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/rcupdate_wait.h>
  11 #include <linux/random.h>
  12 #include <uapi/linux/btf.h>
  13 #include <linux/rcupdate_trace.h>
  14 #include <linux/btf_ids.h>
  15 #include "percpu_freelist.h"
  16 #include "bpf_lru_list.h"
  17 #include "map_in_map.h"
  18 #include <linux/bpf_mem_alloc.h>
  19
  20 #define HTAB_CREATE_FLAG_MASK                                           \
  21         (BPF_F_NO_PREALLOC | BPF_F_NO_COMMON_LRU | BPF_F_NUMA_NODE |    \
  22          BPF_F_ACCESS_MASK | BPF_F_ZERO_SEED)
  23
  24 #define BATCH_OPS(_name)                        \
  25         .map_lookup_batch =                     \
  26         _name##_map_lookup_batch,               \
  27         .map_lookup_and_delete_batch =          \
  28         _name##_map_lookup_and_delete_batch,    \
  29         .map_update_batch =                     \
  30         generic_map_update_batch,               \
  31         .map_delete_batch =                     \
  32         generic_map_delete_batch
  33
  34 /*
  35  * The bucket lock has two protection scopes:
  36  *
  37  * 1) Serializing concurrent operations from BPF programs on different
  38  *    CPUs
  39  *
  40  * 2) Serializing concurrent operations from BPF programs and sys_bpf()
  41  *
  42  * BPF programs can execute in any context including perf, kprobes and
  43  * tracing. As there are almost no limits where perf, kprobes and tracing
  44  * can be invoked from the lock operations need to be protected against
  45  * deadlocks. Deadlocks can be caused by recursion and by an invocation in
  46  * the lock held section when functions which acquire this lock are invoked
  47  * from sys_bpf(). BPF recursion is prevented by incrementing the per CPU
  48  * variable bpf_prog_active, which prevents BPF programs attached to perf
  49  * events, kprobes and tracing to be invoked before the prior invocation
  50  * from one of these contexts completed. sys_bpf() uses the same mechanism
  51  * by pinning the task to the current CPU and incrementing the recursion
  52  * protection across the map operation.
  53  *
  54  * This has subtle implications on PREEMPT_RT. PREEMPT_RT forbids certain
  55  * operations like memory allocations (even with GFP_ATOMIC) from atomic
  56  * contexts. This is required because even with GFP_ATOMIC the memory
  57  * allocator calls into code paths which acquire locks with long held lock
  58  * sections. To ensure the deterministic behaviour these locks are regular
  59  * spinlocks, which are converted to 'sleepable' spinlocks on RT. The only
  60  * true atomic contexts on an RT kernel are the low level hardware
  61  * handling, scheduling, low level interrupt handling, NMIs etc. None of
  62  * these contexts should ever do memory allocations.
  63  *
  64  * As regular device interrupt handlers and soft interrupts are forced into
  65  * thread context, the existing code which does
  66  *   spin_lock*(); alloc(GFP_ATOMIC); spin_unlock*();
  67  * just works.
  68  *
  69  * In theory the BPF locks could be converted to regular spinlocks as well,
  70  * but the bucket locks and percpu_freelist locks can be taken from
  71  * arbitrary contexts (perf, kprobes, tracepoints) which are required to be
  72  * atomic contexts even on RT. Before the introduction of bpf_mem_alloc,
  73  * it is only safe to use raw spinlock for preallocated hash map on a RT kernel,
  74  * because there is no memory allocation within the lock held sections. However
  75  * after hash map was fully converted to use bpf_mem_alloc, there will be
  76  * non-synchronous memory allocation for non-preallocated hash map, so it is
  77  * safe to always use raw spinlock for bucket lock.
  78  */
  79 struct bucket {
  80         struct hlist_nulls_head head;
  81         raw_spinlock_t raw_lock;
  82 };
  83
  84 #define HASHTAB_MAP_LOCK_COUNT 8
  85 #define HASHTAB_MAP_LOCK_MASK (HASHTAB_MAP_LOCK_COUNT - 1)
  86
  87 struct bpf_htab {
  88         struct bpf_map map;
  89         struct bpf_mem_alloc ma;
  90         struct bpf_mem_alloc pcpu_ma;
  91         struct bucket *buckets;
  92         void *elems;
  93         union {
  94                 struct pcpu_freelist freelist;
  95                 struct bpf_lru lru;
  96         };
  97         struct htab_elem *__percpu *extra_elems;
  98         /* number of elements in non-preallocated hashtable are kept
  99          * in either pcount or count
 100          */
 101         struct percpu_counter pcount;
 102         atomic_t count;
 103         bool use_percpu_counter;
 104         u32 n_buckets;  /* number of hash buckets */
 105         u32 elem_size;  /* size of each element in bytes */
 106         u32 hashrnd;
 107         struct lock_class_key lockdep_key;
 108         int __percpu *map_locked[HASHTAB_MAP_LOCK_COUNT];
 109 };
 110
 111 /* each htab element is struct htab_elem + key + value */
 112 struct htab_elem {
 113         union {
 114                 struct hlist_nulls_node hash_node;
 115                 struct {
 116                         void *padding;
 117                         union {
 118                                 struct pcpu_freelist_node fnode;
 119                                 struct htab_elem *batch_flink;
 120                         };
 121                 };
 122         };
 123         union {
 124                 /* pointer to per-cpu pointer */
 125                 void *ptr_to_pptr;
 126                 struct bpf_lru_node lru_node;
 127         };
 128         u32 hash;
 129         char key[] __aligned(8);
 130 };
 131
 132 static inline bool htab_is_prealloc(const struct bpf_htab *htab)
 133 {
 134         return !(htab->map.map_flags & BPF_F_NO_PREALLOC);
 135 }
 136
 137 static void htab_init_buckets(struct bpf_htab *htab)
 138 {
 139         unsigned int i;
 140
 141         for (i = 0; i < htab->n_buckets; i++) {
 142                 INIT_HLIST_NULLS_HEAD(&htab->buckets[i].head, i);
 143                 raw_spin_lock_init(&htab->buckets[i].raw_lock);
 144                 lockdep_set_class(&htab->buckets[i].raw_lock,
 145                                           &htab->lockdep_key);
 146                 cond_resched();
 147         }
 148 }
 149
 150 static inline int htab_lock_bucket(const struct bpf_htab *htab,
 151                                    struct bucket *b, u32 hash,
 152                                    unsigned long *pflags)
 153 {
 154         unsigned long flags;
 155
 156         hash = hash & min_t(u32, HASHTAB_MAP_LOCK_MASK, htab->n_buckets - 1);
 157
 158         preempt_disable();
 159         local_irq_save(flags);
 160         if (unlikely(__this_cpu_inc_return(*(htab->map_locked[hash])) != 1)) {
 161                 __this_cpu_dec(*(htab->map_locked[hash]));
 162                 local_irq_restore(flags);
 163                 preempt_enable();
 164                 return -EBUSY;
 165         }
 166
 167         raw_spin_lock(&b->raw_lock);
 168         *pflags = flags;
 169
 170         return 0;
 171 }
 172
 173 static inline void htab_unlock_bucket(const struct bpf_htab *htab,
 174                                       struct bucket *b, u32 hash,
 175                                       unsigned long flags)
 176 {
 177         hash = hash & min_t(u32, HASHTAB_MAP_LOCK_MASK, htab->n_buckets - 1);
 178         raw_spin_unlock(&b->raw_lock);
 179         __this_cpu_dec(*(htab->map_locked[hash]));
 180         local_irq_restore(flags);
 181         preempt_enable();
 182 }
 183
 184 static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node);
 185
 186 static bool htab_is_lru(const struct bpf_htab *htab)
 187 {
 188         return htab->map.map_type == BPF_MAP_TYPE_LRU_HASH ||
 189                 htab->map.map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH;
 190 }
 191
 192 static bool htab_is_percpu(const struct bpf_htab *htab)
 193 {
 194         return htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH ||
 195                 htab->map.map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH;
 196 }
 197
 198 static inline void htab_elem_set_ptr(struct htab_elem *l, u32 key_size,
 199                                      void __percpu *pptr)
 200 {
 201         *(void __percpu **)(l->key + key_size) = pptr;
 202 }
 203
 204 static inline void __percpu *htab_elem_get_ptr(struct htab_elem *l, u32 key_size)
 205 {
 206         return *(void __percpu **)(l->key + key_size);
 207 }
 208
 209 static void *fd_htab_map_get_ptr(const struct bpf_map *map, struct htab_elem *l)
 210 {
 211         return *(void **)(l->key + roundup(map->key_size, 8));
 212 }
 213
 214 static struct htab_elem *get_htab_elem(struct bpf_htab *htab, int i)
 215 {
 216         return (struct htab_elem *) (htab->elems + i * (u64)htab->elem_size);
 217 }
 218
 219 static bool htab_has_extra_elems(struct bpf_htab *htab)
 220 {
 221         return !htab_is_percpu(htab) && !htab_is_lru(htab);
 222 }
 223
 224 static void htab_free_prealloced_timers(struct bpf_htab *htab)
 225 {
 226         u32 num_entries = htab->map.max_entries;
 227         int i;
 228
 229         if (!btf_record_has_field(htab->map.record, BPF_TIMER))
 230                 return;
 231         if (htab_has_extra_elems(htab))
 232                 num_entries += num_possible_cpus();
 233
 234         for (i = 0; i < num_entries; i++) {
 235                 struct htab_elem *elem;
 236
 237                 elem = get_htab_elem(htab, i);
 238                 bpf_obj_free_timer(htab->map.record, elem->key + round_up(htab->map.key_size, 8));
 239                 cond_resched();
 240         }
 241 }
 242
 243 static void htab_free_prealloced_fields(struct bpf_htab *htab)
 244 {
 245         u32 num_entries = htab->map.max_entries;
 246         int i;
 247
 248         if (IS_ERR_OR_NULL(htab->map.record))
 249                 return;
 250         if (htab_has_extra_elems(htab))
 251                 num_entries += num_possible_cpus();
 252         for (i = 0; i < num_entries; i++) {
 253                 struct htab_elem *elem;
 254
 255                 elem = get_htab_elem(htab, i);
 256                 if (htab_is_percpu(htab)) {
 257                         void __percpu *pptr = htab_elem_get_ptr(elem, htab->map.key_size);
 258                         int cpu;
 259
 260                         for_each_possible_cpu(cpu) {
 261                                 bpf_obj_free_fields(htab->map.record, per_cpu_ptr(pptr, cpu));
 262                                 cond_resched();
 263                         }
 264                 } else {
 265                         bpf_obj_free_fields(htab->map.record, elem->key + round_up(htab->map.key_size, 8));
 266                         cond_resched();
 267                 }
 268                 cond_resched();
 269         }
 270 }
 271
 272 static void htab_free_elems(struct bpf_htab *htab)
 273 {
 274         int i;
 275
 276         if (!htab_is_percpu(htab))
 277                 goto free_elems;
 278
 279         for (i = 0; i < htab->map.max_entries; i++) {
 280                 void __percpu *pptr;
 281
 282                 pptr = htab_elem_get_ptr(get_htab_elem(htab, i),
 283                                          htab->map.key_size);
 284                 free_percpu(pptr);
 285                 cond_resched();
 286         }
 287 free_elems:
 288         bpf_map_area_free(htab->elems);
 289 }
 290
 291 /* The LRU list has a lock (lru_lock). Each htab bucket has a lock
 292  * (bucket_lock). If both locks need to be acquired together, the lock
 293  * order is always lru_lock -> bucket_lock and this only happens in
 294  * bpf_lru_list.c logic. For example, certain code path of
 295  * bpf_lru_pop_free(), which is called by function prealloc_lru_pop(),
 296  * will acquire lru_lock first followed by acquiring bucket_lock.
 297  *
 298  * In hashtab.c, to avoid deadlock, lock acquisition of
 299  * bucket_lock followed by lru_lock is not allowed. In such cases,
 300  * bucket_lock needs to be released first before acquiring lru_lock.
