kernel/bpf/hashtab.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /* Copyright (c) 2011-2014 PLUMgrid, http://plumgrid.com
   3  * Copyright (c) 2016 Facebook
   4  */
   5 #include <linux/bpf.h>
   6 #include <linux/btf.h>
   7 #include <linux/jhash.h>
   8 #include <linux/filter.h>
   9 #include <linux/rculist_nulls.h>
  10 #include <linux/random.h>
  11 #include <uapi/linux/btf.h>
  12 #include <linux/rcupdate_trace.h>
  13 #include "percpu_freelist.h"
  14 #include "bpf_lru_list.h"
  15 #include "map_in_map.h"
  16
  17 #define HTAB_CREATE_FLAG_MASK                                           \
  18         (BPF_F_NO_PREALLOC | BPF_F_NO_COMMON_LRU | BPF_F_NUMA_NODE |    \
  19          BPF_F_ACCESS_MASK | BPF_F_ZERO_SEED)
  20
  21 #define BATCH_OPS(_name)                        \
  22         .map_lookup_batch =                     \
  23         _name##_map_lookup_batch,               \
  24         .map_lookup_and_delete_batch =          \
  25         _name##_map_lookup_and_delete_batch,    \
  26         .map_update_batch =                     \
  27         generic_map_update_batch,               \
  28         .map_delete_batch =                     \
  29         generic_map_delete_batch
  30
  31 /*
  32  * The bucket lock has two protection scopes:
  33  *
  34  * 1) Serializing concurrent operations from BPF programs on differrent
  35  *    CPUs
  36  *
  37  * 2) Serializing concurrent operations from BPF programs and sys_bpf()
  38  *
  39  * BPF programs can execute in any context including perf, kprobes and
  40  * tracing. As there are almost no limits where perf, kprobes and tracing
  41  * can be invoked from the lock operations need to be protected against
  42  * deadlocks. Deadlocks can be caused by recursion and by an invocation in
  43  * the lock held section when functions which acquire this lock are invoked
  44  * from sys_bpf(). BPF recursion is prevented by incrementing the per CPU
  45  * variable bpf_prog_active, which prevents BPF programs attached to perf
  46  * events, kprobes and tracing to be invoked before the prior invocation
  47  * from one of these contexts completed. sys_bpf() uses the same mechanism
  48  * by pinning the task to the current CPU and incrementing the recursion
  49  * protection accross the map operation.
  50  *
  51  * This has subtle implications on PREEMPT_RT. PREEMPT_RT forbids certain
  52  * operations like memory allocations (even with GFP_ATOMIC) from atomic
  53  * contexts. This is required because even with GFP_ATOMIC the memory
  54  * allocator calls into code pathes which acquire locks with long held lock
  55  * sections. To ensure the deterministic behaviour these locks are regular
  56  * spinlocks, which are converted to 'sleepable' spinlocks on RT. The only
  57  * true atomic contexts on an RT kernel are the low level hardware
  58  * handling, scheduling, low level interrupt handling, NMIs etc. None of
  59  * these contexts should ever do memory allocations.
  60  *
  61  * As regular device interrupt handlers and soft interrupts are forced into
  62  * thread context, the existing code which does
  63  *   spin_lock*(); alloc(GPF_ATOMIC); spin_unlock*();
  64  * just works.
  65  *
  66  * In theory the BPF locks could be converted to regular spinlocks as well,
  67  * but the bucket locks and percpu_freelist locks can be taken from
  68  * arbitrary contexts (perf, kprobes, tracepoints) which are required to be
  69  * atomic contexts even on RT. These mechanisms require preallocated maps,
  70  * so there is no need to invoke memory allocations within the lock held
  71  * sections.
  72  *
  73  * BPF maps which need dynamic allocation are only used from (forced)
  74  * thread context on RT and can therefore use regular spinlocks which in
  75  * turn allows to invoke memory allocations from the lock held section.
  76  *
  77  * On a non RT kernel this distinction is neither possible nor required.
  78  * spinlock maps to raw_spinlock and the extra code is optimized out by the
  79  * compiler.
  80  */
  81 struct bucket {
  82         struct hlist_nulls_head head;
  83         union {
  84                 raw_spinlock_t raw_lock;
  85                 spinlock_t     lock;
  86         };
  87 };
  88
  89 #define HASHTAB_MAP_LOCK_COUNT 8
  90 #define HASHTAB_MAP_LOCK_MASK (HASHTAB_MAP_LOCK_COUNT - 1)
  91
  92 struct bpf_htab {
  93         struct bpf_map map;
  94         struct bucket *buckets;
  95         void *elems;
  96         union {
  97                 struct pcpu_freelist freelist;
  98                 struct bpf_lru lru;
  99         };
 100         struct htab_elem *__percpu *extra_elems;
 101         atomic_t count; /* number of elements in this hashtable */
 102         u32 n_buckets;  /* number of hash buckets */
 103         u32 elem_size;  /* size of each element in bytes */
 104         u32 hashrnd;
 105         struct lock_class_key lockdep_key;
 106         int __percpu *map_locked[HASHTAB_MAP_LOCK_COUNT];
 107 };
 108
 109 /* each htab element is struct htab_elem + key + value */
 110 struct htab_elem {
 111         union {
 112                 struct hlist_nulls_node hash_node;
 113                 struct {
 114                         void *padding;
 115                         union {
 116                                 struct bpf_htab *htab;
 117                                 struct pcpu_freelist_node fnode;
 118                                 struct htab_elem *batch_flink;
 119                         };
 120                 };
 121         };
 122         union {
 123                 struct rcu_head rcu;
 124                 struct bpf_lru_node lru_node;
 125         };
 126         u32 hash;
 127         char key[] __aligned(8);
 128 };
 129
 130 static inline bool htab_is_prealloc(const struct bpf_htab *htab)
 131 {
 132         return !(htab->map.map_flags & BPF_F_NO_PREALLOC);
 133 }
 134
 135 static inline bool htab_use_raw_lock(const struct bpf_htab *htab)
 136 {
 137         return (!IS_ENABLED(CONFIG_PREEMPT_RT) || htab_is_prealloc(htab));
 138 }
 139
 140 static void htab_init_buckets(struct bpf_htab *htab)
 141 {
 142         unsigned i;
 143
 144         for (i = 0; i < htab->n_buckets; i++) {
 145                 INIT_HLIST_NULLS_HEAD(&htab->buckets[i].head, i);
 146                 if (htab_use_raw_lock(htab)) {
 147                         raw_spin_lock_init(&htab->buckets[i].raw_lock);
 148                         lockdep_set_class(&htab->buckets[i].raw_lock,
 149                                           &htab->lockdep_key);
 150                 } else {
 151                         spin_lock_init(&htab->buckets[i].lock);
 152                         lockdep_set_class(&htab->buckets[i].lock,
 153                                           &htab->lockdep_key);
 154                 }
 155         }
 156 }
 157
 158 static inline int htab_lock_bucket(const struct bpf_htab *htab,
 159                                    struct bucket *b, u32 hash,
 160                                    unsigned long *pflags)
 161 {
 162         unsigned long flags;
 163
 164         hash = hash & HASHTAB_MAP_LOCK_MASK;
 165
 166         migrate_disable();
 167         if (unlikely(__this_cpu_inc_return(*(htab->map_locked[hash])) != 1)) {
 168                 __this_cpu_dec(*(htab->map_locked[hash]));
 169                 migrate_enable();
 170                 return -EBUSY;
 171         }
 172
 173         if (htab_use_raw_lock(htab))
 174                 raw_spin_lock_irqsave(&b->raw_lock, flags);
 175         else
 176                 spin_lock_irqsave(&b->lock, flags);
 177         *pflags = flags;
 178
 179         return 0;
 180 }
 181
 182 static inline void htab_unlock_bucket(const struct bpf_htab *htab,
 183                                       struct bucket *b, u32 hash,
 184                                       unsigned long flags)
 185 {
 186         hash = hash & HASHTAB_MAP_LOCK_MASK;
 187         if (htab_use_raw_lock(htab))
 188                 raw_spin_unlock_irqrestore(&b->raw_lock, flags);
 189         else
 190                 spin_unlock_irqrestore(&b->lock, flags);
 191         __this_cpu_dec(*(htab->map_locked[hash]));
 192         migrate_enable();
 193 }
 194
 195 static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node);
 196
 197 static bool htab_is_lru(const struct bpf_htab *htab)
 198 {
 199         return htab->map.map_type == BPF_MAP_TYPE_LRU_HASH ||
 200                 htab->map.map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH;
 201 }
 202
 203 static bool htab_is_percpu(const struct bpf_htab *htab)
 204 {
 205         return htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH ||
 206                 htab->map.map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH;
 207 }
 208
 209 static inline void htab_elem_set_ptr(struct htab_elem *l, u32 key_size,
 210                                      void __percpu *pptr)
 211 {
 212         *(void __percpu **)(l->key + key_size) = pptr;
 213 }
 214
 215 static inline void __percpu *htab_elem_get_ptr(struct htab_elem *l, u32 key_size)
 216 {
 217         return *(void __percpu **)(l->key + key_size);
 218 }
 219
 220 static void *fd_htab_map_get_ptr(const struct bpf_map *map, struct htab_elem *l)
 221 {
 222         return *(void **)(l->key + roundup(map->key_size, 8));
 223 }
 224
 225 static struct htab_elem *get_htab_elem(struct bpf_htab *htab, int i)
 226 {
 227         return (struct htab_elem *) (htab->elems + i * htab->elem_size);
 228 }
 229
 230 static void htab_free_elems(struct bpf_htab *htab)
 231 {
 232         int i;
 233
 234         if (!htab_is_percpu(htab))
 235                 goto free_elems;
 236
 237         for (i = 0; i < htab->map.max_entries; i++) {
 238                 void __percpu *pptr;
 239
 240                 pptr = htab_elem_get_ptr(get_htab_elem(htab, i),
 241                                          htab->map.key_size);
 242                 free_percpu(pptr);
 243                 cond_resched();
 244         }
 245 free_elems:
 246         bpf_map_area_free(htab->elems);
 247 }
 248
 249 /* The LRU list has a lock (lru_lock). Each htab bucket has a lock
 250  * (bucket_lock). If both locks need to be acquired together, the lock
 251  * order is always lru_lock -> bucket_lock and this only happens in
 252  * bpf_lru_list.c logic. For example, certain code path of
 253  * bpf_lru_pop_free(), which is called by function prealloc_lru_pop(),
 254  * will acquire lru_lock first followed by acquiring bucket_lock.
