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_and_wq(struct bpf_htab *htab)
 225 {
 226         u32 num_entries = htab->map.max_entries;
 227         int i;
 228
 229         if (htab_has_extra_elems(htab))
 230                 num_entries += num_possible_cpus();
 231
 232         for (i = 0; i < num_entries; i++) {
 233                 struct htab_elem *elem;
 234
 235                 elem = get_htab_elem(htab, i);
 236                 if (btf_record_has_field(htab->map.record, BPF_TIMER))
 237                         bpf_obj_free_timer(htab->map.record,
 238                                            elem->key + round_up(htab->map.key_size, 8));
 239                 if (btf_record_has_field(htab->map.record, BPF_WORKQUEUE))
 240                         bpf_obj_free_workqueue(htab->map.record,
 241                                                elem->key + round_up(htab->map.key_size, 8));
 242                 cond_resched();
 243         }
 244 }
 245
 246 static void htab_free_prealloced_fields(struct bpf_htab *htab)
 247 {
 248         u32 num_entries = htab->map.max_entries;
 249         int i;
 250
 251         if (IS_ERR_OR_NULL(htab->map.record))
 252                 return;
 253         if (htab_has_extra_elems(htab))
 254                 num_entries += num_possible_cpus();
 255         for (i = 0; i < num_entries; i++) {
 256                 struct htab_elem *elem;
 257
 258                 elem = get_htab_elem(htab, i);
 259                 if (htab_is_percpu(htab)) {
 260                         void __percpu *pptr = htab_elem_get_ptr(elem, htab->map.key_size);
 261                         int cpu;
 262
 263                         for_each_possible_cpu(cpu) {
 264                                 bpf_obj_free_fields(htab->map.record, per_cpu_ptr(pptr, cpu));
 265                                 cond_resched();
 266                         }
 267                 } else {
 268                         bpf_obj_free_fields(htab->map.record, elem->key + round_up(htab->map.key_size, 8));
 269                         cond_resched();
 270                 }
 271                 cond_resched();
 272         }
 273 }
 274
 275 static void htab_free_elems(struct bpf_htab *htab)
 276 {
 277         int i;
 278
 279         if (!htab_is_percpu(htab))
 280                 goto free_elems;
 281
 282         for (i = 0; i < htab->map.max_entries; i++) {
 283                 void __percpu *pptr;
 284
 285                 pptr = htab_elem_get_ptr(get_htab_elem(htab, i),
 286                                          htab->map.key_size);
 287                 free_percpu(pptr);
 288                 cond_resched();
 289         }
 290 free_elems:
 291         bpf_map_area_free(htab->elems);
 292 }
 293
 294 /* The LRU list has a lock (lru_lock). Each htab bucket has a lock
 295  * (bucket_lock). If both locks need to be acquired together, the lock
 296  * order is always lru_lock -> bucket_lock and this only happens in
 297  * bpf_lru_list.c logic. For example, certain code path of
 298  * bpf_lru_pop_free(), which is called by function prealloc_lru_pop(),
 299  * will acquire lru_lock first followed by acquiring bucket_lock.
 300  *
 301  * In hashtab.c, to avoid deadlock, lock acquisition of
 302  * bucket_lock followed by lru_lock is not allowed. In such cases,
 303  * bucket_lock needs to be released first before acquiring lru_lock.
