTER_BATCH 32
 569         if (attr->max_entries / 2 > num_online_cpus() * PERCPU_COUNTER_BATCH)
 570                 htab->use_percpu_counter = true;
 571
 572         if (htab->use_percpu_counter) {
 573                 err = percpu_counter_init(&htab->pcount, 0, GFP_KERNEL);
 574                 if (err)
 575                         goto free_map_locked;
 576         }
 577
 578         if (prealloc) {
 579                 err = prealloc_init(htab);
 580                 if (err)
 581                         goto free_map_locked;
 582
 583                 if (!percpu && !lru) {
 584                         /* lru itself can remove the least used element, so
 585                          * there is no need for an extra elem during map_update.
 586                          */
 587                         err = alloc_extra_elems(htab);
 588                         if (err)
 589                                 goto free_prealloc;
 590                 }
 591         } else {
 592                 err = bpf_mem_alloc_init(&htab->ma, htab->elem_size, false);
 593                 if (err)
 594                         goto free_map_locked;
 595                 if (percpu) {
 596                         err = bpf_mem_alloc_init(&htab->pcpu_ma,
 597                                                  round_up(htab->map.value_size, 8), true);
 598                         if (err)
 599                                 goto free_map_locked;
 600                 }
 601         }
 602
 603         return &htab->map;
 604
 605 free_prealloc:
 606         prealloc_destroy(htab);
 607 free_map_locked:
 608         if (htab->use_percpu_counter)
 609                 percpu_counter_destroy(&htab->pcount);
 610         for (i = 0; i < HASHTAB_MAP_LOCK_COUNT; i++)
 611                 free_percpu(htab->map_locked[i]);
 612         bpf_map_area_free(htab->buckets);
 613         bpf_mem_alloc_destroy(&htab->pcpu_ma);
 614         bpf_mem_alloc_destroy(&htab->ma);
 615 free_elem_count:
 616         bpf_map_free_elem_count(&htab->map);
 617 free_htab:
 618         lockdep_unregister_key(&htab->lockdep_key);
 619         bpf_map_area_free(htab);
 620         return ERR_PTR(err);
 621 }
 622
 623 static inline u32 htab_map_hash(const void *key, u32 key_len, u32 hashrnd)
 624 {
 625         if (likely(key_len % 4 == 0))
 626                 return jhash2(key, key_len / 4, hashrnd);
 627         return jhash(key, key_len, hashrnd);
 628 }
 629
 630 static inline struct bucket *__select_bucket(struct bpf_htab *htab, u32 hash)
 631 {
 632         return &htab->buckets[hash & (htab->n_buckets - 1)];
 633 }
 634
 635 static inline struct hlist_nulls_head *select_bucket(struct bpf_htab *htab, u32 hash)
 636 {
 637         return &__select_bucket(htab, hash)->head;
 638 }
 639
 640 /* this lookup function can only be called with bucket lock taken */
 641 static struct htab_elem *lookup_elem_raw(struct hlist_nulls_head *head, u32 hash,
 642                                          void *key, u32 key_size)
 643 {
 644         struct hlist_nulls_node *n;
 645         struct htab_elem *l;
 646
 647         hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
 648                 if (l->hash == hash && !memcmp(&l->key, key, key_size))
 649                         return l;
 650
 651         return NULL;
 652 }
 653
 654 /* can be called without bucket lock. it will repeat the loop in
 655  * the unlikely event when elements moved from one bucket into another
 656  * while link list is being walked
 657  */
 658 static struct htab_elem *lookup_nulls_elem_raw(struct hlist_nulls_head *head,
 659                                                u32 hash, void *key,
 660                                                u32 key_size, u32 n_buckets)
 661 {
 662         struct hlist_nulls_node *n;
 663         struct htab_elem *l;
 664
 665 again:
 666         hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
 667                 if (l->hash == hash && !memcmp(&l->key, key, key_size))
 668                         return l;
 669
 670         if (unlikely(get_nulls_value(n) != (hash & (n_buckets - 1))))
 671                 goto again;
 672
 673         return NULL;
 674 }
 675
 676 /* Called from syscall or from eBPF program directly, so
 677  * arguments have to match bpf_map_lookup_elem() exactly.
 678  * The return value is adjusted by BPF instructions
 679  * in htab_map_gen_lookup().
 680  */
 681 static void *__htab_map_lookup_elem(struct bpf_map *map, void *key)
 682 {
 683         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 684         struct hlist_nulls_head *head;
 685         struct htab_elem *l;
 686         u32 hash, key_size;
 687
 688         WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
 689                      !rcu_read_lock_bh_held());
 690
 691         key_size = map->key_size;
 692
 693         hash = htab_map_hash(key, key_size, htab->hashrnd);
 694
 695         head = select_bucket(htab, hash);
 696
 697         l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
 698
 699         return l;
 700 }
 701
 702 static void *htab_map_lookup_elem(struct bpf_map *map, void *key)
 703 {
 704         struct htab_elem *l = __htab_map_lookup_elem(map, key);
 705
 706         if (l)
 707                 return l->key + round_up(map->key_size, 8);
 708
 709         return NULL;
 710 }
 711
 712 /* inline bpf_map_lookup_elem() call.
 713  * Instead of:
 714  * bpf_prog
 715  *   bpf_map_lookup_elem
 716  *     map->ops->map_lookup_elem
 717  *       htab_map_lookup_elem
 718  *         __htab_map_lookup_elem
 719  * do:
 720  * bpf_prog
 721  *   __htab_map_lookup_elem
 722  */
 723 static int htab_map_gen_lookup(struct bpf_map *map, struct bpf_insn *insn_buf)
 724 {
 725         struct bpf_insn *insn = insn_buf;
 726         const int ret = BPF_REG_0;
 727
 728         BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
 729                      (void *(*)(struct bpf_map *map, void *key))NULL));
 730         *insn++ = BPF_EMIT_CALL(__htab_map_lookup_elem);
 731         *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 1);
 732         *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
 733                                 offsetof(struct htab_elem, key) +
 734                                 round_up(map->key_size, 8));
 735         return insn - insn_buf;
 736 }
 737
 738 static __always_inline void *__htab_lru_map_lookup_elem(struct bpf_map *map,
 739                                                         void *key, const bool mark)
 740 {
 741         struct htab_elem *l = __htab_map_lookup_elem(map, key);
 742
 743         if (l) {
 744                 if (mark)
 745                         bpf_lru_node_set_ref(&l->lru_node);
 746                 return l->key + round_up(map->key_size, 8);
 747         }
 748
 749         return NULL;
 750 }
 751
 752 static void *htab_lru_map_lookup_elem(struct bpf_map *map, void *key)
 753 {
 754         return __htab_lru_map_lookup_elem(map, key, true);
 755 }
 756
 757 static void *htab_lru_map_lookup_elem_sys(struct bpf_map *map, void *key)
 758 {
 759         return __htab_lru_map_lookup_elem(map, key, false);
 760 }
 761
 762 static int htab_lru_map_gen_lookup(struct bpf_map *map,
 763                                    struct bpf_insn *insn_buf)
 764 {
 765         struct bpf_insn *insn = insn_buf;
 766         const int ret = BPF_REG_0;
 767         const int ref_reg = BPF_REG_1;
 768
 769         BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
 770                      (void *(*)(struct bpf_map *map, void *key))NULL));
 771         *insn++ = BPF_EMIT_CALL(__htab_map_lookup_elem);
 772         *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 4);
 773         *insn++ = BPF_LDX_MEM(BPF_B, ref_reg, ret,
 774                               offsetof(struct htab_elem, lru_node) +
 775                               offsetof(struct bpf_lru_node, ref));
 776         *insn++ = BPF_JMP_IMM(BPF_JNE, ref_reg, 0, 1);
 777         *insn++ = BPF_ST_MEM(BPF_B, ret,
 778                              offsetof(struct htab_elem, lru_node) +
 779                              offsetof(struct bpf_lru_node, ref),
 780                              1);
 781         *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
 782                                 offsetof(struct htab_elem, key) +
 783                                 round_up(map->key_size, 8));
 784         return insn - insn_buf;
 785 }
 786
 787 static void check_and_free_fields(struct bpf_htab *htab,
 788                                   struct htab_elem *elem)
 789 {
 790         if (htab_is_percpu(htab)) {
 791                 void __percpu *pptr = htab_elem_get_ptr(elem, htab->map.key_size);
 792                 int cpu;
 793
 794                 for_each_possible_cpu(cpu)
 795                         bpf_obj_free_fields(htab->map.record, per_cpu_ptr(pptr, cpu));
 796         } else {
 797                 void *map_value = elem->key + round_up(htab->map.key_size, 8);
 798
 799                 bpf_obj_free_fields(htab->map.record, map_value);
 800         }
 801 }
 802
 803 /* It is called from the bpf_lru_list when the LRU needs to delete
 804  * older elements from the htab.