 255  *
 256  * In hashtab.c, to avoid deadlock, lock acquisition of
 257  * bucket_lock followed by lru_lock is not allowed. In such cases,
 258  * bucket_lock needs to be released first before acquiring lru_lock.
 259  */
 260 static struct htab_elem *prealloc_lru_pop(struct bpf_htab *htab, void *key,
 261                                           u32 hash)
 262 {
 263         struct bpf_lru_node *node = bpf_lru_pop_free(&htab->lru, hash);
 264         struct htab_elem *l;
 265
 266         if (node) {
 267                 l = container_of(node, struct htab_elem, lru_node);
 268                 memcpy(l->key, key, htab->map.key_size);
 269                 return l;
 270         }
 271
 272         return NULL;
 273 }
 274
 275 static int prealloc_init(struct bpf_htab *htab)
 276 {
 277         u32 num_entries = htab->map.max_entries;
 278         int err = -ENOMEM, i;
 279
 280         if (!htab_is_percpu(htab) && !htab_is_lru(htab))
 281                 num_entries += num_possible_cpus();
 282
 283         htab->elems = bpf_map_area_alloc(htab->elem_size * num_entries,
 284                                          htab->map.numa_node);
 285         if (!htab->elems)
 286                 return -ENOMEM;
 287
 288         if (!htab_is_percpu(htab))
 289                 goto skip_percpu_elems;
 290
 291         for (i = 0; i < num_entries; i++) {
 292                 u32 size = round_up(htab->map.value_size, 8);
 293                 void __percpu *pptr;
 294
 295                 pptr = __alloc_percpu_gfp(size, 8, GFP_USER | __GFP_NOWARN);
 296                 if (!pptr)
 297                         goto free_elems;
 298                 htab_elem_set_ptr(get_htab_elem(htab, i), htab->map.key_size,
 299                                   pptr);
 300                 cond_resched();
 301         }
 302
 303 skip_percpu_elems:
 304         if (htab_is_lru(htab))
 305                 err = bpf_lru_init(&htab->lru,
 306                                    htab->map.map_flags & BPF_F_NO_COMMON_LRU,
 307                                    offsetof(struct htab_elem, hash) -
 308                                    offsetof(struct htab_elem, lru_node),
 309                                    htab_lru_map_delete_node,
 310                                    htab);
 311         else
 312                 err = pcpu_freelist_init(&htab->freelist);
 313
 314         if (err)
 315                 goto free_elems;
 316
 317         if (htab_is_lru(htab))
 318                 bpf_lru_populate(&htab->lru, htab->elems,
 319                                  offsetof(struct htab_elem, lru_node),
 320                                  htab->elem_size, num_entries);
 321         else
 322                 pcpu_freelist_populate(&htab->freelist,
 323                                        htab->elems + offsetof(struct htab_elem, fnode),
 324                                        htab->elem_size, num_entries);
 325
 326         return 0;
 327
 328 free_elems:
 329         htab_free_elems(htab);
 330         return err;
 331 }
 332
 333 static void prealloc_destroy(struct bpf_htab *htab)
 334 {
 335         htab_free_elems(htab);
 336
 337         if (htab_is_lru(htab))
 338                 bpf_lru_destroy(&htab->lru);
 339         else
 340                 pcpu_freelist_destroy(&htab->freelist);
 341 }
 342
 343 static int alloc_extra_elems(struct bpf_htab *htab)
 344 {
 345         struct htab_elem *__percpu *pptr, *l_new;
 346         struct pcpu_freelist_node *l;
 347         int cpu;
 348
 349         pptr = __alloc_percpu_gfp(sizeof(struct htab_elem *), 8,
 350                                   GFP_USER | __GFP_NOWARN);
 351         if (!pptr)
 352                 return -ENOMEM;
 353
 354         for_each_possible_cpu(cpu) {
 355                 l = pcpu_freelist_pop(&htab->freelist);
 356                 /* pop will succeed, since prealloc_init()
 357                  * preallocated extra num_possible_cpus elements
 358                  */
 359                 l_new = container_of(l, struct htab_elem, fnode);
 360                 *per_cpu_ptr(pptr, cpu) = l_new;
 361         }
 362         htab->extra_elems = pptr;
 363         return 0;
 364 }
 365
 366 /* Called from syscall */
 367 static int htab_map_alloc_check(union bpf_attr *attr)
 368 {
 369         bool percpu = (attr->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
 370                        attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
 371         bool lru = (attr->map_type == BPF_MAP_TYPE_LRU_HASH ||
 372                     attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
 373         /* percpu_lru means each cpu has its own LRU list.
 374          * it is different from BPF_MAP_TYPE_PERCPU_HASH where
 375          * the map's value itself is percpu.  percpu_lru has
 376          * nothing to do with the map's value.
 377          */
 378         bool percpu_lru = (attr->map_flags & BPF_F_NO_COMMON_LRU);
 379         bool prealloc = !(attr->map_flags & BPF_F_NO_PREALLOC);
 380         bool zero_seed = (attr->map_flags & BPF_F_ZERO_SEED);
 381         int numa_node = bpf_map_attr_numa_node(attr);
 382
 383         BUILD_BUG_ON(offsetof(struct htab_elem, htab) !=
 384                      offsetof(struct htab_elem, hash_node.pprev));
 385         BUILD_BUG_ON(offsetof(struct htab_elem, fnode.next) !=
 386                      offsetof(struct htab_elem, hash_node.pprev));
 387
 388         if (lru && !bpf_capable())
 389                 /* LRU implementation is much complicated than other
 390                  * maps.  Hence, limit to CAP_BPF.
 391                  */
 392                 return -EPERM;
 393
 394         if (zero_seed && !capable(CAP_SYS_ADMIN))
 395                 /* Guard against local DoS, and discourage production use. */
 396                 return -EPERM;
 397
 398         if (attr->map_flags & ~HTAB_CREATE_FLAG_MASK ||
 399             !bpf_map_flags_access_ok(attr->map_flags))
 400                 return -EINVAL;
 401
 402         if (!lru && percpu_lru)
 403                 return -EINVAL;
 404
 405         if (lru && !prealloc)
 406                 return -ENOTSUPP;
 407
 408         if (numa_node != NUMA_NO_NODE && (percpu || percpu_lru))
 409                 return -EINVAL;
 410
 411         /* check sanity of attributes.
 412          * value_size == 0 may be allowed in the future to use map as a set
 413          */
 414         if (attr->max_entries == 0 || attr->key_size == 0 ||
 415             attr->value_size == 0)
 416                 return -EINVAL;
 417
 418         if (attr->key_size > MAX_BPF_STACK)
 419                 /* eBPF programs initialize keys on stack, so they cannot be
 420                  * larger than max stack size
 421                  */
 422                 return -E2BIG;
 423
 424         if (attr->value_size >= KMALLOC_MAX_SIZE -
 425             MAX_BPF_STACK - sizeof(struct htab_elem))
 426                 /* if value_size is bigger, the user space won't be able to
 427                  * access the elements via bpf syscall. This check also makes
 428                  * sure that the elem_size doesn't overflow and it's
 429                  * kmalloc-able later in htab_map_update_elem()
 430                  */
 431                 return -E2BIG;
 432
 433         return 0;
 434 }
 435
 436 static struct bpf_map *htab_map_alloc(union bpf_attr *attr)
 437 {
 438         bool percpu = (attr->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
 439                        attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
 440         bool lru = (attr->map_type == BPF_MAP_TYPE_LRU_HASH ||
 441                     attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
 442         /* percpu_lru means each cpu has its own LRU list.
 443          * it is different from BPF_MAP_TYPE_PERCPU_HASH where
 444          * the map's value itself is percpu.  percpu_lru has
 445          * nothing to do with the map's value.
 446          */
 447         bool percpu_lru = (attr->map_flags & BPF_F_NO_COMMON_LRU);
 448         bool prealloc = !(attr->map_flags & BPF_F_NO_PREALLOC);
 449         struct bpf_htab *htab;
 450         int err, i;
 451         u64 cost;
 452
 453         htab = kzalloc(sizeof(*htab), GFP_USER);
 454         if (!htab)
 455                 return ERR_PTR(-ENOMEM);
 456
 457         lockdep_register_key(&htab->lockdep_key);
 458
 459         bpf_map_init_from_attr(&htab->map, attr);
 460
 461         if (percpu_lru) {
 462                 /* ensure each CPU's lru list has >=1 elements.
 463                  * since we are at it, make each lru list has the same
 464                  * number of elements.
