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