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 <linux/btf_ids.h>
  14 #include "percpu_freelist.h"
  15 #include "bpf_lru_list.h"
  16 #include "map_in_map.h"
  17 #include <linux/bpf_mem_alloc.h>
  18
  19 #define HTAB_CREATE_FLAG_MASK                                           \
  20         (BPF_F_NO_PREALLOC | BPF_F_NO_COMMON_LRU | BPF_F_NUMA_NODE |    \
  21          BPF_F_ACCESS_MASK | BPF_F_ZERO_SEED)
  22
  23 #define BATCH_OPS(_name)                        \
  24         .map_lookup_batch =                     \
  25         _name##_map_lookup_batch,               \
  26         .map_lookup_and_delete_batch =          \
  27         _name##_map_lookup_and_delete_batch,    \
  28         .map_update_batch =                     \
  29         generic_map_update_batch,               \
  30         .map_delete_batch =                     \
  31         generic_map_delete_batch
  32
  33 /*
  34  * The bucket lock has two protection scopes:
  35  *
  36  * 1) Serializing concurrent operations from BPF programs on different
  37  *    CPUs
  38  *
  39  * 2) Serializing concurrent operations from BPF programs and sys_bpf()
  40  *
  41  * BPF programs can execute in any context including perf, kprobes and
  42  * tracing. As there are almost no limits where perf, kprobes and tracing
  43  * can be invoked from the lock operations need to be protected against
  44  * deadlocks. Deadlocks can be caused by recursion and by an invocation in
  45  * the lock held section when functions which acquire this lock are invoked
  46  * from sys_bpf(). BPF recursion is prevented by incrementing the per CPU
  47  * variable bpf_prog_active, which prevents BPF programs attached to perf
  48  * events, kprobes and tracing to be invoked before the prior invocation
  49  * from one of these contexts completed. sys_bpf() uses the same mechanism
  50  * by pinning the task to the current CPU and incrementing the recursion
  51  * protection across the map operation.
  52  *
  53  * This has subtle implications on PREEMPT_RT. PREEMPT_RT forbids certain
  54  * operations like memory allocations (even with GFP_ATOMIC) from atomic
  55  * contexts. This is required because even with GFP_ATOMIC the memory
  56  * allocator calls into code paths which acquire locks with long held lock
  57  * sections. To ensure the deterministic behaviour these locks are regular
  58  * spinlocks, which are converted to 'sleepable' spinlocks on RT. The only
  59  * true atomic contexts on an RT kernel are the low level hardware
  60  * handling, scheduling, low level interrupt handling, NMIs etc. None of
  61  * these contexts should ever do memory allocations.
  62  *
  63  * As regular device interrupt handlers and soft interrupts are forced into
  64  * thread context, the existing code which does
  65  *   spin_lock*(); alloc(GFP_ATOMIC); spin_unlock*();
  66  * just works.
  67  *
  68  * In theory the BPF locks could be converted to regular spinlocks as well,
  69  * but the bucket locks and percpu_freelist locks can be taken from
  70  * arbitrary contexts (perf, kprobes, tracepoints) which are required to be
  71  * atomic contexts even on RT. Before the introduction of bpf_mem_alloc,
  72  * it is only safe to use raw spinlock for preallocated hash map on a RT kernel,
  73  * because there is no memory allocation within the lock held sections. However
  74  * after hash map was fully converted to use bpf_mem_alloc, there will be
  75  * non-synchronous memory allocation for non-preallocated hash map, so it is
  76  * safe to always use raw spinlock for bucket lock.
  77  */
  78 struct bucket {
  79         struct hlist_nulls_head head;
  80         raw_spinlock_t raw_lock;
  81 };
  82
  83 #define HASHTAB_MAP_LOCK_COUNT 8
  84 #define HASHTAB_MAP_LOCK_MASK (HASHTAB_MAP_LOCK_COUNT - 1)
  85
  86 struct bpf_htab {
  87         struct bpf_map map;
  88         struct bpf_mem_alloc ma;
  89         struct bpf_mem_alloc pcpu_ma;
  90         struct bucket *buckets;
  91         void *elems;
  92         union {
  93                 struct pcpu_freelist freelist;
  94                 struct bpf_lru lru;
  95         };
  96         struct htab_elem *__percpu *extra_elems;
  97         /* number of elements in non-preallocated hashtable are kept
  98          * in either pcount or count
  99          */
 100         struct percpu_counter pcount;
 101         atomic_t count;
 102         bool use_percpu_counter;
 103         u32 n_buckets;  /* number of hash buckets */
 104         u32 elem_size;  /* size of each element in bytes */
 105         u32 hashrnd;
 106         struct lock_class_key lockdep_key;
 107         int __percpu *map_locked[HASHTAB_MAP_LOCK_COUNT];
 108 };
 109
 110 /* each htab element is struct htab_elem + key + value */
 111 struct htab_elem {
 112         union {
 113                 struct hlist_nulls_node hash_node;
 114                 struct {
 115                         void *padding;
 116                         union {
 117                                 struct pcpu_freelist_node fnode;
 118                                 struct htab_elem *batch_flink;
 119                         };
 120                 };
 121         };
 122         union {
 123                 /* pointer to per-cpu pointer */
 124                 void *ptr_to_pptr;
 125                 struct bpf_lru_node lru_node;
 126         };
 127         u32 hash;
 128         char key[] __aligned(8);
 129 };
 130
 131 static inline bool htab_is_prealloc(const struct bpf_htab *htab)
 132 {
 133         return !(htab->map.map_flags & BPF_F_NO_PREALLOC);
 134 }
 135
 136 static void htab_init_buckets(struct bpf_htab *htab)
 137 {
 138         unsigned int i;
 139
 140         for (i = 0; i < htab->n_buckets; i++) {
 141                 INIT_HLIST_NULLS_HEAD(&htab->buckets[i].head, i);
 142                 raw_spin_lock_init(&htab->buckets[i].raw_lock);
 143                 lockdep_set_class(&htab->buckets[i].raw_lock,
 144                                           &htab->lockdep_key);
 145                 cond_resched();
 146         }
 147 }
 148
 149 static inline int htab_lock_bucket(const struct bpf_htab *htab,
 150                                    struct bucket *b, u32 hash,
 151                                    unsigned long *pflags)
 152 {
 153         unsigned long flags;
 154
 155         hash = hash & min_t(u32, HASHTAB_MAP_LOCK_MASK, htab->n_buckets - 1);
 156
 157         preempt_disable();
 158         if (unlikely(__this_cpu_inc_return(*(htab->map_locked[hash])) != 1)) {
 159                 __this_cpu_dec(*(htab->map_locked[hash]));
 160                 preempt_enable();
 161                 return -EBUSY;
 162         }
 163
 164         raw_spin_lock_irqsave(&b->raw_lock, flags);
 165         *pflags = flags;
 166
 167         return 0;
 168 }
 169
 170 static inline void htab_unlock_bucket(const struct bpf_htab *htab,
 171                                       struct bucket *b, u32 hash,
 172                                       unsigned long flags)
 173 {
 174         hash = hash & min_t(u32, HASHTAB_MAP_LOCK_MASK, htab->n_buckets - 1);
 175         raw_spin_unlock_irqrestore(&b->raw_lock, flags);
 176         __this_cpu_dec(*(htab->map_locked[hash]));
 177         preempt_enable();
 178 }
 179
 180 static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node);
 181
 182 static bool htab_is_lru(const struct bpf_htab *htab)
 183 {
 184         return htab->map.map_type == BPF_MAP_TYPE_LRU_HASH ||
 185                 htab->map.map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH;
 186 }
 187
 188 static bool htab_is_percpu(const struct bpf_htab *htab)
 189 {
 190         return htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH ||
 191                 htab->map.map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH;
 192 }
 193
 194 static inline void htab_elem_set_ptr(struct htab_elem *l, u32 key_size,
 195                                      void __percpu *pptr)
 196 {
 197         *(void __percpu **)(l->key + key_size) = pptr;
 198 }
 199
 200 static inline void __percpu *htab_elem_get_ptr(struct htab_elem *l, u32 key_size)
 201 {
 202         return *(void __percpu **)(l->key + key_size);
 203 }
 204
 205 static void *fd_htab_map_get_ptr(const struct bpf_map *map, struct htab_elem *l)
 206 {
 207         return *(void **)(l->key + roundup(map->key_size, 8));
 208 }
 209
 210 static struct htab_elem *get_htab_elem(struct bpf_htab *htab, int i)
 211 {
 212         return (struct htab_elem *) (htab->elems + i * (u64)htab->elem_size);
 213 }
 214
 215 static bool htab_has_extra_elems(struct bpf_htab *htab)
 216 {
 217         return !htab_is_percpu(htab) && !htab_is_lru(htab);
 218 }
 219
 220 static void htab_free_prealloced_timers(struct bpf_htab *htab)
 221 {
 222         u32 num_entries = htab->map.max_entries;
 223         int i;
 224
 225         if (!btf_record_has_field(htab->map.record, BPF_TIMER))
 226                 return;
 227         if (htab_has_extra_elems(htab))
 228                 num_entries += num_possible_cpus();
 229
 230         for (i = 0; i < num_entries; i++) {
 231                 struct htab_elem *elem;
 232
 233                 elem = get_htab_elem(htab, i);
 234                 bpf_obj_free_timer(htab->map.record, elem->key + round_up(htab->map.key_size, 8));
 235                 cond_resched();
 236         }
 237 }
 238
 239 static void htab_free_prealloced_fields(struct bpf_htab *htab)
 240 {
 241         u32 num_entries = htab->map.max_entries;
 242         int i;
 243
 244         if (IS_ERR_OR_NULL(htab->map.record))
 245                 return;
 246         if (htab_has_extra_elems(htab))
 247                 num_entries += num_possible_cpus();
 248         for (i = 0; i < num_entries; i++) {
 249                 struct htab_elem *elem;
 250
 251                 elem = get_htab_elem(htab, i);
 252                 if (htab_is_percpu(htab)) {
 253                         void __percpu *pptr = htab_elem_get_ptr(elem, htab->map.key_size);
 254                         int cpu;
 255
 256                         for_each_possible_cpu(cpu) {
 257                                 bpf_obj_free_fields(htab->map.record, per_cpu_ptr(pptr, cpu));
 258                                 cond_resched();
 259                         }
 260                 } else {
 261                         bpf_obj_free_fields(htab->map.record, elem->key + round_up(htab->map.key_size, 8));
 262                         cond_resched();
 263                 }
 264                 cond_resched();
 265         }
 266 }
 267
 268 static void htab_free_elems(struct bpf_htab *htab)
 269 {
 270         int i;
 271
 272         if (!htab_is_percpu(htab))
 273                 goto free_elems;
 274
 275         for (i = 0; i < htab->map.max_entries; i++) {
 276                 void __percpu *pptr;
 277
 278                 pptr = htab_elem_get_ptr(get_htab_elem(htab, i),
 279                                          htab->map.key_size);
 280                 free_percpu(pptr);
 281                 cond_resched();
 282         }
 283 free_elems:
 284         bpf_map_area_free(htab->elems);
 285 }
 286
 287 /* The LRU list has a lock (lru_lock). Each htab bucket has a lock
 288  * (bucket_lock). If both locks need to be acquired together, the lock
 289  * order is always lru_lock -> bucket_lock and this only happens in
 290  * bpf_lru_list.c logic. For example, certain code path of
 291  * bpf_lru_pop_free(), which is called by function prealloc_lru_pop(),
 292  * will acquire lru_lock first followed by acquiring bucket_lock.
 293  *
 294  * In hashtab.c, to avoid deadlock, lock acquisition of
 295  * bucket_lock followed by lru_lock is not allowed. In such cases,
 296  * bucket_lock needs to be released first before acquiring lru_lock.