 304  */
 305 static struct htab_elem *prealloc_lru_pop(struct bpf_htab *htab, void *key,
 306                                           u32 hash)
 307 {
 308         struct bpf_lru_node *node = bpf_lru_pop_free(&htab->lru, hash);
 309         struct htab_elem *l;
 310
 311         if (node) {
 312                 bpf_map_inc_elem_count(&htab->map);
 313                 l = container_of(node, struct htab_elem, lru_node);
 314                 memcpy(l->key, key, htab->map.key_size);
 315                 return l;
 316         }
 317
 318         return NULL;
 319 }
 320
 321 static int prealloc_init(struct bpf_htab *htab)
 322 {
 323         u32 num_entries = htab->map.max_entries;
 324         int err = -ENOMEM, i;
 325
 326         if (htab_has_extra_elems(htab))
 327                 num_entries += num_possible_cpus();
 328
 329         htab->elems = bpf_map_area_alloc((u64)htab->elem_size * num_entries,
 330                                          htab->map.numa_node);
 331         if (!htab->elems)
 332                 return -ENOMEM;
 333
 334         if (!htab_is_percpu(htab))
 335                 goto skip_percpu_elems;
 336
 337         for (i = 0; i < num_entries; i++) {
 338                 u32 size = round_up(htab->map.value_size, 8);
 339                 void __percpu *pptr;
 340
 341                 pptr = bpf_map_alloc_percpu(&htab->map, size, 8,
 342                                             GFP_USER | __GFP_NOWARN);
 343                 if (!pptr)
 344                         goto free_elems;
 345                 htab_elem_set_ptr(get_htab_elem(htab, i), htab->map.key_size,
 346                                   pptr);
 347                 cond_resched();
 348         }
 349
 350 skip_percpu_elems:
 351         if (htab_is_lru(htab))
 352                 err = bpf_lru_init(&htab->lru,
 353                                    htab->map.map_flags & BPF_F_NO_COMMON_LRU,
 354                                    offsetof(struct htab_elem, hash) -
 355                                    offsetof(struct htab_elem, lru_node),
 356                                    htab_lru_map_delete_node,
 357                                    htab);
 358         else
 359                 err = pcpu_freelist_init(&htab->freelist);
 360
 361         if (err)
 362                 goto free_elems;
 363
 364         if (htab_is_lru(htab))
 365                 bpf_lru_populate(&htab->lru, htab->elems,
 366                                  offsetof(struct htab_elem, lru_node),
 367                                  htab->elem_size, num_entries);
 368         else
 369                 pcpu_freelist_populate(&htab->freelist,
 370                                        htab->elems + offsetof(struct htab_elem, fnode),
 371                                        htab->elem_size, num_entries);
 372
 373         return 0;
 374
 375 free_elems:
 376         htab_free_elems(htab);
 377         return err;
 378 }
 379
 380 static void prealloc_destroy(struct bpf_htab *htab)
 381 {
 382         htab_free_elems(htab);
 383
 384         if (htab_is_lru(htab))
 385                 bpf_lru_destroy(&htab->lru);
 386         else
 387                 pcpu_freelist_destroy(&htab->freelist);
 388 }
 389
 390 static int alloc_extra_elems(struct bpf_htab *htab)
 391 {
 392         struct htab_elem *__percpu *pptr, *l_new;
 393         struct pcpu_freelist_node *l;
 394         int cpu;
 395
 396         pptr = bpf_map_alloc_percpu(&htab->map, sizeof(struct htab_elem *), 8,
 397                                     GFP_USER | __GFP_NOWARN);
 398         if (!pptr)
 399                 return -ENOMEM;
 400
 401         for_each_possible_cpu(cpu) {
 402                 l = pcpu_freelist_pop(&htab->freelist);
 403                 /* pop will succeed, since prealloc_init()
 404                  * preallocated extra num_possible_cpus elements
 405                  */
 406                 l_new = container_of(l, struct htab_elem, fnode);
 407                 *per_cpu_ptr(pptr, cpu) = l_new;
 408         }
 409         htab->extra_elems = pptr;
 410         return 0;
 411 }
 412
 413 /* Called from syscall */
 414 static int htab_map_alloc_check(union bpf_attr *attr)
 415 {
 416         bool percpu = (attr->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
 417                        attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
 418         bool lru = (attr->map_type == BPF_MAP_TYPE_LRU_HASH ||
 419                     attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
 420         /* percpu_lru means each cpu has its own LRU list.
 421          * it is different from BPF_MAP_TYPE_PERCPU_HASH where
 422          * the map's value itself is percpu.  percpu_lru has
 423          * nothing to do with the map's value.
 424          */
 425         bool percpu_lru = (attr->map_flags & BPF_F_NO_COMMON_LRU);
 426         bool prealloc = !(attr->map_flags & BPF_F_NO_PREALLOC);
 427         bool zero_seed = (attr->map_flags & BPF_F_ZERO_SEED);
 428         int numa_node = bpf_map_attr_numa_node(attr);
 429
 430         BUILD_BUG_ON(offsetof(struct htab_elem, fnode.next) !=
 431                      offsetof(struct htab_elem, hash_node.pprev));
 432
 433         if (zero_seed && !capable(CAP_SYS_ADMIN))
 434                 /* Guard against local DoS, and discourage production use. */
 435                 return -EPERM;
 436
 437         if (attr->map_flags & ~HTAB_CREATE_FLAG_MASK ||
 438             !bpf_map_flags_access_ok(attr->map_flags))
 439                 return -EINVAL;
 440
 441         if (!lru && percpu_lru)
 442                 return -EINVAL;
 443
 444         if (lru && !prealloc)
 445                 return -ENOTSUPP;
 446
 447         if (numa_node != NUMA_NO_NODE && (percpu || percpu_lru))
 448                 return -EINVAL;
 449
 450         /* check sanity of attributes.