 805  */
 806 static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node)
 807 {
 808         struct bpf_htab *htab = arg;
 809         struct htab_elem *l = NULL, *tgt_l;
 810         struct hlist_nulls_head *head;
 811         struct hlist_nulls_node *n;
 812         unsigned long flags;
 813         struct bucket *b;
 814         int ret;
 815
 816         tgt_l = container_of(node, struct htab_elem, lru_node);
 817         b = __select_bucket(htab, tgt_l->hash);
 818         head = &b->head;
 819
 820         ret = htab_lock_bucket(htab, b, tgt_l->hash, &flags);
 821         if (ret)
 822                 return false;
 823
 824         hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
 825                 if (l == tgt_l) {
 826                         hlist_nulls_del_rcu(&l->hash_node);
 827                         check_and_free_fields(htab, l);
 828                         bpf_map_dec_elem_count(&htab->map);
 829                         break;
 830                 }
 831
 832         htab_unlock_bucket(htab, b, tgt_l->hash, flags);
 833
 834         return l == tgt_l;
 835 }
 836
 837 /* Called from syscall */
 838 static int htab_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
 839 {
 840         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 841         struct hlist_nulls_head *head;
 842         struct htab_elem *l, *next_l;
 843         u32 hash, key_size;
 844         int i = 0;
 845
 846         WARN_ON_ONCE(!rcu_read_lock_held());
 847
 848         key_size = map->key_size;
 849
 850         if (!key)
 851                 goto find_first_elem;
 852
 853         hash = htab_map_hash(key, key_size, htab->hashrnd);
 854
 855         head = select_bucket(htab, hash);
 856
 857         /* lookup the key */
 858         l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
 859
 860         if (!l)
 861                 goto find_first_elem;
 862
 863         /* key was found, get next key in the same bucket */
 864         next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_next_rcu(&l->hash_node)),
 865                                   struct htab_elem, hash_node);
 866
 867         if (next_l) {
 868                 /* if next elem in this hash list is non-zero, just return it */
 869                 memcpy(next_key, next_l->key, key_size);
 870                 return 0;
 871         }
 872
 873         /* no more elements in this hash list, go to the next bucket */
 874         i = hash & (htab->n_buckets - 1);
 875         i++;
 876
 877 find_first_elem:
 878         /* iterate over buckets */
 879         for (; i < htab->n_buckets; i++) {
 880                 head = select_bucket(htab, i);
 881
 882                 /* pick first element in the bucket */
 883                 next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_first_rcu(head)),
 884                                           struct htab_elem, hash_node);
 885                 if (next_l) {
 886                         /* if it's not empty, just return it */
 887                         memcpy(next_key, next_l->key, key_size);
 888                         return 0;
 889                 }
 890         }
 891
 892         /* iterated over all buckets and all elements */
 893         return -ENOENT;
 894 }
 895
 896 static void htab_elem_free(struct bpf_htab *htab, struct htab_elem *l)
 897 {
 898         check_and_free_fields(htab, l);
 899         if (htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH)
 900                 bpf_mem_cache_free(&htab->pcpu_ma, l->ptr_to_pptr);
 901         bpf_mem_cache_free(&htab->ma, l);
 902 }
 903
 904 static void htab_put_fd_value(struct bpf_htab *htab, struct htab_elem *l)
 905 {
 906         struct bpf_map *map = &htab->map;
 907         void *ptr;
 908
 909         if (map->ops->map_fd_put_ptr) {
 910                 ptr = fd_htab_map_get_ptr(map, l);
 911                 map->ops->map_fd_put_ptr(map, ptr, true);
 912         }
 913 }
 914
 915 static bool is_map_full(struct bpf_htab *htab)
 916 {
 917         if (htab->use_percpu_counter)
 918                 return __percpu_counter_compare(&htab->pcount, htab->map.max_entries,
 919                                                 PERCPU_COUNTER_BATCH) >= 0;
 920         return atomic_read(&htab->count) >= htab->map.max_entries;
 921 }
 922
 923 static void inc_elem_count(struct bpf_htab *htab)
 924 {
 925         bpf_map_inc_elem_count(&htab->map);
 926
 927         if (htab->use_percpu_counter)
 928                 percpu_counter_add_batch(&htab->pcount, 1, PERCPU_COUNTER_BATCH);
 929         else
 930                 atomic_inc(&htab->count);
 931 }
 932
 933 static void dec_elem_count(struct bpf_htab *htab)
 934 {
 935         bpf_map_dec_elem_count(&htab->map);
 936
 937         if (htab->use_percpu_counter)
 938                 percpu_counter_add_batch(&htab->pcount, -1, PERCPU_COUNTER_BATCH);
 939         else
 940                 atomic_dec(&htab->count);
 941 }
 942
 943
 944 static void free_htab_elem(struct bpf_htab *htab, struct htab_elem *l)
 945 {
 946         htab_put_fd_value(htab, l);
 947
 948         if (htab_is_prealloc(htab)) {
 949                 bpf_map_dec_elem_count(&htab->map);
 950                 check_and_free_fields(htab, l);
 951                 __pcpu_freelist_push(&htab->freelist, &l->fnode);
 952         } else {
 953                 dec_elem_count(htab);
 954                 htab_elem_free(htab, l);
 955         }
 956 }
 957
 958 static void pcpu_copy_value(struct bpf_htab *htab, void __percpu *pptr,
 959                             void *value, bool onallcpus)
 960 {
 961         if (!onallcpus) {
 962                 /* copy true value_size bytes */
 963                 copy_map_value(&htab->map, this_cpu_ptr(pptr), value);
 964         } else {
 965                 u32 size = round_up(htab->map.value_size, 8);
 966                 int off = 0, cpu;
 967
 968                 for_each_possible_cpu(cpu) {
 969                         copy_map_value_long(&htab->map, per_cpu_ptr(pptr, cpu), value + off);
 970                         off += size;
 971                 }
 972         }
 973 }
 974
 975 static void pcpu_init_value(struct bpf_htab *htab, void __percpu *pptr,
 976                             void *value, bool onallcpus)
 977 {
 978         /* When not setting the initial value on all cpus, zero-fill element
 979          * values for other cpus. Otherwise, bpf program has no way to ensure
 980          * known initial values for cpus other than current one
 981          * (onallcpus=false always when coming from bpf prog).
 982          */
 983         if (!onallcpus) {
 984                 int current_cpu = raw_smp_processor_id();
 985                 int cpu;
 986
 987                 for_each_possible_cpu(cpu) {
 988                         if (cpu == current_cpu)
 989                                 copy_map_value_long(&htab->map, per_cpu_ptr(pptr, cpu), value);
 990                         else /* Since elem is preallocated, we cannot touch special fields */
 991                                 zero_map_value(&htab->map, per_cpu_ptr(pptr, cpu));
 992                 }
 993         } else {
 994                 pcpu_copy_value(htab, pptr, value, onallcpus);
 995         }
 996 }
 997
 998 static bool fd_htab_map_needs_adjust(const struct bpf_htab *htab)
 999 {
1000         return htab->map.map_type == BPF_MAP_TYPE_HASH_OF_MAPS &&
1001                BITS_PER_LONG == 64;
1002 }
1003
1004 static struct htab_elem *alloc_htab_elem(struct bpf_htab *htab, void *key,
1005                                          void *value, u32 key_size, u32 hash,
1006                                          bool percpu, bool onallcpus,
1007                                          struct htab_elem *old_elem)
1008 {
1009         u32 size = htab->map.value_size;
1010         bool prealloc = htab_is_prealloc(htab);
1011         struct htab_elem *l_new, **pl_new;
1012         void __percpu *pptr;
1013
1014         if (prealloc) {
1015                 if (old_elem) {
1016                         /* if we're updating the existing element,
1017                          * use per-cpu extra elems to avoid freelist_pop/push
1018                          */
1019                         pl_new = this_cpu_ptr(htab->extra_elems);
1020                         l_new = *pl_new;
1021                         htab_put_fd_value(htab, old_elem);
1022                         *pl_new = old_elem;
1023                 } else {
1024                         struct pcpu_freelist_node *l;
1025
1026                         l = __pcpu_freelist_pop(&htab->freelist);
1027                         if (!l)
1028                                 return ERR_PTR(-E2BIG);
1029                         l_new = container_of(l, struct htab_elem, fnode);
1030                         bpf_map_inc_elem_count(&htab->map);
1031                 }
1032         } else {
1033                 if (is_map_full(htab))
1034                         if (!old_elem)
1035                                 /* when map is full and update() is replacing
1036                                  * old element, it's ok to allocate, since
1037                                  * old element will be freed immediately.
1038                                  * Otherwise return an error
1039                                  */
1040                                 return ERR_PTR(-E2BIG);
1041                 inc_elem_count(htab);
1042                 l_new = bpf_mem_cache_alloc(&htab->ma);
1043                 if (!l_new) {
1044                         l_new = ERR_PTR(-ENOMEM);
1045                         goto dec_count;
1046                 }
1047         }
1048
1049         memcpy(l_new->key, key, key_size);
1050         if (percpu) {
1051                 if (prealloc) {
1052                         pptr = htab_elem_get_ptr(l_new, key_size);
1053                 } else {
1054                         /* alloc_percpu zero-fills */
1055                         void *ptr = bpf_mem_cache_alloc(&htab->pcpu_ma);
1056
1057                         if (!ptr) {
1058                                 bpf_mem_cache_free(&htab->ma, l_new);
1059                                 l_new = ERR_PTR(-ENOMEM);
1060                                 goto dec_count;
1061                         }
1062                         l_new->ptr_to_pptr = ptr;
1063                         pptr = *(void __percpu **)ptr;
1064                 }
1065
1066                 pcpu_init_value(htab, pptr, value, onallcpus);
1067
1068                 if (!prealloc)
1069                         htab_elem_set_ptr(l_new, key_size, pptr);
1070         } else if (fd_htab_map_needs_adjust(htab)) {
1071                 size = round_up(size, 8);
1072                 memcpy(l_new->key + round_up(key_size, 8), value, size);
1073         } else {
1074                 copy_map_value(&htab->map,
1075                                l_new->key + round_up(key_size, 8),
1076                                value);
1077         }
1078
1079         l_new->hash = hash;
1080         return l_new;
1081 dec_count:
1082         dec_elem_count(htab);
1083         return l_new;
1084 }
1085
1086 static int check_flags(struct bpf_htab *htab, struct htab_elem *l_old,
1087                        u64 map_flags)
1088 {
1089         if (l_old && (map_flags & ~BPF_F_LOCK) == BPF_NOEXIST)
1090                 /* elem already exists */
1091                 return -EEXIST;
1092
1093         if (!l_old && (map_flags & ~BPF_F_LOCK) == BPF_EXIST)
1094                 /* elem doesn't exist, cannot update it */
1095                 return -ENOENT;
1096
1097         return 0;
1098 }
1099
1100 /* Called from syscall or from eBPF program */
1101 static long htab_map_update_elem(struct bpf_map *map, void *key, void *value,
1102                                  u64 map_flags)
1103 {
1104         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1105         struct htab_elem *l_new = NULL, *l_old;
1106         struct hlist_nulls_head *head;
1107         unsigned long flags;
1108         struct bucket *b;
1109         u32 key_size, hash;
1110         int ret;
1111
1112         if (unlikely((map_flags & ~BPF_F_LOCK) > BPF_EXIST))
1113                 /* unknown flags */
1114                 return -EINVAL;
1115
1116         WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
1117                      !rcu_read_lock_bh_held());
1118
1119         key_size = map->key_size;
1120
1121         hash = htab_map_hash(key, key_size, htab->hashrnd);
1122
1123         b = __select_bucket(htab, hash);
1124         head = &b->head;
1125
1126         if (unlikely(map_flags & BPF_F_LOCK)) {
1127                 if (unlikely(!btf_record_has_field(map->record, BPF_SPIN_LOCK)))
1128                         return -EINVAL;
1129                 /* find an element without taking the bucket lock */
1130                 l_old = lookup_nulls_elem_raw(head, hash, key, key_size,
1131                                               htab->n_buckets);
1132                 ret = check_flags(htab, l_old, map_flags);
1133                 if (ret)
1134                         return ret;
1135                 if (l_old) {
1136                         /* grab the element lock and update value in place */
1137                         copy_map_value_locked(map,
1138                                               l_old->key + round_up(key_size, 8),
1139                                               value, false);
1140                         return 0;
1141                 }
1142                 /* fall through, grab the bucket lock and lookup again.