 297  */
 298 static struct htab_elem *prealloc_lru_pop(struct bpf_htab *htab, void *key,
 299                                           u32 hash)
 300 {
 301         struct bpf_lru_node *node = bpf_lru_pop_free(&htab->lru, hash);
 302         struct htab_elem *l;
 303
 304         if (node) {
 305                 l = container_of(node, struct htab_elem, lru_node);
 306                 memcpy(l->key, key, htab->map.key_size);
 307                 return l;
 308         }
 309
 310         return NULL;
 311 }
 312
 313 static int prealloc_init(struct bpf_htab *htab)
 314 {
 315         u32 num_entries = htab->map.max_entries;
 316         int err = -ENOMEM, i;
 317
 318         if (htab_has_extra_elems(htab))
 319                 num_entries += num_possible_cpus();
 320
 321         htab->elems = bpf_map_area_alloc((u64)htab->elem_size * num_entries,
 322                                          htab->map.numa_node);
 323         if (!htab->elems)
 324                 return -ENOMEM;
 325
 326         if (!htab_is_percpu(htab))
 327                 goto skip_percpu_elems;
 328
 329         for (i = 0; i < num_entries; i++) {
 330                 u32 size = round_up(htab->map.value_size, 8);
 331                 void __percpu *pptr;
 332
 333                 pptr = bpf_map_alloc_percpu(&htab->map, size, 8,
 334                                             GFP_USER | __GFP_NOWARN);
 335                 if (!pptr)
 336                         goto free_elems;
 337                 htab_elem_set_ptr(get_htab_elem(htab, i), htab->map.key_size,
 338                                   pptr);
 339                 cond_resched();
 340         }
 341
 342 skip_percpu_elems:
 343         if (htab_is_lru(htab))
 344                 err = bpf_lru_init(&htab->lru,
 345                                    htab->map.map_flags & BPF_F_NO_COMMON_LRU,
 346                                    offsetof(struct htab_elem, hash) -
 347                                    offsetof(struct htab_elem, lru_node),
 348                                    htab_lru_map_delete_node,
 349                                    htab);
 350         else
 351                 err = pcpu_freelist_init(&htab->freelist);
 352
 353         if (err)
 354                 goto free_elems;
 355
 356         if (htab_is_lru(htab))
 357                 bpf_lru_populate(&htab->lru, htab->elems,
 358                                  offsetof(struct htab_elem, lru_node),
 359                                  htab->elem_size, num_entries);
 360         else
 361                 pcpu_freelist_populate(&htab->freelist,
 362                                        htab->elems + offsetof(struct htab_elem, fnode),
 363                                        htab->elem_size, num_entries);
 364
 365         return 0;
 366
 367 free_elems:
 368         htab_free_elems(htab);
 369         return err;
 370 }
 371
 372 static void prealloc_destroy(struct bpf_htab *htab)
 373 {
 374         htab_free_elems(htab);
 375
 376         if (htab_is_lru(htab))
 377                 bpf_lru_destroy(&htab->lru);
 378         else
 379                 pcpu_freelist_destroy(&htab->freelist);
 380 }
 381
 382 static int alloc_extra_elems(struct bpf_htab *htab)
 383 {
 384         struct htab_elem *__percpu *pptr, *l_new;
 385         struct pcpu_freelist_node *l;
 386         int cpu;
 387
 388         pptr = bpf_map_alloc_percpu(&htab->map, sizeof(struct htab_elem *), 8,
 389                                     GFP_USER | __GFP_NOWARN);
 390         if (!pptr)
 391                 return -ENOMEM;
 392
 393         for_each_possible_cpu(cpu) {
 394                 l = pcpu_freelist_pop(&htab->freelist);
 395                 /* pop will succeed, since prealloc_init()
 396                  * preallocated extra num_possible_cpus elements
 397                  */
 398                 l_new = container_of(l, struct htab_elem, fnode);
 399                 *per_cpu_ptr(pptr, cpu) = l_new;
 400         }
 401         htab->extra_elems = pptr;
 402         return 0;
 403 }
 404
 405 /* Called from syscall */
 406 static int htab_map_alloc_check(union bpf_attr *attr)
 407 {
 408         bool percpu = (attr->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
 409                        attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
 410         bool lru = (attr->map_type == BPF_MAP_TYPE_LRU_HASH ||
 411                     attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
 412         /* percpu_lru means each cpu has its own LRU list.
 413          * it is different from BPF_MAP_TYPE_PERCPU_HASH where
 414          * the map's value itself is percpu.  percpu_lru has
 415          * nothing to do with the map's value.
 416          */
 417         bool percpu_lru = (attr->map_flags & BPF_F_NO_COMMON_LRU);
 418         bool prealloc = !(attr->map_flags & BPF_F_NO_PREALLOC);
 419         bool zero_seed = (attr->map_flags & BPF_F_ZERO_SEED);
 420         int numa_node = bpf_map_attr_numa_node(attr);
 421
 422         BUILD_BUG_ON(offsetof(struct htab_elem, fnode.next) !=
 423                      offsetof(struct htab_elem, hash_node.pprev));
 424
 425         if (zero_seed && !capable(CAP_SYS_ADMIN))
 426                 /* Guard against local DoS, and discourage production use. */
 427                 return -EPERM;
 428
 429         if (attr->map_flags & ~HTAB_CREATE_FLAG_MASK ||
 430             !bpf_map_flags_access_ok(attr->map_flags))
 431                 return -EINVAL;
 432
 433         if (!lru && percpu_lru)
 434                 return -EINVAL;
 435
 436         if (lru && !prealloc)
 437                 return -ENOTSUPP;
 438
 439         if (numa_node != NUMA_NO_NODE && (percpu || percpu_lru))
 440                 return -EINVAL;
 441
 442         /* check sanity of attributes.
 443          * value_size == 0 may be allowed in the future to use map as a set
 444          */
 445         if (attr->max_entries == 0 || attr->key_size == 0 ||
 446             attr->value_size == 0)
 447                 return -EINVAL;
 448
 449         if ((u64)attr->key_size + attr->value_size >= KMALLOC_MAX_SIZE -
 450            sizeof(struct htab_elem))
 451                 /* if key_size + value_size is bigger, the user space won't be
 452                  * able to access the elements via bpf syscall. This check
 453                  * also makes sure that the elem_size doesn't overflow and it's
 454                  * kmalloc-able later in htab_map_update_elem()
 455                  */
 456                 return -E2BIG;
 457
 458         return 0;
 459 }
 460
 461 static struct bpf_map *htab_map_alloc(union bpf_attr *attr)
 462 {
 463         bool percpu = (attr->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
 464                        attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
 465         bool lru = (attr->map_type == BPF_MAP_TYPE_LRU_HASH ||
 466                     attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
 467         /* percpu_lru means each cpu has its own LRU list.
 468          * it is different from BPF_MAP_TYPE_PERCPU_HASH where
 469          * the map's value itself is percpu.  percpu_lru has
 470          * nothing to do with the map's value.
 471          */
 472         bool percpu_lru = (attr->map_flags & BPF_F_NO_COMMON_LRU);
 473         bool prealloc = !(attr->map_flags & BPF_F_NO_PREALLOC);
 474         struct bpf_htab *htab;
 475         int err, i;
 476
 477         htab = bpf_map_area_alloc(sizeof(*htab), NUMA_NO_NODE);
 478         if (!htab)
 479                 return ERR_PTR(-ENOMEM);
 480
 481         lockdep_register_key(&htab->lockdep_key);
 482
 483         bpf_map_init_from_attr(&htab->map, attr);
 484
 485         if (percpu_lru) {
 486                 /* ensure each CPU's lru list has >=1 elements.
 487                  * since we are at it, make each lru list has the same
 488                  * number of elements.
 489                  */
 490                 htab->map.max_entries = roundup(attr->max_entries,
 491                                                 num_possible_cpus());
 492                 if (htab->map.max_entries < attr->max_entries)
 493                         htab->map.max_entries = rounddown(attr->max_entries,
 494                                                           num_possible_cpus());
 495         }
 496
 497         /* hash table size must be power of 2 */
 498         htab->n_buckets = roundup_pow_of_two(htab->map.max_entries);
 499
 500         htab->elem_size = sizeof(struct htab_elem) +
 501                           round_up(htab->map.key_size, 8);
 502         if (percpu)
 503                 htab->elem_size += sizeof(void *);
 504         else
 505                 htab->elem_size += round_up(htab->map.value_size, 8);
 506
 507         err = -E2BIG;
 508         /* prevent zero size kmalloc and check for u32 overflow */
 509         if (htab->n_buckets == 0 ||
 510             htab->n_buckets > U32_MAX / sizeof(struct bucket))
 511                 goto free_htab;
 512
 513         err = -ENOMEM;
 514         htab->buckets = bpf_map_area_alloc(htab->n_buckets *
 515                                            sizeof(struct bucket),
 516                                            htab->map.numa_node);
 517         if (!htab->buckets)
 518                 goto free_htab;
 519
 520         for (i = 0; i < HASHTAB_MAP_LOCK_COUNT; i++) {
 521                 htab->map_locked[i] = bpf_map_alloc_percpu(&htab->map,
 522                                                            sizeof(int),
 523                                                            sizeof(int),
 524                                                            GFP_USER);
 525                 if (!htab->map_locked[i])
 526                         goto free_map_locked;
 527         }
 528
 529         if (htab->map.map_flags & BPF_F_ZERO_SEED)
 530                 htab->hashrnd = 0;
 531         else
 532                 htab->hashrnd = get_random_u32();
 533
 534         htab_init_buckets(htab);
 535
 536 /* compute_batch_value() computes batch value as num_online_cpus() * 2
 537  * and __percpu_counter_compare() needs
 538  * htab->max_entries - cur_number_of_elems to be more than batch * num_online_cpus()
 539  * for percpu_counter to be faster than atomic_t. In practice the average bpf
 540  * hash map size is 10k, which means that a system with 64 cpus will fill
 541  * hashmap to 20% of 10k before percpu_counter becomes ineffective. Therefore
 542  * define our own batch count as 32 then 10k hash map can be filled up to 80%:
 543  * 10k - 8k > 32 _batch_ * 64 _cpus_
 544  * and __percpu_counter_compare() will still be fast. At that point hash map
 545  * collisions will dominate its performance anyway. Assume that hash map filled
 546  * to 50+% isn't going to be O(1) and use the following formula to choose
 547  * between percpu_counter and atomic_t.
 548  */
 549 #define PERCPU_COUNTER_BATCH 32
 550         if (attr->max_entries / 2 > num_online_cpus() * PERCPU_COUNTER_BATCH)
 551                 htab->use_percpu_counter = true;
 552
 553         if (htab->use_percpu_counter) {
 554                 err = percpu_counter_init(&htab->pcount, 0, GFP_KERNEL);
 555                 if (err)
 556                         goto free_map_locked;
 557         }
 558
 559         if (prealloc) {
 560                 err = prealloc_init(htab);
 561                 if (err)
 562                         goto free_map_locked;
 563
 564                 if (!percpu && !lru) {
 565                         /* lru itself can remove the least used element, so
 566                          * there is no need for an extra elem during map_update.