 465                  */
 466                 htab->map.max_entries = roundup(attr->max_entries,
 467                                                 num_possible_cpus());
 468                 if (htab->map.max_entries < attr->max_entries)
 469                         htab->map.max_entries = rounddown(attr->max_entries,
 470                                                           num_possible_cpus());
 471         }
 472
 473         /* hash table size must be power of 2 */
 474         htab->n_buckets = roundup_pow_of_two(htab->map.max_entries);
 475
 476         htab->elem_size = sizeof(struct htab_elem) +
 477                           round_up(htab->map.key_size, 8);
 478         if (percpu)
 479                 htab->elem_size += sizeof(void *);
 480         else
 481                 htab->elem_size += round_up(htab->map.value_size, 8);
 482
 483         err = -E2BIG;
 484         /* prevent zero size kmalloc and check for u32 overflow */
 485         if (htab->n_buckets == 0 ||
 486             htab->n_buckets > U32_MAX / sizeof(struct bucket))
 487                 goto free_htab;
 488
 489         cost = (u64) htab->n_buckets * sizeof(struct bucket) +
 490                (u64) htab->elem_size * htab->map.max_entries;
 491
 492         if (percpu)
 493                 cost += (u64) round_up(htab->map.value_size, 8) *
 494                         num_possible_cpus() * htab->map.max_entries;
 495         else
 496                cost += (u64) htab->elem_size * num_possible_cpus();
 497
 498         /* if map size is larger than memlock limit, reject it */
 499         err = bpf_map_charge_init(&htab->map.memory, cost);
 500         if (err)
 501                 goto free_htab;
 502
 503         err = -ENOMEM;
 504         htab->buckets = bpf_map_area_alloc(htab->n_buckets *
 505                                            sizeof(struct bucket),
 506                                            htab->map.numa_node);
 507         if (!htab->buckets)
 508                 goto free_charge;
 509
 510         for (i = 0; i < HASHTAB_MAP_LOCK_COUNT; i++) {
 511                 htab->map_locked[i] = __alloc_percpu_gfp(sizeof(int),
 512                                                          sizeof(int), GFP_USER);
 513                 if (!htab->map_locked[i])
 514                         goto free_map_locked;
 515         }
 516
 517         if (htab->map.map_flags & BPF_F_ZERO_SEED)
 518                 htab->hashrnd = 0;
 519         else
 520                 htab->hashrnd = get_random_int();
 521
 522         htab_init_buckets(htab);
 523
 524         if (prealloc) {
 525                 err = prealloc_init(htab);
 526                 if (err)
 527                         goto free_map_locked;
 528
 529                 if (!percpu && !lru) {
 530                         /* lru itself can remove the least used element, so
 531                          * there is no need for an extra elem during map_update.
 532                          */
 533                         err = alloc_extra_elems(htab);
 534                         if (err)
 535                                 goto free_prealloc;
 536                 }
 537         }
 538
 539         return &htab->map;
 540
 541 free_prealloc:
 542         prealloc_destroy(htab);
 543 free_map_locked:
 544         for (i = 0; i < HASHTAB_MAP_LOCK_COUNT; i++)
 545                 free_percpu(htab->map_locked[i]);
 546         bpf_map_area_free(htab->buckets);
 547 free_charge:
 548         bpf_map_charge_finish(&htab->map.memory);
 549 free_htab:
 550         lockdep_unregister_key(&htab->lockdep_key);
 551         kfree(htab);
 552         return ERR_PTR(err);
 553 }
 554
 555 static inline u32 htab_map_hash(const void *key, u32 key_len, u32 hashrnd)
 556 {
 557         return jhash(key, key_len, hashrnd);
 558 }
 559
 560 static inline struct bucket *__select_bucket(struct bpf_htab *htab, u32 hash)
 561 {
 562         return &htab->buckets[hash & (htab->n_buckets - 1)];
 563 }
 564
 565 static inline struct hlist_nulls_head *select_bucket(struct bpf_htab *htab, u32 hash)
 566 {
 567         return &__select_bucket(htab, hash)->head;
 568 }
 569
 570 /* this lookup function can only be called with bucket lock taken */
 571 static struct htab_elem *lookup_elem_raw(struct hlist_nulls_head *head, u32 hash,
 572                                          void *key, u32 key_size)
 573 {
 574         struct hlist_nulls_node *n;
 575         struct htab_elem *l;
 576
 577         hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
 578                 if (l->hash == hash && !memcmp(&l->key, key, key_size))
 579                         return l;
 580
 581         return NULL;
 582 }
 583
 584 /* can be called without bucket lock. it will repeat the loop in
 585  * the unlikely event when elements moved from one bucket into another
 586  * while link list is being walked
 587  */
 588 static struct htab_elem *lookup_nulls_elem_raw(struct hlist_nulls_head *head,
 589                                                u32 hash, void *key,
 590                                                u32 key_size, u32 n_buckets)
 591 {
 592         struct hlist_nulls_node *n;
 593         struct htab_elem *l;
 594
 595 again:
 596         hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
 597                 if (l->hash == hash && !memcmp(&l->key, key, key_size))
 598                         return l;
 599
 600         if (unlikely(get_nulls_value(n) != (hash & (n_buckets - 1))))
 601                 goto again;
 602
 603         return NULL;
 604 }
 605
 606 /* Called from syscall or from eBPF program directly, so
 607  * arguments have to match bpf_map_lookup_elem() exactly.
 608  * The return value is adjusted by BPF instructions
 609  * in htab_map_gen_lookup().
 610  */
 611 static void *__htab_map_lookup_elem(struct bpf_map *map, void *key)
 612 {
 613         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 614         struct hlist_nulls_head *head;
 615         struct htab_elem *l;
 616         u32 hash, key_size;
 617
 618         WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held());
 619
 620         key_size = map->key_size;
 621
 622         hash = htab_map_hash(key, key_size, htab->hashrnd);
 623
 624         head = select_bucket(htab, hash);
 625
 626         l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
 627
 628         return l;
 629 }
 630
 631 static void *htab_map_lookup_elem(struct bpf_map *map, void *key)
 632 {
 633         struct htab_elem *l = __htab_map_lookup_elem(map, key);
 634
 635         if (l)
 636                 return l->key + round_up(map->key_size, 8);
 637
 638         return NULL;
 639 }
 640
 641 /* inline bpf_map_lookup_elem() call.
 642  * Instead of:
 643  * bpf_prog
 644  *   bpf_map_lookup_elem
 645  *     map->ops->map_lookup_elem
 646  *       htab_map_lookup_elem
 647  *         __htab_map_lookup_elem
 648  * do:
 649  * bpf_prog
 650  *   __htab_map_lookup_elem
 651  */
 652 static int htab_map_gen_lookup(struct bpf_map *map, struct bpf_insn *insn_buf)
 653 {
 654         struct bpf_insn *insn = insn_buf;
 655         const int ret = BPF_REG_0;
 656
 657         BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
 658                      (void *(*)(struct bpf_map *map, void *key))NULL));
 659         *insn++ = BPF_EMIT_CALL(BPF_CAST_CALL(__htab_map_lookup_elem));
 660         *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 1);
 661         *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
 662                                 offsetof(struct htab_elem, key) +
 663                                 round_up(map->key_size, 8));
 664         return insn - insn_buf;
 665 }
 666
 667 static __always_inline void *__htab_lru_map_lookup_elem(struct bpf_map *map,
 668                                                         void *key, const bool mark)
 669 {
 670         struct htab_elem *l = __htab_map_lookup_elem(map, key);
 671
 672         if (l) {
 673                 if (mark)
 674                         bpf_lru_node_set_ref(&l->lru_node);
 675                 return l->key + round_up(map->key_size, 8);
 676         }
 677
 678         return NULL;
 679 }
 680
 681 static void *htab_lru_map_lookup_elem(struct bpf_map *map, void *key)
 682 {
 683         return __htab_lru_map_lookup_elem(map, key, true);
 684 }
 685
 686 static void *htab_lru_map_lookup_elem_sys(struct bpf_map *map, void *key)
 687 {
 688         return __htab_lru_map_lookup_elem(map, key, false);
 689 }
 690
 691 static int htab_lru_map_gen_lookup(struct bpf_map *map,
 692                                    struct bpf_insn *insn_buf)
 693 {
 694         struct bpf_insn *insn = insn_buf;
 695         const int ret = BPF_REG_0;
 696         const int ref_reg = BPF_REG_1;
 697
 698         BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
 699                      (void *(*)(struct bpf_map *map, void *key))NULL));
 700         *insn++ = BPF_EMIT_CALL(BPF_CAST_CALL(__htab_map_lookup_elem));
 701         *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 4);
 702         *insn++ = BPF_LDX_MEM(BPF_B, ref_reg, ret,
 703                               offsetof(struct htab_elem, lru_node) +
 704                               offsetof(struct bpf_lru_node, ref));
 705         *insn++ = BPF_JMP_IMM(BPF_JNE, ref_reg, 0, 1);
 706         *insn++ = BPF_ST_MEM(BPF_B, ret,
 707                              offsetof(struct htab_elem, lru_node) +
 708                              offsetof(struct bpf_lru_node, ref),
 709                              1);
 710         *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
 711                                 offsetof(struct htab_elem, key) +
 712                                 round_up(map->key_size, 8));
 713         return insn - insn_buf;
 714 }
 715
 716 /* It is called from the bpf_lru_list when the LRU needs to delete
 717  * older elements from the htab.