1143                  * 99.9% chance that the element won't be found,
1144                  * but second lookup under lock has to be done.
1145                  */
1146         }
1147
1148         ret = htab_lock_bucket(htab, b, hash, &flags);
1149         if (ret)
1150                 return ret;
1151
1152         l_old = lookup_elem_raw(head, hash, key, key_size);
1153
1154         ret = check_flags(htab, l_old, map_flags);
1155         if (ret)
1156                 goto err;
1157
1158         if (unlikely(l_old && (map_flags & BPF_F_LOCK))) {
1159                 /* first lookup without the bucket lock didn't find the element,
1160                  * but second lookup with the bucket lock found it.
1161                  * This case is highly unlikely, but has to be dealt with:
1162                  * grab the element lock in addition to the bucket lock
1163                  * and update element in place
1164                  */
1165                 copy_map_value_locked(map,
1166                                       l_old->key + round_up(key_size, 8),
1167                                       value, false);
1168                 ret = 0;
1169                 goto err;
1170         }
1171
1172         l_new = alloc_htab_elem(htab, key, value, key_size, hash, false, false,
1173                                 l_old);
1174         if (IS_ERR(l_new)) {
1175                 /* all pre-allocated elements are in use or memory exhausted */
1176                 ret = PTR_ERR(l_new);
1177                 goto err;
1178         }
1179
1180         /* add new element to the head of the list, so that
1181          * concurrent search will find it before old elem
1182          */
1183         hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1184         if (l_old) {
1185                 hlist_nulls_del_rcu(&l_old->hash_node);
1186                 if (!htab_is_prealloc(htab))
1187                         free_htab_elem(htab, l_old);
1188                 else
1189                         check_and_free_fields(htab, l_old);
1190         }
1191         ret = 0;
1192 err:
1193         htab_unlock_bucket(htab, b, hash, flags);
1194         return ret;
1195 }
1196
1197 static void htab_lru_push_free(struct bpf_htab *htab, struct htab_elem *elem)
1198 {
1199         check_and_free_fields(htab, elem);
1200         bpf_map_dec_elem_count(&htab->map);
1201         bpf_lru_push_free(&htab->lru, &elem->lru_node);
1202 }
1203
1204 static long htab_lru_map_update_elem(struct bpf_map *map, void *key, void *value,
1205                                      u64 map_flags)
1206 {
1207         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1208         struct htab_elem *l_new, *l_old = NULL;
1209         struct hlist_nulls_head *head;
1210         unsigned long flags;
1211         struct bucket *b;
1212         u32 key_size, hash;
1213         int ret;
1214
1215         if (unlikely(map_flags > BPF_EXIST))
1216                 /* unknown flags */
1217                 return -EINVAL;
1218
1219         WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
1220                      !rcu_read_lock_bh_held());
1221
1222         key_size = map->key_size;
1223
1224         hash = htab_map_hash(key, key_size, htab->hashrnd);
1225
1226         b = __select_bucket(htab, hash);
1227         head = &b->head;
1228
1229         /* For LRU, we need to alloc before taking bucket's
1230          * spinlock because getting free nodes from LRU may need
1231          * to remove older elements from htab and this removal
1232          * operation will need a bucket lock.
1233          */
1234         l_new = prealloc_lru_pop(htab, key, hash);
1235         if (!l_new)
1236                 return -ENOMEM;
1237         copy_map_value(&htab->map,
1238                        l_new->key + round_up(map->key_size, 8), value);
1239
1240         ret = htab_lock_bucket(htab, b, hash, &flags);
1241         if (ret)
1242                 goto err_lock_bucket;
1243
1244         l_old = lookup_elem_raw(head, hash, key, key_size);
1245
1246         ret = check_flags(htab, l_old, map_flags);
1247         if (ret)
1248                 goto err;
1249
1250         /* add new element to the head of the list, so that
1251          * concurrent search will find it before old elem
1252          */
1253         hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1254         if (l_old) {
1255                 bpf_lru_node_set_ref(&l_new->lru_node);
1256                 hlist_nulls_del_rcu(&l_old->hash_node);
1257         }
1258         ret = 0;
1259
1260 err:
1261         htab_unlock_bucket(htab, b, hash, flags);
1262
1263 err_lock_bucket:
1264         if (ret)
1265                 htab_lru_push_free(htab, l_new);
1266         else if (l_old)
1267                 htab_lru_push_free(htab, l_old);
1268
1269         return ret;
1270 }
1271
1272 static long __htab_percpu_map_update_elem(struct bpf_map *map, void *key,
1273                                           void *value, u64 map_flags,
1274                                           bool onallcpus)
1275 {
1276         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1277         struct htab_elem *l_new = NULL, *l_old;
1278         struct hlist_nulls_head *head;
1279         unsigned long flags;
1280         struct bucket *b;
1281         u32 key_size, hash;
1282         int ret;
1283
1284         if (unlikely(map_flags > BPF_EXIST))
1285                 /* unknown flags */
1286                 return -EINVAL;
1287
1288         WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
1289                      !rcu_read_lock_bh_held());
1290
1291         key_size = map->key_size;
1292
1293         hash = htab_map_hash(key, key_size, htab->hashrnd);
1294
1295         b = __select_bucket(htab, hash);
1296         head = &b->head;
1297
1298         ret = htab_lock_bucket(htab, b, hash, &flags);
1299         if (ret)
1300                 return ret;
1301
1302         l_old = lookup_elem_raw(head, hash, key, key_size);
1303
1304         ret = check_flags(htab, l_old, map_flags);
1305         if (ret)
1306                 goto err;
1307
1308         if (l_old) {
1309                 /* per-cpu hash map can update value in-place */
1310                 pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
1311                                 value, onallcpus);
1312         } else {
1313                 l_new = alloc_htab_elem(htab, key, value, key_size,
1314                                         hash, true, onallcpus, NULL);
1315                 if (IS_ERR(l_new)) {
1316                         ret = PTR_ERR(l_new);
1317                         goto err;
1318                 }
1319                 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1320         }
1321         ret = 0;
1322 err:
1323         htab_unlock_bucket(htab, b, hash, flags);
1324         return ret;
1325 }
1326
1327 static long __htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
1328                                               void *value, u64 map_flags,
1329                                               bool onallcpus)
1330 {
1331         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1332         struct htab_elem *l_new = NULL, *l_old;
1333         struct hlist_nulls_head *head;
1334         unsigned long flags;
1335         struct bucket *b;
1336         u32 key_size, hash;
1337         int ret;
1338
1339         if (unlikely(map_flags > BPF_EXIST))
1340                 /* unknown flags */
1341                 return -EINVAL;
1342
1343         WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
1344                      !rcu_read_lock_bh_held());
1345
1346         key_size = map->key_size;
1347
1348         hash = htab_map_hash(key, key_size, htab->hashrnd);
1349
1350         b = __select_bucket(htab, hash);
1351         head = &b->head;
1352
1353         /* For LRU, we need to alloc before taking bucket's
1354          * spinlock because LRU's elem alloc may need
1355          * to remove older elem from htab and this removal
1356          * operation will need a bucket lock.