 567                          */
 568                         err = alloc_extra_elems(htab);
 569                         if (err)
 570                                 goto free_prealloc;
 571                 }
 572         } else {
 573                 err = bpf_mem_alloc_init(&htab->ma, htab->elem_size, false);
 574                 if (err)
 575                         goto free_map_locked;
 576                 if (percpu) {
 577                         err = bpf_mem_alloc_init(&htab->pcpu_ma,
 578                                                  round_up(htab->map.value_size, 8), true);
 579                         if (err)
 580                                 goto free_map_locked;
 581                 }
 582         }
 583
 584         return &htab->map;
 585
 586 free_prealloc:
 587         prealloc_destroy(htab);
 588 free_map_locked:
 589         if (htab->use_percpu_counter)
 590                 percpu_counter_destroy(&htab->pcount);
 591         for (i = 0; i < HASHTAB_MAP_LOCK_COUNT; i++)
 592                 free_percpu(htab->map_locked[i]);
 593         bpf_map_area_free(htab->buckets);
 594         bpf_mem_alloc_destroy(&htab->pcpu_ma);
 595         bpf_mem_alloc_destroy(&htab->ma);
 596 free_htab:
 597         lockdep_unregister_key(&htab->lockdep_key);
 598         bpf_map_area_free(htab);
 599         return ERR_PTR(err);
 600 }
 601
 602 static inline u32 htab_map_hash(const void *key, u32 key_len, u32 hashrnd)
 603 {
 604         if (likely(key_len % 4 == 0))
 605                 return jhash2(key, key_len / 4, hashrnd);
 606         return jhash(key, key_len, hashrnd);
 607 }
 608
 609 static inline struct bucket *__select_bucket(struct bpf_htab *htab, u32 hash)
 610 {
 611         return &htab->buckets[hash & (htab->n_buckets - 1)];
 612 }
 613
 614 static inline struct hlist_nulls_head *select_bucket(struct bpf_htab *htab, u32 hash)
 615 {
 616         return &__select_bucket(htab, hash)->head;
 617 }
 618
 619 /* this lookup function can only be called with bucket lock taken */
 620 static struct htab_elem *lookup_elem_raw(struct hlist_nulls_head *head, u32 hash,
 621                                          void *key, u32 key_size)
 622 {
 623         struct hlist_nulls_node *n;
 624         struct htab_elem *l;
 625
 626         hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
 627                 if (l->hash == hash && !memcmp(&l->key, key, key_size))
 628                         return l;
 629
 630         return NULL;
 631 }
 632
 633 /* can be called without bucket lock. it will repeat the loop in
 634  * the unlikely event when elements moved from one bucket into another
 635  * while link list is being walked
 636  */
 637 static struct htab_elem *lookup_nulls_elem_raw(struct hlist_nulls_head *head,
 638                                                u32 hash, void *key,
 639                                                u32 key_size, u32 n_buckets)
 640 {
 641         struct hlist_nulls_node *n;
 642         struct htab_elem *l;
 643
 644 again:
 645         hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
 646                 if (l->hash == hash && !memcmp(&l->key, key, key_size))
 647                         return l;
 648
 649         if (unlikely(get_nulls_value(n) != (hash & (n_buckets - 1))))
 650                 goto again;
 651
 652         return NULL;
 653 }
 654
 655 /* Called from syscall or from eBPF program directly, so
 656  * arguments have to match bpf_map_lookup_elem() exactly.
 657  * The return value is adjusted by BPF instructions
 658  * in htab_map_gen_lookup().
 659  */
 660 static void *__htab_map_lookup_elem(struct bpf_map *map, void *key)
 661 {
 662         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 663         struct hlist_nulls_head *head;
 664         struct htab_elem *l;
 665         u32 hash, key_size;
 666
 667         WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
 668                      !rcu_read_lock_bh_held());
 669
 670         key_size = map->key_size;
 671
 672         hash = htab_map_hash(key, key_size, htab->hashrnd);
 673
 674         head = select_bucket(htab, hash);
 675
 676         l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
 677
 678         return l;
 679 }
 680
 681 static void *htab_map_lookup_elem(struct bpf_map *map, void *key)
 682 {
 683         struct htab_elem *l = __htab_map_lookup_elem(map, key);
 684
 685         if (l)
 686                 return l->key + round_up(map->key_size, 8);
 687
 688         return NULL;
 689 }
 690
 691 /* inline bpf_map_lookup_elem() call.
 692  * Instead of:
 693  * bpf_prog
 694  *   bpf_map_lookup_elem
 695  *     map->ops->map_lookup_elem
 696  *       htab_map_lookup_elem
 697  *         __htab_map_lookup_elem
 698  * do:
 699  * bpf_prog
 700  *   __htab_map_lookup_elem
 701  */
 702 static int htab_map_gen_lookup(struct bpf_map *map, struct bpf_insn *insn_buf)
 703 {
 704         struct bpf_insn *insn = insn_buf;
 705         const int ret = BPF_REG_0;
 706
 707         BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
 708                      (void *(*)(struct bpf_map *map, void *key))NULL));
 709         *insn++ = BPF_EMIT_CALL(__htab_map_lookup_elem);
 710         *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 1);
 711         *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
 712                                 offsetof(struct htab_elem, key) +
 713                                 round_up(map->key_size, 8));
 714         return insn - insn_buf;
 715 }
 716
 717 static __always_inline void *__htab_lru_map_lookup_elem(struct bpf_map *map,
 718                                                         void *key, const bool mark)
 719 {
 720         struct htab_elem *l = __htab_map_lookup_elem(map, key);
 721
 722         if (l) {
 723                 if (mark)
 724                         bpf_lru_node_set_ref(&l->lru_node);
 725                 return l->key + round_up(map->key_size, 8);
 726         }
 727
 728         return NULL;
 729 }
 730
 731 static void *htab_lru_map_lookup_elem(struct bpf_map *map, void *key)
 732 {
 733         return __htab_lru_map_lookup_elem(map, key, true);
 734 }
 735
 736 static void *htab_lru_map_lookup_elem_sys(struct bpf_map *map, void *key)
 737 {
 738         return __htab_lru_map_lookup_elem(map, key, false);
 739 }
 740
 741 static int htab_lru_map_gen_lookup(struct bpf_map *map,
 742                                    struct bpf_insn *insn_buf)
 743 {
 744         struct bpf_insn *insn = insn_buf;
 745         const int ret = BPF_REG_0;
 746         const int ref_reg = BPF_REG_1;
 747
 748         BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
 749                      (void *(*)(struct bpf_map *map, void *key))NULL));
 750         *insn++ = BPF_EMIT_CALL(__htab_map_lookup_elem);
 751         *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 4);
 752         *insn++ = BPF_LDX_MEM(BPF_B, ref_reg, ret,
 753                               offsetof(struct htab_elem, lru_node) +
 754                               offsetof(struct bpf_lru_node, ref));
 755         *insn++ = BPF_JMP_IMM(BPF_JNE, ref_reg, 0, 1);
 756         *insn++ = BPF_ST_MEM(BPF_B, ret,
 757                              offsetof(struct htab_elem, lru_node) +
 758                              offsetof(struct bpf_lru_node, ref),
 759                              1);
 760         *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
 761                                 offsetof(struct htab_elem, key) +
 762                                 round_up(map->key_size, 8));
 763         return insn - insn_buf;
 764 }
 765
 766 static void check_and_free_fields(struct bpf_htab *htab,
 767                                   struct htab_elem *elem)
 768 {
 769         if (htab_is_percpu(htab)) {
 770                 void __percpu *pptr = htab_elem_get_ptr(elem, htab->map.key_size);
 771                 int cpu;
 772
 773                 for_each_possible_cpu(cpu)
 774                         bpf_obj_free_fields(htab->map.record, per_cpu_ptr(pptr, cpu));
 775         } else {
 776                 void *map_value = elem->key + round_up(htab->map.key_size, 8);
 777
 778                 bpf_obj_free_fields(htab->map.record, map_value);
 779         }
 780 }
 781
 782 /* It is called from the bpf_lru_list when the LRU needs to delete
 783  * older elements from the htab.
 784  */
 785 static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node)
 786 {
 787         struct bpf_htab *htab = arg;
 788         struct htab_elem *l = NULL, *tgt_l;
 789         struct hlist_nulls_head *head;
 790         struct hlist_nulls_node *n;
 791         unsigned long flags;
 792         struct bucket *b;
 793         int ret;
 794
 795         tgt_l = container_of(node, struct htab_elem, lru_node);
 796         b = __select_bucket(htab, tgt_l->hash);
 797         head = &b->head;
 798
 799         ret = htab_lock_bucket(htab, b, tgt_l->hash, &flags);
 800         if (ret)
 801                 return false;
 802
 803         hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
 804                 if (l == tgt_l) {
 805                         hlist_nulls_del_rcu(&l->hash_node);
 806                         check_and_free_fields(htab, l);
 807                         break;
 808                 }
 809
 810         htab_unlock_bucket(htab, b, tgt_l->hash, flags);
 811
 812         return l == tgt_l;
 813 }
 814
 815 /* Called from syscall */
 816 static int htab_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
 817 {
 818         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 819         struct hlist_nulls_head *head;
 820         struct htab_elem *l, *next_l;
 821         u32 hash, key_size;
 822         int i = 0;
 823
 824         WARN_ON_ONCE(!rcu_read_lock_held());
 825
 826         key_size = map->key_size;
 827
 828         if (!key)
 829                 goto find_first_elem;
 830
 831         hash = htab_map_hash(key, key_size, htab->hashrnd);
 832
 833         head = select_bucket(htab, hash);
 834
 835         /* lookup the key */
 836         l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
 837
 838         if (!l)
 839                 goto find_first_elem;
 840
 841         /* key was found, get next key in the same bucket */
 842         next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_next_rcu(&l->hash_node)),
 843                                   struct htab_elem, hash_node);
 844
 845         if (next_l) {
 846                 /* if next elem in this hash list is non-zero, just return it */
 847                 memcpy(next_key, next_l->key, key_size);
 848                 return 0;
 849         }
 850
 851         /* no more elements in this hash list, go to the next bucket */
 852         i = hash & (htab->n_buckets - 1);
 853         i++;
 854
 855 find_first_elem:
 856         /* iterate over buckets */
 857         for (; i < htab->n_buckets; i++) {
 858                 head = select_bucket(htab, i);
 859
 860                 /* pick first element in the bucket */
 861                 next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_first_rcu(head)),
 862                                           struct htab_elem, hash_node);
 863                 if (next_l) {
 864                         /* if it's not empty, just return it */
 865                         memcpy(next_key, next_l->key, key_size);
 866                         return 0;
 867                 }
 868         }
 869
 870         /* iterated over all buckets and all elements */
 871         return -ENOENT;
 872 }
 873
 874 static void htab_elem_free(struct bpf_htab *htab, struct htab_elem *l)
 875 {
 876         check_and_free_fields(htab, l);
 877         if (htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH)
 878                 bpf_mem_cache_free(&htab->pcpu_ma, l->ptr_to_pptr);
 879         bpf_mem_cache_free(&htab->ma, l);
 880 }
 881
 882 static void htab_put_fd_value(struct bpf_htab *htab, struct htab_elem *l)
 883 {
 884         struct bpf_map *map = &htab->map;
 885         void *ptr;
 886
 887         if (map->ops->map_fd_put_ptr) {
 888                 ptr = fd_htab_map_get_ptr(map, l);
 889                 map->ops->map_fd_put_ptr(ptr);
 890         }
 891 }
 892
 893 static bool is_map_full(struct bpf_htab *htab)
 894 {
 895         if (htab->use_percpu_counter)
 896                 return __percpu_counter_compare(&htab->pcount, htab->map.max_entries,
 897                                                 PERCPU_COUNTER_BATCH) >= 0;
 898         return atomic_read(&htab->count) >= htab->map.max_entries;
 899 }
 900
 901 static void inc_elem_count(struct bpf_htab *htab)
 902 {
 903         if (htab->use_percpu_counter)
 904                 percpu_counter_add_batch(&htab->pcount, 1, PERCPU_COUNTER_BATCH);
 905         else
 906                 atomic_inc(&htab->count);
 907 }
 908
 909 static void dec_elem_count(struct bpf_htab *htab)
 910 {
 911         if (htab->use_percpu_counter)
 912                 percpu_counter_add_batch(&htab->pcount, -1, PERCPU_COUNTER_BATCH);
 913         else
 914                 atomic_dec(&htab->count);
 915 }
 916
 917
 918 static void free_htab_elem(struct bpf_htab *htab, struct htab_elem *l)
 919 {
 920         htab_put_fd_value(htab, l);
 921
 922         if (htab_is_prealloc(htab)) {
 923                 check_and_free_fields(htab, l);
 924                 __pcpu_freelist_push(&htab->freelist, &l->fnode);
 925         } else {
 926                 dec_elem_count(htab);
 927                 htab_elem_free(htab, l);
 928         }
 929 }
 930
 931 static void pcpu_copy_value(struct bpf_htab *htab, void __percpu *pptr,
 932                             void *value, bool onallcpus)
 933 {
 934         if (!onallcpus) {
 935                 /* copy true value_size bytes */
 936                 copy_map_value(&htab->map, this_cpu_ptr(pptr), value);
 937         } else {
 938                 u32 size = round_up(htab->map.value_size, 8);
 939                 int off = 0, cpu;
 940
 941                 for_each_possible_cpu(cpu) {
 942                         copy_map_value_long(&htab->map, per_cpu_ptr(pptr, cpu), value + off);
 943                         off += size;
 944                 }
 945         }
 946 }
 947
 948 static void pcpu_init_value(struct bpf_htab *htab, void __percpu *pptr,
 949                             void *value, bool onallcpus)
 950 {
 951         /* When not setting the initial value on all cpus, zero-fill element
 952          * values for other cpus. Otherwise, bpf program has no way to ensure
 953          * known initial values for cpus other than current one
 954          * (onallcpus=false always when coming from bpf prog).