 301  */
 302 static struct htab_elem *prealloc_lru_pop(struct bpf_htab *htab, void *key,
 303                                           u32 hash)
 304 {
 305         struct bpf_lru_node *node = bpf_lru_pop_free(&htab->lru, hash);
 306         struct htab_elem *l;
 307
 308         if (node) {
 309                 bpf_map_inc_elem_count(&htab->map);
 310                 l = container_of(node, struct htab_elem, lru_node);
 311                 memcpy(l->key, key, htab->map.key_size);
 312                 return l;
 313         }
 314
 315         return NULL;
 316 }
 317
 318 static int prealloc_init(struct bpf_htab *htab)
 319 {
 320         u32 num_entries = htab->map.max_entries;
 321         int err = -ENOMEM, i;
 322
 323         if (htab_has_extra_elems(htab))
 324                 num_entries += num_possible_cpus();
 325
 326         htab->elems = bpf_map_area_alloc((u64)htab->elem_size * num_entries,
 327                                          htab->map.numa_node);
 328         if (!htab->elems)
 329                 return -ENOMEM;
 330
 331         if (!htab_is_percpu(htab))
 332                 goto skip_percpu_elems;
 333
 334         for (i = 0; i < num_entries; i++) {
 335                 u32 size = round_up(htab->map.value_size, 8);
 336                 void __percpu *pptr;
 337
 338                 pptr = bpf_map_alloc_percpu(&htab->map, size, 8,
 339                                             GFP_USER | __GFP_NOWARN);
 340                 if (!pptr)
 341                         goto free_elems;
 342                 htab_elem_set_ptr(get_htab_elem(htab, i), htab->map.key_size,
 343                                   pptr);
 344                 cond_resched();
 345         }
 346
 347 skip_percpu_elems:
 348         if (htab_is_lru(htab))
 349                 err = bpf_lru_init(&htab->lru,
 350                                    htab->map.map_flags & BPF_F_NO_COMMON_LRU,
 351                                    offsetof(struct htab_elem, hash) -
 352                                    offsetof(struct htab_elem, lru_node),
 353                                    htab_lru_map_delete_node,
 354                                    htab);
 355         else
 356                 err = pcpu_freelist_init(&htab->freelist);
 357
 358         if (err)
 359                 goto free_elems;
 360
 361         if (htab_is_lru(htab))
 362                 bpf_lru_populate(&htab->lru, htab->elems,
 363                                  offsetof(struct htab_elem, lru_node),
 364                                  htab->elem_size, num_entries);
 365         else
 366                 pcpu_freelist_populate(&htab->freelist,
 367                                        htab->elems + offsetof(struct htab_elem, fnode),
 368                                        htab->elem_size, num_entries);
 369
 370         return 0;
 371
 372 free_elems:
 373         htab_free_elems(htab);
 374         return err;
 375 }
 376
 377 static void prealloc_destroy(struct bpf_htab *htab)
 378 {
 379         htab_free_elems(htab);
 380
 381         if (htab_is_lru(htab))
 382                 bpf_lru_destroy(&htab->lru);
 383         else
 384                 pcpu_freelist_destroy(&htab->freelist);
 385 }
 386
 387 static int alloc_extra_elems(struct bpf_htab *htab)
 388 {
 389         struct htab_elem *__percpu *pptr, *l_new;
 390         struct pcpu_freelist_node *l;
 391         int cpu;
 392
 393         pptr = bpf_map_alloc_percpu(&htab->map, sizeof(struct htab_elem *), 8,
 394                                     GFP_USER | __GFP_NOWARN);
 395         if (!pptr)
 396                 return -ENOMEM;
 397
 398         for_each_possible_cpu(cpu) {
 399                 l = pcpu_freelist_pop(&htab->freelist);
 400                 /* pop will succeed, since prealloc_init()
 401                  * preallocated extra num_possible_cpus elements
 402                  */
 403                 l_new = container_of(l, struct htab_elem, fnode);
 404                 *per_cpu_ptr(pptr, cpu) = l_new;
 405         }
 406         htab->extra_elems = pptr;
 407         return 0;
 408 }
 409
 410 /* Called from syscall */
 411 static int htab_map_alloc_check(union bpf_attr *attr)
 412 {
 413         bool percpu = (attr->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
 414                        attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
 415         bool lru = (attr->map_type == BPF_MAP_TYPE_LRU_HASH ||
 416                     attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
 417         /* percpu_lru means each cpu has its own LRU list.
 418          * it is different from BPF_MAP_TYPE_PERCPU_HASH where
 419          * the map's value itself is percpu.  percpu_lru has
 420          * nothing to do with the map's value.
 421          */
 422         bool percpu_lru = (attr->map_flags & BPF_F_NO_COMMON_LRU);
 423         bool prealloc = !(attr->map_flags & BPF_F_NO_PREALLOC);
 424         bool zero_seed = (attr->map_flags & BPF_F_ZERO_SEED);
 425         int numa_node = bpf_map_attr_numa_node(attr);
 426
 427         BUILD_BUG_ON(offsetof(struct htab_elem, fnode.next) !=
 428                      offsetof(struct htab_elem, hash_node.pprev));
 429
 430         if (zero_seed && !capable(CAP_SYS_ADMIN))
 431                 /* Guard against local DoS, and discourage production use. */
 432                 return -EPERM;
 433
 434         if (attr->map_flags & ~HTAB_CREATE_FLAG_MASK ||
 435             !bpf_map_flags_access_ok(attr->map_flags))
 436                 return -EINVAL;
 437
 438         if (!lru && percpu_lru)
 439                 return -EINVAL;
 440
 441         if (lru && !prealloc)
 442                 return -ENOTSUPP;
 443
 444         if (numa_node != NUMA_NO_NODE && (percpu || percpu_lru))
 445                 return -EINVAL;
 446
 447         /* check sanity of attributes.
 448          * value_size == 0 may be allowed in the future to use map as a set
 449          */
 450         if (attr->max_entries == 0 || attr->key_size == 0 ||
 451             attr->value_size == 0)
 452                 return -EINVAL;
 453
 454         if ((u64)attr->key_size + attr->value_size >= KMALLOC_MAX_SIZE -
 455            sizeof(struct htab_elem))
 456                 /* if key_size + value_size is bigger, the user space won't be
 457                  * able to access the elements via bpf syscall. This check
 458                  * also makes sure that the elem_size doesn't overflow and it's
 459                  * kmalloc-able later in htab_map_update_elem()
 460                  */
 461                 return -E2BIG;
 462
 463         return 0;
 464 }
 465
 466 static struct bpf_map *htab_map_alloc(union bpf_attr *attr)
 467 {
 468         bool percpu = (attr->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
 469                        attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
 470         bool lru = (attr->map_type == BPF_MAP_TYPE_LRU_HASH ||
 471                     attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
 472         /* percpu_lru means each cpu has its own LRU list.
 473          * it is different from BPF_MAP_TYPE_PERCPU_HASH where
 474          * the map's value itself is percpu.  percpu_lru has
 475          * nothing to do with the map's value.
 476          */
 477         bool percpu_lru = (attr->map_flags & BPF_F_NO_COMMON_LRU);
 478         bool prealloc = !(attr->map_flags & BPF_F_NO_PREALLOC);
 479         struct bpf_htab *htab;
 480         int err, i;
 481
 482         htab = bpf_map_area_alloc(sizeof(*htab), NUMA_NO_NODE);
 483         if (!htab)
 484                 return ERR_PTR(-ENOMEM);
 485
 486         lockdep_register_key(&htab->lockdep_key);
 487
 488         bpf_map_init_from_attr(&htab->map, attr);
 489
 490         if (percpu_lru) {
 491                 /* ensure each CPU's lru list has >=1 elements.
 492                  * since we are at it, make each lru list has the same
 493                  * number of elements.
 494                  */
 495                 htab->map.max_entries = roundup(attr->max_entries,
 496                                                 num_possible_cpus());
 497                 if (htab->map.max_entries < attr->max_entries)
 498                         htab->map.max_entries = rounddown(attr->max_entries,
 499                                                           num_possible_cpus());
 500         }
 501
 502         /* hash table size must be power of 2; roundup_pow_of_two() can overflow
 503          * into UB on 32-bit arches, so check that first
 504          */
 505         err = -E2BIG;
 506         if (htab->map.max_entries > 1UL << 31)
 507                 goto free_htab;
 508
 509         htab->n_buckets = roundup_pow_of_two(htab->map.max_entries);
 510
 511         htab->elem_size = sizeof(struct htab_elem) +
 512                           round_up(htab->map.key_size, 8);
 513         if (percpu)
 514                 htab->elem_size += sizeof(void *);
 515         else
 516                 htab->elem_size += round_up(htab->map.value_size, 8);
 517
 518         /* check for u32 overflow */
 519         if (htab->n_buckets > U32_MAX / sizeof(struct bucket))
 520                 goto free_htab;
 521
 522         err = bpf_map_init_elem_count(&htab->map);
 523         if (err)
 524                 goto free_htab;
 525
 526         err = -ENOMEM;
 527         htab->buckets = bpf_map_area_alloc(htab->n_buckets *
 528                                            sizeof(struct bucket),
 529                                            htab->map.numa_node);
 530         if (!htab->buckets)
 531                 goto free_elem_count;
 532
 533         for (i = 0; i < HASHTAB_MAP_LOCK_COUNT; i++) {
 534                 htab->map_locked[i] = bpf_map_alloc_percpu(&htab->map,
 535                                                            sizeof(int),
 536                                                            sizeof(int),
 537                                                            GFP_USER);
 538                 if (!htab->map_locked[i])
 539                         goto free_map_locked;
 540         }
 541
 542         if (htab->map.map_flags & BPF_F_ZERO_SEED)
 543                 htab->hashrnd = 0;
 544         else
 545                 htab->hashrnd = get_random_u32();
 546
 547         htab_init_buckets(htab);
 548
 549 /* compute_batch_value() computes batch value as num_online_cpus() * 2
 550  * and __percpu_counter_compare() needs
 551  * htab->max_entries - cur_number_of_elems to be more than batch * num_online_cpus()
 552  * for percpu_counter to be faster than atomic_t. In practice the average bpf
 553  * hash map size is 10k, which means that a system with 64 cpus will fill
 554  * hashmap to 20% of 10k before percpu_counter becomes ineffective. Therefore
 555  * define our own batch count as 32 then 10k hash map can be filled up to 80%:
 556  * 10k - 8k > 32 _batch_ * 64 _cpus_
 557  * and __percpu_counter_compare() will still be fast. At that point hash map
 558  * collisions will dominate its performance anyway. Assume that hash map filled
 559  * to 50+% isn't going to be O(1) and use the following formula to choose
 560  * between percpu_counter and atomic_t.
 561  */
 562 #define PERCPU_COUNTER_BATCH 32
 563         if (attr->max_entries / 2 > num_online_cpus() * PERCPU_COUNTER_BATCH)
 564                 htab->use_percpu_counter = true;
 565
 566         if (htab->use_percpu_counter) {
 567                 err = percpu_counter_init(&htab->pcount, 0, GFP_KERNEL);
 568                 if (err)
 569                         goto free_map_locked;
 570         }
 571
 572         if (prealloc) {
 573                 err = prealloc_init(htab);
 574                 if (err)
 575                         goto free_map_locked;
 576
 577                 if (!percpu && !lru) {
 578                         /* lru itself can remove the least used element, so
 579                          * there is no need for an extra elem during map_update.