1357          */
1358         if (map_flags != BPF_EXIST) {
1359                 l_new = prealloc_lru_pop(htab, key, hash);
1360                 if (!l_new)
1361                         return -ENOMEM;
1362         }
1363
1364         ret = htab_lock_bucket(htab, b, hash, &flags);
1365         if (ret)
1366                 goto err_lock_bucket;
1367
1368         l_old = lookup_elem_raw(head, hash, key, key_size);
1369
1370         ret = check_flags(htab, l_old, map_flags);
1371         if (ret)
1372                 goto err;
1373
1374         if (l_old) {
1375                 bpf_lru_node_set_ref(&l_old->lru_node);
1376
1377                 /* per-cpu hash map can update value in-place */
1378                 pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
1379                                 value, onallcpus);
1380         } else {
1381                 pcpu_init_value(htab, htab_elem_get_ptr(l_new, key_size),
1382                                 value, onallcpus);
1383                 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1384                 l_new = NULL;
1385         }
1386         ret = 0;
1387 err:
1388         htab_unlock_bucket(htab, b, hash, flags);
1389 err_lock_bucket:
1390         if (l_new) {
1391                 bpf_map_dec_elem_count(&htab->map);
1392                 bpf_lru_push_free(&htab->lru, &l_new->lru_node);
1393         }
1394         return ret;
1395 }
1396
1397 static long htab_percpu_map_update_elem(struct bpf_map *map, void *key,
1398                                         void *value, u64 map_flags)
1399 {
1400         return __htab_percpu_map_update_elem(map, key, value, map_flags, false);
1401 }
1402
1403 static long htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
1404                                             void *value, u64 map_flags)
1405 {
1406         return __htab_lru_percpu_map_update_elem(map, key, value, map_flags,
1407                                                  false);
1408 }
1409
1410 /* Called from syscall or from eBPF program */
1411 static long htab_map_delete_elem(struct bpf_map *map, void *key)
1412 {
1413         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1414         struct hlist_nulls_head *head;
1415         struct bucket *b;
1416         struct htab_elem *l;
1417         unsigned long flags;
1418         u32 hash, key_size;
1419         int ret;
1420
1421         WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
1422                      !rcu_read_lock_bh_held());
1423
1424         key_size = map->key_size;
1425
1426         hash = htab_map_hash(key, key_size, htab->hashrnd);
1427         b = __select_bucket(htab, hash);
1428         head = &b->head;
1429
1430         ret = htab_lock_bucket(htab, b, hash, &flags);
1431         if (ret)
1432                 return ret;
1433
1434         l = lookup_elem_raw(head, hash, key, key_size);
1435
1436         if (l) {
1437                 hlist_nulls_del_rcu(&l->hash_node);
1438                 free_htab_elem(htab, l);
1439         } else {
1440                 ret = -ENOENT;
1441         }
1442
1443         htab_unlock_bucket(htab, b, hash, flags);
1444         return ret;
1445 }
1446
1447 static long htab_lru_map_delete_elem(struct bpf_map *map, void *key)
1448 {
1449         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1450         struct hlist_nulls_head *head;
1451         struct bucket *b;
1452         struct htab_elem *l;
1453         unsigned long flags;
1454         u32 hash, key_size;
1455         int ret;
1456
1457         WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
1458                      !rcu_read_lock_bh_held());
1459
1460         key_size = map->key_size;
1461
1462         hash = htab_map_hash(key, key_size, htab->hashrnd);
1463         b = __select_bucket(htab, hash);
1464         head = &b->head;
1465
1466         ret = htab_lock_bucket(htab, b, hash, &flags);
1467         if (ret)
1468                 return ret;
1469
1470         l = lookup_elem_raw(head, hash, key, key_size);
1471
1472         if (l)
1473                 hlist_nulls_del_rcu(&l->hash_node);
1474         else
1475                 ret = -ENOENT;
1476
1477         htab_unlock_bucket(htab, b, hash, flags);
1478         if (l)
1479                 htab_lru_push_free(htab, l);
1480         return ret;
1481 }
1482
1483 static void delete_all_elements(struct bpf_htab *htab)
1484 {
1485         int i;
1486
1487         /* It's called from a worker thread, so disable migration here,
1488          * since bpf_mem_cache_free() relies on that.
1489          */
1490         migrate_disable();
1491         for (i = 0; i < htab->n_buckets; i++) {
1492                 struct hlist_nulls_head *head = select_bucket(htab, i);
1493                 struct hlist_nulls_node *n;
1494                 struct htab_elem *l;
1495
1496                 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1497                         hlist_nulls_del_rcu(&l->hash_node);
1498                         htab_elem_free(htab, l);
1499                 }
1500                 cond_resched();
1501         }
1502         migrate_enable();
1503 }
1504
1505 static void htab_free_malloced_timers_and_wq(struct bpf_htab *htab)
1506 {
1507         int i;
1508
1509         rcu_read_lock();
1510         for (i = 0; i < htab->n_buckets; i++) {
1511                 struct hlist_nulls_head *head = select_bucket(htab, i);
1512                 struct hlist_nulls_node *n;
1513                 struct htab_elem *l;
1514
1515                 hlist_nulls_for_each_entry(l, n, head, hash_node) {
1516                         /* We only free timer on uref dropping to zero */
1517                         if (btf_record_has_field(htab->map.record, BPF_TIMER))
1518                                 bpf_obj_free_timer(htab->map.record,
1519                                                    l->key + round_up(htab->map.key_size, 8));
1520                         if (btf_record_has_field(htab->map.record, BPF_WORKQUEUE))
1521                                 bpf_obj_free_workqueue(htab->map.record,
1522                                                        l->key + round_up(htab->map.key_size, 8));
1523                 }
1524                 cond_resched_rcu();
1525         }
1526         rcu_read_unlock();
1527 }
1528
1529 static void htab_map_free_timers_and_wq(struct bpf_map *map)
1530 {
1531         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1532
1533         /* We only free timer and workqueue on uref dropping to zero */
1534         if (btf_record_has_field(htab->map.record, BPF_TIMER | BPF_WORKQUEUE)) {
1535                 if (!htab_is_prealloc(htab))
1536                         htab_free_malloced_timers_and_wq(htab);
1537                 else
1538                         htab_free_prealloced_timers_and_wq(htab);
1539         }
1540 }
1541
1542 /* Called when map->refcnt goes to zero, either from workqueue or from syscall */
1543 static void htab_map_free(struct bpf_map *map)
1544 {
1545         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1546         int i;
1547
1548         /* bpf_free_used_maps() or close(map_fd) will trigger this map_free callback.
1549          * bpf_free_used_maps() is called after bpf prog is no longer executing.
1550          * There is no need to synchronize_rcu() here to protect map elements.
1551          */
1552
1553         /* htab no longer uses call_rcu() directly. bpf_mem_alloc does it
1554          * underneath and is responsible for waiting for callbacks to finish
1555          * during bpf_mem_alloc_destroy().
1556          */
1557         if (!htab_is_prealloc(htab)) {
1558                 delete_all_elements(htab);
1559         } else {
1560                 htab_free_prealloced_fields(htab);
1561                 prealloc_destroy(htab);
1562         }
1563
1564         bpf_map_free_elem_count(map);
1565         free_percpu(htab->extra_elems);
1566         bpf_map_area_free(htab->buckets);
1567         bpf_mem_alloc_destroy(&htab->pcpu_ma);
1568         bpf_mem_alloc_destroy(&htab->ma);
1569         if (htab->use_percpu_counter)
1570                 percpu_counter_destroy(&htab->pcount);
1571         for (i = 0; i < HASHTAB_MAP_LOCK_COUNT; i++)
1572                 free_percpu(htab->map_locked[i]);
1573         lockdep_unregister_key(&htab->lockdep_key);
1574         bpf_map_area_free(htab);
1575 }
1576
1577 static void htab_map_seq_show_elem(struct bpf_map *map, void *key,
1578                                    struct seq_file *m)
1579 {
1580         void *value;
1581
1582         rcu_read_lock();
1583
1584         value = htab_map_lookup_elem(map, key);
1585         if (!value) {
1586                 rcu_read_unlock();
1587                 return;
1588         }
1589
1590         btf_type_seq_show(map->btf, map->btf_key_type_id, key, m);
1591         seq_puts(m, ": ");
1592         btf_type_seq_show(map->btf, map->btf_value_type_id, value, m);
1593         seq_putc(m, '\n');
1594
1595         rcu_read_unlock();
1596 }
1597
1598 static int __htab_map_lookup_and_delete_elem(struct bpf_map *map, void *key,
1599                                              void *value, bool is_lru_map,
1600                                              bool is_percpu, u64 flags)
1601 {
1602         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1603         struct hlist_nulls_head *head;
1604         unsigned long bflags;
1605         struct htab_elem *l;
1606         u32 hash, key_size;
1607         struct bucket *b;
1608         int ret;
1609
1610         key_size = map->key_size;
1611
1612         hash = htab_map_hash(key, key_size, htab->hashrnd);
1613         b = __select_bucket(htab, hash);
1614         head = &b->head;
1615
1616         ret = htab_lock_bucket(htab, b, hash, &bflags);
1617         if (ret)
1618                 return ret;
1619
1620         l = lookup_elem_raw(head, hash, key, key_size);
1621         if (!l) {
1622                 ret = -ENOENT;
1623         } else {
1624                 if (is_percpu) {
1625                         u32 roundup_value_size = round_up(map->value_size, 8);
1626                         void __percpu *pptr;
1627                         int off = 0, cpu;
1628
1629                         pptr = htab_elem_get_ptr(l, key_size);
1630                         for_each_possible_cpu(cpu) {
1631                                 copy_map_value_long(&htab->map, value + off, per_cpu_ptr(pptr, cpu));
1632                                 check_and_init_map_value(&htab->map, value + off);
1633                                 off += roundup_value_size;