 955          */
 956         if (!onallcpus) {
 957                 int current_cpu = raw_smp_processor_id();
 958                 int cpu;
 959
 960                 for_each_possible_cpu(cpu) {
 961                         if (cpu == current_cpu)
 962                                 copy_map_value_long(&htab->map, per_cpu_ptr(pptr, cpu), value);
 963                         else /* Since elem is preallocated, we cannot touch special fields */
 964                                 zero_map_value(&htab->map, per_cpu_ptr(pptr, cpu));
 965                 }
 966         } else {
 967                 pcpu_copy_value(htab, pptr, value, onallcpus);
 968         }
 969 }
 970
 971 static bool fd_htab_map_needs_adjust(const struct bpf_htab *htab)
 972 {
 973         return htab->map.map_type == BPF_MAP_TYPE_HASH_OF_MAPS &&
 974                BITS_PER_LONG == 64;
 975 }
 976
 977 static struct htab_elem *alloc_htab_elem(struct bpf_htab *htab, void *key,
 978                                          void *value, u32 key_size, u32 hash,
 979                                          bool percpu, bool onallcpus,
 980                                          struct htab_elem *old_elem)
 981 {
 982         u32 size = htab->map.value_size;
 983         bool prealloc = htab_is_prealloc(htab);
 984         struct htab_elem *l_new, **pl_new;
 985         void __percpu *pptr;
 986
 987         if (prealloc) {
 988                 if (old_elem) {
 989                         /* if we're updating the existing element,
 990                          * use per-cpu extra elems to avoid freelist_pop/push
 991                          */
 992                         pl_new = this_cpu_ptr(htab->extra_elems);
 993                         l_new = *pl_new;
 994                         htab_put_fd_value(htab, old_elem);
 995                         *pl_new = old_elem;
 996                 } else {
 997                         struct pcpu_freelist_node *l;
 998
 999                         l = __pcpu_freelist_pop(&htab->freelist);
1000                         if (!l)
1001                                 return ERR_PTR(-E2BIG);
1002                         l_new = container_of(l, struct htab_elem, fnode);
1003                 }
1004         } else {
1005                 if (is_map_full(htab))
1006                         if (!old_elem)
1007                                 /* when map is full and update() is replacing
1008                                  * old element, it's ok to allocate, since
1009                                  * old element will be freed immediately.
1010                                  * Otherwise return an error
1011                                  */
1012                                 return ERR_PTR(-E2BIG);
1013                 inc_elem_count(htab);
1014                 l_new = bpf_mem_cache_alloc(&htab->ma);
1015                 if (!l_new) {
1016                         l_new = ERR_PTR(-ENOMEM);
1017                         goto dec_count;
1018                 }
1019         }
1020
1021         memcpy(l_new->key, key, key_size);
1022         if (percpu) {
1023                 if (prealloc) {
1024                         pptr = htab_elem_get_ptr(l_new, key_size);
1025                 } else {
1026                         /* alloc_percpu zero-fills */
1027                         pptr = bpf_mem_cache_alloc(&htab->pcpu_ma);
1028                         if (!pptr) {
1029                                 bpf_mem_cache_free(&htab->ma, l_new);
1030                                 l_new = ERR_PTR(-ENOMEM);
1031                                 goto dec_count;
1032                         }
1033                         l_new->ptr_to_pptr = pptr;
1034                         pptr = *(void **)pptr;
1035                 }
1036
1037                 pcpu_init_value(htab, pptr, value, onallcpus);
1038
1039                 if (!prealloc)
1040                         htab_elem_set_ptr(l_new, key_size, pptr);
1041         } else if (fd_htab_map_needs_adjust(htab)) {
1042                 size = round_up(size, 8);
1043                 memcpy(l_new->key + round_up(key_size, 8), value, size);
1044         } else {
1045                 copy_map_value(&htab->map,
1046                                l_new->key + round_up(key_size, 8),
1047                                value);
1048         }
1049
1050         l_new->hash = hash;
1051         return l_new;
1052 dec_count:
1053         dec_elem_count(htab);
1054         return l_new;
1055 }
1056
1057 static int check_flags(struct bpf_htab *htab, struct htab_elem *l_old,
1058                        u64 map_flags)
1059 {
1060         if (l_old && (map_flags & ~BPF_F_LOCK) == BPF_NOEXIST)
1061                 /* elem already exists */
1062                 return -EEXIST;
1063
1064         if (!l_old && (map_flags & ~BPF_F_LOCK) == BPF_EXIST)
1065                 /* elem doesn't exist, cannot update it */
1066                 return -ENOENT;
1067
1068         return 0;
1069 }
1070
1071 /* Called from syscall or from eBPF program */
1072 static long htab_map_update_elem(struct bpf_map *map, void *key, void *value,
1073                                  u64 map_flags)
1074 {
1075         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1076         struct htab_elem *l_new = NULL, *l_old;
1077         struct hlist_nulls_head *head;
1078         unsigned long flags;
1079         struct bucket *b;
1080         u32 key_size, hash;
1081         int ret;
1082
1083         if (unlikely((map_flags & ~BPF_F_LOCK) > BPF_EXIST))
1084                 /* unknown flags */
1085                 return -EINVAL;
1086
1087         WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
1088                      !rcu_read_lock_bh_held());
1089
1090         key_size = map->key_size;
1091
1092         hash = htab_map_hash(key, key_size, htab->hashrnd);
1093
1094         b = __select_bucket(htab, hash);
1095         head = &b->head;
1096
1097         if (unlikely(map_flags & BPF_F_LOCK)) {
1098                 if (unlikely(!btf_record_has_field(map->record, BPF_SPIN_LOCK)))
1099                         return -EINVAL;
1100                 /* find an element without taking the bucket lock */
1101                 l_old = lookup_nulls_elem_raw(head, hash, key, key_size,
1102                                               htab->n_buckets);
1103                 ret = check_flags(htab, l_old, map_flags);
1104                 if (ret)
1105                         return ret;
1106                 if (l_old) {
1107                         /* grab the element lock and update value in place */
1108                         copy_map_value_locked(map,
1109                                               l_old->key + round_up(key_size, 8),
1110                                               value, false);
1111                         return 0;
1112                 }
1113                 /* fall through, grab the bucket lock and lookup again.
1114                  * 99.9% chance that the element won't be found,
1115                  * but second lookup under lock has to be done.
1116                  */
1117         }
1118
1119         ret = htab_lock_bucket(htab, b, hash, &flags);
1120         if (ret)
1121                 return ret;
1122
1123         l_old = lookup_elem_raw(head, hash, key, key_size);
1124
1125         ret = check_flags(htab, l_old, map_flags);
1126         if (ret)
1127                 goto err;
1128
1129         if (unlikely(l_old && (map_flags & BPF_F_LOCK))) {
1130                 /* first lookup without the bucket lock didn't find the element,
1131                  * but second lookup with the bucket lock found it.
1132                  * This case is highly unlikely, but has to be dealt with:
1133                  * grab the element lock in addition to the bucket lock
1134                  * and update element in place
1135                  */
1136                 copy_map_value_locked(map,
1137                                       l_old->key + round_up(key_size, 8),
1138                                       value, false);
1139                 ret = 0;
1140                 goto err;
1141         }
1142
1143         l_new = alloc_htab_elem(htab, key, value, key_size, hash, false, false,
1144                                 l_old);
1145         if (IS_ERR(l_new)) {
1146                 /* all pre-allocated elements are in use or memory exhausted */
1147                 ret = PTR_ERR(l_new);
1148                 goto err;
1149         }
1150
1151         /* add new element to the head of the list, so that
1152          * concurrent search will find it before old elem
1153          */
1154         hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1155         if (l_old) {
1156                 hlist_nulls_del_rcu(&l_old->hash_node);
1157                 if (!htab_is_prealloc(htab))
1158                         free_htab_elem(htab, l_old);
1159                 else
1160                         check_and_free_fields(htab, l_old);
1161         }
1162         ret = 0;
1163 err:
1164         htab_unlock_bucket(htab, b, hash, flags);
1165         return ret;
1166 }
1167
1168 static void htab_lru_push_free(struct bpf_htab *htab, struct htab_elem *elem)
1169 {
1170         check_and_free_fields(htab, elem);
1171         bpf_lru_push_free(&htab->lru, &elem->lru_node);
1172 }
1173
1174 static long htab_lru_map_update_elem(struct bpf_map *map, void *key, void *value,
1175                                      u64 map_flags)
1176 {
1177         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1178         struct htab_elem *l_new, *l_old = NULL;
1179         struct hlist_nulls_head *head;
1180         unsigned long flags;
1181         struct bucket *b;
1182         u32 key_size, hash;
1183         int ret;
1184
1185         if (unlikely(map_flags > BPF_EXIST))
1186                 /* unknown flags */
1187                 return -EINVAL;
1188
1189         WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
1190                      !rcu_read_lock_bh_held());
1191
1192         key_size = map->key_size;
1193
1194         hash = htab_map_hash(key, key_size, htab->hashrnd);
1195
1196         b = __select_bucket(htab, hash);
1197         head = &b->head;
1198
1199         /* For LRU, we need to alloc before taking bucket's
1200          * spinlock because getting free nodes from LRU may need
1201          * to remove older elements from htab and this removal
1202          * operation will need a bucket lock.
1203          */
1204         l_new = prealloc_lru_pop(htab, key, hash);
1205         if (!l_new)
1206                 return -ENOMEM;
1207         copy_map_value(&htab->map,
1208                        l_new->key + round_up(map->key_size, 8), value);
1209
1210         ret = htab_lock_bucket(htab, b, hash, &flags);
1211         if (ret)
1212                 goto err_lock_bucket;
1213
1214         l_old = lookup_elem_raw(head, hash, key, key_size);
1215
1216         ret = check_flags(htab, l_old, map_flags);
1217         if (ret)
1218                 goto err;
1219
1220         /* add new element to the head of the list, so that
1221          * concurrent search will find it before old elem
1222          */
1223         hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1224         if (l_old) {
1225                 bpf_lru_node_set_ref(&l_new->lru_node);
1226                 hlist_nulls_del_rcu(&l_old->hash_node);
1227         }
1228         ret = 0;
1229
1230 err:
1231         htab_unlock_bucket(htab, b, hash, flags);
1232
1233 err_lock_bucket:
1234         if (ret)
1235                 htab_lru_push_free(htab, l_new);
1236         else if (l_old)
1237                 htab_lru_push_free(htab, l_old);
1238
1239         return ret;
1240 }
1241
1242 static long __htab_percpu_map_update_elem(struct bpf_map *map, void *key,
1243                                           void *value, u64 map_flags,
1244                                           bool onallcpus)
1245 {
1246         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1247         struct htab_elem *l_new = NULL, *l_old;
1248         struct hlist_nulls_head *head;
1249         unsigned long flags;
1250         struct bucket *b;
1251         u32 key_size, hash;
1252         int ret;
1253
1254         if (unlikely(map_flags > BPF_EXIST))
1255                 /* unknown flags */
1256                 return -EINVAL;
1257
1258         WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
1259                      !rcu_read_lock_bh_held());
1260
1261         key_size = map->key_size;
1262
1263         hash = htab_map_hash(key, key_size, htab->hashrnd);
1264
1265         b = __select_bucket(htab, hash);
1266         head = &b->head;
1267
1268         ret = htab_lock_bucket(htab, b, hash, &flags);
1269         if (ret)
1270                 return ret;
1271
1272         l_old = lookup_elem_raw(head, hash, key, key_size);
1273
1274         ret = check_flags(htab, l_old, map_flags);
1275         if (ret)
1276                 goto err;
1277
1278         if (l_old) {
1279                 /* per-cpu hash map can update value in-place */
1280                 pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
1281                                 value, onallcpus);
1282         } else {
1283                 l_new = alloc_htab_elem(htab, key, value, key_size,
1284                                         hash, true, onallcpus, NULL);
1285                 if (IS_ERR(l_new)) {
1286                         ret = PTR_ERR(l_new);
1287                         goto err;
1288                 }
1289                 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1290         }
1291         ret = 0;
1292 err:
1293         htab_unlock_bucket(htab, b, hash, flags);
1294         return ret;
1295 }
1296
1297 static long __htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
1298                                               void *value, u64 map_flags,
1299                                               bool onallcpus)
1300 {
1301         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1302         struct htab_elem *l_new = NULL, *l_old;
1303         struct hlist_nulls_head *head;
1304         unsigned long flags;
1305         struct bucket *b;
1306         u32 key_size, hash;
1307         int ret;
1308
1309         if (unlikely(map_flags > BPF_EXIST))
1310                 /* unknown flags */
1311                 return -EINVAL;
1312
1313         WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
1314                      !rcu_read_lock_bh_held());
1315
1316         key_size = map->key_size;
1317
1318         hash = htab_map_hash(key, key_size, htab->hashrnd);
1319
1320         b = __select_bucket(htab, hash);
1321         head = &b->head;
1322
1323         /* For LRU, we need to alloc before taking bucket's
1324          * spinlock because LRU's elem alloc may need
1325          * to remove older elem from htab and this removal
1326          * operation will need a bucket lock.