 718  */
 719 static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node)
 720 {
 721         struct bpf_htab *htab = (struct bpf_htab *)arg;
 722         struct htab_elem *l = NULL, *tgt_l;
 723         struct hlist_nulls_head *head;
 724         struct hlist_nulls_node *n;
 725         unsigned long flags;
 726         struct bucket *b;
 727         int ret;
 728
 729         tgt_l = container_of(node, struct htab_elem, lru_node);
 730         b = __select_bucket(htab, tgt_l->hash);
 731         head = &b->head;
 732
 733         ret = htab_lock_bucket(htab, b, tgt_l->hash, &flags);
 734         if (ret)
 735                 return false;
 736
 737         hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
 738                 if (l == tgt_l) {
 739                         hlist_nulls_del_rcu(&l->hash_node);
 740                         break;
 741                 }
 742
 743         htab_unlock_bucket(htab, b, tgt_l->hash, flags);
 744
 745         return l == tgt_l;
 746 }
 747
 748 /* Called from syscall */
 749 static int htab_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
 750 {
 751         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 752         struct hlist_nulls_head *head;
 753         struct htab_elem *l, *next_l;
 754         u32 hash, key_size;
 755         int i = 0;
 756
 757         WARN_ON_ONCE(!rcu_read_lock_held());
 758
 759         key_size = map->key_size;
 760
 761         if (!key)
 762                 goto find_first_elem;
 763
 764         hash = htab_map_hash(key, key_size, htab->hashrnd);
 765
 766         head = select_bucket(htab, hash);
 767
 768         /* lookup the key */
 769         l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
 770
 771         if (!l)
 772                 goto find_first_elem;
 773
 774         /* key was found, get next key in the same bucket */
 775         next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_next_rcu(&l->hash_node)),
 776                                   struct htab_elem, hash_node);
 777
 778         if (next_l) {
 779                 /* if next elem in this hash list is non-zero, just return it */
 780                 memcpy(next_key, next_l->key, key_size);
 781                 return 0;
 782         }
 783
 784         /* no more elements in this hash list, go to the next bucket */
 785         i = hash & (htab->n_buckets - 1);
 786         i++;
 787
 788 find_first_elem:
 789         /* iterate over buckets */
 790         for (; i < htab->n_buckets; i++) {
 791                 head = select_bucket(htab, i);
 792
 793                 /* pick first element in the bucket */
 794                 next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_first_rcu(head)),
 795                                           struct htab_elem, hash_node);
 796                 if (next_l) {
 797                         /* if it's not empty, just return it */
 798                         memcpy(next_key, next_l->key, key_size);
 799                         return 0;
 800                 }
 801         }
 802
 803         /* iterated over all buckets and all elements */
 804         return -ENOENT;
 805 }
 806
 807 static void htab_elem_free(struct bpf_htab *htab, struct htab_elem *l)
 808 {
 809         if (htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH)
 810                 free_percpu(htab_elem_get_ptr(l, htab->map.key_size));
 811         kfree(l);
 812 }
 813
 814 static void htab_elem_free_rcu(struct rcu_head *head)
 815 {
 816         struct htab_elem *l = container_of(head, struct htab_elem, rcu);
 817         struct bpf_htab *htab = l->htab;
 818
 819         htab_elem_free(htab, l);
 820 }
 821
 822 static void htab_put_fd_value(struct bpf_htab *htab, struct htab_elem *l)
 823 {
 824         struct bpf_map *map = &htab->map;
 825         void *ptr;
 826
 827         if (map->ops->map_fd_put_ptr) {
 828                 ptr = fd_htab_map_get_ptr(map, l);
 829                 map->ops->map_fd_put_ptr(ptr);
 830         }
 831 }
 832
 833 static void free_htab_elem(struct bpf_htab *htab, struct htab_elem *l)
 834 {
 835         htab_put_fd_value(htab, l);
 836
 837         if (htab_is_prealloc(htab)) {
 838                 __pcpu_freelist_push(&htab->freelist, &l->fnode);
 839         } else {
 840                 atomic_dec(&htab->count);
 841                 l->htab = htab;
 842                 call_rcu(&l->rcu, htab_elem_free_rcu);
 843         }
 844 }
 845
 846 static void pcpu_copy_value(struct bpf_htab *htab, void __percpu *pptr,
 847                             void *value, bool onallcpus)
 848 {
 849         if (!onallcpus) {
 850                 /* copy true value_size bytes */
 851                 memcpy(this_cpu_ptr(pptr), value, htab->map.value_size);
 852         } else {
 853                 u32 size = round_up(htab->map.value_size, 8);
 854                 int off = 0, cpu;
 855
 856                 for_each_possible_cpu(cpu) {
 857                         bpf_long_memcpy(per_cpu_ptr(pptr, cpu),
 858                                         value + off, size);
 859                         off += size;
 860                 }
 861         }
 862 }
 863
 864 static bool fd_htab_map_needs_adjust(const struct bpf_htab *htab)
 865 {
 866         return htab->map.map_type == BPF_MAP_TYPE_HASH_OF_MAPS &&
 867                BITS_PER_LONG == 64;
 868 }
 869
 870 static struct htab_elem *alloc_htab_elem(struct bpf_htab *htab, void *key,
 871                                          void *value, u32 key_size, u32 hash,
 872                                          bool percpu, bool onallcpus,
 873                                          struct htab_elem *old_elem)
 874 {
 875         u32 size = htab->map.value_size;
 876         bool prealloc = htab_is_prealloc(htab);
 877         struct htab_elem *l_new, **pl_new;
 878         void __percpu *pptr;
 879
 880         if (prealloc) {
 881                 if (old_elem) {
 882                         /* if we're updating the existing element,
 883                          * use per-cpu extra elems to avoid freelist_pop/push
 884                          */
 885                         pl_new = this_cpu_ptr(htab->extra_elems);
 886                         l_new = *pl_new;
 887                         htab_put_fd_value(htab, old_elem);
 888                         *pl_new = old_elem;
 889                 } else {
 890                         struct pcpu_freelist_node *l;
 891
 892                         l = __pcpu_freelist_pop(&htab->freelist);
 893                         if (!l)
 894                                 return ERR_PTR(-E2BIG);
 895                         l_new = container_of(l, struct htab_elem, fnode);
 896                 }
 897         } else {
 898                 if (atomic_inc_return(&htab->count) > htab->map.max_entries)
 899                         if (!old_elem) {
 900                                 /* when map is full and update() is replacing
 901                                  * old element, it's ok to allocate, since
 902                                  * old element will be freed immediately.
 903                                  * Otherwise return an error
 904                                  */
 905                                 l_new = ERR_PTR(-E2BIG);
 906                                 goto dec_count;
 907                         }
 908                 l_new = kmalloc_node(htab->elem_size, GFP_ATOMIC | __GFP_NOWARN,
 909                                      htab->map.numa_node);
 910                 if (!l_new) {
 911                         l_new = ERR_PTR(-ENOMEM);
 912                         goto dec_count;
 913                 }
 914                 check_and_init_map_lock(&htab->map,
 915                                         l_new->key + round_up(key_size, 8));
 916         }
 917
 918         memcpy(l_new->key, key, key_size);
 919         if (percpu) {
 920                 size = round_up(size, 8);
 921                 if (prealloc) {
 922                         pptr = htab_elem_get_ptr(l_new, key_size);
 923                 } else {
 924                         /* alloc_percpu zero-fills */
 925                         pptr = __alloc_percpu_gfp(size, 8,
 926                                                   GFP_ATOMIC | __GFP_NOWARN);
 927                         if (!pptr) {
 928                                 kfree(l_new);
 929                                 l_new = ERR_PTR(-ENOMEM);
 930                                 goto dec_count;
 931                         }
 932                 }
 933
 934                 pcpu_copy_value(htab, pptr, value, onallcpus);
 935
 936                 if (!prealloc)
 937                         htab_elem_set_ptr(l_new, key_size, pptr);
 938         } else if (fd_htab_map_needs_adjust(htab)) {
 939                 size = round_up(size, 8);
 940                 memcpy(l_new->key + round_up(key_size, 8), value, size);
 941         } else {
 942                 copy_map_value(&htab->map,
 943                                l_new->key + round_up(key_size, 8),
 944                                value);
 945         }
 946
 947         l_new->hash = hash;
 948         return l_new;
 949 dec_count:
 950         atomic_dec(&htab->count);
 951         return l_new;
 952 }
 953
 954 static int check_flags(struct bpf_htab *htab, struct htab_elem *l_old,
 955                        u64 map_flags)
 956 {
 957         if (l_old && (map_flags & ~BPF_F_LOCK) == BPF_NOEXIST)
 958                 /* elem already exists */
 959                 return -EEXIST;
 960
 961         if (!l_old && (map_flags & ~BPF_F_LOCK) == BPF_EXIST)
 962                 /* elem doesn't exist, cannot update it */
 963                 return -ENOENT;
 964
 965         return 0;
 966 }
 967
 968 /* Called from syscall or from eBPF program */
 969 static int htab_map_update_elem(struct bpf_map *map, void *key, void *value,
 970                                 u64 map_flags)
 971 {
 972         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 973         struct htab_elem *l_new = NULL, *l_old;
 974         struct hlist_nulls_head *head;
 975         unsigned long flags;
 976         struct bucket *b;
 977         u32 key_size, hash;
 978         int ret;
 979
 980         if (unlikely((map_flags & ~BPF_F_LOCK) > BPF_EXIST))
 981                 /* unknown flags */
 982                 return -EINVAL;
 983
 984         WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held());
 985
 986         key_size = map->key_size;
 987
 988         hash = htab_map_hash(key, key_size, htab->hashrnd);
 989
 990         b = __select_bucket(htab, hash);
 991         head = &b->head;
 992
 993         if (unlikely(map_flags & BPF_F_LOCK)) {
 994                 if (unlikely(!map_value_has_spin_lock(map)))
 995                         return -EINVAL;
 996                 /* find an element without taking the bucket lock */
 997                 l_old = lookup_nulls_elem_raw(head, hash, key, key_size,
 998                                               htab->n_buckets);
 999                 ret = check_flags(htab, l_old, map_flags);
1000                 if (ret)
1001                         return ret;
1002                 if (l_old) {
1003                         /* grab the element lock and update value in place */
1004                         copy_map_value_locked(map,
1005                                               l_old->key + round_up(key_size, 8),
1006                                               value, false);
1007                         return 0;
1008                 }
1009                 /* fall through, grab the bucket lock and lookup again.
1010                  * 99.9% chance that the element won't be found,
1011                  * but second lookup under lock has to be done.
1012                  */
1013         }
1014
1015         ret = htab_lock_bucket(htab, b, hash, &flags);
1016         if (ret)
1017                 return ret;
1018
1019         l_old = lookup_elem_raw(head, hash, key, key_size);
1020
1021         ret = check_flags(htab, l_old, map_flags);
1022         if (ret)
1023                 goto err;
1024
1025         if (unlikely(l_old && (map_flags & BPF_F_LOCK))) {
1026                 /* first lookup without the bucket lock didn't find the element,
1027                  * but second lookup with the bucket lock found it.