1634                         }
1635                 } else {
1636                         u32 roundup_key_size = round_up(map->key_size, 8);
1637
1638                         if (flags & BPF_F_LOCK)
1639                                 copy_map_value_locked(map, value, l->key +
1640                                                       roundup_key_size,
1641                                                       true);
1642                         else
1643                                 copy_map_value(map, value, l->key +
1644                                                roundup_key_size);
1645                         /* Zeroing special fields in the temp buffer */
1646                         check_and_init_map_value(map, value);
1647                 }
1648
1649                 hlist_nulls_del_rcu(&l->hash_node);
1650                 if (!is_lru_map)
1651                         free_htab_elem(htab, l);
1652         }
1653
1654         htab_unlock_bucket(htab, b, hash, bflags);
1655
1656         if (is_lru_map && l)
1657                 htab_lru_push_free(htab, l);
1658
1659         return ret;
1660 }
1661
1662 static int htab_map_lookup_and_delete_elem(struct bpf_map *map, void *key,
1663                                            void *value, u64 flags)
1664 {
1665         return __htab_map_lookup_and_delete_elem(map, key, value, false, false,
1666                                                  flags);
1667 }
1668
1669 static int htab_percpu_map_lookup_and_delete_elem(struct bpf_map *map,
1670                                                   void *key, void *value,
1671                                                   u64 flags)
1672 {
1673         return __htab_map_lookup_and_delete_elem(map, key, value, false, true,
1674                                                  flags);
1675 }
1676
1677 static int htab_lru_map_lookup_and_delete_elem(struct bpf_map *map, void *key,
1678                                                void *value, u64 flags)
1679 {
1680         return __htab_map_lookup_and_delete_elem(map, key, value, true, false,
1681                                                  flags);
1682 }
1683
1684 static int htab_lru_percpu_map_lookup_and_delete_elem(struct bpf_map *map,
1685                                                       void *key, void *value,
1686                                                       u64 flags)
1687 {
1688         return __htab_map_lookup_and_delete_elem(map, key, value, true, true,
1689                                                  flags);
1690 }
1691
1692 static int
1693 __htab_map_lookup_and_delete_batch(struct bpf_map *map,
1694                                    const union bpf_attr *attr,
1695                                    union bpf_attr __user *uattr,
1696                                    bool do_delete, bool is_lru_map,
1697                                    bool is_percpu)
1698 {
1699         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1700         u32 bucket_cnt, total, key_size, value_size, roundup_key_size;
1701         void *keys = NULL, *values = NULL, *value, *dst_key, *dst_val;
1702         void __user *uvalues = u64_to_user_ptr(attr->batch.values);
1703         void __user *ukeys = u64_to_user_ptr(attr->batch.keys);
1704         void __user *ubatch = u64_to_user_ptr(attr->batch.in_batch);
1705         u32 batch, max_count, size, bucket_size, map_id;
1706         struct htab_elem *node_to_free = NULL;
1707         u64 elem_map_flags, map_flags;
1708         struct hlist_nulls_head *head;
1709         struct hlist_nulls_node *n;
1710         unsigned long flags = 0;
1711         bool locked = false;
1712         struct htab_elem *l;
1713         struct bucket *b;
1714         int ret = 0;
1715
1716         elem_map_flags = attr->batch.elem_flags;
1717         if ((elem_map_flags & ~BPF_F_LOCK) ||
1718             ((elem_map_flags & BPF_F_LOCK) && !btf_record_has_field(map->record, BPF_SPIN_LOCK)))
1719                 return -EINVAL;
1720
1721         map_flags = attr->batch.flags;
1722         if (map_flags)
1723                 return -EINVAL;
1724
1725         max_count = attr->batch.count;
1726         if (!max_count)
1727                 return 0;
1728
1729         if (put_user(0, &uattr->batch.count))
1730                 return -EFAULT;
1731
1732         batch = 0;
1733         if (ubatch && copy_from_user(&batch, ubatch, sizeof(batch)))
1734                 return -EFAULT;
1735
1736         if (batch >= htab->n_buckets)
1737                 return -ENOENT;
1738
1739         key_size = htab->map.key_size;
1740         roundup_key_size = round_up(htab->map.key_size, 8);
1741         value_size = htab->map.value_size;
1742         size = round_up(value_size, 8);
1743         if (is_percpu)
1744                 value_size = size * num_possible_cpus();
1745         total = 0;
1746         /* while experimenting with hash tables with sizes ranging from 10 to
1747          * 1000, it was observed that a bucket can have up to 5 entries.
1748          */
1749         bucket_size = 5;
1750
1751 alloc:
1752         /* We cannot do copy_from_user or copy_to_user inside
1753          * the rcu_read_lock. Allocate enough space here.
1754          */
1755         keys = kvmalloc_array(key_size, bucket_size, GFP_USER | __GFP_NOWARN);
1756         values = kvmalloc_array(value_size, bucket_size, GFP_USER | __GFP_NOWARN);
1757         if (!keys || !values) {
1758                 ret = -ENOMEM;
1759                 goto after_loop;
1760         }
1761
1762 again:
1763         bpf_disable_instrumentation();
1764         rcu_read_lock();
1765 again_nocopy:
1766         dst_key = keys;
1767         dst_val = values;
1768         b = &htab->buckets[batch];
1769         head = &b->head;
1770         /* do not grab the lock unless need it (bucket_cnt > 0). */
1771         if (locked) {
1772                 ret = htab_lock_bucket(htab, b, batch, &flags);
1773                 if (ret) {
1774                         rcu_read_unlock();
1775                         bpf_enable_instrumentation();
1776                         goto after_loop;
1777                 }
1778         }
1779
1780         bucket_cnt = 0;
1781         hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
1782                 bucket_cnt++;
1783
1784         if (bucket_cnt && !locked) {
1785                 locked = true;
1786                 goto again_nocopy;
1787         }
1788
1789         if (bucket_cnt > (max_count - total)) {
1790                 if (total == 0)
1791                         ret = -ENOSPC;
1792                 /* Note that since bucket_cnt > 0 here, it is implicit
1793                  * that the locked was grabbed, so release it.
1794                  */
1795                 htab_unlock_bucket(htab, b, batch, flags);
1796                 rcu_read_unlock();
1797                 bpf_enable_instrumentation();
1798                 goto after_loop;
1799         }
1800
1801         if (bucket_cnt > bucket_size) {
1802                 bucket_size = bucket_cnt;
1803                 /* Note that since bucket_cnt > 0 here, it is implicit
1804                  * that the locked was grabbed, so release it.
1805                  */
1806                 htab_unlock_bucket(htab, b, batch, flags);
1807                 rcu_read_unlock();
1808                 bpf_enable_instrumentation();
1809                 kvfree(keys);
1810                 kvfree(values);
1811                 goto alloc;
1812         }
1813
1814         /* Next block is only safe to run if you have grabbed the lock */
1815         if (!locked)
1816                 goto next_batch;
1817
1818         hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1819                 memcpy(dst_key, l->key, key_size);
1820
1821                 if (is_percpu) {
1822                         int off = 0, cpu;
1823                         void __percpu *pptr;
1824
1825                         pptr = htab_elem_get_ptr(l, map->key_size);
1826                         for_each_possible_cpu(cpu) {
1827                                 copy_map_value_long(&htab->map, dst_val + off, per_cpu_ptr(pptr, cpu));
1828                                 check_and_init_map_value(&htab->map, dst_val + off);
1829                                 off += size;
1830                         }
1831                 } else {
1832                         value = l->key + roundup_key_size;
1833                         if (map->map_type == BPF_MAP_TYPE_HASH_OF_MAPS) {
1834                                 struct bpf_map **inner_map = value;
1835
1836                                  /* Actual value is the id of the inner map */
1837                                 map_id = map->ops->map_fd_sys_lookup_elem(*inner_map);
1838                                 value = &map_id;
1839                         }
1840
1841                         if (elem_map_flags & BPF_F_LOCK)
1842                                 copy_map_value_locked(map, dst_val, value,
1843                                                       true);
1844                         else
1845                                 copy_map_value(map, dst_val, value);
1846                         /* Zeroing special fields in the temp buffer */
1847                         check_and_init_map_value(map, dst_val);
1848                 }
1849                 if (do_delete) {
1850                         hlist_nulls_del_rcu(&l->hash_node);
1851
1852                         /* bpf_lru_push_free() will acquire lru_lock, which
1853                          * may cause deadlock. See comments in function
1854                          * prealloc_lru_pop(). Let us do bpf_lru_push_free()
1855                          * after releasing the bucket lock.
1856                          */
1857                         if (is_lru_map) {
1858                                 l->batch_flink = node_to_free;
1859                                 node_to_free = l;
1860                         } else {
1861                                 free_htab_elem(htab, l);
1862                         }
1863                 }
1864                 dst_key += key_size;
1865                 dst_val += value_size;
1866         }
1867
1868         htab_unlock_bucket(htab, b, batch, flags);
1869         locked = false;
1870
1871         while (node_to_free) {
1872                 l = node_to_free;
1873                 node_to_free = node_to_free->batch_flink;
1874                 htab_lru_push_free(htab, l);
1875         }
1876
1877 next_batch:
1878         /* If we are not copying data, we can go to next bucket and avoid
1879          * unlocking the rcu.