1327          */
1328         if (map_flags != BPF_EXIST) {
1329                 l_new = prealloc_lru_pop(htab, key, hash);
1330                 if (!l_new)
1331                         return -ENOMEM;
1332         }
1333
1334         ret = htab_lock_bucket(htab, b, hash, &flags);
1335         if (ret)
1336                 goto err_lock_bucket;
1337
1338         l_old = lookup_elem_raw(head, hash, key, key_size);
1339
1340         ret = check_flags(htab, l_old, map_flags);
1341         if (ret)
1342                 goto err;
1343
1344         if (l_old) {
1345                 bpf_lru_node_set_ref(&l_old->lru_node);
1346
1347                 /* per-cpu hash map can update value in-place */
1348                 pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
1349                                 value, onallcpus);
1350         } else {
1351                 pcpu_init_value(htab, htab_elem_get_ptr(l_new, key_size),
1352                                 value, onallcpus);
1353                 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1354                 l_new = NULL;
1355         }
1356         ret = 0;
1357 err:
1358         htab_unlock_bucket(htab, b, hash, flags);
1359 err_lock_bucket:
1360         if (l_new)
1361                 bpf_lru_push_free(&htab->lru, &l_new->lru_node);
1362         return ret;
1363 }
1364
1365 static long htab_percpu_map_update_elem(struct bpf_map *map, void *key,
1366                                         void *value, u64 map_flags)
1367 {
1368         return __htab_percpu_map_update_elem(map, key, value, map_flags, false);
1369 }
1370
1371 static long htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
1372                                             void *value, u64 map_flags)
1373 {
1374         return __htab_lru_percpu_map_update_elem(map, key, value, map_flags,
1375                                                  false);
1376 }
1377
1378 /* Called from syscall or from eBPF program */
1379 static long htab_map_delete_elem(struct bpf_map *map, void *key)
1380 {
1381         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1382         struct hlist_nulls_head *head;
1383         struct bucket *b;
1384         struct htab_elem *l;
1385         unsigned long flags;
1386         u32 hash, key_size;
1387         int ret;
1388
1389         WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
1390                      !rcu_read_lock_bh_held());
1391
1392         key_size = map->key_size;
1393
1394         hash = htab_map_hash(key, key_size, htab->hashrnd);
1395         b = __select_bucket(htab, hash);
1396         head = &b->head;
1397
1398         ret = htab_lock_bucket(htab, b, hash, &flags);
1399         if (ret)
1400                 return ret;
1401
1402         l = lookup_elem_raw(head, hash, key, key_size);
1403
1404         if (l) {
1405                 hlist_nulls_del_rcu(&l->hash_node);
1406                 free_htab_elem(htab, l);
1407         } else {
1408                 ret = -ENOENT;
1409         }
1410
1411         htab_unlock_bucket(htab, b, hash, flags);
1412         return ret;
1413 }
1414
1415 static long htab_lru_map_delete_elem(struct bpf_map *map, void *key)
1416 {
1417         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1418         struct hlist_nulls_head *head;
1419         struct bucket *b;
1420         struct htab_elem *l;
1421         unsigned long flags;
1422         u32 hash, key_size;
1423         int ret;
1424
1425         WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
1426                      !rcu_read_lock_bh_held());
1427
1428         key_size = map->key_size;
1429
1430         hash = htab_map_hash(key, key_size, htab->hashrnd);
1431         b = __select_bucket(htab, hash);
1432         head = &b->head;
1433
1434         ret = htab_lock_bucket(htab, b, hash, &flags);
1435         if (ret)
1436                 return ret;
1437
1438         l = lookup_elem_raw(head, hash, key, key_size);
1439
1440         if (l)
1441                 hlist_nulls_del_rcu(&l->hash_node);
1442         else
1443                 ret = -ENOENT;
1444
1445         htab_unlock_bucket(htab, b, hash, flags);
1446         if (l)
1447                 htab_lru_push_free(htab, l);
1448         return ret;
1449 }
1450
1451 static void delete_all_elements(struct bpf_htab *htab)
1452 {
1453         int i;
1454
1455         /* It's called from a worker thread, so disable migration here,
1456          * since bpf_mem_cache_free() relies on that.
1457          */
1458         migrate_disable();
1459         for (i = 0; i < htab->n_buckets; i++) {
1460                 struct hlist_nulls_head *head = select_bucket(htab, i);
1461                 struct hlist_nulls_node *n;
1462                 struct htab_elem *l;
1463
1464                 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1465                         hlist_nulls_del_rcu(&l->hash_node);
1466                         htab_elem_free(htab, l);
1467                 }
1468         }
1469         migrate_enable();
1470 }
1471
1472 static void htab_free_malloced_timers(struct bpf_htab *htab)
1473 {
1474         int i;
1475
1476         rcu_read_lock();
1477         for (i = 0; i < htab->n_buckets; i++) {
1478                 struct hlist_nulls_head *head = select_bucket(htab, i);
1479                 struct hlist_nulls_node *n;
1480                 struct htab_elem *l;
1481
1482                 hlist_nulls_for_each_entry(l, n, head, hash_node) {
1483                         /* We only free timer on uref dropping to zero */
1484                         bpf_obj_free_timer(htab->map.record, l->key + round_up(htab->map.key_size, 8));
1485                 }
1486                 cond_resched_rcu();
1487         }
1488         rcu_read_unlock();
1489 }
1490
1491 static void htab_map_free_timers(struct bpf_map *map)
1492 {
1493         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1494
1495         /* We only free timer on uref dropping to zero */
1496         if (!btf_record_has_field(htab->map.record, BPF_TIMER))
1497                 return;
1498         if (!htab_is_prealloc(htab))
1499                 htab_free_malloced_timers(htab);
1500         else
1501                 htab_free_prealloced_timers(htab);
1502 }
1503
1504 /* Called when map->refcnt goes to zero, either from workqueue or from syscall */
1505 static void htab_map_free(struct bpf_map *map)
1506 {
1507         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1508         int i;
1509
1510         /* bpf_free_used_maps() or close(map_fd) will trigger this map_free callback.
1511          * bpf_free_used_maps() is called after bpf prog is no longer executing.
1512          * There is no need to synchronize_rcu() here to protect map elements.
1513          */
1514
1515         /* htab no longer uses call_rcu() directly. bpf_mem_alloc does it
1516          * underneath and is reponsible for waiting for callbacks to finish
1517          * during bpf_mem_alloc_destroy().
1518          */
1519         if (!htab_is_prealloc(htab)) {
1520                 delete_all_elements(htab);
1521         } else {
1522                 htab_free_prealloced_fields(htab);
1523                 prealloc_destroy(htab);
1524         }
1525
1526         free_percpu(htab->extra_elems);
1527         bpf_map_area_free(htab->buckets);
1528         bpf_mem_alloc_destroy(&htab->pcpu_ma);
1529         bpf_mem_alloc_destroy(&htab->ma);
1530         if (htab->use_percpu_counter)
1531                 percpu_counter_destroy(&htab->pcount);
1532         for (i = 0; i < HASHTAB_MAP_LOCK_COUNT; i++)
1533                 free_percpu(htab->map_locked[i]);
1534         lockdep_unregister_key(&htab->lockdep_key);
1535         bpf_map_area_free(htab);
1536 }
1537
1538 static void htab_map_seq_show_elem(struct bpf_map *map, void *key,
1539                                    struct seq_file *m)
1540 {
1541         void *value;
1542
1543         rcu_read_lock();
1544
1545         value = htab_map_lookup_elem(map, key);
1546         if (!value) {
1547                 rcu_read_unlock();
1548                 return;
1549         }
1550
1551         btf_type_seq_show(map->btf, map->btf_key_type_id, key, m);
1552         seq_puts(m, ": ");
1553         btf_type_seq_show(map->btf, map->btf_value_type_id, value, m);
1554         seq_puts(m, "\n");
1555
1556         rcu_read_unlock();
1557 }
1558
1559 static int __htab_map_lookup_and_delete_elem(struct bpf_map *map, void *key,
1560                                              void *value, bool is_lru_map,
1561                                              bool is_percpu, u64 flags)
1562 {
1563         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1564         struct hlist_nulls_head *head;
1565         unsigned long bflags;
1566         struct htab_elem *l;
1567         u32 hash, key_size;
1568         struct bucket *b;
1569         int ret;
1570
1571         key_size = map->key_size;
1572
1573         hash = htab_map_hash(key, key_size, htab->hashrnd);
1574         b = __select_bucket(htab, hash);
1575         head = &b->head;
1576
1577         ret = htab_lock_bucket(htab, b, hash, &bflags);
1578         if (ret)
1579                 return ret;
1580
1581         l = lookup_elem_raw(head, hash, key, key_size);
1582         if (!l) {
1583                 ret = -ENOENT;
1584         } else {
1585                 if (is_percpu) {
1586                         u32 roundup_value_size = round_up(map->value_size, 8);
1587                         void __percpu *pptr;
1588                         int off = 0, cpu;
1589
1590                         pptr = htab_elem_get_ptr(l, key_size);
1591                         for_each_possible_cpu(cpu) {
1592                                 copy_map_value_long(&htab->map, value + off, per_cpu_ptr(pptr, cpu));
1593                                 check_and_init_map_value(&htab->map, value + off);
1594                                 off += roundup_value_size;
1595                         }
1596                 } else {
1597                         u32 roundup_key_size = round_up(map->key_size, 8);
1598
1599                         if (flags & BPF_F_LOCK)
1600                                 copy_map_value_locked(map, value, l->key +
1601                                                       roundup_key_size,
1602                                                       true);
1603                         else
1604                                 copy_map_value(map, value, l->key +
1605                                                roundup_key_size);
1606                         /* Zeroing special fields in the temp buffer */
1607                         check_and_init_map_value(map, value);
1608                 }
1609
1610                 hlist_nulls_del_rcu(&l->hash_node);
1611                 if (!is_lru_map)
1612                         free_htab_elem(htab, l);
1613         }
1614
1615         htab_unlock_bucket(htab, b, hash, bflags);
1616
1617         if (is_lru_map && l)
1618                 htab_lru_push_free(htab, l);
1619
1620         return ret;
1621 }
1622
1623 static int htab_map_lookup_and_delete_elem(struct bpf_map *map, void *key,
1624                                            void *value, u64 flags)
1625 {
1626         return __htab_map_lookup_and_delete_elem(map, key, value, false, false,
1627                                                  flags);
1628 }
1629
1630 static int htab_percpu_map_lookup_and_delete_elem(struct bpf_map *map,
1631                                                   void *key, void *value,
1632                                                   u64 flags)
1633 {
1634         return __htab_map_lookup_and_delete_elem(map, key, value, false, true,
1635                                                  flags);
1636 }
1637
1638 static int htab_lru_map_lookup_and_delete_elem(struct bpf_map *map, void *key,
1639                                                void *value, u64 flags)
1640 {
1641         return __htab_map_lookup_and_delete_elem(map, key, value, true, false,
1642                                                  flags);
1643 }
1644
1645 static int htab_lru_percpu_map_lookup_and_delete_elem(struct bpf_map *map,
1646                                                       void *key, void *value,
1647                                                       u64 flags)
1648 {
1649         return __htab_map_lookup_and_delete_elem(map, key, value, true, true,
1650                                                  flags);
1651 }
1652
1653 static int
1654 __htab_map_lookup_and_delete_batch(struct bpf_map *map,
1655                                    const union bpf_attr *attr,
1656                                    union bpf_attr __user *uattr,
1657                                    bool do_delete, bool is_lru_map,
1658                                    bool is_percpu)
1659 {
1660         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1661         u32 bucket_cnt, total, key_size, value_size, roundup_key_size;
1662         void *keys = NULL, *values = NULL, *value, *dst_key, *dst_val;
1663         void __user *uvalues = u64_to_user_ptr(attr->batch.values);
1664         void __user *ukeys = u64_to_user_ptr(attr->batch.keys);
1665         void __user *ubatch = u64_to_user_ptr(attr->batch.in_batch);
1666         u32 batch, max_count, size, bucket_size, map_id;
1667         struct htab_elem *node_to_free = NULL;
1668         u64 elem_map_flags, map_flags;
1669         struct hlist_nulls_head *head;
1670         struct hlist_nulls_node *n;
1671         unsigned long flags = 0;
1672         bool locked = false;
1673         struct htab_elem *l;
1674         struct bucket *b;
1675         int ret = 0;
1676
1677         elem_map_flags = attr->batch.elem_flags;
1678         if ((elem_map_flags & ~BPF_F_LOCK) ||
1679             ((elem_map_flags & BPF_F_LOCK) && !btf_record_has_field(map->record, BPF_SPIN_LOCK)))
1680                 return -EINVAL;
1681
1682         map_flags = attr->batch.flags;
1683         if (map_flags)
1684                 return -EINVAL;
1685
1686         max_count = attr->batch.count;
1687         if (!max_count)
1688                 return 0;
1689
1690         if (put_user(0, &uattr->batch.count))
1691                 return -EFAULT;
1692
1693         batch = 0;
1694         if (ubatch && copy_from_user(&batch, ubatch, sizeof(batch)))
1695                 return -EFAULT;
1696
1697         if (batch >= htab->n_buckets)
1698                 return -ENOENT;
1699
1700         key_size = htab->map.key_size;
1701         roundup_key_size = round_up(htab->map.key_size, 8);
1702         value_size = htab->map.value_size;
1703         size = round_up(value_size, 8);
1704         if (is_percpu)
1705                 value_size = size * num_possible_cpus();
1706         total = 0;
1707         /* while experimenting with hash tables with sizes ranging from 10 to
1708          * 1000, it was observed that a bucket can have up to 5 entries.