 451          * value_size == 0 may be allowed in the future to use map as a set
 452          */
 453         if (attr->max_entries == 0 || attr->key_size == 0 ||
 454             attr->value_size == 0)
 455                 return -EINVAL;
 456
 457         if ((u64)attr->key_size + attr->value_size >= KMALLOC_MAX_SIZE -
 458            sizeof(struct htab_elem))
 459                 /* if key_size + value_size is bigger, the user space won't be
 460                  * able to access the elements via bpf syscall. This check
 461                  * also makes sure that the elem_size doesn't overflow and it's
 462                  * kmalloc-able later in htab_map_update_elem()
 463                  */
 464                 return -E2BIG;
 465
 466         return 0;
 467 }
 468
 469 static struct bpf_map *htab_map_alloc(union bpf_attr *attr)
 470 {
 471         bool percpu = (attr->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
 472                        attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
 473         bool lru = (attr->map_type == BPF_MAP_TYPE_LRU_HASH ||
 474                     attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
 475         /* percpu_lru means each cpu has its own LRU list.
 476          * it is different from BPF_MAP_TYPE_PERCPU_HASH where
 477          * the map's value itself is percpu.  percpu_lru has
 478          * nothing to do with the map's value.
 479          */
 480         bool percpu_lru = (attr->map_flags & BPF_F_NO_COMMON_LRU);
 481         bool prealloc = !(attr->map_flags & BPF_F_NO_PREALLOC);
 482         struct bpf_htab *htab;
 483         int err, i;
 484
 485         htab = bpf_map_area_alloc(sizeof(*htab), NUMA_NO_NODE);
 486         if (!htab)
 487                 return ERR_PTR(-ENOMEM);
 488
 489         lockdep_register_key(&htab->lockdep_key);
 490
 491         bpf_map_init_from_attr(&htab->map, attr);
 492
 493         if (percpu_lru) {
 494                 /* ensure each CPU's lru list has >=1 elements.
 495                  * since we are at it, make each lru list has the same
 496                  * number of elements.
 497                  */
 498                 htab->map.max_entries = roundup(attr->max_entries,
 499                                                 num_possible_cpus());
 500                 if (htab->map.max_entries < attr->max_entries)
 501                         htab->map.max_entries = rounddown(attr->max_entries,
 502                                                           num_possible_cpus());
 503         }
 504
 505         /* hash table size must be power of 2; roundup_pow_of_two() can overflow
 506          * into UB on 32-bit arches, so check that first
 507          */
 508         err = -E2BIG;
 509         if (htab->map.max_entries > 1UL << 31)
 510                 goto free_htab;
 511
 512         htab->n_buckets = roundup_pow_of_two(htab->map.max_entries);
 513
 514         htab->elem_size = sizeof(struct htab_elem) +
 515                           round_up(htab->map.key_size, 8);
 516         if (percpu)
 517                 htab->elem_size += sizeof(void *);
 518         else
 519                 htab->elem_size += round_up(htab->map.value_size, 8);
 520
 521         /* check for u32 overflow */
 522         if (htab->n_buckets > U32_MAX / sizeof(struct bucket))
 523                 goto free_htab;
 524
 525         err = bpf_map_init_elem_count(&htab->map);
 526         if (err)
 527                 goto free_htab;
 528
 529         err = -ENOMEM;
 530         htab->buckets = bpf_map_area_alloc(htab->n_buckets *
 531                                            sizeof(struct bucket),
 532                                            htab->map.numa_node);
 533         if (!htab->buckets)
 534                 goto free_elem_count;
 535
 536         for (i = 0; i < HASHTAB_MAP_LOCK_COUNT; i++) {
 537                 htab->map_locked[i] = bpf_map_alloc_percpu(&htab->map,
 538                                                            sizeof(int),
 539                                                            sizeof(int),
 540                                                            GFP_USER);
 541                 if (!htab->map_locked[i])
 542                         goto free_map_locked;
 543         }
 544
 545         if (htab->map.map_flags & BPF_F_ZERO_SEED)
 546                 htab->hashrnd = 0;
 547         else
 548                 htab->hashrnd = get_random_u32();
 549
 550         htab_init_buckets(htab);
 551
 552 /* compute_batch_value() computes batch value as num_online_cpus() * 2
 553  * and __percpu_counter_compare() needs
 554  * htab->max_entries - cur_number_of_elems to be more than batch * num_online_cpus()
 555  * for percpu_counter to be faster than atomic_t. In practice the average bpf
 556  * hash map size is 10k, which means that a system with 64 cpus will fill
 557  * hashmap to 20% of 10k before percpu_counter becomes ineffective. Therefore
 558  * define our own batch count as 32 then 10k hash map can be filled up to 80%:
 559  * 10k - 8k > 32 _batch_ * 64 _cpus_
 560  * and __percpu_counter_compare() will still be fast. At that point hash map
 561  * collisions will dominate its performance anyway. Assume that hash map filled
 562  * to 50+% isn't going to be O(1) and use the following formula to choose
 563  * between percpu_counter and atomic_t.