1028                  * This case is highly unlikely, but has to be dealt with:
1029                  * grab the element lock in addition to the bucket lock
1030                  * and update element in place
1031                  */
1032                 copy_map_value_locked(map,
1033                                       l_old->key + round_up(key_size, 8),
1034                                       value, false);
1035                 ret = 0;
1036                 goto err;
1037         }
1038
1039         l_new = alloc_htab_elem(htab, key, value, key_size, hash, false, false,
1040                                 l_old);
1041         if (IS_ERR(l_new)) {
1042                 /* all pre-allocated elements are in use or memory exhausted */
1043                 ret = PTR_ERR(l_new);
1044                 goto err;
1045         }
1046
1047         /* add new element to the head of the list, so that
1048          * concurrent search will find it before old elem
1049          */
1050         hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1051         if (l_old) {
1052                 hlist_nulls_del_rcu(&l_old->hash_node);
1053                 if (!htab_is_prealloc(htab))
1054                         free_htab_elem(htab, l_old);
1055         }
1056         ret = 0;
1057 err:
1058         htab_unlock_bucket(htab, b, hash, flags);
1059         return ret;
1060 }
1061
1062 static int htab_lru_map_update_elem(struct bpf_map *map, void *key, void *value,
1063                                     u64 map_flags)
1064 {
1065         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1066         struct htab_elem *l_new, *l_old = NULL;
1067         struct hlist_nulls_head *head;
1068         unsigned long flags;
1069         struct bucket *b;
1070         u32 key_size, hash;
1071         int ret;
1072
1073         if (unlikely(map_flags > BPF_EXIST))
1074                 /* unknown flags */
1075                 return -EINVAL;
1076
1077         WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held());
1078
1079         key_size = map->key_size;
1080
1081         hash = htab_map_hash(key, key_size, htab->hashrnd);
1082
1083         b = __select_bucket(htab, hash);
1084         head = &b->head;
1085
1086         /* For LRU, we need to alloc before taking bucket's
1087          * spinlock because getting free nodes from LRU may need
1088          * to remove older elements from htab and this removal
1089          * operation will need a bucket lock.
1090          */
1091         l_new = prealloc_lru_pop(htab, key, hash);
1092         if (!l_new)
1093                 return -ENOMEM;
1094         memcpy(l_new->key + round_up(map->key_size, 8), value, map->value_size);
1095
1096         ret = htab_lock_bucket(htab, b, hash, &flags);
1097         if (ret)
1098                 return ret;
1099
1100         l_old = lookup_elem_raw(head, hash, key, key_size);
1101
1102         ret = check_flags(htab, l_old, map_flags);
1103         if (ret)
1104                 goto err;
1105
1106         /* add new element to the head of the list, so that
1107          * concurrent search will find it before old elem
1108          */
1109         hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1110         if (l_old) {
1111                 bpf_lru_node_set_ref(&l_new->lru_node);
1112                 hlist_nulls_del_rcu(&l_old->hash_node);
1113         }
1114         ret = 0;
1115
1116 err:
1117         htab_unlock_bucket(htab, b, hash, flags);
1118
1119         if (ret)
1120                 bpf_lru_push_free(&htab->lru, &l_new->lru_node);
1121         else if (l_old)
1122                 bpf_lru_push_free(&htab->lru, &l_old->lru_node);
1123
1124         return ret;
1125 }
1126
1127 static int __htab_percpu_map_update_elem(struct bpf_map *map, void *key,
1128                                          void *value, u64 map_flags,
1129                                          bool onallcpus)
1130 {
1131         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1132         struct htab_elem *l_new = NULL, *l_old;
1133         struct hlist_nulls_head *head;
1134         unsigned long flags;
1135         struct bucket *b;
1136         u32 key_size, hash;
1137         int ret;
1138
1139         if (unlikely(map_flags > BPF_EXIST))
1140                 /* unknown flags */
1141                 return -EINVAL;
1142
1143         WARN_ON_ONCE(!rcu_read_lock_held());
1144
1145         key_size = map->key_size;
1146
1147         hash = htab_map_hash(key, key_size, htab->hashrnd);
1148
1149         b = __select_bucket(htab, hash);
1150         head = &b->head;
1151
1152         ret = htab_lock_bucket(htab, b, hash, &flags);
1153         if (ret)
1154                 return ret;
1155
1156         l_old = lookup_elem_raw(head, hash, key, key_size);
1157
1158         ret = check_flags(htab, l_old, map_flags);
1159         if (ret)
1160                 goto err;
1161
1162         if (l_old) {
1163                 /* per-cpu hash map can update value in-place */
1164                 pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
1165                                 value, onallcpus);
1166         } else {
1167                 l_new = alloc_htab_elem(htab, key, value, key_size,
1168                                         hash, true, onallcpus, NULL);
1169                 if (IS_ERR(l_new)) {
1170                         ret = PTR_ERR(l_new);
1171                         goto err;
1172                 }
1173                 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1174         }
1175         ret = 0;
1176 err:
1177         htab_unlock_bucket(htab, b, hash, flags);
1178         return ret;
1179 }
1180
1181 static int __htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
1182                                              void *value, u64 map_flags,
1183                                              bool onallcpus)
1184 {
1185         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1186         struct htab_elem *l_new = NULL, *l_old;
1187         struct hlist_nulls_head *head;
1188         unsigned long flags;
1189         struct bucket *b;
1190         u32 key_size, hash;
1191         int ret;
1192
1193         if (unlikely(map_flags > BPF_EXIST))
1194                 /* unknown flags */
1195                 return -EINVAL;
1196
1197         WARN_ON_ONCE(!rcu_read_lock_held());
1198
1199         key_size = map->key_size;
1200
1201         hash = htab_map_hash(key, key_size, htab->hashrnd);
1202
1203         b = __select_bucket(htab, hash);
1204         head = &b->head;
1205
1206         /* For LRU, we need to alloc before taking bucket's
1207          * spinlock because LRU's elem alloc may need
1208          * to remove older elem from htab and this removal
1209          * operation will need a bucket lock.
1210          */
1211         if (map_flags != BPF_EXIST) {
1212                 l_new = prealloc_lru_pop(htab, key, hash);
1213                 if (!l_new)
1214                         return -ENOMEM;
1215         }
1216
1217         ret = htab_lock_bucket(htab, b, hash, &flags);
1218         if (ret)
1219                 return ret;
1220
1221         l_old = lookup_elem_raw(head, hash, key, key_size);
1222
1223         ret = check_flags(htab, l_old, map_flags);
1224         if (ret)
1225                 goto err;
1226
1227         if (l_old) {
1228                 bpf_lru_node_set_ref(&l_old->lru_node);
1229
1230                 /* per-cpu hash map can update value in-place */
1231                 pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
1232                                 value, onallcpus);
1233         } else {
1234                 pcpu_copy_value(htab, htab_elem_get_ptr(l_new, key_size),
1235                                 value, onallcpus);
1236                 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1237                 l_new = NULL;
1238         }
1239         ret = 0;
1240 err:
1241         htab_unlock_bucket(htab, b, hash, flags);
1242         if (l_new)
1243                 bpf_lru_push_free(&htab->lru, &l_new->lru_node);
1244         return ret;
1245 }
1246
1247 static int htab_percpu_map_update_elem(struct bpf_map *map, void *key,
1248                                        void *value, u64 map_flags)
1249 {
1250         return __htab_percpu_map_update_elem(map, key, value, map_flags, false);
1251 }
1252
1253 static int htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
1254                                            void *value, u64 map_flags)
1255 {
1256         return __htab_lru_percpu_map_update_elem(map, key, value, map_flags,
1257                                                  false);
1258 }
1259
1260 /* Called from syscall or from eBPF program */
1261 static int htab_map_delete_elem(struct bpf_map *map, void *key)
1262 {
1263         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1264         struct hlist_nulls_head *head;
1265         struct bucket *b;
1266         struct htab_elem *l;
1267         unsigned long flags;
1268         u32 hash, key_size;
1269         int ret;
1270
1271         WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held());
1272
1273         key_size = map->key_size;
1274
1275         hash = htab_map_hash(key, key_size, htab->hashrnd);
1276         b = __select_bucket(htab, hash);
1277         head = &b->head;
1278
1279         ret = htab_lock_bucket(htab, b, hash, &flags);
1280         if (ret)
1281                 return ret;
1282
1283         l = lookup_elem_raw(head, hash, key, key_size);
1284
1285         if (l) {
1286                 hlist_nulls_del_rcu(&l->hash_node);
1287                 free_htab_elem(htab, l);
1288         } else {
1289                 ret = -ENOENT;
1290         }
1291
1292         htab_unlock_bucket(htab, b, hash, flags);
1293         return ret;
1294 }
1295
1296 static int htab_lru_map_delete_elem(struct bpf_map *map, void *key)
1297 {
1298         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1299         struct hlist_nulls_head *head;
1300         struct bucket *b;
1301         struct htab_elem *l;
1302         unsigned long flags;
1303         u32 hash, key_size;
1304         int ret;
1305
1306         WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held());
1307
1308         key_size = map->key_size;
1309
1310         hash = htab_map_hash(key, key_size, htab->hashrnd);
1311         b = __select_bucket(htab, hash);
1312         head = &b->head;
1313
1314         ret = htab_lock_bucket(htab, b, hash, &flags);
1315         if (ret)
1316                 return ret;
1317
1318         l = lookup_elem_raw(head, hash, key, key_size);
1319
1320         if (l)
1321                 hlist_nulls_del_rcu(&l->hash_node);
1322         else
1323                 ret = -ENOENT;
1324
1325         htab_unlock_bucket(htab, b, hash, flags);
1326         if (l)
1327                 bpf_lru_push_free(&htab->lru, &l->lru_node);
1328         return ret;
1329 }
1330
1331 static void delete_all_elements(struct bpf_htab *htab)
1332 {
1333         int i;
1334
1335         for (i = 0; i < htab->n_buckets; i++) {
1336                 struct hlist_nulls_head *head = select_bucket(htab, i);
1337                 struct hlist_nulls_node *n;
1338                 struct htab_elem *l;
1339
1340                 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1341                         hlist_nulls_del_rcu(&l->hash_node);
1342                         htab_elem_free(htab, l);
1343                 }
1344         }
1345 }
1346
1347 /* Called when map->refcnt goes to zero, either from workqueue or from syscall */
1348 static void htab_map_free(struct bpf_map *map)
1349 {
1350         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1351         int i;
1352
1353         /* bpf_free_used_maps() or close(map_fd) will trigger this map_free callback.
1354          * bpf_free_used_maps() is called after bpf prog is no longer executing.
1355          * There is no need to synchronize_rcu() here to protect map elements.
1356          */
1357
1358         /* some of free_htab_elem() callbacks for elements of this map may
1359          * not have executed. Wait for them.