1709          */
1710         bucket_size = 5;
1711
1712 alloc:
1713         /* We cannot do copy_from_user or copy_to_user inside
1714          * the rcu_read_lock. Allocate enough space here.
1715          */
1716         keys = kvmalloc_array(key_size, bucket_size, GFP_USER | __GFP_NOWARN);
1717         values = kvmalloc_array(value_size, bucket_size, GFP_USER | __GFP_NOWARN);
1718         if (!keys || !values) {
1719                 ret = -ENOMEM;
1720                 goto after_loop;
1721         }
1722
1723 again:
1724         bpf_disable_instrumentation();
1725         rcu_read_lock();
1726 again_nocopy:
1727         dst_key = keys;
1728         dst_val = values;
1729         b = &htab->buckets[batch];
1730         head = &b->head;
1731         /* do not grab the lock unless need it (bucket_cnt > 0). */
1732         if (locked) {
1733                 ret = htab_lock_bucket(htab, b, batch, &flags);
1734                 if (ret) {
1735                         rcu_read_unlock();
1736                         bpf_enable_instrumentation();
1737                         goto after_loop;
1738                 }
1739         }
1740
1741         bucket_cnt = 0;
1742         hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
1743                 bucket_cnt++;
1744
1745         if (bucket_cnt && !locked) {
1746                 locked = true;
1747                 goto again_nocopy;
1748         }
1749
1750         if (bucket_cnt > (max_count - total)) {
1751                 if (total == 0)
1752                         ret = -ENOSPC;
1753                 /* Note that since bucket_cnt > 0 here, it is implicit
1754                  * that the locked was grabbed, so release it.
1755                  */
1756                 htab_unlock_bucket(htab, b, batch, flags);
1757                 rcu_read_unlock();
1758                 bpf_enable_instrumentation();
1759                 goto after_loop;
1760         }
1761
1762         if (bucket_cnt > bucket_size) {
1763                 bucket_size = bucket_cnt;
1764                 /* Note that since bucket_cnt > 0 here, it is implicit
1765                  * that the locked was grabbed, so release it.
1766                  */
1767                 htab_unlock_bucket(htab, b, batch, flags);
1768                 rcu_read_unlock();
1769                 bpf_enable_instrumentation();
1770                 kvfree(keys);
1771                 kvfree(values);
1772                 goto alloc;
1773         }
1774
1775         /* Next block is only safe to run if you have grabbed the lock */
1776         if (!locked)
1777                 goto next_batch;
1778
1779         hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1780                 memcpy(dst_key, l->key, key_size);
1781
1782                 if (is_percpu) {
1783                         int off = 0, cpu;
1784                         void __percpu *pptr;
1785
1786                         pptr = htab_elem_get_ptr(l, map->key_size);
1787                         for_each_possible_cpu(cpu) {
1788                                 copy_map_value_long(&htab->map, dst_val + off, per_cpu_ptr(pptr, cpu));
1789                                 check_and_init_map_value(&htab->map, dst_val + off);
1790                                 off += size;
1791                         }
1792                 } else {
1793                         value = l->key + roundup_key_size;
1794                         if (map->map_type == BPF_MAP_TYPE_HASH_OF_MAPS) {
1795                                 struct bpf_map **inner_map = value;
1796
1797                                  /* Actual value is the id of the inner map */
1798                                 map_id = map->ops->map_fd_sys_lookup_elem(*inner_map);
1799                                 value = &map_id;
1800                         }
1801
1802                         if (elem_map_flags & BPF_F_LOCK)
1803                                 copy_map_value_locked(map, dst_val, value,
1804                                                       true);
1805                         else
1806                                 copy_map_value(map, dst_val, value);
1807                         /* Zeroing special fields in the temp buffer */
1808                         check_and_init_map_value(map, dst_val);
1809                 }
1810                 if (do_delete) {
1811                         hlist_nulls_del_rcu(&l->hash_node);
1812
1813                         /* bpf_lru_push_free() will acquire lru_lock, which
1814                          * may cause deadlock. See comments in function
1815                          * prealloc_lru_pop(). Let us do bpf_lru_push_free()
1816                          * after releasing the bucket lock.
1817                          */
1818                         if (is_lru_map) {
1819                                 l->batch_flink = node_to_free;
1820                                 node_to_free = l;
1821                         } else {
1822                                 free_htab_elem(htab, l);
1823                         }
1824                 }
1825                 dst_key += key_size;
1826                 dst_val += value_size;
1827         }
1828
1829         htab_unlock_bucket(htab, b, batch, flags);
1830         locked = false;
1831
1832         while (node_to_free) {
1833                 l = node_to_free;
1834                 node_to_free = node_to_free->batch_flink;
1835                 htab_lru_push_free(htab, l);
1836         }
1837
1838 next_batch:
1839         /* If we are not copying data, we can go to next bucket and avoid
1840          * unlocking the rcu.
1841          */
1842         if (!bucket_cnt && (batch + 1 < htab->n_buckets)) {
1843                 batch++;
1844                 goto again_nocopy;
1845         }
1846
1847         rcu_read_unlock();
1848         bpf_enable_instrumentation();
1849         if (bucket_cnt && (copy_to_user(ukeys + total * key_size, keys,
1850             key_size * bucket_cnt) ||
1851             copy_to_user(uvalues + total * value_size, values,
1852             value_size * bucket_cnt))) {
1853                 ret = -EFAULT;
1854                 goto after_loop;
1855         }
1856
1857         total += bucket_cnt;
1858         batch++;
1859         if (batch >= htab->n_buckets) {
1860                 ret = -ENOENT;
1861                 goto after_loop;
1862         }
1863         goto again;
1864
1865 after_loop:
1866         if (ret == -EFAULT)
1867                 goto out;
1868
1869         /* copy # of entries and next batch */
1870         ubatch = u64_to_user_ptr(attr->batch.out_batch);
1871         if (copy_to_user(ubatch, &batch, sizeof(batch)) ||
1872             put_user(total, &uattr->batch.count))
1873                 ret = -EFAULT;
1874
1875 out:
1876         kvfree(keys);
1877         kvfree(values);
1878         return ret;
1879 }
1880
1881 static int
1882 htab_percpu_map_lookup_batch(struct bpf_map *map, const union bpf_attr *attr,
1883                              union bpf_attr __user *uattr)
1884 {
1885         return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1886                                                   false, true);
1887 }
1888
1889 static int
1890 htab_percpu_map_lookup_and_delete_batch(struct bpf_map *map,
1891                                         const union bpf_attr *attr,
1892                                         union bpf_attr __user *uattr)
1893 {
1894         return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1895                                                   false, true);
1896 }
1897
1898 static int
1899 htab_map_lookup_batch(struct bpf_map *map, const union bpf_attr *attr,
1900                       union bpf_attr __user *uattr)
1901 {
1902         return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1903                                                   false, false);
1904 }
1905
1906 static int
1907 htab_map_lookup_and_delete_batch(struct bpf_map *map,
1908                                  const union bpf_attr *attr,
1909                                  union bpf_attr __user *uattr)
1910 {
1911         return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1912                                                   false, false);
1913 }
1914
1915 static int
1916 htab_lru_percpu_map_lookup_batch(struct bpf_map *map,
1917                                  const union bpf_attr *attr,
1918                                  union bpf_attr __user *uattr)
1919 {
1920         return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1921                                                   true, true);
1922 }
1923
1924 static int
1925 htab_lru_percpu_map_lookup_and_delete_batch(struct bpf_map *map,
1926                                             const union bpf_attr *attr,
1927                                             union bpf_attr __user *uattr)
1928 {
1929         return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1930                                                   true, true);
1931 }
1932
1933 static int
1934 htab_lru_map_lookup_batch(struct bpf_map *map, const union bpf_attr *attr,
1935                           union bpf_attr __user *uattr)
1936 {
1937         return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1938                                                   true, false);
1939 }
1940
1941 static int
1942 htab_lru_map_lookup_and_delete_batch(struct bpf_map *map,
1943                                      const union bpf_attr *attr,
1944                                      union bpf_attr __user *uattr)
1945 {
1946         return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1947                                                   true, false);
1948 }
1949
1950 struct bpf_iter_seq_hash_map_info {
1951         struct bpf_map *map;
1952         struct bpf_htab *htab;
1953         void *percpu_value_buf; // non-zero means percpu hash
1954         u32 bucket_id;
1955         u32 skip_elems;
1956 };
1957
1958 static struct htab_elem *
1959 bpf_hash_map_seq_find_next(struct bpf_iter_seq_hash_map_info *info,
1960                            struct htab_elem *prev_elem)
1961 {
1962         const struct bpf_htab *htab = info->htab;
1963         u32 skip_elems = info->skip_elems;
1964         u32 bucket_id = info->bucket_id;
1965         struct hlist_nulls_head *head;
1966         struct hlist_nulls_node *n;
1967         struct htab_elem *elem;
1968         struct bucket *b;
1969         u32 i, count;
1970
1971         if (bucket_id >= htab->n_buckets)
1972                 return NULL;
1973
1974         /* try to find next elem in the same bucket */
1975         if (prev_elem) {
1976                 /* no update/deletion on this bucket, prev_elem should be still valid
1977                  * and we won't skip elements.