 564  */
 565 #define PERCPU_COUNTER_BATCH 32
 566         if (attr->max_entries / 2 > num_online_cpus() * PERCPU_COUNTER_BATCH)
 567                 htab->use_percpu_counter = true;
 568
 569         if (htab->use_percpu_counter) {
 570                 err = percpu_counter_init(&htab->pcount, 0, GFP_KERNEL);
 571                 if (err)
 572                         goto free_map_locked;
 573         }
 574
 575         if (prealloc) {
 576                 err = prealloc_init(htab);
 577                 if (err)
 578                         goto free_map_locked;
 579
 580                 if (!percpu && !lru) {
 581                         /* lru itself can remove the least used element, so
 582                          * there is no need for an extra elem during map_update.
 583                          */
 584                         err = alloc_extra_elems(htab);
 585                         if (err)
 586                                 goto free_prealloc;
 587                 }
 588         } else {
 589                 err = bpf_mem_alloc_init(&htab->ma, htab->elem_size, false);
 590                 if (err)
 591                         goto free_map_locked;
 592                 if (percpu) {
 593                         err = bpf_mem_alloc_init(&htab->pcpu_ma,
 594                                                  round_up(htab->map.value_size, 8), true);
 595                         if (err)
 596                                 goto free_map_locked;
 597                 }
 598         }
 599
 600         return &htab->map;
 601
 602 free_prealloc:
 603         prealloc_destroy(htab);
 604 free_map_locked:
 605         if (htab->use_percpu_counter)
 606                 percpu_counter_destroy(&htab->pcount);
 607         for (i = 0; i < HASHTAB_MAP_LOCK_COUNT; i++)
 608                 free_percpu(htab->map_locked[i]);
 609         bpf_map_area_free(htab->buckets);
 610         bpf_mem_alloc_destroy(&htab->pcpu_ma);
 611         bpf_mem_alloc_destroy(&htab->ma);
 612 free_elem_count:
 613         bpf_map_free_elem_count(&htab->map);
 614 free_htab:
 615         lockdep_unregister_key(&htab->lockdep_key);
 616         bpf_map_area_free(htab);
 617         return ERR_PTR(err);
 618 }
 619
 620 static inline u32 htab_map_hash(const void *key, u32 key_len, u32 hashrnd)
 621 {
 622         if (likely(key_len % 4 == 0))
 623                 return jhash2(key, key_len / 4, hashrnd);
 624         return jhash(key, key_len, hashrnd);
 625 }
 626
 627 static inline struct bucket *__select_bucket(struct bpf_htab *htab, u32 hash)
 628 {
 629         return &htab->buckets[hash & (htab->n_buckets - 1)];
 630 }
 631
 632 static inline struct hlist_nulls_head *select_bucket(struct bpf_htab *htab, u32 hash)
 633 {
 634         return &__select_bucket(htab, hash)->head;
 635 }
 636
 637 /* this lookup function can only be called with bucket lock taken */
 638 static struct htab_elem *lookup_elem_raw(struct hlist_nulls_head *head, u32 hash,
 639                                          void *key, u32 key_size)
 640 {
 641         struct hlist_nulls_node *n;
 642         struct htab_elem *l;
 643
 644         hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
 645                 if (l->hash == hash && !memcmp(&l->key, key, key_size))
 646                         return l;
 647
 648         return NULL;
 649 }
 650
 651 /* can be called without bucket lock. it will repeat the loop in
 652  * the unlikely event when elements moved from one bucket into another
 653  * while link list is being walked
 654  */
 655 static struct htab_elem *lookup_nulls_elem_raw(struct hlist_nulls_head *head,
 656                                                u32 hash, void *key,
 657                                                u32 key_size, u32 n_buckets)
 658 {
 659         struct hlist_nulls_node *n;
 660         struct htab_elem *l;
 661
 662 again:
 663         hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
 664                 if (l->hash == hash && !memcmp(&l->key, key, key_size))
 665                         return l;
 666
 667         if (unlikely(get_nulls_value(n) != (hash & (n_buckets - 1))))
 668                 goto again;
 669
 670         return NULL;
 671 }
 672
 673 /* Called from syscall or from eBPF program directly, so
 674  * arguments have to match bpf_map_lookup_elem() exactly.