 580                          */
 581                         err = alloc_extra_elems(htab);
 582                         if (err)
 583                                 goto free_prealloc;
 584                 }
 585         } else {
 586                 err = bpf_mem_alloc_init(&htab->ma, htab->elem_size, false);
 587                 if (err)
 588                         goto free_map_locked;
 589                 if (percpu) {
 590                         err = bpf_mem_alloc_init(&htab->pcpu_ma,
 591                                                  round_up(htab->map.value_size, 8), true);
 592                         if (err)
 593                                 goto free_map_locked;
 594                 }
 595         }
 596
 597         return &htab->map;
 598
 599 free_prealloc:
 600         prealloc_destroy(htab);
 601 free_map_locked:
 602         if (htab->use_percpu_counter)
 603                 percpu_counter_destroy(&htab->pcount);
 604         for (i = 0; i < HASHTAB_MAP_LOCK_COUNT; i++)
 605                 free_percpu(htab->map_locked[i]);
 606         bpf_map_area_free(htab->buckets);
 607         bpf_mem_alloc_destroy(&htab->pcpu_ma);
 608         bpf_mem_alloc_destroy(&htab->ma);
 609 free_elem_count:
 610         bpf_map_free_elem_count(&htab->map);
 611 free_htab:
 612         lockdep_unregister_key(&htab->lockdep_key);
 613         bpf_map_area_free(htab);
 614         return ERR_PTR(err);
 615 }
 616
 617 static inline u32 htab_map_hash(const void *key, u32 key_len, u32 hashrnd)
 618 {
 619         if (likely(key_len % 4 == 0))
 620                 return jhash2(key, key_len / 4, hashrnd);
 621         return jhash(key, key_len, hashrnd);
 622 }
 623
 624 static inline struct bucket *__select_bucket(struct bpf_htab *htab, u32 hash)
 625 {
 626         return &htab->buckets[hash & (htab->n_buckets - 1)];
 627 }
 628
 629 static inline struct hlist_nulls_head *select_bucket(struct bpf_htab *htab, u32 hash)
 630 {
 631         return &__select_bucket(htab, hash)->head;
 632 }
 633
 634 /* this lookup function can only be called with bucket lock taken */
 635 static struct htab_elem *lookup_elem_raw(struct hlist_nulls_head *head, u32 hash,
 636                                          void *key, u32 key_size)
 637 {
 638         struct hlist_nulls_node *n;
 639         struct htab_elem *l;
 640
 641         hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
 642                 if (l->hash == hash && !memcmp(&l->key, key, key_size))
 643                         return l;
 644
 645         return NULL;
 646 }
 647
 648 /* can be called without bucket lock. it will repeat the loop in
 649  * the unlikely event when elements moved from one bucket into another
 650  * while link list is being walked
 651  */
 652 static struct htab_elem *lookup_nulls_elem_raw(struct hlist_nulls_head *head,
 653                                                u32 hash, void *key,
 654                                                u32 key_size, u32 n_buckets)
 655 {
 656         struct hlist_nulls_node *n;
 657         struct htab_elem *l;
 658
 659 again:
 660         hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
 661                 if (l->hash == hash && !memcmp(&l->key, key, key_size))
 662                         return l;
 663
 664         if (unlikely(get_nulls_value(n) != (hash & (n_buckets - 1))))
 665                 goto again;
 666
 667         return NULL;
 668 }
 669
 670 /* Called from syscall or from eBPF program directly, so
 671  * arguments have to match bpf_map_lookup_elem() exactly.
 672  * The return value is adjusted by BPF instructions
 673  * in htab_map_gen_lookup().
 674  */
 675 static void *__htab_map_lookup_elem(struct bpf_map *map, void *key)
 676 {
 677         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 678         struct hlist_nulls_head *head;
 679         struct htab_elem *l;
 680         u32 hash, key_size;
 681
 682         WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
 683                      !rcu_read_lock_bh_held());
 684
 685         key_size = map->key_size;
 686
 687         hash = htab_map_hash(key, key_size, htab->hashrnd);
 688
 689         head = select_bucket(htab, hash);
 690
 691         l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
 692
 693         return l;
 694 }
 695
 696 static void *htab_map_lookup_elem(struct bpf_map *map, void *key)
 697 {
 698         struct htab_elem *l = __htab_map_lookup_elem(map, key);
 699
 700         if (l)
 701                 return l->key + round_up(map->key_size, 8);
 702
 703         return NULL;
 704 }
 705
 706 /* inline bpf_map_lookup_elem() call.
 707  * Instead of:
 708  * bpf_prog
 709  *   bpf_map_lookup_elem
 710  *     map->ops->map_lookup_elem
 711  *       htab_map_lookup_elem
 712  *         __htab_map_lookup_elem
 713  * do:
 714  * bpf_prog
 715  *   __htab_map_lookup_elem
 716  */
 717 static int htab_map_gen_lookup(struct bpf_map *map, struct bpf_insn *insn_buf)
 718 {
 719         struct bpf_insn *insn = insn_buf;
 720         const int ret = BPF_REG_0;
 721
 722         BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
 723                      (void *(*)(struct bpf_map *map, void *key))NULL));
 724         *insn++ = BPF_EMIT_CALL(__htab_map_lookup_elem);
 725         *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 1);
 726         *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
 727                                 offsetof(struct htab_elem, key) +
 728                                 round_up(map->key_size, 8));
 729         return insn - insn_buf;
 730 }
 731
 732 static __always_inline void *__htab_lru_map_lookup_elem(struct bpf_map *map,
 733                                                         void *key, const bool mark)
 734 {
 735         struct htab_elem *l = __htab_map_lookup_elem(map, key);
 736
 737         if (l) {
 738                 if (mark)
 739                         bpf_lru_node_set_ref(&l->lru_node);
 740                 return l->key + round_up(map->key_size, 8);
 741         }
 742
 743         return NULL;
 744 }
 745
 746 static void *htab_lru_map_lookup_elem(struct bpf_map *map, void *key)
 747 {
 748         return __htab_lru_map_lookup_elem(map, key, true);
 749 }
 750
 751 static void *htab_lru_map_lookup_elem_sys(struct bpf_map *map, void *key)
 752 {
 753         return __htab_lru_map_lookup_elem(map, key, false);
 754 }
 755
 756 static int htab_lru_map_gen_lookup(struct bpf_map *map,
 757                                    struct bpf_insn *insn_buf)
 758 {
 759         struct bpf_insn *insn = insn_buf;
 760         const int ret = BPF_REG_0;
 761         const int ref_reg = BPF_REG_1;
 762
 763         BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
 764                      (void *(*)(struct bpf_map *map, void *key))NULL));
 765         *insn++ = BPF_EMIT_CALL(__htab_map_lookup_elem);
 766         *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 4);
 767         *insn++ = BPF_LDX_MEM(BPF_B, ref_reg, ret,
 768                               offsetof(struct htab_elem, lru_node) +
 769                               offsetof(struct bpf_lru_node, ref));
 770         *insn++ = BPF_JMP_IMM(BPF_JNE, ref_reg, 0, 1);
 771         *insn++ = BPF_ST_MEM(BPF_B, ret,
 772                              offsetof(struct htab_elem, lru_node) +
 773                              offsetof(struct bpf_lru_node, ref),
 774                              1);
 775         *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
 776                                 offsetof(struct htab_elem, key) +
 777                                 round_up(map->key_size, 8));
 778         return insn - insn_buf;
 779 }
 780
 781 static void check_and_free_fields(struct bpf_htab *htab,
 782                                   struct htab_elem *elem)
 783 {
 784         if (htab_is_percpu(htab)) {
 785                 void __percpu *pptr = htab_elem_get_ptr(elem, htab->map.key_size);
 786                 int cpu;
 787
 788                 for_each_possible_cpu(cpu)
 789                         bpf_obj_free_fields(htab->map.record, per_cpu_ptr(pptr, cpu));
 790         } else {
 791                 void *map_value = elem->key + round_up(htab->map.key_size, 8);
 792
 793                 bpf_obj_free_fields(htab->map.record, map_value);
 794         }
 795 }
 796
 797 /* It is called from the bpf_lru_list when the LRU needs to delete
 798  * older elements from the htab.
 799  */
 800 static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node)
 801 {
 802         struct bpf_htab *htab = arg;
 803         struct htab_elem *l = NULL, *tgt_l;
 804         struct hlist_nulls_head *head;
 805         struct hlist_nulls_node *n;
 806         unsigned long flags;
 807         struct bucket *b;
 808         int ret;
 809
 810         tgt_l = container_of(node, struct htab_elem, lru_node);
 811         b = __select_bucket(htab, tgt_l->hash);
 812         head = &b->head;
 813
 814         ret = htab_lock_bucket(htab, b, tgt_l->hash, &flags);
 815         if (ret)
 816                 return false;
 817
 818         hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
 819                 if (l == tgt_l) {
 820                         hlist_nulls_del_rcu(&l->hash_node);
 821                         check_and_free_fields(htab, l);
 822                         bpf_map_dec_elem_count(&htab->map);
 823                         break;
 824                 }
 825
 826         htab_unlock_bucket(htab, b, tgt_l->hash, flags);
 827
 828         return l == tgt_l;
 829 }
 830
 831 /* Called from syscall */
 832 static int htab_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
 833 {
 834         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 835         struct hlist_nulls_head *head;
 836         struct htab_elem *l, *next_l;
 837         u32 hash, key_size;
 838         int i = 0;
 839
 840         WARN_ON_ONCE(!rcu_read_lock_held());
 841
 842         key_size = map->key_size;
 843
 844         if (!key)
 845                 goto find_first_elem;
 846
 847         hash = htab_map_hash(key, key_size, htab->hashrnd);
 848
 849         head = select_bucket(htab, hash);
 850
 851         /* lookup the key */
 852         l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
 853
 854         if (!l)
 855                 goto find_first_elem;
 856
 857         /* key was found, get next key in the same bucket */
 858         next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_next_rcu(&l->hash_node)),
 859                                   struct htab_elem, hash_node);
 860
 861         if (next_l) {
 862                 /* if next elem in this hash list is non-zero, just return it */
 863                 memcpy(next_key, next_l->key, key_size);
 864                 return 0;
 865         }
 866
 867         /* no more elements in this hash list, go to the next bucket */
 868         i = hash & (htab->n_buckets - 1);
 869         i++;
 870
 871 find_first_elem:
 872         /* iterate over buckets */
 873         for (; i < htab->n_buckets; i++) {
 874                 head = select_bucket(htab, i);
 875
 876                 /* pick first element in the bucket */
 877                 next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_first_rcu(head)),
 878                                           struct htab_elem, hash_node);
 879                 if (next_l) {
 880                         /* if it's not empty, just return it */
 881                         memcpy(next_key, next_l->key, key_size);
 882                         return 0;
 883                 }
 884         }
 885
 886         /* iterated over all buckets and all elements */
 887         return -ENOENT;
 888 }
 889
 890 static void htab_elem_free(struct bpf_htab *htab, struct htab_elem *l)
 891 {
 892         check_and_free_fields(htab, l);
 893         if (htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH)
 894                 bpf_mem_cache_free(&htab->pcpu_ma, l->ptr_to_pptr);
 895         bpf_mem_cache_free(&htab->ma, l);
 896 }
 897
 898 static void htab_put_fd_value(struct bpf_htab *htab, struct htab_elem *l)
 899 {
 900         struct bpf_map *map = &htab->map;
 901         void *ptr;
 902
 903         if (map->ops->map_fd_put_ptr) {
 904                 ptr = fd_htab_map_get_ptr(map, l);
 905                 map->ops->map_fd_put_ptr(map, ptr, true);
 906         }
 907 }
 908
 909 static bool is_map_full(struct bpf_htab *htab)
 910 {
 911         if (htab->use_percpu_counter)
 912                 return __percpu_counter_compare(&htab->pcount, htab->map.max_entries,
 913                                                 PERCPU_COUNTER_BATCH) >= 0;
 914         return atomic_read(&htab->count) >= htab->map.max_entries;
 915 }
 916
 917 static void inc_elem_count(struct bpf_htab *htab)
 918 {
 919         bpf_map_inc_elem_count(&htab->map);
 920
 921         if (htab->use_percpu_counter)
 922                 percpu_counter_add_batch(&htab->pcount, 1, PERCPU_COUNTER_BATCH);
 923         else
 924                 atomic_inc(&htab->count);
 925 }
 926
 927 static void dec_elem_count(struct bpf_htab *htab)
 928 {
 929         bpf_map_dec_elem_count(&htab->map);
 930
 931         if (htab->use_percpu_counter)
 932                 percpu_counter_add_batch(&htab->pcount, -1, PERCPU_COUNTER_BATCH);
 933         else
 934                 atomic_dec(&htab->count);
 935 }
 936
 937
 938 static void free_htab_elem(struct bpf_htab *htab, struct htab_elem *l)
 939 {
 940         htab_put_fd_value(htab, l);
 941
 942         if (htab_is_prealloc(htab)) {
 943                 bpf_map_dec_elem_count(&htab->map);
 944                 check_and_free_fields(htab, l);
 945                 __pcpu_freelist_push(&htab->freelist, &l->fnode);
 946         } else {
 947                 dec_elem_count(htab);
 948                 htab_elem_free(htab, l);
 949         }
 950 }
 951
 952 static void pcpu_copy_value(struct bpf_htab *htab, void __percpu *pptr,
 953                             void *value, bool onallcpus)
 954 {
 955         if (!onallcpus) {
 956                 /* copy true value_size bytes */
 957                 copy_map_value(&htab->map, this_cpu_ptr(pptr), value);
 958         } else {
 959                 u32 size = round_up(htab->map.value_size, 8);
 960                 int off = 0, cpu;
 961
 962                 for_each_possible_cpu(cpu) {
 963                         copy_map_value_long(&htab->map, per_cpu_ptr(pptr, cpu), value + off);
 964                         off += size;
 965                 }
 966         }
 967 }
 968
 969 static void pcpu_init_value(struct bpf_htab *htab, void __percpu *pptr,
 970                             void *value, bool onallcpus)
 971 {
 972         /* When not setting the initial value on all cpus, zero-fill element
 973          * values for other cpus. Otherwise, bpf program has no way to ensure
 974          * known initial values for cpus other than current one
 975          * (onallcpus=false always when coming from bpf prog).