1978                  */
1979                 n = rcu_dereference_raw(hlist_nulls_next_rcu(&prev_elem->hash_node));
1980                 elem = hlist_nulls_entry_safe(n, struct htab_elem, hash_node);
1981                 if (elem)
1982                         return elem;
1983
1984                 /* not found, unlock and go to the next bucket */
1985                 b = &htab->buckets[bucket_id++];
1986                 rcu_read_unlock();
1987                 skip_elems = 0;
1988         }
1989
1990         for (i = bucket_id; i < htab->n_buckets; i++) {
1991                 b = &htab->buckets[i];
1992                 rcu_read_lock();
1993
1994                 count = 0;
1995                 head = &b->head;
1996                 hlist_nulls_for_each_entry_rcu(elem, n, head, hash_node) {
1997                         if (count >= skip_elems) {
1998                                 info->bucket_id = i;
1999                                 info->skip_elems = count;
2000                                 return elem;
2001                         }
2002                         count++;
2003                 }
2004
2005                 rcu_read_unlock();
2006                 skip_elems = 0;
2007         }
2008
2009         info->bucket_id = i;
2010         info->skip_elems = 0;
2011         return NULL;
2012 }
2013
2014 static void *bpf_hash_map_seq_start(struct seq_file *seq, loff_t *pos)
2015 {
2016         struct bpf_iter_seq_hash_map_info *info = seq->private;
2017         struct htab_elem *elem;
2018
2019         elem = bpf_hash_map_seq_find_next(info, NULL);
2020         if (!elem)
2021                 return NULL;
2022
2023         if (*pos == 0)
2024                 ++*pos;
2025         return elem;
2026 }
2027
2028 static void *bpf_hash_map_seq_next(struct seq_file *seq, void *v, loff_t *pos)
2029 {
2030         struct bpf_iter_seq_hash_map_info *info = seq->private;
2031
2032         ++*pos;
2033         ++info->skip_elems;
2034         return bpf_hash_map_seq_find_next(info, v);
2035 }
2036
2037 static int __bpf_hash_map_seq_show(struct seq_file *seq, struct htab_elem *elem)
2038 {
2039         struct bpf_iter_seq_hash_map_info *info = seq->private;
2040         u32 roundup_key_size, roundup_value_size;
2041         struct bpf_iter__bpf_map_elem ctx = {};
2042         struct bpf_map *map = info->map;
2043         struct bpf_iter_meta meta;
2044         int ret = 0, off = 0, cpu;
2045         struct bpf_prog *prog;
2046         void __percpu *pptr;
2047
2048         meta.seq = seq;
2049         prog = bpf_iter_get_info(&meta, elem == NULL);
2050         if (prog) {
2051                 ctx.meta = &meta;
2052                 ctx.map = info->map;
2053                 if (elem) {
2054                         roundup_key_size = round_up(map->key_size, 8);
2055                         ctx.key = elem->key;
2056                         if (!info->percpu_value_buf) {
2057                                 ctx.value = elem->key + roundup_key_size;
2058                         } else {
2059                                 roundup_value_size = round_up(map->value_size, 8);
2060                                 pptr = htab_elem_get_ptr(elem, map->key_size);
2061                                 for_each_possible_cpu(cpu) {
2062                                         copy_map_value_long(map, info->percpu_value_buf + off,
2063                                                             per_cpu_ptr(pptr, cpu));
2064                                         check_and_init_map_value(map, info->percpu_value_buf + off);
2065                                         off += roundup_value_size;
2066                                 }
2067                                 ctx.value = info->percpu_value_buf;
2068                         }
2069                 }
2070                 ret = bpf_iter_run_prog(prog, &ctx);
2071         }
2072
2073         return ret;
2074 }
2075
2076 static int bpf_hash_map_seq_show(struct seq_file *seq, void *v)
2077 {
2078         return __bpf_hash_map_seq_show(seq, v);
2079 }
2080
2081 static void bpf_hash_map_seq_stop(struct seq_file *seq, void *v)
2082 {
2083         if (!v)
2084                 (void)__bpf_hash_map_seq_show(seq, NULL);
2085         else
2086                 rcu_read_unlock();
2087 }
2088
2089 static int bpf_iter_init_hash_map(void *priv_data,
2090                                   struct bpf_iter_aux_info *aux)
2091 {
2092         struct bpf_iter_seq_hash_map_info *seq_info = priv_data;
2093         struct bpf_map *map = aux->map;
2094         void *value_buf;
2095         u32 buf_size;
2096
2097         if (map->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
2098             map->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH) {
2099                 buf_size = round_up(map->value_size, 8) * num_possible_cpus();
2100                 value_buf = kmalloc(buf_size, GFP_USER | __GFP_NOWARN);
2101                 if (!value_buf)
2102                         return -ENOMEM;
2103
2104                 seq_info->percpu_value_buf = value_buf;
2105         }
2106
2107         bpf_map_inc_with_uref(map);
2108         seq_info->map = map;
2109         seq_info->htab = container_of(map, struct bpf_htab, map);
2110         return 0;
2111 }
2112
2113 static void bpf_iter_fini_hash_map(void *priv_data)
2114 {
2115         struct bpf_iter_seq_hash_map_info *seq_info = priv_data;
2116
2117         bpf_map_put_with_uref(seq_info->map);
2118         kfree(seq_info->percpu_value_buf);
2119 }
2120
2121 static const struct seq_operations bpf_hash_map_seq_ops = {
2122         .start  = bpf_hash_map_seq_start,
2123         .next   = bpf_hash_map_seq_next,
2124         .stop   = bpf_hash_map_seq_stop,
2125         .show   = bpf_hash_map_seq_show,
2126 };
2127
2128 static const struct bpf_iter_seq_info iter_seq_info = {
2129         .seq_ops                = &bpf_hash_map_seq_ops,
2130         .init_seq_private       = bpf_iter_init_hash_map,
2131         .fini_seq_private       = bpf_iter_fini_hash_map,
2132         .seq_priv_size          = sizeof(struct bpf_iter_seq_hash_map_info),
2133 };
2134
2135 static long bpf_for_each_hash_elem(struct bpf_map *map, bpf_callback_t callback_fn,
2136                                    void *callback_ctx, u64 flags)
2137 {
2138         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
2139         struct hlist_nulls_head *head;
2140         struct hlist_nulls_node *n;
2141         struct htab_elem *elem;
2142         u32 roundup_key_size;
2143         int i, num_elems = 0;
2144         void __percpu *pptr;
2145         struct bucket *b;
2146         void *key, *val;
2147         bool is_percpu;
2148         u64 ret = 0;
2149
2150         if (flags != 0)
2151                 return -EINVAL;
2152
2153         is_percpu = htab_is_percpu(htab);
2154
2155         roundup_key_size = round_up(map->key_size, 8);
2156         /* disable migration so percpu value prepared here will be the
2157          * same as the one seen by the bpf program with bpf_map_lookup_elem().
2158          */
2159         if (is_percpu)
2160                 migrate_disable();
2161         for (i = 0; i < htab->n_buckets; i++) {
2162                 b = &htab->buckets[i];
2163                 rcu_read_lock();
2164                 head = &b->head;
2165                 hlist_nulls_for_each_entry_rcu(elem, n, head, hash_node) {
2166                         key = elem->key;
2167                         if (is_percpu) {
2168                                 /* current cpu value for percpu map */
2169                                 pptr = htab_elem_get_ptr(elem, map->key_size);
2170                                 val = this_cpu_ptr(pptr);
2171                         } else {
2172                                 val = elem->key + roundup_key_size;
2173                         }
2174                         num_elems++;
2175                         ret = callback_fn((u64)(long)map, (u64)(long)key,
2176                                           (u64)(long)val, (u64)(long)callback_ctx, 0);
2177                         /* return value: 0 - continue, 1 - stop and return */
2178                         if (ret) {
2179                                 rcu_read_unlock();
2180                                 goto out;
2181                         }
2182                 }
2183                 rcu_read_unlock();
2184         }
2185 out:
2186         if (is_percpu)
2187                 migrate_enable();
2188         return num_elems;
2189 }
2190
2191 static u64 htab_map_mem_usage(const struct bpf_map *map)
2192 {
2193         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
2194         u32 value_size = round_up(htab->map.value_size, 8);
2195         bool prealloc = htab_is_prealloc(htab);
2196         bool percpu = htab_is_percpu(htab);
2197         bool lru = htab_is_lru(htab);
2198         u64 num_entries;
2199         u64 usage = sizeof(struct bpf_htab);
2200
2201         usage += sizeof(struct bucket) * htab->n_buckets;
2202         usage += sizeof(int) * num_possible_cpus() * HASHTAB_MAP_LOCK_COUNT;
2203         if (prealloc) {
2204                 num_entries = map->max_entries;
2205                 if (htab_has_extra_elems(htab))
2206                         num_entries += num_possible_cpus();
2207
2208                 usage += htab->elem_size * num_entries;
2209
2210                 if (percpu)
2211                         usage += value_size * num_possible_cpus() * num_entries;
2212                 else if (!lru)
2213                         usage += sizeof(struct htab_elem *) * num_possible_cpus();
2214         } else {
2215 #define LLIST_NODE_SZ sizeof(struct llist_node)
2216
2217                 num_entries = htab->use_percpu_counter ?
2218                                           percpu_counter_sum(&htab->pcount) :
2219                                           atomic_read(&htab->count);
2220                 usage += (htab->elem_size + LLIST_NODE_SZ) * num_entries;
2221                 if (percpu) {
2222                         usage += (LLIST_NODE_SZ + sizeof(void *)) * num_entries;
2223                         usage += value_size * num_possible_cpus() * num_entries;
2224                 }
2225         }
2226         return usage;
2227 }
2228
2229 BTF_ID_LIST_SINGLE(htab_map_btf_ids, struct, bpf_htab)
2230 const struct bpf_map_ops htab_map_ops = {
2231         .map_meta_equal = bpf_map_meta_equal,
2232         .map_alloc_check = htab_map_alloc_check,
2233         .map_alloc = htab_map_alloc,
2234         .map_free = htab_map_free,
2235         .map_get_next_key = htab_map_get_next_key,
2236         .map_release_uref = htab_map_free_timers,
2237         .map_lookup_elem = htab_map_lookup_elem,
2238         .map_lookup_and_delete_elem = htab_map_lookup_and_delete_elem,
2239         .map_update_elem = htab_map_update_elem,
2240         .map_delete_elem = htab_map_delete_elem,
2241         .map_gen_lookup = htab_map_gen_lookup,
2242         .map_seq_show_elem = htab_map_seq_show_elem,
2243         .map_set_for_each_callback_args = map_set_for_each_callback_args,
2244         .map_for_each_callback = bpf_for_each_hash_elem,
2245         .map_mem_usage = htab_map_mem_usage,
2246         BATCH_OPS(htab),
2247         .map_btf_id = &htab_map_btf_ids[0],
2248         .iter_seq_info = &iter_seq_info,
2249 };
2250
2251 const struct bpf_map_ops htab_lru_map_ops = {
2252         .map_meta_equal = bpf_map_meta_equal,
2253         .map_alloc_check = htab_map_alloc_check,
2254         .map_alloc = htab_map_alloc,
2255         .map_free = htab_map_free,
2256         .map_get_next_key = htab_map_get_next_key,
2257         .map_release_uref = htab_map_free_timers,
2258         .map_lookup_elem = htab_lru_map_lookup_elem,
2259         .map_lookup_and_delete_elem = htab_lru_map_lookup_and_delete_elem,
2260         .map_lookup_elem_sys_only = htab_lru_map_lookup_elem_sys,
2261         .map_update_elem = htab_lru_map_update_elem,
2262         .map_delete_elem = htab_lru_map_delete_elem,
2263         .map_gen_lookup = htab_lru_map_gen_lookup,
2264         .map_seq_show_elem = htab_map_seq_show_elem,
2265         .map_set_for_each_callback_args = map_set_for_each_callback_args,
2266         .map_for_each_callback = bpf_for_each_hash_elem,
2267         .map_mem_usage = htab_map_mem_usage,
2268         BATCH_OPS(htab_lru),
2269         .map_btf_id = &htab_map_btf_ids[0],
2270         .iter_seq_info = &iter_seq_info,
2271 };
2272
2273 /* Called from eBPF program */
2274 static void *htab_percpu_map_lookup_elem(struct bpf_map *map, void *key)
2275 {
2276         struct htab_elem *l = __htab_map_lookup_elem(map, key);
2277
2278         if (l)
2279                 return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
2280         else
2281                 return NULL;
2282 }
2283
2284 static void *htab_percpu_map_lookup_percpu_elem(struct bpf_map *map, void *key, u32 cpu)
2285 {
2286         struct htab_elem *l;
2287
2288         if (cpu >= nr_cpu_ids)
2289                 return NULL;
2290
2291         l = __htab_map_lookup_elem(map, key);
2292         if (l)
2293                 return per_cpu_ptr(htab_elem_get_ptr(l, map->key_size), cpu);
2294         else
2295                 return NULL;
2296 }
2297
2298 static void *htab_lru_percpu_map_lookup_elem(struct bpf_map *map, void *key)
2299 {
2300         struct htab_elem *l = __htab_map_lookup_elem(map, key);
2301
2302         if (l) {
2303                 bpf_lru_node_set_ref(&l->lru_node);
2304                 return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
2305         }
2306
2307         return NULL;
2308 }
2309
2310 static void *htab_lru_percpu_map_lookup_percpu_elem(struct bpf_map *map, void *key, u32 cpu)
2311 {
2312         struct htab_elem *l;
2313
2314         if (cpu >= nr_cpu_ids)
2315                 return NULL;
2316
2317         l = __htab_map_lookup_elem(map, key);
2318         if (l) {
2319                 bpf_lru_node_set_ref(&l->lru_node);
2320                 return per_cpu_ptr(htab_elem_get_ptr(l, map->key_size), cpu);
2321         }
2322
2323         return NULL;
2324 }
2325
2326 int bpf_percpu_hash_copy(struct bpf_map *map, void *key, void *value)
2327 {
2328         struct htab_elem *l;
2329         void __percpu *pptr;
2330         int ret = -ENOENT;
2331         int cpu, off = 0;
2332         u32 size;
2333
2334         /* per_cpu areas are zero-filled and bpf programs can only
2335          * access 'value_size' of them, so copying rounded areas
2336          * will not leak any kernel data
2337          */
2338         size = round_up(map->value_size, 8);
2339         rcu_read_lock();
2340         l = __htab_map_lookup_elem(map, key);
2341         if (!l)
2342                 goto out;
2343         /* We do not mark LRU map element here in order to not mess up
2344          * eviction heuristics when user space does a map walk.