1360          */
1361         rcu_barrier();
1362         if (!htab_is_prealloc(htab))
1363                 delete_all_elements(htab);
1364         else
1365                 prealloc_destroy(htab);
1366
1367         free_percpu(htab->extra_elems);
1368         bpf_map_area_free(htab->buckets);
1369         for (i = 0; i < HASHTAB_MAP_LOCK_COUNT; i++)
1370                 free_percpu(htab->map_locked[i]);
1371         lockdep_unregister_key(&htab->lockdep_key);
1372         kfree(htab);
1373 }
1374
1375 static void htab_map_seq_show_elem(struct bpf_map *map, void *key,
1376                                    struct seq_file *m)
1377 {
1378         void *value;
1379
1380         rcu_read_lock();
1381
1382         value = htab_map_lookup_elem(map, key);
1383         if (!value) {
1384                 rcu_read_unlock();
1385                 return;
1386         }
1387
1388         btf_type_seq_show(map->btf, map->btf_key_type_id, key, m);
1389         seq_puts(m, ": ");
1390         btf_type_seq_show(map->btf, map->btf_value_type_id, value, m);
1391         seq_puts(m, "\n");
1392
1393         rcu_read_unlock();
1394 }
1395
1396 static int
1397 __htab_map_lookup_and_delete_batch(struct bpf_map *map,
1398                                    const union bpf_attr *attr,
1399                                    union bpf_attr __user *uattr,
1400                                    bool do_delete, bool is_lru_map,
1401                                    bool is_percpu)
1402 {
1403         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1404         u32 bucket_cnt, total, key_size, value_size, roundup_key_size;
1405         void *keys = NULL, *values = NULL, *value, *dst_key, *dst_val;
1406         void __user *uvalues = u64_to_user_ptr(attr->batch.values);
1407         void __user *ukeys = u64_to_user_ptr(attr->batch.keys);
1408         void *ubatch = u64_to_user_ptr(attr->batch.in_batch);
1409         u32 batch, max_count, size, bucket_size;
1410         struct htab_elem *node_to_free = NULL;
1411         u64 elem_map_flags, map_flags;
1412         struct hlist_nulls_head *head;
1413         struct hlist_nulls_node *n;
1414         unsigned long flags = 0;
1415         bool locked = false;
1416         struct htab_elem *l;
1417         struct bucket *b;
1418         int ret = 0;
1419
1420         elem_map_flags = attr->batch.elem_flags;
1421         if ((elem_map_flags & ~BPF_F_LOCK) ||
1422             ((elem_map_flags & BPF_F_LOCK) && !map_value_has_spin_lock(map)))
1423                 return -EINVAL;
1424
1425         map_flags = attr->batch.flags;
1426         if (map_flags)
1427                 return -EINVAL;
1428
1429         max_count = attr->batch.count;
1430         if (!max_count)
1431                 return 0;
1432
1433         if (put_user(0, &uattr->batch.count))
1434                 return -EFAULT;
1435
1436         batch = 0;
1437         if (ubatch && copy_from_user(&batch, ubatch, sizeof(batch)))
1438                 return -EFAULT;
1439
1440         if (batch >= htab->n_buckets)
1441                 return -ENOENT;
1442
1443         key_size = htab->map.key_size;
1444         roundup_key_size = round_up(htab->map.key_size, 8);
1445         value_size = htab->map.value_size;
1446         size = round_up(value_size, 8);
1447         if (is_percpu)
1448                 value_size = size * num_possible_cpus();
1449         total = 0;
1450         /* while experimenting with hash tables with sizes ranging from 10 to
1451          * 1000, it was observed that a bucket can have upto 5 entries.
1452          */
1453         bucket_size = 5;
1454
1455 alloc:
1456         /* We cannot do copy_from_user or copy_to_user inside
1457          * the rcu_read_lock. Allocate enough space here.
1458          */
1459         keys = kvmalloc(key_size * bucket_size, GFP_USER | __GFP_NOWARN);
1460         values = kvmalloc(value_size * bucket_size, GFP_USER | __GFP_NOWARN);
1461         if (!keys || !values) {
1462                 ret = -ENOMEM;
1463                 goto after_loop;
1464         }
1465
1466 again:
1467         bpf_disable_instrumentation();
1468         rcu_read_lock();
1469 again_nocopy:
1470         dst_key = keys;
1471         dst_val = values;
1472         b = &htab->buckets[batch];
1473         head = &b->head;
1474         /* do not grab the lock unless need it (bucket_cnt > 0). */
1475         if (locked) {
1476                 ret = htab_lock_bucket(htab, b, batch, &flags);
1477                 if (ret)
1478                         goto next_batch;
1479         }
1480
1481         bucket_cnt = 0;
1482         hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
1483                 bucket_cnt++;
1484
1485         if (bucket_cnt && !locked) {
1486                 locked = true;
1487                 goto again_nocopy;
1488         }
1489
1490         if (bucket_cnt > (max_count - total)) {
1491                 if (total == 0)
1492                         ret = -ENOSPC;
1493                 /* Note that since bucket_cnt > 0 here, it is implicit
1494                  * that the locked was grabbed, so release it.
1495                  */
1496                 htab_unlock_bucket(htab, b, batch, flags);
1497                 rcu_read_unlock();
1498                 bpf_enable_instrumentation();
1499                 goto after_loop;
1500         }
1501
1502         if (bucket_cnt > bucket_size) {
1503                 bucket_size = bucket_cnt;
1504                 /* Note that since bucket_cnt > 0 here, it is implicit
1505                  * that the locked was grabbed, so release it.
1506                  */
1507                 htab_unlock_bucket(htab, b, batch, flags);
1508                 rcu_read_unlock();
1509                 bpf_enable_instrumentation();
1510                 kvfree(keys);
1511                 kvfree(values);
1512                 goto alloc;
1513         }
1514
1515         /* Next block is only safe to run if you have grabbed the lock */
1516         if (!locked)
1517                 goto next_batch;
1518
1519         hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1520                 memcpy(dst_key, l->key, key_size);
1521
1522                 if (is_percpu) {
1523                         int off = 0, cpu;
1524                         void __percpu *pptr;
1525
1526                         pptr = htab_elem_get_ptr(l, map->key_size);
1527                         for_each_possible_cpu(cpu) {
1528                                 bpf_long_memcpy(dst_val + off,
1529                                                 per_cpu_ptr(pptr, cpu), size);
1530                                 off += size;
1531                         }
1532                 } else {
1533                         value = l->key + roundup_key_size;
1534                         if (elem_map_flags & BPF_F_LOCK)
1535                                 copy_map_value_locked(map, dst_val, value,
1536                                                       true);
1537                         else
1538                                 copy_map_value(map, dst_val, value);
1539                         check_and_init_map_lock(map, dst_val);
1540                 }
1541                 if (do_delete) {
1542                         hlist_nulls_del_rcu(&l->hash_node);
1543
1544                         /* bpf_lru_push_free() will acquire lru_lock, which
1545                          * may cause deadlock. See comments in function
1546                          * prealloc_lru_pop(). Let us do bpf_lru_push_free()
1547                          * after releasing the bucket lock.
1548                          */
1549                         if (is_lru_map) {
1550                                 l->batch_flink = node_to_free;
1551                                 node_to_free = l;
1552                         } else {
1553                                 free_htab_elem(htab, l);
1554                         }
1555                 }
1556                 dst_key += key_size;
1557                 dst_val += value_size;
1558         }
1559
1560         htab_unlock_bucket(htab, b, batch, flags);
1561         locked = false;
1562
1563         while (node_to_free) {
1564                 l = node_to_free;
1565                 node_to_free = node_to_free->batch_flink;
1566                 bpf_lru_push_free(&htab->lru, &l->lru_node);
1567         }
1568
1569 next_batch:
1570         /* If we are not copying data, we can go to next bucket and avoid
1571          * unlocking the rcu.
1572          */
1573         if (!bucket_cnt && (batch + 1 < htab->n_buckets)) {
1574                 batch++;
1575                 goto again_nocopy;
1576         }
1577
1578         rcu_read_unlock();
1579         bpf_enable_instrumentation();
1580         if (bucket_cnt && (copy_to_user(ukeys + total * key_size, keys,
1581             key_size * bucket_cnt) ||
1582             copy_to_user(uvalues + total * value_size, values,
1583             value_size * bucket_cnt))) {
1584                 ret = -EFAULT;
1585                 goto after_loop;
1586         }
1587
1588         total += bucket_cnt;
1589         batch++;
1590         if (batch >= htab->n_buckets) {
1591                 ret = -ENOENT;
1592                 goto after_loop;
1593         }
1594         goto again;
1595
1596 after_loop:
1597         if (ret == -EFAULT)
1598                 goto out;
1599
1600         /* copy # of entries and next batch */
1601         ubatch = u64_to_user_ptr(attr->batch.out_batch);
1602         if (copy_to_user(ubatch, &batch, sizeof(batch)) ||
1603             put_user(total, &uattr->batch.count))
1604                 ret = -EFAULT;
1605
1606 out:
1607         kvfree(keys);
1608         kvfree(values);
1609         return ret;
1610 }
1611
1612 static int
1613 htab_percpu_map_lookup_batch(struct bpf_map *map, const union bpf_attr *attr,
1614                              union bpf_attr __user *uattr)
1615 {
1616         return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1617                                                   false, true);
1618 }
1619
1620 static int
1621 htab_percpu_map_lookup_and_delete_batch(struct bpf_map *map,
1622                                         const union bpf_attr *attr,
1623                                         union bpf_attr __user *uattr)
1624 {
1625         return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1626                                                   false, true);
1627 }
1628
1629 static int
1630 htab_map_lookup_batch(struct bpf_map *map, const union bpf_attr *attr,
1631                       union bpf_attr __user *uattr)
1632 {
1633         return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1634                                                   false, false);
1635 }
1636
1637 static int
1638 htab_map_lookup_and_delete_batch(struct bpf_map *map,
1639                                  const union bpf_attr *attr,
1640                                  union bpf_attr __user *uattr)
1641 {
1642         return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1643                                                   false, false);
1644 }
1645
1646 static int
1647 htab_lru_percpu_map_lookup_batch(struct bpf_map *map,
1648                                  const union bpf_attr *attr,
1649                                  union bpf_attr __user *uattr)
1650 {
1651         return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1652                                                   true, true);
1653 }
1654
1655 static int
1656 htab_lru_percpu_map_lookup_and_delete_batch(struct bpf_map *map,
1657                                             const union bpf_attr *attr,
1658                                             union bpf_attr __user *uattr)
1659 {
1660         return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1661                                                   true, true);
1662 }
1663
1664 static int
1665 htab_lru_map_lookup_batch(struct bpf_map *map, const union bpf_attr *attr,
1666                           union bpf_attr __user *uattr)
1667 {
1668         return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1669                                                   true, false);
1670 }
1671
1672 static int
1673 htab_lru_map_lookup_and_delete_batch(struct bpf_map *map,
1674                                      const union bpf_attr *attr,
1675                                      union bpf_attr __user *uattr)
1676 {
1677         return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1678                                                   true, false);
1679 }
1680
1681 struct bpf_iter_seq_hash_map_info {
1682         struct bpf_map *map;
1683         struct bpf_htab *htab;
1684         void *percpu_value_buf; // non-zero means percpu hash
1685         u32 bucket_id;
1686         u32 skip_elems;
1687 };
1688
1689 static struct htab_elem *
1690 bpf_hash_map_seq_find_next(struct bpf_iter_seq_hash_map_info *info,
1691                            struct htab_elem *prev_elem)
1692 {
1693         const struct bpf_htab *htab = info->htab;
1694         u32 skip_elems = info->skip_elems;
1695         u32 bucket_id = info->bucket_id;
1696         struct hlist_nulls_head *head;
1697         struct hlist_nulls_node *n;
1698         struct htab_elem *elem;
1699         struct bucket *b;
1700         u32 i, count;
1701
1702         if (bucket_id >= htab->n_buckets)
1703                 return NULL;
1704
1705         /* try to find next elem in the same bucket */
1706         if (prev_elem) {
1707                 /* no update/deletion on this bucket, prev_elem should be still valid
1708                  * and we won't skip elements.