 976          */
 977         if (!onallcpus) {
 978                 int current_cpu = raw_smp_processor_id();
 979                 int cpu;
 980
 981                 for_each_possible_cpu(cpu) {
 982                         if (cpu == current_cpu)
 983                                 copy_map_value_long(&htab->map, per_cpu_ptr(pptr, cpu), value);
 984                         else /* Since elem is preallocated, we cannot touch special fields */
 985                                 zero_map_value(&htab->map, per_cpu_ptr(pptr, cpu));
 986                 }
 987         } else {
 988                 pcpu_copy_value(htab, pptr, value, onallcpus);
 989         }
 990 }
 991
 992 static bool fd_htab_map_needs_adjust(const struct bpf_htab *htab)
 993 {
 994         return htab->map.map_type == BPF_MAP_TYPE_HASH_OF_MAPS &&
 995                BITS_PER_LONG == 64;
 996 }
 997
 998 static struct htab_elem *alloc_htab_elem(struct bpf_htab *htab, void *key,
 999                                          void *value, u32 key_size, u32 hash,
1000                                          bool percpu, bool onallcpus,
1001                                          struct htab_elem *old_elem)
1002 {
1003         u32 size = htab->map.value_size;
1004         bool prealloc = htab_is_prealloc(htab);
1005         struct htab_elem *l_new, **pl_new;
1006         void __percpu *pptr;
1007
1008         if (prealloc) {
1009                 if (old_elem) {
1010                         /* if we're updating the existing element,
1011                          * use per-cpu extra elems to avoid freelist_pop/push
1012                          */
1013                         pl_new = this_cpu_ptr(htab->extra_elems);
1014                         l_new = *pl_new;
1015                         htab_put_fd_value(htab, old_elem);
1016                         *pl_new = old_elem;
1017                 } else {
1018                         struct pcpu_freelist_node *l;
1019
1020                         l = __pcpu_freelist_pop(&htab->freelist);
1021                         if (!l)
1022                                 return ERR_PTR(-E2BIG);
1023                         l_new = container_of(l, struct htab_elem, fnode);
1024                         bpf_map_inc_elem_count(&htab->map);
1025                 }
1026         } else {
1027                 if (is_map_full(htab))
1028                         if (!old_elem)
1029                                 /* when map is full and update() is replacing
1030                                  * old element, it's ok to allocate, since
1031                                  * old element will be freed immediately.
1032                                  * Otherwise return an error
1033                                  */
1034                                 return ERR_PTR(-E2BIG);
1035                 inc_elem_count(htab);
1036                 l_new = bpf_mem_cache_alloc(&htab->ma);
1037                 if (!l_new) {
1038                         l_new = ERR_PTR(-ENOMEM);
1039                         goto dec_count;
1040                 }
1041         }
1042
1043         memcpy(l_new->key, key, key_size);
1044         if (percpu) {
1045                 if (prealloc) {
1046                         pptr = htab_elem_get_ptr(l_new, key_size);
1047                 } else {
1048                         /* alloc_percpu zero-fills */
1049                         pptr = bpf_mem_cache_alloc(&htab->pcpu_ma);
1050                         if (!pptr) {
1051                                 bpf_mem_cache_free(&htab->ma, l_new);
1052                                 l_new = ERR_PTR(-ENOMEM);
1053                                 goto dec_count;
1054                         }
1055                         l_new->ptr_to_pptr = pptr;
1056                         pptr = *(void **)pptr;
1057                 }
1058
1059                 pcpu_init_value(htab, pptr, value, onallcpus);
1060
1061                 if (!prealloc)
1062                         htab_elem_set_ptr(l_new, key_size, pptr);
1063         } else if (fd_htab_map_needs_adjust(htab)) {
1064                 size = round_up(size, 8);
1065                 memcpy(l_new->key + round_up(key_size, 8), value, size);
1066         } else {
1067                 copy_map_value(&htab->map,
1068                                l_new->key + round_up(key_size, 8),
1069                                value);
1070         }
1071
1072         l_new->hash = hash;
1073         return l_new;
1074 dec_count:
1075         dec_elem_count(htab);
1076         return l_new;
1077 }
1078
1079 static int check_flags(struct bpf_htab *htab, struct htab_elem *l_old,
1080                        u64 map_flags)
1081 {
1082         if (l_old && (map_flags & ~BPF_F_LOCK) == BPF_NOEXIST)
1083                 /* elem already exists */
1084                 return -EEXIST;
1085
1086         if (!l_old && (map_flags & ~BPF_F_LOCK) == BPF_EXIST)
1087                 /* elem doesn't exist, cannot update it */
1088                 return -ENOENT;
1089
1090         return 0;
1091 }
1092
1093 /* Called from syscall or from eBPF program */
1094 static long htab_map_update_elem(struct bpf_map *map, void *key, void *value,
1095                                  u64 map_flags)
1096 {
1097         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1098         struct htab_elem *l_new = NULL, *l_old;
1099         struct hlist_nulls_head *head;
1100         unsigned long flags;
1101         struct bucket *b;
1102         u32 key_size, hash;
1103         int ret;
1104
1105         if (unlikely((map_flags & ~BPF_F_LOCK) > BPF_EXIST))
1106                 /* unknown flags */
1107                 return -EINVAL;
1108
1109         WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
1110                      !rcu_read_lock_bh_held());
1111
1112         key_size = map->key_size;
1113
1114         hash = htab_map_hash(key, key_size, htab->hashrnd);
1115
1116         b = __select_bucket(htab, hash);
1117         head = &b->head;
1118
1119         if (unlikely(map_flags & BPF_F_LOCK)) {
1120                 if (unlikely(!btf_record_has_field(map->record, BPF_SPIN_LOCK)))
1121                         return -EINVAL;
1122                 /* find an element without taking the bucket lock */
1123                 l_old = lookup_nulls_elem_raw(head, hash, key, key_size,
1124                                               htab->n_buckets);
1125                 ret = check_flags(htab, l_old, map_flags);
1126                 if (ret)
1127                         return ret;
1128                 if (l_old) {
1129                         /* grab the element lock and update value in place */
1130                         copy_map_value_locked(map,
1131                                               l_old->key + round_up(key_size, 8),
1132                                               value, false);
1133                         return 0;
1134                 }
1135                 /* fall through, grab the bucket lock and lookup again.
1136                  * 99.9% chance that the element won't be found,
1137                  * but second lookup under lock has to be done.
1138                  */
1139         }
1140
1141         ret = htab_lock_bucket(htab, b, hash, &flags);
1142         if (ret)
1143                 return ret;
1144
1145         l_old = lookup_elem_raw(head, hash, key, key_size);
1146
1147         ret = check_flags(htab, l_old, map_flags);
1148         if (ret)
1149                 goto err;
1150
1151         if (unlikely(l_old && (map_flags & BPF_F_LOCK))) {
1152                 /* first lookup without the bucket lock didn't find the element,
1153                  * but second lookup with the bucket lock found it.
1154                  * This case is highly unlikely, but has to be dealt with:
1155                  * grab the element lock in addition to the bucket lock
1156                  * and update element in place
1157                  */
1158                 copy_map_value_locked(map,
1159                                       l_old->key + round_up(key_size, 8),
1160                                       value, false);
1161                 ret = 0;
1162                 goto err;
1163         }
1164
1165         l_new = alloc_htab_elem(htab, key, value, key_size, hash, false, false,
1166                                 l_old);
1167         if (IS_ERR(l_new)) {
1168                 /* all pre-allocated elements are in use or memory exhausted */
1169                 ret = PTR_ERR(l_new);
1170                 goto err;
1171         }
1172
1173         /* add new element to the head of the list, so that
1174          * concurrent search will find it before old elem
1175          */
1176         hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1177         if (l_old) {
1178                 hlist_nulls_del_rcu(&l_old->hash_node);
1179                 if (!htab_is_prealloc(htab))
1180                         free_htab_elem(htab, l_old);
1181                 else
1182                         check_and_free_fields(htab, l_old);
1183         }
1184         ret = 0;
1185 err:
1186         htab_unlock_bucket(htab, b, hash, flags);
1187         return ret;
1188 }
1189
1190 static void htab_lru_push_free(struct bpf_htab *htab, struct htab_elem *elem)
1191 {
1192         check_and_free_fields(htab, elem);
1193         bpf_map_dec_elem_count(&htab->map);
1194         bpf_lru_push_free(&htab->lru, &elem->lru_node);
1195 }
1196
1197 static long htab_lru_map_update_elem(struct bpf_map *map, void *key, void *value,
1198                                      u64 map_flags)
1199 {
1200         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1201         struct htab_elem *l_new, *l_old = NULL;
1202         struct hlist_nulls_head *head;
1203         unsigned long flags;
1204         struct bucket *b;
1205         u32 key_size, hash;
1206         int ret;
1207
1208         if (unlikely(map_flags > BPF_EXIST))
1209                 /* unknown flags */
1210                 return -EINVAL;
1211
1212         WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
1213                      !rcu_read_lock_bh_held());
1214
1215         key_size = map->key_size;
1216
1217         hash = htab_map_hash(key, key_size, htab->hashrnd);
1218
1219         b = __select_bucket(htab, hash);
1220         head = &b->head;
1221
1222         /* For LRU, we need to alloc before taking bucket's
1223          * spinlock because getting free nodes from LRU may need
1224          * to remove older elements from htab and this removal
1225          * operation will need a bucket lock.