2345          */
2346         pptr = htab_elem_get_ptr(l, map->key_size);
2347         for_each_possible_cpu(cpu) {
2348                 copy_map_value_long(map, value + off, per_cpu_ptr(pptr, cpu));
2349                 check_and_init_map_value(map, value + off);
2350                 off += size;
2351         }
2352         ret = 0;
2353 out:
2354         rcu_read_unlock();
2355         return ret;
2356 }
2357
2358 int bpf_percpu_hash_update(struct bpf_map *map, void *key, void *value,
2359                            u64 map_flags)
2360 {
2361         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
2362         int ret;
2363
2364         rcu_read_lock();
2365         if (htab_is_lru(htab))
2366                 ret = __htab_lru_percpu_map_update_elem(map, key, value,
2367                                                         map_flags, true);
2368         else
2369                 ret = __htab_percpu_map_update_elem(map, key, value, map_flags,
2370                                                     true);
2371         rcu_read_unlock();
2372
2373         return ret;
2374 }
2375
2376 static void htab_percpu_map_seq_show_elem(struct bpf_map *map, void *key,
2377                                           struct seq_file *m)
2378 {
2379         struct htab_elem *l;
2380         void __percpu *pptr;
2381         int cpu;
2382
2383         rcu_read_lock();
2384
2385         l = __htab_map_lookup_elem(map, key);
2386         if (!l) {
2387                 rcu_read_unlock();
2388                 return;
2389         }
2390
2391         btf_type_seq_show(map->btf, map->btf_key_type_id, key, m);
2392         seq_puts(m, ": {\n");
2393         pptr = htab_elem_get_ptr(l, map->key_size);
2394         for_each_possible_cpu(cpu) {
2395                 seq_printf(m, "\tcpu%d: ", cpu);
2396                 btf_type_seq_show(map->btf, map->btf_value_type_id,
2397                                   per_cpu_ptr(pptr, cpu), m);
2398                 seq_puts(m, "\n");
2399         }
2400         seq_puts(m, "}\n");
2401
2402         rcu_read_unlock();
2403 }
2404
2405 const struct bpf_map_ops htab_percpu_map_ops = {
2406         .map_meta_equal = bpf_map_meta_equal,
2407         .map_alloc_check = htab_map_alloc_check,
2408         .map_alloc = htab_map_alloc,
2409         .map_free = htab_map_free,
2410         .map_get_next_key = htab_map_get_next_key,
2411         .map_lookup_elem = htab_percpu_map_lookup_elem,
2412         .map_lookup_and_delete_elem = htab_percpu_map_lookup_and_delete_elem,
2413         .map_update_elem = htab_percpu_map_update_elem,
2414         .map_delete_elem = htab_map_delete_elem,
2415         .map_lookup_percpu_elem = htab_percpu_map_lookup_percpu_elem,
2416         .map_seq_show_elem = htab_percpu_map_seq_show_elem,
2417         .map_set_for_each_callback_args = map_set_for_each_callback_args,
2418         .map_for_each_callback = bpf_for_each_hash_elem,
2419         .map_mem_usage = htab_map_mem_usage,
2420         BATCH_OPS(htab_percpu),
2421         .map_btf_id = &htab_map_btf_ids[0],
2422         .iter_seq_info = &iter_seq_info,
2423 };
2424
2425 const struct bpf_map_ops htab_lru_percpu_map_ops = {
2426         .map_meta_equal = bpf_map_meta_equal,
2427         .map_alloc_check = htab_map_alloc_check,
2428         .map_alloc = htab_map_alloc,
2429         .map_free = htab_map_free,
2430         .map_get_next_key = htab_map_get_next_key,
2431         .map_lookup_elem = htab_lru_percpu_map_lookup_elem,
2432         .map_lookup_and_delete_elem = htab_lru_percpu_map_lookup_and_delete_elem,
2433         .map_update_elem = htab_lru_percpu_map_update_elem,
2434         .map_delete_elem = htab_lru_map_delete_elem,
2435         .map_lookup_percpu_elem = htab_lru_percpu_map_lookup_percpu_elem,
2436         .map_seq_show_elem = htab_percpu_map_seq_show_elem,
2437         .map_set_for_each_callback_args = map_set_for_each_callback_args,
2438         .map_for_each_callback = bpf_for_each_hash_elem,
2439         .map_mem_usage = htab_map_mem_usage,
2440         BATCH_OPS(htab_lru_percpu),
2441         .map_btf_id = &htab_map_btf_ids[0],
2442         .iter_seq_info = &iter_seq_info,
2443 };
2444
2445 static int fd_htab_map_alloc_check(union bpf_attr *attr)
2446 {
2447         if (attr->value_size != sizeof(u32))
2448                 return -EINVAL;
2449         return htab_map_alloc_check(attr);
2450 }
2451
2452 static void fd_htab_map_free(struct bpf_map *map)
2453 {
2454         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
2455         struct hlist_nulls_node *n;
2456         struct hlist_nulls_head *head;
2457         struct htab_elem *l;
2458         int i;
2459
2460         for (i = 0; i < htab->n_buckets; i++) {
2461                 head = select_bucket(htab, i);
2462
2463                 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
2464                         void *ptr = fd_htab_map_get_ptr(map, l);
2465
2466                         map->ops->map_fd_put_ptr(ptr);
2467                 }
2468         }
2469
2470         htab_map_free(map);
2471 }
2472
2473 /* only called from syscall */
2474 int bpf_fd_htab_map_lookup_elem(struct bpf_map *map, void *key, u32 *value)
2475 {
2476         void **ptr;
2477         int ret = 0;
2478
2479         if (!map->ops->map_fd_sys_lookup_elem)
2480                 return -ENOTSUPP;
2481
2482         rcu_read_lock();
2483         ptr = htab_map_lookup_elem(map, key);
2484         if (ptr)
2485                 *value = map->ops->map_fd_sys_lookup_elem(READ_ONCE(*ptr));
2486         else
2487                 ret = -ENOENT;
2488         rcu_read_unlock();
2489
2490         return ret;
2491 }
2492
2493 /* only called from syscall */
2494 int bpf_fd_htab_map_update_elem(struct bpf_map *map, struct file *map_file,
2495                                 void *key, void *value, u64 map_flags)
2496 {
2497         void *ptr;
2498         int ret;
2499         u32 ufd = *(u32 *)value;
2500
2501         ptr = map->ops->map_fd_get_ptr(map, map_file, ufd);
2502         if (IS_ERR(ptr))
2503                 return PTR_ERR(ptr);
2504
2505         ret = htab_map_update_elem(map, key, &ptr, map_flags);
2506         if (ret)
2507                 map->ops->map_fd_put_ptr(ptr);
2508
2509         return ret;
2510 }
2511
2512 static struct bpf_map *htab_of_map_alloc(union bpf_attr *attr)
2513 {
2514         struct bpf_map *map, *inner_map_meta;
2515
2516         inner_map_meta = bpf_map_meta_alloc(attr->inner_map_fd);
2517         if (IS_ERR(inner_map_meta))
2518                 return inner_map_meta;
2519
2520         map = htab_map_alloc(attr);
2521         if (IS_ERR(map)) {
2522                 bpf_map_meta_free(inner_map_meta);
2523                 return map;
2524         }
2525
2526         map->inner_map_meta = inner_map_meta;
2527
2528         return map;
2529 }
2530
2531 static void *htab_of_map_lookup_elem(struct bpf_map *map, void *key)
2532 {
2533         struct bpf_map **inner_map  = htab_map_lookup_elem(map, key);
2534
2535         if (!inner_map)
2536                 return NULL;
2537
2538         return READ_ONCE(*inner_map);
2539 }
2540
2541 static int htab_of_map_gen_lookup(struct bpf_map *map,
2542                                   struct bpf_insn *insn_buf)
2543 {
2544         struct bpf_insn *insn = insn_buf;
2545         const int ret = BPF_REG_0;
2546
2547         BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
2548                      (void *(*)(struct bpf_map *map, void *key))NULL));
2549         *insn++ = BPF_EMIT_CALL(__htab_map_lookup_elem);
2550         *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 2);
2551         *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
2552                                 offsetof(struct htab_elem, key) +
2553                                 round_up(map->key_size, 8));
2554         *insn++ = BPF_LDX_MEM(BPF_DW, ret, ret, 0);
2555
2556         return insn - insn_buf;
2557 }
2558
2559 static void htab_of_map_free(struct bpf_map *map)
2560 {
2561         bpf_map_meta_free(map->inner_map_meta);
2562         fd_htab_map_free(map);
2563 }
2564
2565 const struct bpf_map_ops htab_of_maps_map_ops = {
2566         .map_alloc_check = fd_htab_map_alloc_check,
2567         .map_alloc = htab_of_map_alloc,
2568         .map_free = htab_of_map_free,
2569         .map_get_next_key = htab_map_get_next_key,
2570         .map_lookup_elem = htab_of_map_lookup_elem,
2571         .map_delete_elem = htab_map_delete_elem,
2572         .map_fd_get_ptr = bpf_map_fd_get_ptr,
2573         .map_fd_put_ptr = bpf_map_fd_put_ptr,
2574         .map_fd_sys_lookup_elem = bpf_map_fd_sys_lookup_elem,
2575         .map_gen_lookup = htab_of_map_gen_lookup,
2576         .map_check_btf = map_check_no_btf,
2577         .map_mem_usage = htab_map_mem_usage,
2578         BATCH_OPS(htab),
2579         .map_btf_id = &htab_map_btf_ids[0],
2580 };