1709                  */
1710                 n = rcu_dereference_raw(hlist_nulls_next_rcu(&prev_elem->hash_node));
1711                 elem = hlist_nulls_entry_safe(n, struct htab_elem, hash_node);
1712                 if (elem)
1713                         return elem;
1714
1715                 /* not found, unlock and go to the next bucket */
1716                 b = &htab->buckets[bucket_id++];
1717                 rcu_read_unlock();
1718                 skip_elems = 0;
1719         }
1720
1721         for (i = bucket_id; i < htab->n_buckets; i++) {
1722                 b = &htab->buckets[i];
1723                 rcu_read_lock();
1724
1725                 count = 0;
1726                 head = &b->head;
1727                 hlist_nulls_for_each_entry_rcu(elem, n, head, hash_node) {
1728                         if (count >= skip_elems) {
1729                                 info->bucket_id = i;
1730                                 info->skip_elems = count;
1731                                 return elem;
1732                         }
1733                         count++;
1734                 }
1735
1736                 rcu_read_unlock();
1737                 skip_elems = 0;
1738         }
1739
1740         info->bucket_id = i;
1741         info->skip_elems = 0;
1742         return NULL;
1743 }
1744
1745 static void *bpf_hash_map_seq_start(struct seq_file *seq, loff_t *pos)
1746 {
1747         struct bpf_iter_seq_hash_map_info *info = seq->private;
1748         struct htab_elem *elem;
1749
1750         elem = bpf_hash_map_seq_find_next(info, NULL);
1751         if (!elem)
1752                 return NULL;
1753
1754         if (*pos == 0)
1755                 ++*pos;
1756         return elem;
1757 }
1758
1759 static void *bpf_hash_map_seq_next(struct seq_file *seq, void *v, loff_t *pos)
1760 {
1761         struct bpf_iter_seq_hash_map_info *info = seq->private;
1762
1763         ++*pos;
1764         ++info->skip_elems;
1765         return bpf_hash_map_seq_find_next(info, v);
1766 }
1767
1768 static int __bpf_hash_map_seq_show(struct seq_file *seq, struct htab_elem *elem)
1769 {
1770         struct bpf_iter_seq_hash_map_info *info = seq->private;
1771         u32 roundup_key_size, roundup_value_size;
1772         struct bpf_iter__bpf_map_elem ctx = {};
1773         struct bpf_map *map = info->map;
1774         struct bpf_iter_meta meta;
1775         int ret = 0, off = 0, cpu;
1776         struct bpf_prog *prog;
1777         void __percpu *pptr;
1778
1779         meta.seq = seq;
1780         prog = bpf_iter_get_info(&meta, elem == NULL);
1781         if (prog) {
1782                 ctx.meta = &meta;
1783                 ctx.map = info->map;
1784                 if (elem) {
1785                         roundup_key_size = round_up(map->key_size, 8);
1786                         ctx.key = elem->key;
1787                         if (!info->percpu_value_buf) {
1788                                 ctx.value = elem->key + roundup_key_size;
1789                         } else {
1790                                 roundup_value_size = round_up(map->value_size, 8);
1791                                 pptr = htab_elem_get_ptr(elem, map->key_size);
1792                                 for_each_possible_cpu(cpu) {
1793                                         bpf_long_memcpy(info->percpu_value_buf + off,
1794                                                         per_cpu_ptr(pptr, cpu),
1795                                                         roundup_value_size);
1796                                         off += roundup_value_size;
1797                                 }
1798                                 ctx.value = info->percpu_value_buf;
1799                         }
1800                 }
1801                 ret = bpf_iter_run_prog(prog, &ctx);
1802         }
1803
1804         return ret;
1805 }
1806
1807 static int bpf_hash_map_seq_show(struct seq_file *seq, void *v)
1808 {
1809         return __bpf_hash_map_seq_show(seq, v);
1810 }
1811
1812 static void bpf_hash_map_seq_stop(struct seq_file *seq, void *v)
1813 {
1814         if (!v)
1815                 (void)__bpf_hash_map_seq_show(seq, NULL);
1816         else
1817                 rcu_read_unlock();
1818 }
1819
1820 static int bpf_iter_init_hash_map(void *priv_data,
1821                                   struct bpf_iter_aux_info *aux)
1822 {
1823         struct bpf_iter_seq_hash_map_info *seq_info = priv_data;
1824         struct bpf_map *map = aux->map;
1825         void *value_buf;
1826         u32 buf_size;
1827
1828         if (map->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
1829             map->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH) {
1830                 buf_size = round_up(map->value_size, 8) * num_possible_cpus();
1831                 value_buf = kmalloc(buf_size, GFP_USER | __GFP_NOWARN);
1832                 if (!value_buf)
1833                         return -ENOMEM;
1834
1835                 seq_info->percpu_value_buf = value_buf;
1836         }
1837
1838         seq_info->map = map;
1839         seq_info->htab = container_of(map, struct bpf_htab, map);
1840         return 0;
1841 }
1842
1843 static void bpf_iter_fini_hash_map(void *priv_data)
1844 {
1845         struct bpf_iter_seq_hash_map_info *seq_info = priv_data;
1846
1847         kfree(seq_info->percpu_value_buf);
1848 }
1849
1850 static const struct seq_operations bpf_hash_map_seq_ops = {
1851         .start  = bpf_hash_map_seq_start,
1852         .next   = bpf_hash_map_seq_next,
1853         .stop   = bpf_hash_map_seq_stop,
1854         .show   = bpf_hash_map_seq_show,
1855 };
1856
1857 static const struct bpf_iter_seq_info iter_seq_info = {
1858         .seq_ops                = &bpf_hash_map_seq_ops,
1859         .init_seq_private       = bpf_iter_init_hash_map,
1860         .fini_seq_private       = bpf_iter_fini_hash_map,
1861         .seq_priv_size          = sizeof(struct bpf_iter_seq_hash_map_info),
1862 };
1863
1864 static int htab_map_btf_id;
1865 const struct bpf_map_ops htab_map_ops = {
1866         .map_meta_equal = bpf_map_meta_equal,
1867         .map_alloc_check = htab_map_alloc_check,
1868         .map_alloc = htab_map_alloc,
1869         .map_free = htab_map_free,
1870         .map_get_next_key = htab_map_get_next_key,
1871         .map_lookup_elem = htab_map_lookup_elem,
1872         .map_update_elem = htab_map_update_elem,
1873         .map_delete_elem = htab_map_delete_elem,
1874         .map_gen_lookup = htab_map_gen_lookup,
1875         .map_seq_show_elem = htab_map_seq_show_elem,
1876         BATCH_OPS(htab),
1877         .map_btf_name = "bpf_htab",
1878         .map_btf_id = &htab_map_btf_id,
1879         .iter_seq_info = &iter_seq_info,
1880 };
1881
1882 static int htab_lru_map_btf_id;
1883 const struct bpf_map_ops htab_lru_map_ops = {
1884         .map_meta_equal = bpf_map_meta_equal,
1885         .map_alloc_check = htab_map_alloc_check,
1886         .map_alloc = htab_map_alloc,
1887         .map_free = htab_map_free,
1888         .map_get_next_key = htab_map_get_next_key,
1889         .map_lookup_elem = htab_lru_map_lookup_elem,
1890         .map_lookup_elem_sys_only = htab_lru_map_lookup_elem_sys,
1891         .map_update_elem = htab_lru_map_update_elem,
1892         .map_delete_elem = htab_lru_map_delete_elem,
1893         .map_gen_lookup = htab_lru_map_gen_lookup,
1894         .map_seq_show_elem = htab_map_seq_show_elem,
1895         BATCH_OPS(htab_lru),
1896         .map_btf_name = "bpf_htab",
1897         .map_btf_id = &htab_lru_map_btf_id,
1898         .iter_seq_info = &iter_seq_info,
1899 };
1900
1901 /* Called from eBPF program */
1902 static void *htab_percpu_map_lookup_elem(struct bpf_map *map, void *key)
1903 {
1904         struct htab_elem *l = __htab_map_lookup_elem(map, key);
1905
1906         if (l)
1907                 return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
1908         else
1909                 return NULL;
1910 }
1911
1912 static void *htab_lru_percpu_map_lookup_elem(struct bpf_map *map, void *key)
1913 {
1914         struct htab_elem *l = __htab_map_lookup_elem(map, key);
1915
1916         if (l) {
1917                 bpf_lru_node_set_ref(&l->lru_node);
1918                 return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
1919         }
1920
1921         return NULL;
1922 }
1923
1924 int bpf_percpu_hash_copy(struct bpf_map *map, void *key, void *value)
1925 {
1926         struct htab_elem *l;
1927         void __percpu *pptr;
1928         int ret = -ENOENT;
1929         int cpu, off = 0;
1930         u32 size;
1931
1932         /* per_cpu areas are zero-filled and bpf programs can only
1933          * access 'value_size' of them, so copying rounded areas
1934          * will not leak any kernel data
1935          */
1936         size = round_up(map->value_size, 8);
1937         rcu_read_lock();
1938         l = __htab_map_lookup_elem(map, key);
1939         if (!l)
1940                 goto out;
1941         /* We do not mark LRU map element here in order to not mess up
1942          * eviction heuristics when user space does a map walk.
