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