1880          */
1881         if (!bucket_cnt && (batch + 1 < htab->n_buckets)) {
1882                 batch++;
1883                 goto again_nocopy;
1884         }
1885
1886         rcu_read_unlock();
1887         bpf_enable_instrumentation();
1888         if (bucket_cnt && (copy_to_user(ukeys + total * key_size, keys,
1889             key_size * bucket_cnt) ||
1890             copy_to_user(uvalues + total * value_size, values,
1891             value_size * bucket_cnt))) {
1892                 ret = -EFAULT;
1893                 goto after_loop;
1894         }
1895
1896         total += bucket_cnt;
1897         batch++;
1898         if (batch >= htab->n_buckets) {
1899                 ret = -ENOENT;
1900                 goto after_loop;
1901         }
1902         goto again;
1903
1904 after_loop:
1905         if (ret == -EFAULT)
1906                 goto out;
1907
1908         /* copy # of entries and next batch */
1909         ubatch = u64_to_user_ptr(attr->batch.out_batch);
1910         if (copy_to_user(ubatch, &batch, sizeof(batch)) ||
1911             put_user(total, &uattr->batch.count))
1912                 ret = -EFAULT;
1913
1914 out:
1915         kvfree(keys);
1916         kvfree(values);
1917         return ret;
1918 }
1919
1920 static int
1921 htab_percpu_map_lookup_batch(struct bpf_map *map, const union bpf_attr *attr,
1922                              union bpf_attr __user *uattr)
1923 {
1924         return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1925                                                   false, true);
1926 }
1927
1928 static int
1929 htab_percpu_map_lookup_and_delete_batch(struct bpf_map *map,
1930                                         const union bpf_attr *attr,
1931                                         union bpf_attr __user *uattr)
1932 {
1933         return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1934                                                   false, true);
1935 }
1936
1937 static int
1938 htab_map_lookup_batch(struct bpf_map *map, const union bpf_attr *attr,
1939                       union bpf_attr __user *uattr)
1940 {
1941         return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1942                                                   false, false);
1943 }
1944
1945 static int
1946 htab_map_lookup_and_delete_batch(struct bpf_map *map,
1947                                  const union bpf_attr *attr,
1948                                  union bpf_attr __user *uattr)
1949 {
1950         return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1951                                                   false, false);
1952 }
1953
1954 static int
1955 htab_lru_percpu_map_lookup_batch(struct bpf_map *map,
1956                                  const union bpf_attr *attr,
1957                                  union bpf_attr __user *uattr)
1958 {
1959         return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1960                                                   true, true);
1961 }
1962
1963 static int
1964 htab_lru_percpu_map_lookup_and_delete_batch(struct bpf_map *map,
1965                                             const union bpf_attr *attr,
1966                                             union bpf_attr __user *uattr)
1967 {
1968         return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1969                                                   true, true);
1970 }
1971
1972 static int
1973 htab_lru_map_lookup_batch(struct bpf_map *map, const union bpf_attr *attr,
1974                           union bpf_attr __user *uattr)
1975 {
1976         return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1977                                                   true, false);
1978 }
1979
1980 static int
1981 htab_lru_map_lookup_and_delete_batch(struct bpf_map *map,
1982                                      const union bpf_attr *attr,
1983                                      union bpf_attr __user *uattr)
1984 {
1985         return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1986                                                   true, false);
1987 }
1988
1989 struct bpf_iter_seq_hash_map_info {
1990         struct bpf_map *map;
1991         struct bpf_htab *htab;
1992         void *percpu_value_buf; // non-zero means percpu hash
1993         u32 bucket_id;
1994         u32 skip_elems;
1995 };
1996
1997 static struct htab_elem *
1998 bpf_hash_map_seq_find_next(struct bpf_iter_seq_hash_map_info *info,
1999                            struct htab_elem *prev_elem)
2000 {
2001         const struct bpf_htab *htab = info->htab;
2002         u32 skip_elems = info->skip_elems;
2003         u32 bucket_id = info->bucket_id;
2004         struct hlist_nulls_head *head;
2005         struct hlist_nulls_node *n;
2006         struct htab_elem *elem;
2007         struct bucket *b;
2008         u32 i, count;
2009
2010         if (bucket_id >= htab->n_buckets)
2011                 return NULL;
2012
2013         /* try to find next elem in the same bucket */
2014         if (prev_elem) {
2015                 /* no update/deletion on this bucket, prev_elem should be still valid
2016                  * and we won't skip elements.
2017                  */
2018                 n = rcu_dereference_raw(hlist_nulls_next_rcu(&prev_elem->hash_node));
2019                 elem = hlist_nulls_entry_safe(n, struct htab_elem, hash_node);
2020                 if (elem)
2021                         return elem;
2022
2023                 /* not found, unlock and go to the next bucket */
2024                 b = &htab->buckets[bucket_id++];
2025                 rcu_read_unlock();
2026                 skip_elems = 0;
2027         }
2028
2029         for (i = bucket_id; i < htab->n_buckets; i++) {
2030                 b = &htab->buckets[i];
2031                 rcu_read_lock();
2032
2033                 count = 0;
2034                 head = &b->head;
2035                 hlist_nulls_for_each_entry_rcu(elem, n, head, hash_node) {
2036                         if (count >= skip_elems) {
2037                                 info->bucket_id = i;
2038                                 info->skip_elems = count;
2039                                 return elem;
2040                         }
2041                         count++;
2042                 }
2043
2044                 rcu_read_unlock();
2045                 skip_elems = 0;
2046         }
2047
2048         info->bucket_id = i;
2049         info->skip_elems = 0;
2050         return NULL;
2051 }
2052
2053 static void *bpf_hash_map_seq_start(struct seq_file *seq, loff_t *pos)
2054 {
2055         struct bpf_iter_seq_hash_map_info *info = seq->private;
2056         struct htab_elem *elem;
2057
2058         elem = bpf_hash_map_seq_find_next(info, NULL);
2059         if (!elem)
2060                 return NULL;
2061
2062         if (*pos == 0)
2063                 ++*pos;
2064         return elem;
2065 }
2066
2067 static void *bpf_hash_map_seq_next(struct seq_file *seq, void *v, loff_t *pos)
2068 {
2069         struct bpf_iter_seq_hash_map_info *info = seq->private;
2070
2071         ++*pos;
2072         ++info->skip_elems;
2073         return bpf_hash_map_seq_find_next(info, v);
2074 }
2075
2076 static int __bpf_hash_map_seq_show(struct seq_file *seq, struct htab_elem *elem)
2077 {
2078         struct bpf_iter_seq_hash_map_info *info = seq->private;
2079         u32 roundup_key_size, roundup_value_size;
2080         struct bpf_iter__bpf_map_elem ctx = {};
2081         struct bpf_map *map = info->map;
2082         struct bpf_iter_meta meta;
2083         int ret = 0, off = 0, cpu;
2084         struct bpf_prog *prog;
2085         void __percpu *pptr;
2086
2087         meta.seq = seq;
2088         prog = bpf_iter_get_info(&meta, elem == NULL);
2089         if (prog) {
2090                 ctx.meta = &meta;
2091                 ctx.map = info->map;
2092                 if (elem) {
2093                         roundup_key_size = round_up(map->key_size, 8);
2094                         ctx.key = elem->key;
2095                         if (!info->percpu_value_buf) {
2096                                 ctx.value = elem->key + roundup_key_size;
2097                         } else {
2098                                 roundup_value_size = round_up(map->value_size, 8);
2099                                 pptr = htab_elem_get_ptr(elem, map->key_size);
2100                                 for_each_possible_cpu(cpu) {
2101                                         copy_map_value_long(map, info->percpu_value_buf + off,
2102                                                             per_cpu_ptr(pptr, cpu));
2103                                         check_and_init_map_value(map, info->percpu_value_buf + off);
2104                                         off += roundup_value_size;
2105                                 }
2106                                 ctx.value = info->percpu_value_buf;
2107                         }
2108                 }
2109                 ret = bpf_iter_run_prog(prog, &ctx);
2110         }
2111
2112         return ret;
2113 }
2114
2115 static int bpf_hash_map_seq_show(struct seq_file *seq, void *v)
2116 {
2117         return __bpf_hash_map_seq_show(seq, v);
2118 }
2119
2120 static void bpf_hash_map_seq_stop(struct seq_file *seq, void *v)
2121 {
2122         if (!v)
2123                 (void)__bpf_hash_map_seq_show(seq, NULL);
2124         else
2125                 rcu_read_unlock();
2126 }
2127
2128 static int bpf_iter_init_hash_map(void *priv_data,
2129                                   struct bpf_iter_aux_info *aux)
2130 {
2131         struct bpf_iter_seq_hash_map_info *seq_info = priv_data;
2132         struct bpf_map *map = aux->map;
2133         void *value_buf;
2134         u32 buf_size;
2135
2136         if (map->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
2137             map->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH) {
2138                 buf_size = round_up(map->value_size, 8) * num_possible_cpus();
2139                 value_buf = kmalloc(buf_size, GFP_USER | __GFP_NOWARN);
2140                 if (!value_buf)
2141                         return -ENOMEM;
2142
2143                 seq_info->percpu_value_buf = value_buf;
2144         }
2145
2146         bpf_map_inc_with_uref(map);
2147         seq_info->map = map;
2148         seq_info->htab = container_of(map, struct bpf_htab, map);
2149         return 0;
2150 }
2151
2152 static void bpf_iter_fini_hash_map(void *priv_data)
2153 {
2154         struct bpf_iter_seq_hash_map_info *seq_info = priv_data;
2155
2156         bpf_map_put_with_uref(seq_info->map);
2157         kfree(seq_info->percpu_value_buf);
2158 }
2159
2160 static const struct seq_operations bpf_hash_map_seq_ops = {
2161         .start  = bpf_hash_map_seq_start,
2162         .next   = bpf_hash_map_seq_next,
2163         .stop   = bpf_hash_map_seq_stop,
2164         .show   = bpf_hash_map_seq_show,
2165 };
2166
2167 static const struct bpf_iter_seq_info iter_seq_info = {
2168         .seq_ops                = &bpf_hash_map_seq_ops,
2169         .init_seq_private       = bpf_iter_init_hash_map,
2170         .fini_seq_private       = bpf_iter_fini_hash_map,
2171         .seq_priv_size          = sizeof(struct bpf_iter_seq_hash_map_info),
2172 };
2173
2174 static long bpf_for_each_hash_elem(struct bpf_map *map, bpf_callback_t callback_fn,
2175                                    void *callback_ctx, u64 flags)
2176 {
2177         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
2178         struct hlist_nulls_head *head;
2179         struct hlist_nulls_node *n;
2180         struct htab_elem *elem;
2181         u32 roundup_key_size;
2182         int i, num_elems = 0;
2183         void __percpu *pptr;
2184         struct bucket *b;
2185         void *key, *val;
2186         bool is_percpu;
2187         u64 ret = 0;
2188
2189         if (flags != 0)
2190                 return -EINVAL;
2191
2192         is_percpu = htab_is_percpu(htab);
2193
2194         roundup_key_size = round_up(map->key_size, 8);
2195         /* disable migration so percpu value prepared here will be the
2196          * same as the one seen by the bpf program with bpf_map_lookup_elem().