1226          */
1227         l_new = prealloc_lru_pop(htab, key, hash);
1228         if (!l_new)
1229                 return -ENOMEM;
1230         copy_map_value(&htab->map,
1231                        l_new->key + round_up(map->key_size, 8), value);
1232
1233         ret = htab_lock_bucket(htab, b, hash, &flags);
1234         if (ret)
1235                 goto err_lock_bucket;
1236
1237         l_old = lookup_elem_raw(head, hash, key, key_size);
1238
1239         ret = check_flags(htab, l_old, map_flags);
1240         if (ret)
1241                 goto err;
1242
1243         /* add new element to the head of the list, so that
1244          * concurrent search will find it before old elem
1245          */
1246         hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1247         if (l_old) {
1248                 bpf_lru_node_set_ref(&l_new->lru_node);
1249                 hlist_nulls_del_rcu(&l_old->hash_node);
1250         }
1251         ret = 0;
1252
1253 err:
1254         htab_unlock_bucket(htab, b, hash, flags);
1255
1256 err_lock_bucket:
1257         if (ret)
1258                 htab_lru_push_free(htab, l_new);
1259         else if (l_old)
1260                 htab_lru_push_free(htab, l_old);
1261
1262         return ret;
1263 }
1264
1265 static long __htab_percpu_map_update_elem(struct bpf_map *map, void *key,
1266                                           void *value, u64 map_flags,
1267                                           bool onallcpus)
1268 {
1269         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1270         struct htab_elem *l_new = NULL, *l_old;
1271         struct hlist_nulls_head *head;
1272         unsigned long flags;
1273         struct bucket *b;
1274         u32 key_size, hash;
1275         int ret;
1276
1277         if (unlikely(map_flags > BPF_EXIST))
1278                 /* unknown flags */
1279                 return -EINVAL;
1280
1281         WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
1282                      !rcu_read_lock_bh_held());
1283
1284         key_size = map->key_size;
1285
1286         hash = htab_map_hash(key, key_size, htab->hashrnd);
1287
1288         b = __select_bucket(htab, hash);
1289         head = &b->head;
1290
1291         ret = htab_lock_bucket(htab, b, hash, &flags);
1292         if (ret)
1293                 return ret;
1294
1295         l_old = lookup_elem_raw(head, hash, key, key_size);
1296
1297         ret = check_flags(htab, l_old, map_flags);
1298         if (ret)
1299                 goto err;
1300
1301         if (l_old) {
1302                 /* per-cpu hash map can update value in-place */
1303                 pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
1304                                 value, onallcpus);
1305         } else {
1306                 l_new = alloc_htab_elem(htab, key, value, key_size,
1307                                         hash, true, onallcpus, NULL);
1308                 if (IS_ERR(l_new)) {
1309                         ret = PTR_ERR(l_new);
1310                         goto err;
1311                 }
1312                 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1313         }
1314         ret = 0;
1315 err:
1316         htab_unlock_bucket(htab, b, hash, flags);
1317         return ret;
1318 }
1319
1320 static long __htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
1321                                               void *value, u64 map_flags,
1322                                               bool onallcpus)
1323 {
1324         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1325         struct htab_elem *l_new = NULL, *l_old;
1326         struct hlist_nulls_head *head;
1327         unsigned long flags;
1328         struct bucket *b;
1329         u32 key_size, hash;
1330         int ret;
1331
1332         if (unlikely(map_flags > BPF_EXIST))
1333                 /* unknown flags */
1334                 return -EINVAL;
1335
1336         WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
1337                      !rcu_read_lock_bh_held());
1338
1339         key_size = map->key_size;
1340
1341         hash = htab_map_hash(key, key_size, htab->hashrnd);
1342
1343         b = __select_bucket(htab, hash);
1344         head = &b->head;
1345
1346         /* For LRU, we need to alloc before taking bucket's
1347          * spinlock because LRU's elem alloc may need
1348          * to remove older elem from htab and this removal
1349          * operation will need a bucket lock.
1350          */
1351         if (map_flags != BPF_EXIST) {
1352                 l_new = prealloc_lru_pop(htab, key, hash);
1353                 if (!l_new)
1354                         return -ENOMEM;
1355         }
1356
1357         ret = htab_lock_bucket(htab, b, hash, &flags);
1358         if (ret)
1359                 goto err_lock_bucket;
1360
1361         l_old = lookup_elem_raw(head, hash, key, key_size);
1362
1363         ret = check_flags(htab, l_old, map_flags);
1364         if (ret)
1365                 goto err;
1366
1367         if (l_old) {
1368                 bpf_lru_node_set_ref(&l_old->lru_node);
1369
1370                 /* per-cpu hash map can update value in-place */
1371                 pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
1372                                 value, onallcpus);
1373         } else {
1374                 pcpu_init_value(htab, htab_elem_get_ptr(l_new, key_size),
1375                                 value, onallcpus);
1376                 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1377                 l_new = NULL;
1378         }
1379         ret = 0;
1380 err:
1381         htab_unlock_bucket(htab, b, hash, flags);
1382 err_lock_bucket:
1383         if (l_new) {
1384                 bpf_map_dec_elem_count(&htab->map);
1385                 bpf_lru_push_free(&htab->lru, &l_new->lru_node);
1386         }
1387         return ret;
1388 }
1389
1390 static long htab_percpu_map_update_elem(struct bpf_map *map, void *key,
1391                                         void *value, u64 map_flags)
1392 {
1393         return __htab_percpu_map_update_elem(map, key, value, map_flags, false);
1394 }
1395
1396 static long htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
1397                                             void *value, u64 map_flags)
1398 {
1399         return __htab_lru_percpu_map_update_elem(map, key, value, map_flags,
1400                                                  false);
1401 }
1402
1403 /* Called from syscall or from eBPF program */
1404 static long htab_map_delete_elem(struct bpf_map *map, void *key)
1405 {
1406         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1407         struct hlist_nulls_head *head;
1408         struct bucket *b;
1409         struct htab_elem *l;
1410         unsigned long flags;
1411         u32 hash, key_size;
1412         int ret;
1413
1414         WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
1415                      !rcu_read_lock_bh_held());
1416
1417         key_size = map->key_size;
1418
1419         hash = htab_map_hash(key, key_size, htab->hashrnd);
1420         b = __select_bucket(htab, hash);
1421         head = &b->head;
1422
1423         ret = htab_lock_bucket(htab, b, hash, &flags);
1424         if (ret)
1425                 return ret;
1426
1427         l = lookup_elem_raw(head, hash, key, key_size);
1428
1429         if (l) {
1430                 hlist_nulls_del_rcu(&l->hash_node);
1431                 free_htab_elem(htab, l);
1432         } else {
1433                 ret = -ENOENT;
1434         }
1435
1436         htab_unlock_bucket(htab, b, hash, flags);
1437         return ret;
1438 }
1439
1440 static long htab_lru_map_delete_elem(struct bpf_map *map, void *key)
1441 {
1442         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1443         struct hlist_nulls_head *head;
1444         struct bucket *b;
1445         struct htab_elem *l;
1446         unsigned long flags;
1447         u32 hash, key_size;
1448         int ret;
1449
1450         WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
1451                      !rcu_read_lock_bh_held());
1452
1453         key_size = map->key_size;
1454
1455         hash = htab_map_hash(key, key_size, htab->hashrnd);
1456         b = __select_bucket(htab, hash);
1457         head = &b->head;
1458
1459         ret = htab_lock_bucket(htab, b, hash, &flags);
1460         if (ret)
1461                 return ret;
1462
1463         l = lookup_elem_raw(head, hash, key, key_size);
1464
1465         if (l)
1466                 hlist_nulls_del_rcu(&l->hash_node);
1467         else
1468                 ret = -ENOENT;
1469
1470         htab_unlock_bucket(htab, b, hash, flags);
1471         if (l)
1472                 htab_lru_push_free(htab, l);
1473         return ret;
1474 }
1475
1476 static void delete_all_elements(struct bpf_htab *htab)
1477 {
1478         int i;
1479
1480         /* It's called from a worker thread, so disable migration here,
1481          * since bpf_mem_cache_free() relies on that.
1482          */
1483         migrate_disable();
1484         for (i = 0; i < htab->n_buckets; i++) {
1485                 struct hlist_nulls_head *head = select_bucket(htab, i);
1486                 struct hlist_nulls_node *n;
1487                 struct htab_elem *l;
1488
1489                 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1490                         hlist_nulls_del_rcu(&l->hash_node);
1491                         htab_elem_free(htab, l);
1492                 }
1493         }
1494         migrate_enable();
1495 }
1496
1497 static void htab_free_malloced_timers(struct bpf_htab *htab)
1498 {
1499         int i;
1500
1501         rcu_read_lock();
1502         for (i = 0; i < htab->n_buckets; i++) {
1503                 struct hlist_nulls_head *head = select_bucket(htab, i);
1504                 struct hlist_nulls_node *n;
1505                 struct htab_elem *l;
1506
1507                 hlist_nulls_for_each_entry(l, n, head, hash_node) {
1508                         /* We only free timer on uref dropping to zero */
1509                         bpf_obj_free_timer(htab->map.record, l->key + round_up(htab->map.key_size, 8));
1510                 }
1511                 cond_resched_rcu();
1512         }
1513         rcu_read_unlock();
1514 }
1515
1516 static void htab_map_free_timers(struct bpf_map *map)
1517 {
1518         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1519
1520         /* We only free timer on uref dropping to zero */
1521         if (!btf_record_has_field(htab->map.record, BPF_TIMER))
1522                 return;
1523         if (!htab_is_prealloc(htab))
1524                 htab_free_malloced_timers(htab);
1525         else
1526                 htab_free_prealloced_timers(htab);
1527 }
1528
1529 /* Called when map->refcnt goes to zero, either from workqueue or from syscall */
1530 static void htab_map_free(struct bpf_map *map)
1531 {
1532         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1533         int i;
1534
1535         /* bpf_free_used_maps() or close(map_fd) will trigger this map_free callback.
1536          * bpf_free_used_maps() is called after bpf prog is no longer executing.
1537          * There is no need to synchronize_rcu() here to protect map elements.
1538          */
1539
1540         /* htab no longer uses call_rcu() directly. bpf_mem_alloc does it
1541          * underneath and is reponsible for waiting for callbacks to finish
1542          * during bpf_mem_alloc_destroy().
1543          */
1544         if (!htab_is_prealloc(htab)) {
1545                 delete_all_elements(htab);
1546         } else {
1547                 htab_free_prealloced_fields(htab);
1548                 prealloc_destroy(htab);
1549         }
1550
1551         bpf_map_free_elem_count(map);
1552         free_percpu(htab->extra_elems);
1553         bpf_map_area_free(htab->buckets);
1554         bpf_mem_alloc_destroy(&htab->pcpu_ma);
1555         bpf_mem_alloc_destroy(&htab->ma);
1556         if (htab->use_percpu_counter)
1557                 percpu_counter_destroy(&htab->pcount);
1558         for (i = 0; i < HASHTAB_MAP_LOCK_COUNT; i++)
1559                 free_percpu(htab->map_locked[i]);
1560         lockdep_unregister_key(&htab->lockdep_key);
1561         bpf_map_area_free(htab);
1562 }
1563
1564 static void htab_map_seq_show_elem(struct bpf_map *map, void *key,
1565                                    struct seq_file *m)
1566 {
1567         void *value;
1568
1569         rcu_read_lock();
1570
1571         value = htab_map_lookup_elem(map, key);
1572         if (!value) {
1573                 rcu_read_unlock();
1574                 return;
1575         }
1576
1577         btf_type_seq_show(map->btf, map->btf_key_type_id, key, m);
1578         seq_puts(m, ": ");
1579         btf_type_seq_show(map->btf, map->btf_value_type_id, value, m);
1580         seq_puts(m, "\n");
1581
1582         rcu_read_unlock();
1583 }
1584
1585 static int __htab_map_lookup_and_delete_elem(struct bpf_map *map, void *key,
1586                                              void *value, bool is_lru_map,
1587                                              bool is_percpu, u64 flags)
1588 {
1589         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1590         struct hlist_nulls_head *head;
1591         unsigned long bflags;
1592         struct htab_elem *l;
1593         u32 hash, key_size;
1594         struct bucket *b;
1595         int ret;
1596
1597         key_size = map->key_size;
1598
1599         hash = htab_map_hash(key, key_size, htab->hashrnd);
1600         b = __select_bucket(htab, hash);
1601         head = &b->head;
1602
1603         ret = htab_lock_bucket(htab, b, hash, &bflags);
1604         if (ret)
1605                 return ret;
1606
1607         l = lookup_elem_raw(head, hash, key, key_size);
1608         if (!l) {
1609                 ret = -ENOENT;
1610         } else {
1611                 if (is_percpu) {
1612                         u32 roundup_value_size = round_up(map->value_size, 8);
1613                         void __percpu *pptr;
1614                         int off = 0, cpu;
1615
1616                         pptr = htab_elem_get_ptr(l, key_size);
1617                         for_each_possible_cpu(cpu) {
1618                                 copy_map_value_long(&htab->map, value + off, per_cpu_ptr(pptr, cpu));
1619                                 check_and_init_map_value(&htab->map, value + off);
1620                                 off += roundup_value_size;
1621                         }
1622                 } else {
1623                         u32 roundup_key_size = round_up(map->key_size, 8);
1624
1625                         if (flags & BPF_F_LOCK)
1626                                 copy_map_value_locked(map, value, l->key +
1627                                                       roundup_key_size,
1628                                                       true);
1629                         else
1630                                 copy_map_value(map, value, l->key +
1631                                                roundup_key_size);
1632                         /* Zeroing special fields in the temp buffer */
1633                         check_and_init_map_value(map, value);
1634                 }
1635
1636                 hlist_nulls_del_rcu(&l->hash_node);
1637                 if (!is_lru_map)
1638                         free_htab_elem(htab, l);
1639         }
1640
1641         htab_unlock_bucket(htab, b, hash, bflags);
1642
1643         if (is_lru_map && l)
1644                 htab_lru_push_free(htab, l);
1645
1646         return ret;
1647 }
1648
1649 static int htab_map_lookup_and_delete_elem(struct bpf_map *map, void *key,
1650                                            void *value, u64 flags)
1651 {
1652         return __htab_map_lookup_and_delete_elem(map, key, value, false, false,
1653                                                  flags);
1654 }
1655
1656 static int htab_percpu_map_lookup_and_delete_elem(struct bpf_map *map,
1657                                                   void *key, void *value,
1658                                                   u64 flags)
1659 {
1660         return __htab_map_lookup_and_delete_elem(map, key, value, false, true,
1661                                                  flags);
1662 }
1663
1664 static int htab_lru_map_lookup_and_delete_elem(struct bpf_map *map, void *key,
1665                                                void *value, u64 flags)
1666 {
1667         return __htab_map_lookup_and_delete_elem(map, key, value, true, false,
1668                                                  flags);
1669 }
1670
1671 static int htab_lru_percpu_map_lookup_and_delete_elem(struct bpf_map *map,
1672                                                       void *key, void *value,
1673                                                       u64 flags)
1674 {
1675         return __htab_map_lookup_and_delete_elem(map, key, value, true, true,
1676                                                  flags);
1677 }
1678
1679 static int
1680 __htab_map_lookup_and_delete_batch(struct bpf_map *map,
1681                                    const union bpf_attr *attr,
1682                                    union bpf_attr __user *uattr,
1683                                    bool do_delete, bool is_lru_map,
1684                                    bool is_percpu)
1685 {
1686         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1687         u32 bucket_cnt, total, key_size, value_size, roundup_key_size;
1688         void *keys = NULL, *values = NULL, *value, *dst_key, *dst_val;
1689         void __user *uvalues = u64_to_user_ptr(attr->batch.values);
1690         void __user *ukeys = u64_to_user_ptr(attr->batch.keys);
1691         void __user *ubatch = u64_to_user_ptr(attr->batch.in_batch);
1692         u32 batch, max_count, size, bucket_size, map_id;
1693         struct htab_elem *node_to_free = NULL;
1694         u64 elem_map_flags, map_flags;
1695         struct hlist_nulls_head *head;
1696         struct hlist_nulls_node *n;
1697         unsigned long flags = 0;
1698         bool locked = false;
1699         struct htab_elem *l;
1700         struct bucket *b;
1701         int ret = 0;
1702
1703         elem_map_flags = attr->batch.elem_flags;
1704         if ((elem_map_flags & ~BPF_F_LOCK) ||
1705             ((elem_map_flags & BPF_F_LOCK) && !btf_record_has_field(map->record, BPF_SPIN_LOCK)))
1706                 return -EINVAL;
1707
1708         map_flags = attr->batch.flags;
1709         if (map_flags)
1710                 return -EINVAL;
1711
1712         max_count = attr->batch.count;
1713         if (!max_count)
1714                 return 0;
1715
1716         if (put_user(0, &uattr->batch.count))
1717                 return -EFAULT;
1718
1719         batch = 0;
1720         if (ubatch && copy_from_user(&batch, ubatch, sizeof(batch)))
1721                 return -EFAULT;
1722
1723         if (batch >= htab->n_buckets)
1724                 return -ENOENT;
1725
1726         key_size = htab->map.key_size;
1727         roundup_key_size = round_up(htab->map.key_size, 8);
1728         value_size = htab->map.value_size;
1729         size = round_up(value_size, 8);
1730         if (is_percpu)
1731                 value_size = size * num_possible_cpus();
1732         total = 0;
1733         /* while experimenting with hash tables with sizes ranging from 10 to
1734          * 1000, it was observed that a bucket can have up to 5 entries.
