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