2197          */
2198         if (is_percpu)
2199                 migrate_disable();
2200         for (i = 0; i < htab->n_buckets; i++) {
2201                 b = &htab->buckets[i];
2202                 rcu_read_lock();
2203                 head = &b->head;
2204                 hlist_nulls_for_each_entry_rcu(elem, n, head, hash_node) {
2205                         key = elem->key;
2206                         if (is_percpu) {
2207                                 /* current cpu value for percpu map */
2208                                 pptr = htab_elem_get_ptr(elem, map->key_size);
2209                                 val = this_cpu_ptr(pptr);
2210                         } else {
2211                                 val = elem->key + roundup_key_size;
2212                         }
2213                         num_elems++;
2214                         ret = callback_fn((u64)(long)map, (u64)(long)key,
2215                                           (u64)(long)val, (u64)(long)callback_ctx, 0);
2216                         /* return value: 0 - continue, 1 - stop and return */
2217                         if (ret) {
2218                                 rcu_read_unlock();
2219                                 goto out;
2220                         }
2221                 }
2222                 rcu_read_unlock();
2223         }
2224 out:
2225         if (is_percpu)
2226                 migrate_enable();
2227         return num_elems;
2228 }
2229
2230 static u64 htab_map_mem_usage(const struct bpf_map *map)
2231 {
2232         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
2233         u32 value_size = round_up(htab->map.value_size, 8);
2234         bool prealloc = htab_is_prealloc(htab);
2235         bool percpu = htab_is_percpu(htab);
2236         bool lru = htab_is_lru(htab);
2237         u64 num_entries;
2238         u64 usage = sizeof(struct bpf_htab);
2239
2240         usage += sizeof(struct bucket) * htab->n_buckets;
2241         usage += sizeof(int) * num_possible_cpus() * HASHTAB_MAP_LOCK_COUNT;
2242         if (prealloc) {
2243                 num_entries = map->max_entries;
2244                 if (htab_has_extra_elems(htab))
2245                         num_entries += num_possible_cpus();
2246
2247                 usage += htab->elem_size * num_entries;
2248
2249                 if (percpu)
2250                         usage += value_size * num_possible_cpus() * num_entries;
2251                 else if (!lru)
2252                         usage += sizeof(struct htab_elem *) * num_possible_cpus();
2253         } else {
2254 #define LLIST_NODE_SZ sizeof(struct llist_node)
2255
2256                 num_entries = htab->use_percpu_counter ?
2257                                           percpu_counter_sum(&htab->pcount) :
2258                                           atomic_read(&htab->count);
2259                 usage += (htab->elem_size + LLIST_NODE_SZ) * num_entries;
2260                 if (percpu) {
2261                         usage += (LLIST_NODE_SZ + sizeof(void *)) * num_entries;
2262                         usage += value_size * num_possible_cpus() * num_entries;
2263                 }
2264         }
2265         return usage;
2266 }
2267
2268 BTF_ID_LIST_SINGLE(htab_map_btf_ids, struct, bpf_htab)
2269 const struct bpf_map_ops htab_map_ops = {
2270         .map_meta_equal = bpf_map_meta_equal,
2271         .map_alloc_check = htab_map_alloc_check,
2272         .map_alloc = htab_map_alloc,
2273         .map_free = htab_map_free,
2274         .map_get_next_key = htab_map_get_next_key,
2275         .map_release_uref = htab_map_free_timers_and_wq,
2276         .map_lookup_elem = htab_map_lookup_elem,
2277         .map_lookup_and_delete_elem = htab_map_lookup_and_delete_elem,
2278         .map_update_elem = htab_map_update_elem,
2279         .map_delete_elem = htab_map_delete_elem,
2280         .map_gen_lookup = htab_map_gen_lookup,
2281         .map_seq_show_elem = htab_map_seq_show_elem,
2282         .map_set_for_each_callback_args = map_set_for_each_callback_args,
2283         .map_for_each_callback = bpf_for_each_hash_elem,
2284         .map_mem_usage = htab_map_mem_usage,
2285         BATCH_OPS(htab),
2286         .map_btf_id = &htab_map_btf_ids[0],
2287         .iter_seq_info = &iter_seq_info,
2288 };
2289
2290 const struct bpf_map_ops htab_lru_map_ops = {
2291         .map_meta_equal = bpf_map_meta_equal,
2292         .map_alloc_check = htab_map_alloc_check,
2293         .map_alloc = htab_map_alloc,
2294         .map_free = htab_map_free,
2295         .map_get_next_key = htab_map_get_next_key,
2296         .map_release_uref = htab_map_free_timers_and_wq,
2297         .map_lookup_elem = htab_lru_map_lookup_elem,
2298         .map_lookup_and_delete_elem = htab_lru_map_lookup_and_delete_elem,
2299         .map_lookup_elem_sys_only = htab_lru_map_lookup_elem_sys,
2300         .map_update_elem = htab_lru_map_update_elem,
2301         .map_delete_elem = htab_lru_map_delete_elem,
2302         .map_gen_lookup = htab_lru_map_gen_lookup,
2303         .map_seq_show_elem = htab_map_seq_show_elem,
2304         .map_set_for_each_callback_args = map_set_for_each_callback_args,
2305         .map_for_each_callback = bpf_for_each_hash_elem,
2306         .map_mem_usage = htab_map_mem_usage,
2307         BATCH_OPS(htab_lru),
2308         .map_btf_id = &htab_map_btf_ids[0],
2309         .iter_seq_info = &iter_seq_info,
2310 };
2311
2312 /* Called from eBPF program */
2313 static void *htab_percpu_map_lookup_elem(struct bpf_map *map, void *key)
2314 {
2315         struct htab_elem *l = __htab_map_lookup_elem(map, key);
2316
2317         if (l)
2318                 return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
2319         else
2320                 return NULL;
2321 }
2322
2323 /* inline bpf_map_lookup_elem() call for per-CPU hashmap */
2324 static int htab_percpu_map_gen_lookup(struct bpf_map *map, struct bpf_insn *insn_buf)
2325 {
2326         struct bpf_insn *insn = insn_buf;
2327
2328         if (!bpf_jit_supports_percpu_insn())
2329                 return -EOPNOTSUPP;
2330
2331         BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
2332                      (void *(*)(struct bpf_map *map, void *key))NULL));
2333         *insn++ = BPF_EMIT_CALL(__htab_map_lookup_elem);
2334         *insn++ = BPF_JMP_IMM(BPF_JEQ, BPF_REG_0, 0, 3);
2335         *insn++ = BPF_ALU64_IMM(BPF_ADD, BPF_REG_0,
2336                                 offsetof(struct htab_elem, key) + map->key_size);
2337         *insn++ = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_0, 0);
2338         *insn++ = BPF_MOV64_PERCPU_REG(BPF_REG_0, BPF_REG_0);
2339
2340         return insn - insn_buf;
2341 }
2342
2343 static void *htab_percpu_map_lookup_percpu_elem(struct bpf_map *map, void *key, u32 cpu)
2344 {
2345         struct htab_elem *l;
2346
2347         if (cpu >= nr_cpu_ids)
2348                 return NULL;
2349
2350         l = __htab_map_lookup_elem(map, key);
2351         if (l)
2352                 return per_cpu_ptr(htab_elem_get_ptr(l, map->key_size), cpu);
2353         else
2354                 return NULL;
2355 }
2356
2357 static void *htab_lru_percpu_map_lookup_elem(struct bpf_map *map, void *key)
2358 {
2359         struct htab_elem *l = __htab_map_lookup_elem(map, key);
2360
2361         if (l) {
2362                 bpf_lru_node_set_ref(&l->lru_node);
2363                 return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
2364         }
2365
2366         return NULL;
2367 }
2368
2369 static void *htab_lru_percpu_map_lookup_percpu_elem(struct bpf_map *map, void *key, u32 cpu)
2370 {
2371         struct htab_elem *l;
2372
2373         if (cpu >= nr_cpu_ids)
2374                 return NULL;
2375
2376         l = __htab_map_lookup_elem(map, key);
2377         if (l) {
2378                 bpf_lru_node_set_ref(&l->lru_node);
2379                 return per_cpu_ptr(htab_elem_get_ptr(l, map->key_size), cpu);
2380         }
2381
2382         return NULL;
2383 }
2384
2385 int bpf_percpu_hash_copy(struct bpf_map *map, void *key, void *value)
2386 {
2387         struct htab_elem *l;
2388         void __percpu *pptr;
2389         int ret = -ENOENT;
2390         int cpu, off = 0;
2391         u32 size;
2392
2393         /* per_cpu areas are zero-filled and bpf programs can only
2394          * access 'value_size' of them, so copying rounded areas
2395          * will not leak any kernel data
2396          */
2397         size = round_up(map->value_size, 8);
2398         rcu_read_lock();
2399         l = __htab_map_lookup_elem(map, key);
2400         if (!l)
2401                 goto out;
2402         /* We do not mark LRU map element here in order to not mess up
2403          * eviction heuristics when user space does a map walk.