1735          */
1736         bucket_size = 5;
1737
1738 alloc:
1739         /* We cannot do copy_from_user or copy_to_user inside
1740          * the rcu_read_lock. Allocate enough space here.
1741          */
1742         keys = kvmalloc_array(key_size, bucket_size, GFP_USER | __GFP_NOWARN);
1743         values = kvmalloc_array(value_size, bucket_size, GFP_USER | __GFP_NOWARN);
1744         if (!keys || !values) {
1745                 ret = -ENOMEM;
1746                 goto after_loop;
1747         }
1748
1749 again:
1750         bpf_disable_instrumentation();
1751         rcu_read_lock();
1752 again_nocopy:
1753         dst_key = keys;
1754         dst_val = values;
1755         b = &htab->buckets[batch];
1756         head = &b->head;
1757         /* do not grab the lock unless need it (bucket_cnt > 0). */
1758         if (locked) {
1759                 ret = htab_lock_bucket(htab, b, batch, &flags);
1760                 if (ret) {
1761                         rcu_read_unlock();
1762                         bpf_enable_instrumentation();
1763                         goto after_loop;
1764                 }
1765         }
1766
1767         bucket_cnt = 0;
1768         hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
1769                 bucket_cnt++;
1770
1771         if (bucket_cnt && !locked) {
1772                 locked = true;
1773                 goto again_nocopy;
1774         }
1775
1776         if (bucket_cnt > (max_count - total)) {
1777                 if (total == 0)
1778                         ret = -ENOSPC;
1779                 /* Note that since bucket_cnt > 0 here, it is implicit
1780                  * that the locked was grabbed, so release it.
1781                  */
1782                 htab_unlock_bucket(htab, b, batch, flags);
1783                 rcu_read_unlock();
1784                 bpf_enable_instrumentation();
1785                 goto after_loop;
1786         }
1787
1788         if (bucket_cnt > bucket_size) {
1789                 bucket_size = bucket_cnt;
1790                 /* Note that since bucket_cnt > 0 here, it is implicit
1791                  * that the locked was grabbed, so release it.
1792                  */
1793                 htab_unlock_bucket(htab, b, batch, flags);
1794                 rcu_read_unlock();
1795                 bpf_enable_instrumentation();
1796                 kvfree(keys);
1797                 kvfree(values);
1798                 goto alloc;
1799         }
1800
1801         /* Next block is only safe to run if you have grabbed the lock */
1802         if (!locked)
1803                 goto next_batch;
1804
1805         hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1806                 memcpy(dst_key, l->key, key_size);
1807
1808                 if (is_percpu) {
1809                         int off = 0, cpu;
1810                         void __percpu *pptr;
1811
1812                         pptr = htab_elem_get_ptr(l, map->key_size);
1813                         for_each_possible_cpu(cpu) {
1814                                 copy_map_value_long(&htab->map, dst_val + off, per_cpu_ptr(pptr, cpu));
1815                                 check_and_init_map_value(&htab->map, dst_val + off);
1816                                 off += size;
1817                         }
1818                 } else {
1819                         value = l->key + roundup_key_size;
1820                         if (map->map_type == BPF_MAP_TYPE_HASH_OF_MAPS) {
1821                                 struct bpf_map **inner_map = value;
1822
1823                                  /* Actual value is the id of the inner map */
1824                                 map_id = map->ops->map_fd_sys_lookup_elem(*inner_map);
1825                                 value = &map_id;
1826                         }
1827
1828                         if (elem_map_flags & BPF_F_LOCK)
1829                                 copy_map_value_locked(map, dst_val, value,
1830                                                       true);
1831                         else
1832                                 copy_map_value(map, dst_val, value);
1833                         /* Zeroing special fields in the temp buffer */
1834                         check_and_init_map_value(map, dst_val);
1835                 }
1836                 if (do_delete) {
1837                         hlist_nulls_del_rcu(&l->hash_node);
1838
1839                         /* bpf_lru_push_free() will acquire lru_lock, which
1840                          * may cause deadlock. See comments in function
1841                          * prealloc_lru_pop(). Let us do bpf_lru_push_free()
1842                          * after releasing the bucket lock.
1843                          */
1844                         if (is_lru_map) {
1845                                 l->batch_flink = node_to_free;
1846                                 node_to_free = l;
1847                         } else {
1848                                 free_htab_elem(htab, l);
1849                         }
1850                 }
1851                 dst_key += key_size;
1852                 dst_val += value_size;
1853         }
1854
1855         htab_unlock_bucket(htab, b, batch, flags);
1856         locked = false;
1857
1858         while (node_to_free) {
1859                 l = node_to_free;
1860                 node_to_free = node_to_free->batch_flink;
1861                 htab_lru_push_free(htab, l);
1862         }
1863
1864 next_batch:
1865         /* If we are not copying data, we can go to next bucket and avoid
1866          * unlocking the rcu.
1867          */
1868         if (!bucket_cnt && (batch + 1 < htab->n_buckets)) {
1869                 batch++;
1870                 goto again_nocopy;
1871         }
1872
1873         rcu_read_unlock();
1874         bpf_enable_instrumentation();
1875         if (bucket_cnt && (copy_to_user(ukeys + total * key_size, keys,
1876             key_size * bucket_cnt) ||
1877             copy_to_user(uvalues + total * value_size, values,
1878             value_size * bucket_cnt))) {
1879                 ret = -EFAULT;
1880                 goto after_loop;
1881         }
1882
1883         total += bucket_cnt;
1884         batch++;
1885         if (batch >= htab->n_buckets) {
1886                 ret = -ENOENT;
1887                 goto after_loop;
1888         }
1889         goto again;
1890
1891 after_loop:
1892         if (ret == -EFAULT)
1893                 goto out;
1894
1895         /* copy # of entries and next batch */
1896         ubatch = u64_to_user_ptr(attr->batch.out_batch);
1897         if (copy_to_user(ubatch, &batch, sizeof(batch)) ||
1898             put_user(total, &uattr->batch.count))
1899                 ret = -EFAULT;
1900
1901 out:
1902         kvfree(keys);
1903         kvfree(values);
1904         return ret;
1905 }
1906
1907 static int
1908 htab_percpu_map_lookup_batch(struct bpf_map *map, const union bpf_attr *attr,
1909                              union bpf_attr __user *uattr)
1910 {
1911         return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1912                                                   false, true);
1913 }
1914
1915 static int
1916 htab_percpu_map_lookup_and_delete_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, true,
1921                                                   false, true);
1922 }
1923
1924 static int
1925 htab_map_lookup_batch(struct bpf_map *map, const union bpf_attr *attr,
1926                       union bpf_attr __user *uattr)
1927 {
1928         return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1929                                                   false, false);
1930 }
1931
1932 static int
1933 htab_map_lookup_and_delete_batch(struct bpf_map *map,
1934                                  const union bpf_attr *attr,
1935                                  union bpf_attr __user *uattr)
1936 {
1937         return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1938                                                   false, false);
1939 }
1940
1941 static int
1942 htab_lru_percpu_map_lookup_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, false,
1947                                                   true, true);
1948 }
1949
1950 static int
1951 htab_lru_percpu_map_lookup_and_delete_batch(struct bpf_map *map,
1952                                             const union bpf_attr *attr,
1953                                             union bpf_attr __user *uattr)
1954 {
1955         return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1956                                                   true, true);
1957 }
1958
1959 static int
1960 htab_lru_map_lookup_batch(struct bpf_map *map, const union bpf_attr *attr,
1961                           union bpf_attr __user *uattr)
1962 {
1963         return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1964                                                   true, false);
1965 }
1966
1967 static int
1968 htab_lru_map_lookup_and_delete_batch(struct bpf_map *map,
1969                                      const union bpf_attr *attr,
1970                                      union bpf_attr __user *uattr)
1971 {
1972         return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1973                                                   true, false);
1974 }
1975
1976 struct bpf_iter_seq_hash_map_info {
1977         struct bpf_map *map;
1978         struct bpf_htab *htab;
1979         void *percpu_value_buf; // non-zero means percpu hash
1980         u32 bucket_id;
1981         u32 skip_elems;
1982 };
1983
1984 static struct htab_elem *
1985 bpf_hash_map_seq_find_next(struct bpf_iter_seq_hash_map_info *info,
1986                            struct htab_elem *prev_elem)
1987 {
1988         const struct bpf_htab *htab = info->htab;
1989         u32 skip_elems = info->skip_elems;
1990         u32 bucket_id = info->bucket_id;
1991         struct hlist_nulls_head *head;
1992         struct hlist_nulls_node *n;
1993         struct htab_elem *elem;
1994         struct bucket *b;
1995         u32 i, count;
1996
1997         if (bucket_id >= htab->n_buckets)
1998                 return NULL;
1999
2000         /* try to find next elem in the same bucket */
2001         if (prev_elem) {
2002                 /* no update/deletion on this bucket, prev_elem should be still valid
2003                  * and we won't skip elements.
