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