2404          */
2405         pptr = htab_elem_get_ptr(l, map->key_size);
2406         for_each_possible_cpu(cpu) {
2407                 copy_map_value_long(map, value + off, per_cpu_ptr(pptr, cpu));
2408                 check_and_init_map_value(map, value + off);
2409                 off += size;
2410         }
2411         ret = 0;
2412 out:
2413         rcu_read_unlock();
2414         return ret;
2415 }
2416
2417 int bpf_percpu_hash_update(struct bpf_map *map, void *key, void *value,
2418                            u64 map_flags)
2419 {
2420         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
2421         int ret;
2422
2423         rcu_read_lock();
2424         if (htab_is_lru(htab))
2425                 ret = __htab_lru_percpu_map_update_elem(map, key, value,
2426                                                         map_flags, true);
2427         else
2428                 ret = __htab_percpu_map_update_elem(map, key, value, map_flags,
2429                                                     true);
2430         rcu_read_unlock();
2431
2432         return ret;
2433 }
2434
2435 static void htab_percpu_map_seq_show_elem(struct bpf_map *map, void *key,
2436                                           struct seq_file *m)
2437 {
2438         struct htab_elem *l;
2439         void __percpu *pptr;
2440         int cpu;
2441
2442         rcu_read_lock();
2443
2444         l = __htab_map_lookup_elem(map, key);
2445         if (!l) {
2446                 rcu_read_unlock();
2447                 return;
2448         }
2449
2450         btf_type_seq_show(map->btf, map->btf_key_type_id, key, m);
2451         seq_puts(m, ": {\n");
2452         pptr = htab_elem_get_ptr(l, map->key_size);
2453         for_each_possible_cpu(cpu) {
2454                 seq_printf(m, "\tcpu%d: ", cpu);
2455                 btf_type_seq_show(map->btf, map->btf_value_type_id,
2456                                   per_cpu_ptr(pptr, cpu), m);
2457                 seq_putc(m, '\n');
2458         }
2459         seq_puts(m, "}\n");
2460
2461         rcu_read_unlock();
2462 }
2463
2464 const struct bpf_map_ops htab_percpu_map_ops = {
2465         .map_meta_equal = bpf_map_meta_equal,
2466         .map_alloc_check = htab_map_alloc_check,
2467         .map_alloc = htab_map_alloc,
2468         .map_free = htab_map_free,
2469         .map_get_next_key = htab_map_get_next_key,
2470         .map_lookup_elem = htab_percpu_map_lookup_elem,
2471         .map_gen_lookup = htab_percpu_map_gen_lookup,
2472         .map_lookup_and_delete_elem = htab_percpu_map_lookup_and_delete_elem,
2473         .map_update_elem = htab_percpu_map_update_elem,
2474         .map_delete_elem = htab_map_delete_elem,
2475         .map_lookup_percpu_elem = htab_percpu_map_lookup_percpu_elem,
2476         .map_seq_show_elem = htab_percpu_map_seq_show_elem,
2477         .map_set_for_each_callback_args = map_set_for_each_callback_args,
2478         .map_for_each_callback = bpf_for_each_hash_elem,
2479         .map_mem_usage = htab_map_mem_usage,
2480         BATCH_OPS(htab_percpu),
2481         .map_btf_id = &htab_map_btf_ids[0],
2482         .iter_seq_info = &iter_seq_info,
2483 };
2484
2485 const struct bpf_map_ops htab_lru_percpu_map_ops = {
2486         .map_meta_equal = bpf_map_meta_equal,
2487         .map_alloc_check = htab_map_alloc_check,
2488         .map_alloc = htab_map_alloc,
2489         .map_free = htab_map_free,
2490         .map_get_next_key = htab_map_get_next_key,
2491         .map_lookup_elem = htab_lru_percpu_map_lookup_elem,
2492         .map_lookup_and_delete_elem = htab_lru_percpu_map_lookup_and_delete_elem,
2493         .map_update_elem = htab_lru_percpu_map_update_elem,
2494         .map_delete_elem = htab_lru_map_delete_elem,
2495         .map_lookup_percpu_elem = htab_lru_percpu_map_lookup_percpu_elem,
2496         .map_seq_show_elem = htab_percpu_map_seq_show_elem,
2497         .map_set_for_each_callback_args = map_set_for_each_callback_args,
2498         .map_for_each_callback = bpf_for_each_hash_elem,
2499         .map_mem_usage = htab_map_mem_usage,
2500         BATCH_OPS(htab_lru_percpu),
2501         .map_btf_id = &htab_map_btf_ids[0],
2502         .iter_seq_info = &iter_seq_info,
2503 };
2504
2505 static int fd_htab_map_alloc_check(union bpf_attr *attr)
2506 {
2507         if (attr->value_size != sizeof(u32))
2508                 return -EINVAL;
2509         return htab_map_alloc_check(attr);
2510 }
2511
2512 static void fd_htab_map_free(struct bpf_map *map)
2513 {
2514         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
2515         struct hlist_nulls_node *n;
2516         struct hlist_nulls_head *head;
2517         struct htab_elem *l;
2518         int i;
2519
2520         for (i = 0; i < htab->n_buckets; i++) {
2521                 head = select_bucket(htab, i);
2522
2523                 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
2524                         void *ptr = fd_htab_map_get_ptr(map, l);
2525
2526                         map->ops->map_fd_put_ptr(map, ptr, false);
2527                 }
2528         }
2529
2530         htab_map_free(map);
2531 }
2532
2533 /* only called from syscall */
2534 int bpf_fd_htab_map_lookup_elem(struct bpf_map *map, void *key, u32 *value)
2535 {
2536         void **ptr;
2537         int ret = 0;
2538
2539         if (!map->ops->map_fd_sys_lookup_elem)
2540                 return -ENOTSUPP;
2541
2542         rcu_read_lock();
2543         ptr = htab_map_lookup_elem(map, key);
2544         if (ptr)
2545                 *value = map->ops->map_fd_sys_lookup_elem(READ_ONCE(*ptr));
2546         else
2547                 ret = -ENOENT;
2548         rcu_read_unlock();
2549
2550         return ret;
2551 }
2552
2553 /* only called from syscall */
2554 int bpf_fd_htab_map_update_elem(struct bpf_map *map, struct file *map_file,
2555                                 void *key, void *value, u64 map_flags)
2556 {
2557         void *ptr;
2558         int ret;
2559         u32 ufd = *(u32 *)value;
2560
2561         ptr = map->ops->map_fd_get_ptr(map, map_file, ufd);
2562         if (IS_ERR(ptr))
2563                 return PTR_ERR(ptr);
2564
2565         /* The htab bucket lock is always held during update operations in fd
2566          * htab map, and the following rcu_read_lock() is only used to avoid
2567          * the WARN_ON_ONCE in htab_map_update_elem().
2568          */
2569         rcu_read_lock();
2570         ret = htab_map_update_elem(map, key, &ptr, map_flags);
2571         rcu_read_unlock();
2572         if (ret)
2573                 map->ops->map_fd_put_ptr(map, ptr, false);
2574
2575         return ret;
2576 }
2577
2578 static struct bpf_map *htab_of_map_alloc(union bpf_attr *attr)
2579 {
2580         struct bpf_map *map, *inner_map_meta;
2581
2582         inner_map_meta = bpf_map_meta_alloc(attr->inner_map_fd);
2583         if (IS_ERR(inner_map_meta))
2584                 return inner_map_meta;
2585
2586         map = htab_map_alloc(attr);
2587         if (IS_ERR(map)) {
2588                 bpf_map_meta_free(inner_map_meta);
2589                 return map;
2590         }
2591
2592         map->inner_map_meta = inner_map_meta;
2593
2594         return map;
2595 }
2596
2597 static void *htab_of_map_lookup_elem(struct bpf_map *map, void *key)
2598 {
2599         struct bpf_map **inner_map  = htab_map_lookup_elem(map, key);
2600
2601         if (!inner_map)
2602                 return NULL;
2603
2604         return READ_ONCE(*inner_map);
2605 }
2606
2607 static int htab_of_map_gen_lookup(struct bpf_map *map,
2608                                   struct bpf_insn *insn_buf)
2609 {
2610         struct bpf_insn *insn = insn_buf;
2611         const int ret = BPF_REG_0;
2612
2613         BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
2614                      (void *(*)(struct bpf_map *map, void *key))NULL));
2615         *insn++ = BPF_EMIT_CALL(__htab_map_lookup_elem);
2616         *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 2);
2617         *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
2618                                 offsetof(struct htab_elem, key) +
2619                                 round_up(map->key_size, 8));
2620         *insn++ = BPF_LDX_MEM(BPF_DW, ret, ret, 0);
2621
2622         return insn - insn_buf;
2623 }
2624
2625 static void htab_of_map_free(struct bpf_map *map)
2626 {
2627         bpf_map_meta_free(map->inner_map_meta);
2628         fd_htab_map_free(map);
2629 }
2630
2631 const struct bpf_map_ops htab_of_maps_map_ops = {
2632         .map_alloc_check = fd_htab_map_alloc_check,
2633         .map_alloc = htab_of_map_alloc,
2634         .map_free = htab_of_map_free,
2635         .map_get_next_key = htab_map_get_next_key,
2636         .map_lookup_elem = htab_of_map_lookup_elem,
2637         .map_delete_elem = htab_map_delete_elem,
2638         .map_fd_get_ptr = bpf_map_fd_get_ptr,
2639         .map_fd_put_ptr = bpf_map_fd_put_ptr,
2640         .map_fd_sys_lookup_elem = bpf_map_fd_sys_lookup_elem,
2641         .map_gen_lookup = htab_of_map_gen_lookup,
2642         .map_check_btf = map_check_no_btf,
2643         .map_mem_usage = htab_map_mem_usage,
2644         BATCH_OPS(htab),
2645         .map_btf_id = &htab_map_btf_ids[0],
2646 };