2004                  */
2005                 n = rcu_dereference_raw(hlist_nulls_next_rcu(&prev_elem->hash_node));
2006                 elem = hlist_nulls_entry_safe(n, struct htab_elem, hash_node);
2007                 if (elem)
2008                         return elem;
2009
2010                 /* not found, unlock and go to the next bucket */
2011                 b = &htab->buckets[bucket_id++];
2012                 rcu_read_unlock();
2013                 skip_elems = 0;
2014         }
2015
2016         for (i = bucket_id; i < htab->n_buckets; i++) {
2017                 b = &htab->buckets[i];
2018                 rcu_read_lock();
2019
2020                 count = 0;
2021                 head = &b->head;
2022                 hlist_nulls_for_each_entry_rcu(elem, n, head, hash_node) {
2023                         if (count >= skip_elems) {
2024                                 info->bucket_id = i;
2025                                 info->skip_elems = count;
2026                                 return elem;
2027                         }
2028                         count++;
2029                 }
2030
2031                 rcu_read_unlock();
2032                 skip_elems = 0;
2033         }
2034
2035         info->bucket_id = i;
2036         info->skip_elems = 0;
2037         return NULL;
2038 }
2039
2040 static void *bpf_hash_map_seq_start(struct seq_file *seq, loff_t *pos)
2041 {
2042         struct bpf_iter_seq_hash_map_info *info = seq->private;
2043         struct htab_elem *elem;
2044
2045         elem = bpf_hash_map_seq_find_next(info, NULL);
2046         if (!elem)
2047                 return NULL;
2048
2049         if (*pos == 0)
2050                 ++*pos;
2051         return elem;
2052 }
2053
2054 static void *bpf_hash_map_seq_next(struct seq_file *seq, void *v, loff_t *pos)
2055 {
2056         struct bpf_iter_seq_hash_map_info *info = seq->private;
2057
2058         ++*pos;
2059         ++info->skip_elems;
2060         return bpf_hash_map_seq_find_next(info, v);
2061 }
2062
2063 static int __bpf_hash_map_seq_show(struct seq_file *seq, struct htab_elem *elem)
2064 {
2065         struct bpf_iter_seq_hash_map_info *info = seq->private;
2066         u32 roundup_key_size, roundup_value_size;
2067         struct bpf_iter__bpf_map_elem ctx = {};
2068         struct bpf_map *map = info->map;
2069         struct bpf_iter_meta meta;
2070         int ret = 0, off = 0, cpu;
2071         struct bpf_prog *prog;
2072         void __percpu *pptr;
2073
2074         meta.seq = seq;
2075         prog = bpf_iter_get_info(&meta, elem == NULL);
2076         if (prog) {
2077                 ctx.meta = &meta;
2078                 ctx.map = info->map;
2079                 if (elem) {
2080                         roundup_key_size = round_up(map->key_size, 8);
2081                         ctx.key = elem->key;
2082                         if (!info->percpu_value_buf) {
2083                                 ctx.value = elem->key + roundup_key_size;
2084                         } else {
2085                                 roundup_value_size = round_up(map->value_size, 8);
2086                                 pptr = htab_elem_get_ptr(elem, map->key_size);
2087                                 for_each_possible_cpu(cpu) {
2088                                         copy_map_value_long(map, info->percpu_value_buf + off,
2089                                                             per_cpu_ptr(pptr, cpu));
2090                                         check_and_init_map_value(map, info->percpu_value_buf + off);
2091                                         off += roundup_value_size;
2092                                 }
2093                                 ctx.value = info->percpu_value_buf;
2094                         }
2095                 }
2096                 ret = bpf_iter_run_prog(prog, &ctx);
2097         }
2098
2099         return ret;
2100 }
2101
2102 static int bpf_hash_map_seq_show(struct seq_file *seq, void *v)
2103 {
2104         return __bpf_hash_map_seq_show(seq, v);
2105 }
2106
2107 static void bpf_hash_map_seq_stop(struct seq_file *seq, void *v)
2108 {
2109         if (!v)
2110                 (void)__bpf_hash_map_seq_show(seq, NULL);
2111         else
2112                 rcu_read_unlock();
2113 }
2114
2115 static int bpf_iter_init_hash_map(void *priv_data,
2116                                   struct bpf_iter_aux_info *aux)
2117 {
2118         struct bpf_iter_seq_hash_map_info *seq_info = priv_data;
2119         struct bpf_map *map = aux->map;
2120         void *value_buf;
2121         u32 buf_size;
2122
2123         if (map->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
2124             map->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH) {
2125                 buf_size = round_up(map->value_size, 8) * num_possible_cpus();
2126                 value_buf = kmalloc(buf_size, GFP_USER | __GFP_NOWARN);
2127                 if (!value_buf)
2128                         return -ENOMEM;
2129
2130                 seq_info->percpu_value_buf = value_buf;
2131         }
2132
2133         bpf_map_inc_with_uref(map);
2134         seq_info->map = map;
2135         seq_info->htab = container_of(map, struct bpf_htab, map);
2136         return 0;
2137 }
2138
2139 static void bpf_iter_fini_hash_map(void *priv_data)
2140 {
2141         struct bpf_iter_seq_hash_map_info *seq_info = priv_data;
2142
2143         bpf_map_put_with_uref(seq_info->map);
2144         kfree(seq_info->percpu_value_buf);
2145 }
2146
2147 static const struct seq_operations bpf_hash_map_seq_ops = {
2148         .start  = bpf_hash_map_seq_start,
2149         .next   = bpf_hash_map_seq_next,
2150         .stop   = bpf_hash_map_seq_stop,
2151         .show   = bpf_hash_map_seq_show,
2152 };
2153
2154 static const struct bpf_iter_seq_info iter_seq_info = {
2155         .seq_ops                = &bpf_hash_map_seq_ops,
2156         .init_seq_private       = bpf_iter_init_hash_map,
2157         .fini_seq_private       = bpf_iter_fini_hash_map,
2158         .seq_priv_size          = sizeof(struct bpf_iter_seq_hash_map_info),
2159 };
2160
2161 static long bpf_for_each_hash_elem(struct bpf_map *map, bpf_callback_t callback_fn,
2162                                    void *callback_ctx, u64 flags)
2163 {
2164         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
2165         struct hlist_nulls_head *head;
2166         struct hlist_nulls_node *n;
2167         struct htab_elem *elem;
2168         u32 roundup_key_size;
2169         int i, num_elems = 0;
2170         void __percpu *pptr;
2171         struct bucket *b;
2172         void *key, *val;
2173         bool is_percpu;
2174         u64 ret = 0;
2175
2176         if (flags != 0)
2177                 return -EINVAL;
2178
2179         is_percpu = htab_is_percpu(htab);
2180
2181         roundup_key_size = round_up(map->key_size, 8);
2182         /* disable migration so percpu value prepared here will be the
2183          * same as the one seen by the bpf program with bpf_map_lookup_elem().
2184          */
2185         if (is_percpu)
2186                 migrate_disable();
2187         for (i = 0; i < htab->n_buckets; i++) {
2188                 b = &htab->buckets[i];
2189                 rcu_read_lock();
2190                 head = &b->head;
2191                 hlist_nulls_for_each_entry_rcu(elem, n, head, hash_node) {
2192                         key = elem->key;
2193                         if (is_percpu) {
2194                                 /* current cpu value for percpu map */
2195                                 pptr = htab_elem_get_ptr(elem, map->key_size);
2196                                 val = this_cpu_ptr(pptr);
2197                         } else {
2198                                 val = elem->key + roundup_key_size;
2199                         }
2200                         num_elems++;
2201                         ret = callback_fn((u64)(long)map, (u64)(long)key,
2202                                           (u64)(long)val, (u64)(long)callback_ctx, 0);
2203                         /* return value: 0 - continue, 1 - stop and return */
2204                         if (ret) {
2205                                 rcu_read_unlock();
2206                                 goto out;
2207                         }
2208                 }
2209                 rcu_read_unlock();
2210         }
2211 out:
2212         if (is_percpu)
2213                 migrate_enable();
2214         return num_elems;
2215 }
2216
2217 static u64 htab_map_mem_usage(const struct bpf_map *map)
2218 {
2219         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
2220         u32 value_size = round_up(htab->map.value_size, 8);
2221         bool prealloc = htab_is_prealloc(htab);
2222         bool percpu = htab_is_percpu(htab);
2223         bool lru = htab_is_lru(htab);
2224         u64 num_entries;
2225         u64 usage = sizeof(struct bpf_htab);
2226
2227         usage += sizeof(struct bucket) * htab->n_buckets;
2228         usage += sizeof(int) * num_possible_cpus() * HASHTAB_MAP_LOCK_COUNT;
2229         if (prealloc) {
2230                 num_entries = map->max_entries;
2231                 if (htab_has_extra_elems(htab))
2232                         num_entries += num_possible_cpus();
2233
2234                 usage += htab->elem_size * num_entries;
2235
2236                 if (percpu)
2237                         usage += value_size * num_possible_cpus() * num_entries;
2238                 else if (!lru)
2239                         usage += sizeof(struct htab_elem *) * num_possible_cpus();
2240         } else {
2241 #define LLIST_NODE_SZ sizeof(struct llist_node)
2242
2243                 num_entries = htab->use_percpu_counter ?
2244                                           percpu_counter_sum(&htab->pcount) :
2245                                           atomic_read(&htab->count);
2246                 usage += (htab->elem_size + LLIST_NODE_SZ) * num_entries;
2247                 if (percpu) {
2248                         usage += (LLIST_NODE_SZ + sizeof(void *)) * num_entries;
2249                         usage += value_size * num_possible_cpus() * num_entries;
2250                 }
2251         }
2252         return usage;
2253 }
2254
2255 BTF_ID_LIST_SINGLE(htab_map_btf_ids, struct, bpf_htab)
2256 const struct bpf_map_ops htab_map_ops = {
2257         .map_meta_equal = bpf_map_meta_equal,
2258         .map_alloc_check = htab_map_alloc_check,
2259         .map_alloc = htab_map_alloc,
2260         .map_free = htab_map_free,
2261         .map_get_next_key = htab_map_get_next_key,
2262         .map_release_uref = htab_map_free_timers,
2263         .map_lookup_elem = htab_map_lookup_elem,
2264         .map_lookup_and_delete_elem = htab_map_lookup_and_delete_elem,
2265         .map_update_elem = htab_map_update_elem,
2266         .map_delete_elem = htab_map_delete_elem,
2267         .map_gen_lookup = htab_map_gen_lookup,
2268         .map_seq_show_elem = htab_map_seq_show_elem,
2269         .map_set_for_each_callback_args = map_set_for_each_callback_args,
2270         .map_for_each_callback = bpf_for_each_hash_elem,
2271         .map_mem_usage = htab_map_mem_usage,
2272         BATCH_OPS(htab),
2273         .map_btf_id = &htab_map_btf_ids[0],
2274         .iter_seq_info = &iter_seq_info,
2275 };
2276
2277 const struct bpf_map_ops htab_lru_map_ops = {
2278         .map_meta_equal = bpf_map_meta_equal,
2279         .map_alloc_check = htab_map_alloc_check,
2280         .map_alloc = htab_map_alloc,
2281         .map_free = htab_map_free,
2282         .map_get_next_key = htab_map_get_next_key,
2283         .map_release_uref = htab_map_free_timers,
2284         .map_lookup_elem = htab_lru_map_lookup_elem,
2285         .map_lookup_and_delete_elem = htab_lru_map_lookup_and_delete_elem,
2286         .map_lookup_elem_sys_only = htab_lru_map_lookup_elem_sys,
2287         .map_update_elem = htab_lru_map_update_elem,
2288         .map_delete_elem = htab_lru_map_delete_elem,
2289         .map_gen_lookup = htab_lru_map_gen_lookup,
2290         .map_seq_show_elem = htab_map_seq_show_elem,
2291         .map_set_for_each_callback_args = map_set_for_each_callback_args,
2292         .map_for_each_callback = bpf_for_each_hash_elem,
2293         .map_mem_usage = htab_map_mem_usage,
2294         BATCH_OPS(htab_lru),
2295         .map_btf_id = &htab_map_btf_ids[0],
2296         .iter_seq_info = &iter_seq_info,
2297 };
2298
2299 /* Called from eBPF program */
2300 static void *htab_percpu_map_lookup_elem(struct bpf_map *map, void *key)
2301 {
2302         struct htab_elem *l = __htab_map_lookup_elem(map, key);
2303
2304         if (l)
2305                 return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
2306         else
2307                 return NULL;
2308 }
2309
2310 static void *htab_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                 return per_cpu_ptr(htab_elem_get_ptr(l, map->key_size), cpu);
2320         else
2321                 return NULL;
2322 }
2323
2324 static void *htab_lru_percpu_map_lookup_elem(struct bpf_map *map, void *key)
2325 {
2326         struct htab_elem *l = __htab_map_lookup_elem(map, key);
2327
2328         if (l) {
2329                 bpf_lru_node_set_ref(&l->lru_node);
2330                 return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
2331         }
2332
2333         return NULL;
2334 }
2335
2336 static void *htab_lru_percpu_map_lookup_percpu_elem(struct bpf_map *map, void *key, u32 cpu)
2337 {
2338         struct htab_elem *l;
2339
2340         if (cpu >= nr_cpu_ids)
2341                 return NULL;
2342
2343         l = __htab_map_lookup_elem(map, key);
2344         if (l) {
2345                 bpf_lru_node_set_ref(&l->lru_node);
2346                 return per_cpu_ptr(htab_elem_get_ptr(l, map->key_size), cpu);
2347         }
2348
2349         return NULL;
2350 }
2351
2352 int bpf_percpu_hash_copy(struct bpf_map *map, void *key, void *value)
2353 {
2354         struct htab_elem *l;
2355         void __percpu *pptr;
2356         int ret = -ENOENT;
2357         int cpu, off = 0;
2358         u32 size;
2359
2360         /* per_cpu areas are zero-filled and bpf programs can only
2361          * access 'value_size' of them, so copying rounded areas
2362          * will not leak any kernel data
2363          */
2364         size = round_up(map->value_size, 8);
2365         rcu_read_lock();
2366         l = __htab_map_lookup_elem(map, key);
2367         if (!l)
2368                 goto out;
2369         /* We do not mark LRU map element here in order to not mess up
2370          * eviction heuristics when user space does a map walk.