1943          */
1944         pptr = htab_elem_get_ptr(l, map->key_size);
1945         for_each_possible_cpu(cpu) {
1946                 bpf_long_memcpy(value + off,
1947                                 per_cpu_ptr(pptr, cpu), size);
1948                 off += size;
1949         }
1950         ret = 0;
1951 out:
1952         rcu_read_unlock();
1953         return ret;
1954 }
1955
1956 int bpf_percpu_hash_update(struct bpf_map *map, void *key, void *value,
1957                            u64 map_flags)
1958 {
1959         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1960         int ret;
1961
1962         rcu_read_lock();
1963         if (htab_is_lru(htab))
1964                 ret = __htab_lru_percpu_map_update_elem(map, key, value,
1965                                                         map_flags, true);
1966         else
1967                 ret = __htab_percpu_map_update_elem(map, key, value, map_flags,
1968                                                     true);
1969         rcu_read_unlock();
1970
1971         return ret;
1972 }
1973
1974 static void htab_percpu_map_seq_show_elem(struct bpf_map *map, void *key,
1975                                           struct seq_file *m)
1976 {
1977         struct htab_elem *l;
1978         void __percpu *pptr;
1979         int cpu;
1980
1981         rcu_read_lock();
1982
1983         l = __htab_map_lookup_elem(map, key);
1984         if (!l) {
1985                 rcu_read_unlock();
1986                 return;
1987         }
1988
1989         btf_type_seq_show(map->btf, map->btf_key_type_id, key, m);
1990         seq_puts(m, ": {\n");
1991         pptr = htab_elem_get_ptr(l, map->key_size);
1992         for_each_possible_cpu(cpu) {
1993                 seq_printf(m, "\tcpu%d: ", cpu);
1994                 btf_type_seq_show(map->btf, map->btf_value_type_id,
1995                                   per_cpu_ptr(pptr, cpu), m);
1996                 seq_puts(m, "\n");
1997         }
1998         seq_puts(m, "}\n");
1999
2000         rcu_read_unlock();
2001 }
2002
2003 static int htab_percpu_map_btf_id;
2004 const struct bpf_map_ops htab_percpu_map_ops = {
2005         .map_meta_equal = bpf_map_meta_equal,
2006         .map_alloc_check = htab_map_alloc_check,
2007         .map_alloc = htab_map_alloc,
2008         .map_free = htab_map_free,
2009         .map_get_next_key = htab_map_get_next_key,
2010         .map_lookup_elem = htab_percpu_map_lookup_elem,
2011         .map_update_elem = htab_percpu_map_update_elem,
2012         .map_delete_elem = htab_map_delete_elem,
2013         .map_seq_show_elem = htab_percpu_map_seq_show_elem,
2014         BATCH_OPS(htab_percpu),
2015         .map_btf_name = "bpf_htab",
2016         .map_btf_id = &htab_percpu_map_btf_id,
2017         .iter_seq_info = &iter_seq_info,
2018 };
2019
2020 static int htab_lru_percpu_map_btf_id;
2021 const struct bpf_map_ops htab_lru_percpu_map_ops = {
2022         .map_meta_equal = bpf_map_meta_equal,
2023         .map_alloc_check = htab_map_alloc_check,
2024         .map_alloc = htab_map_alloc,
2025         .map_free = htab_map_free,
2026         .map_get_next_key = htab_map_get_next_key,
2027         .map_lookup_elem = htab_lru_percpu_map_lookup_elem,
2028         .map_update_elem = htab_lru_percpu_map_update_elem,
2029         .map_delete_elem = htab_lru_map_delete_elem,
2030         .map_seq_show_elem = htab_percpu_map_seq_show_elem,
2031         BATCH_OPS(htab_lru_percpu),
2032         .map_btf_name = "bpf_htab",
2033         .map_btf_id = &htab_lru_percpu_map_btf_id,
2034         .iter_seq_info = &iter_seq_info,
2035 };
2036
2037 static int fd_htab_map_alloc_check(union bpf_attr *attr)
2038 {
2039         if (attr->value_size != sizeof(u32))
2040                 return -EINVAL;
2041         return htab_map_alloc_check(attr);
2042 }
2043
2044 static void fd_htab_map_free(struct bpf_map *map)
2045 {
2046         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
2047         struct hlist_nulls_node *n;
2048         struct hlist_nulls_head *head;
2049         struct htab_elem *l;
2050         int i;
2051
2052         for (i = 0; i < htab->n_buckets; i++) {
2053                 head = select_bucket(htab, i);
2054
2055                 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
2056                         void *ptr = fd_htab_map_get_ptr(map, l);
2057
2058                         map->ops->map_fd_put_ptr(ptr);
2059                 }
2060         }
2061
2062         htab_map_free(map);
2063 }
2064
2065 /* only called from syscall */
2066 int bpf_fd_htab_map_lookup_elem(struct bpf_map *map, void *key, u32 *value)
2067 {
2068         void **ptr;
2069         int ret = 0;
2070
2071         if (!map->ops->map_fd_sys_lookup_elem)
2072                 return -ENOTSUPP;
2073
2074         rcu_read_lock();
2075         ptr = htab_map_lookup_elem(map, key);
2076         if (ptr)
2077                 *value = map->ops->map_fd_sys_lookup_elem(READ_ONCE(*ptr));
2078         else
2079                 ret = -ENOENT;
2080         rcu_read_unlock();
2081
2082         return ret;
2083 }
2084
2085 /* only called from syscall */
2086 int bpf_fd_htab_map_update_elem(struct bpf_map *map, struct file *map_file,
2087                                 void *key, void *value, u64 map_flags)
2088 {
2089         void *ptr;
2090         int ret;
2091         u32 ufd = *(u32 *)value;
2092
2093         ptr = map->ops->map_fd_get_ptr(map, map_file, ufd);
2094         if (IS_ERR(ptr))
2095                 return PTR_ERR(ptr);
2096
2097         ret = htab_map_update_elem(map, key, &ptr, map_flags);
2098         if (ret)
2099                 map->ops->map_fd_put_ptr(ptr);
2100
2101         return ret;
2102 }
2103
2104 static struct bpf_map *htab_of_map_alloc(union bpf_attr *attr)
2105 {
2106         struct bpf_map *map, *inner_map_meta;
2107
2108         inner_map_meta = bpf_map_meta_alloc(attr->inner_map_fd);
2109         if (IS_ERR(inner_map_meta))
2110                 return inner_map_meta;
2111
2112         map = htab_map_alloc(attr);
2113         if (IS_ERR(map)) {
2114                 bpf_map_meta_free(inner_map_meta);
2115                 return map;
2116         }
2117
2118         map->inner_map_meta = inner_map_meta;
2119
2120         return map;
2121 }
2122
2123 static void *htab_of_map_lookup_elem(struct bpf_map *map, void *key)
2124 {
2125         struct bpf_map **inner_map  = htab_map_lookup_elem(map, key);
2126
2127         if (!inner_map)
2128                 return NULL;
2129
2130         return READ_ONCE(*inner_map);
2131 }
2132
2133 static int htab_of_map_gen_lookup(struct bpf_map *map,
2134                                   struct bpf_insn *insn_buf)
2135 {
2136         struct bpf_insn *insn = insn_buf;
2137         const int ret = BPF_REG_0;
2138
2139         BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
2140                      (void *(*)(struct bpf_map *map, void *key))NULL));
2141         *insn++ = BPF_EMIT_CALL(BPF_CAST_CALL(__htab_map_lookup_elem));
2142         *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 2);
2143         *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
2144                                 offsetof(struct htab_elem, key) +
2145                                 round_up(map->key_size, 8));
2146         *insn++ = BPF_LDX_MEM(BPF_DW, ret, ret, 0);
2147
2148         return insn - insn_buf;
2149 }
2150
2151 static void htab_of_map_free(struct bpf_map *map)
2152 {
2153         bpf_map_meta_free(map->inner_map_meta);
2154         fd_htab_map_free(map);
2155 }
2156
2157 static int htab_of_maps_map_btf_id;
2158 const struct bpf_map_ops htab_of_maps_map_ops = {
2159         .map_alloc_check = fd_htab_map_alloc_check,
2160         .map_alloc = htab_of_map_alloc,
2161         .map_free = htab_of_map_free,
2162         .map_get_next_key = htab_map_get_next_key,
2163         .map_lookup_elem = htab_of_map_lookup_elem,
2164         .map_delete_elem = htab_map_delete_elem,
2165         .map_fd_get_ptr = bpf_map_fd_get_ptr,
2166         .map_fd_put_ptr = bpf_map_fd_put_ptr,
2167         .map_fd_sys_lookup_elem = bpf_map_fd_sys_lookup_elem,
2168         .map_gen_lookup = htab_of_map_gen_lookup,
2169         .map_check_btf = map_check_no_btf,
2170         .map_btf_name = "bpf_htab",
2171         .map_btf_id = &htab_of_maps_map_btf_id,
2172 };