 675  * The return value is adjusted by BPF instructions
 676  * in htab_map_gen_lookup().
 677  */
 678 static void *__htab_map_lookup_elem(struct bpf_map *map, void *key)
 679 {
 680         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 681         struct hlist_nulls_head *head;
 682         struct htab_elem *l;
 683         u32 hash, key_size;
 684
 685         WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
 686                      !rcu_read_lock_bh_held());
 687
 688         key_size = map->key_size;
 689
 690         hash = htab_map_hash(key, key_size, htab->hashrnd);
 691
 692         head = select_bucket(htab, hash);
 693
 694         l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
 695
 696         return l;
 697 }
 698
 699 static void *htab_map_lookup_elem(struct bpf_map *map, void *key)
 700 {
 701         struct htab_elem *l = __htab_map_lookup_elem(map, key);
 702
 703         if (l)
 704                 return l->key + round_up(map->key_size, 8);
 705
 706         return NULL;
 707 }
 708
 709 /* inline bpf_map_lookup_elem() call.
 710  * Instead of:
 711  * bpf_prog
 712  *   bpf_map_lookup_elem
 713  *     map->ops->map_lookup_elem
 714  *       htab_map_lookup_elem
 715  *         __htab_map_lookup_elem
 716  * do:
 717  * bpf_prog
 718  *   __htab_map_lookup_elem
 719  */
 720 static int htab_map_gen_lookup(struct bpf_map *map, struct bpf_insn *insn_buf)
 721 {
 722         struct bpf_insn *insn = insn_buf;
 723         const int ret = BPF_REG_0;
 724
 725         BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
 726                      (void *(*)(struct bpf_map *map, void *key))NULL));
 727         *insn++ = BPF_EMIT_CALL(__htab_map_lookup_elem);
 728         *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 1);
 729         *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
 730                                 offsetof(struct htab_elem, key) +
 731                                 round_up(map->key_size, 8));
 732         return insn - insn_buf;
 733 }
 734
 735 static __always_inline void *__htab_lru_map_lookup_elem(struct bpf_map *map,
 736                                                         void *key, const bool mark)
 737 {
 738         struct htab_elem *l = __htab_map_lookup_elem(map, key);
 739
 740         if (l) {
 741                 if (mark)
 742                         bpf_lru_node_set_ref(&l->lru_node);
 743                 return l->key + round_up(map->key_size, 8);
 744         }
 745
 746         return NULL;
 747 }
 748
 749 static void *htab_lru_map_lookup_elem(struct bpf_map *map, void *key)
 750 {
 751         return __htab_lru_map_lookup_elem(map, key, true);
 752 }
 753
 754 static void *htab_lru_map_lookup_elem_sys(struct bpf_map *map, void *key)
 755 {
 756         return __htab_lru_map_lookup_elem(map, key, false);
 757 }
 758
 759 static int htab_lru_map_gen_lookup(struct bpf_map *map,
 760                                    struct bpf_insn *insn_buf)
 761 {
 762         struct bpf_insn *insn = insn_buf;
 763         const int ret = BPF_REG_0;
 764         const int ref_reg = BPF_REG_1;
 765
 766         BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
 767                      (void *(*)(struct bpf_map *map, void *key))NULL));
 768         *insn++ = BPF_EMIT_CALL(__htab_map_lookup_elem);
 769         *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 4);
 770         *insn++ = BPF_LDX_MEM(BPF_B, ref_reg, ret,
 771                               offsetof(struct htab_elem, lru_node) +
 772                               offsetof(struct bpf_lru_node, ref));
 773         *insn++ = BPF_JMP_IMM(BPF_JNE, ref_reg, 0, 1);
 774         *insn++ = BPF_ST_MEM(BPF_B, ret,
 775                              offsetof(struct htab_elem, lru_node) +
 776                              offsetof(struct bpf_lru_node, ref),
 777                              1);
 778         *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
 779                                 offsetof(struct htab_elem, key) +
 780                                 round_up(map->key_size, 8));
 781         return insn - insn_buf;
 782 }
 783
 784 static void check_and_free_fields(struct bpf_htab *htab,