2371          */
2372         pptr = htab_elem_get_ptr(l, map->key_size);
2373         for_each_possible_cpu(cpu) {
2374                 copy_map_value_long(map, value + off, per_cpu_ptr(pptr, cpu));
2375                 check_and_init_map_value(map, value + off);
2376                 off += size;
2377         }
2378         ret = 0;
2379 out:
2380         rcu_read_unlock();
2381         return ret;
2382 }
2383
2384 int bpf_percpu_hash_update(struct bpf_map *map, void *key, void *value,
2385                            u64 map_flags)
2386 {
2387         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
2388         int ret;
2389
2390         rcu_read_lock();
2391         if (htab_is_lru(htab))
2392                 ret = __htab_lru_percpu_map_update_elem(map, key, value,
2393                                                         map_flags, true);
2394         else
2395                 ret = __htab_percpu_map_update_elem(map, key, value, map_flags,
2396                                                     true);
2397         rcu_read_unlock();
2398
2399         return ret;
2400 }
2401
2402 static void htab_percpu_map_seq_show_elem(struct bpf_map *map, void *key,
2403                                           struct seq_file *m)
2404 {
2405         struct htab_elem *l;
2406         void __percpu *pptr;
2407         int cpu;
2408
2409         rcu_read_lock();
2410
2411         l = __htab_map_lookup_elem(map, key);
2412         if (!l) {
2413                 rcu_read_unlock();
2414                 return;
2415         }
2416
2417         btf_type_seq_show(map->btf, map->btf_key_type_id, key, m);
2418         seq_puts(m, ": {\n");
2419         pptr = htab_elem_get_ptr(l, map->key_size);
2420         for_each_possible_cpu(cpu) {
2421                 seq_printf(m, "\tcpu%d: ", cpu);
2422                 btf_type_seq_show(map->btf, map->btf_value_type_id,
2423                                   per_cpu_ptr(pptr, cpu), m);
2424                 seq_puts(m, "\n");
2425         }
2426         seq_puts(m, "}\n");
2427
2428         rcu_read_unlock();
2429 }
2430
2431 const struct bpf_map_ops htab_percpu_map_ops = {
2432         .map_meta_equal = bpf_map_meta_equal,
2433         .map_alloc_check = htab_map_alloc_check,
2434         .map_alloc = htab_map_alloc,
2435         .map_free = htab_map_free,
2436         .map_get_next_key = htab_map_get_next_key,
2437         .map_lookup_elem = htab_percpu_map_lookup_elem,
2438         .map_lookup_and_delete_elem = htab_percpu_map_lookup_and_delete_elem,
2439         .map_update_elem = htab_percpu_map_update_elem,
2440         .map_delete_elem = htab_map_delete_elem,
2441         .map_lookup_percpu_elem = htab_percpu_map_lookup_percpu_elem,
2442         .map_seq_show_elem = htab_percpu_map_seq_show_elem,
2443         .map_set_for_each_callback_args = map_set_for_each_callback_args,
2444         .map_for_each_callback = bpf_for_each_hash_elem,
2445         .map_mem_usage = htab_map_mem_usage,
2446         BATCH_OPS(htab_percpu),
2447         .map_btf_id = &htab_map_btf_ids[0],
2448         .iter_seq_info = &iter_seq_info,
2449 };
2450
2451 const struct bpf_map_ops htab_lru_percpu_map_ops = {
2452         .map_meta_equal = bpf_map_meta_equal,
2453         .map_alloc_check = htab_map_alloc_check,
2454         .map_alloc = htab_map_alloc,
2455         .map_free = htab_map_free,
2456         .map_get_next_key = htab_map_get_next_key,
2457         .map_lookup_elem = htab_lru_percpu_map_lookup_elem,
2458         .map_lookup_and_delete_elem = htab_lru_percpu_map_lookup_and_delete_elem,
2459         .map_update_elem = htab_lru_percpu_map_update_elem,
2460         .map_delete_elem = htab_lru_map_delete_elem,
2461         .map_lookup_percpu_elem = htab_lru_percpu_map_lookup_percpu_elem,
2462         .map_seq_show_elem = htab_percpu_map_seq_show_elem,
2463         .map_set_for_each_callback_args = map_set_for_each_callback_args,
2464         .map_for_each_callback = bpf_for_each_hash_elem,
2465         .map_mem_usage = htab_map_mem_usage,
2466         BATCH_OPS(htab_lru_percpu),
2467         .map_btf_id = &htab_map_btf_ids[0],
2468         .iter_seq_info = &iter_seq_info,
2469 };
2470
2471 static int fd_htab_map_alloc_check(union bpf_attr *attr)
2472 {
2473         if (attr->value_size != sizeof(u32))
2474                 return -EINVAL;
2475         return htab_map_alloc_check(attr);
2476 }
2477
2478 static void fd_htab_map_free(struct bpf_map *map)
2479 {
2480         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
2481         struct hlist_nulls_node *n;
2482         struct hlist_nulls_head *head;
2483         struct htab_elem *l;
2484         int i;
2485
2486         for (i = 0; i < htab->n_buckets; i++) {
2487                 head = select_bucket(htab, i);
2488
2489                 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
2490                         void *ptr = fd_htab_map_get_ptr(map, l);
2491
2492                         map->ops->map_fd_put_ptr(map, ptr, false);
2493                 }
2494         }
2495
2496         htab_map_free(map);
2497 }
2498
2499 /* only called from syscall */
2500 int bpf_fd_htab_map_lookup_elem(struct bpf_map *map, void *key, u32 *value)
2501 {
2502         void **ptr;
2503         int ret = 0;
2504
2505         if (!map->ops->map_fd_sys_lookup_elem)
2506                 return -ENOTSUPP;
2507
2508         rcu_read_lock();
2509         ptr = htab_map_lookup_elem(map, key);
2510         if (ptr)
2511                 *value = map->ops->map_fd_sys_lookup_elem(READ_ONCE(*ptr));
2512         else
2513                 ret = -ENOENT;
2514         rcu_read_unlock();
2515
2516         return ret;
2517 }
2518
2519 /* only called from syscall */
2520 int bpf_fd_htab_map_update_elem(struct bpf_map *map, struct file *map_file,
2521                                 void *key, void *value, u64 map_flags)
2522 {
2523         void *ptr;
2524         int ret;
2525         u32 ufd = *(u32 *)value;
2526
2527         ptr = map->ops->map_fd_get_ptr(map, map_file, ufd);
2528         if (IS_ERR(ptr))
2529                 return PTR_ERR(ptr);
2530
2531         /* The htab bucket lock is always held during update operations in fd
2532          * htab map, and the following rcu_read_lock() is only used to avoid
2533          * the WARN_ON_ONCE in htab_map_update_elem().
2534          */
2535         rcu_read_lock();
2536         ret = htab_map_update_elem(map, key, &ptr, map_flags);
2537         rcu_read_unlock();
2538         if (ret)
2539                 map->ops->map_fd_put_ptr(map, ptr, false);
2540
2541         return ret;
2542 }
2543
2544 static struct bpf_map *htab_of_map_alloc(union bpf_attr *attr)
2545 {
2546         struct bpf_map *map, *inner_map_meta;
2547
2548         inner_map_meta = bpf_map_meta_alloc(attr->inner_map_fd);
2549         if (IS_ERR(inner_map_meta))
2550                 return inner_map_meta;
2551
2552         map = htab_map_alloc(attr);
2553         if (IS_ERR(map)) {
2554                 bpf_map_meta_free(inner_map_meta);
2555                 return map;
2556         }
2557
2558         map->inner_map_meta = inner_map_meta;
2559
2560         return map;
2561 }
2562
2563 static void *htab_of_map_lookup_elem(struct bpf_map *map, void *key)
2564 {
2565         struct bpf_map **inner_map  = htab_map_lookup_elem(map, key);
2566
2567         if (!inner_map)
2568                 return NULL;
2569
2570         return READ_ONCE(*inner_map);
2571 }
2572
2573 static int htab_of_map_gen_lookup(struct bpf_map *map,
2574                                   struct bpf_insn *insn_buf)
2575 {
2576         struct bpf_insn *insn = insn_buf;
2577         const int ret = BPF_REG_0;
2578
2579         BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
2580                      (void *(*)(struct bpf_map *map, void *key))NULL));
2581         *insn++ = BPF_EMIT_CALL(__htab_map_lookup_elem);
2582         *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 2);
2583         *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
2584                                 offsetof(struct htab_elem, key) +
2585                                 round_up(map->key_size, 8));
2586         *insn++ = BPF_LDX_MEM(BPF_DW, ret, ret, 0);
2587
2588         return insn - insn_buf;
2589 }
2590
2591 static void htab_of_map_free(struct bpf_map *map)
2592 {
2593         bpf_map_meta_free(map->inner_map_meta);
2594         fd_htab_map_free(map);
2595 }
2596
2597 const struct bpf_map_ops htab_of_maps_map_ops = {
2598         .map_alloc_check = fd_htab_map_alloc_check,
2599         .map_alloc = htab_of_map_alloc,
2600         .map_free = htab_of_map_free,
2601         .map_get_next_key = htab_map_get_next_key,
2602         .map_lookup_elem = htab_of_map_lookup_elem,
2603         .map_delete_elem = htab_map_delete_elem,
2604         .map_fd_get_ptr = bpf_map_fd_get_ptr,
2605         .map_fd_put_ptr = bpf_map_fd_put_ptr,
2606         .map_fd_sys_lookup_elem = bpf_map_fd_sys_lookup_elem,
2607         .map_gen_lookup = htab_of_map_gen_lookup,
2608         .map_check_btf = map_check_no_btf,
2609         .map_mem_usage = htab_map_mem_usage,
2610         BATCH_OPS(htab),
2611         .map_btf_id = &htab_map_btf_ids[0],
2612 };