 785                                   struct htab_elem *elem)
 786 {
 787         if (htab_is_percpu(htab)) {
 788                 void __percpu *pptr = htab_elem_get_ptr(elem, htab->map.key_size);
 789                 int cpu;
 790
 791                 for_each_possible_cpu(cpu)
 792                         bpf_obj_free_fields(htab->map.record, per_cpu_ptr(pptr, cpu));
 793         } else {
 794                 void *map_value = elem->key + round_up(htab->map.key_size, 8);
 795
 796                 bpf_obj_free_fields(htab->map.record, map_value);
 797         }
 798 }
 799
 800 /* It is called from the bpf_lru_list when the LRU needs to delete
 801  * older elements from the htab.
 802  */
 803 static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node)
 804 {
 805         struct bpf_htab *htab = arg;
 806         struct htab_elem *l = NULL, *tgt_l;
 807         struct hlist_nulls_head *head;
 808         struct hlist_nulls_node *n;
 809         unsigned long flags;
 810         struct bucket *b;
 811         int ret;
 812
 813         tgt_l = container_of(node, struct htab_elem, lru_node);
 814         b = __select_bucket(htab, tgt_l->hash);
 815         head = &b->head;
 816
 817         ret = htab_lock_bucket(htab, b, tgt_l->hash, &flags);
 818         if (ret)
 819                 return false;
 820
 821         hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
 822                 if (l == tgt_l) {
 823                         hlist_nulls_del_rcu(&l->hash_node);
 824                         check_and_free_fields(htab, l);
 825                         bpf_map_dec_elem_count(&htab->map);
 826                         break;
 827                 }
 828
 829         htab_unlock_bucket(htab, b, tgt_l->hash, flags);
 830
 831         return l == tgt_l;
 832 }
 833
 834 /* Called from syscall */
 835 static int htab_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
 836 {
 837         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 838         struct hlist_nulls_head *head;
 839         struct htab_elem *l, *next_l;
 840         u32 hash, key_size;
 841         int i = 0;
 842
 843         WARN_ON_ONCE(!rcu_read_lock_held());
 844
 845         key_size = map->key_size;
 846
 847         if (!key)
 848                 goto find_first_elem;
 849
 850         hash = htab_map_hash(key, key_size, htab->hashrnd);
 851
 852         head = select_bucket(htab, hash);
 853
 854         /* lookup the key */
 855         l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
 856
 857         if (!l)
 858                 goto find_first_elem;
 859
 860         /* key was found, get next key in the same bucket */
 861         next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_next_rcu(&l->hash_node)),
 862                                   struct htab_elem, hash_node);
 863
 864         if (next_l) {
 865                 /* if next elem in this hash list is non-zero, just return it */
 866                 memcpy(next_key, next_l->key, key_size);
 867                 return 0;
 868         }
 869
 870         /* no more elements in this hash list, go to the next bucket */
 871         i = hash & (htab->n_buckets - 1);
 872         i++;
 873
 874 find_first_elem:
 875         /* iterate over buckets */
 876         for (; i < htab->n_buckets; i++) {
 877                 head = select_bucket(htab, i);
 878
 879                 /* pick first element in the bucket */
 880                 next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_first_rcu(head)),
 881                                           struct htab_elem, hash_node);
 882                 if (next_l) {
 883                         /* if it's not empty, just return it */
 884                         memcpy(next_key, next_l->key, key_size);
 885                         return 0;
 886                 }
 887         }
 888
 889         /* iterated over all buckets and all elements */
 890         return -ENOENT;
 891 }
 892
 893 static void htab_elem_free(struct bpf_htab *htab, struct htab_elem *l)
 894 {
 895         check_and_free_fields(htab, l);
 896         if (htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH)
 897                 bpf_mem_cache_free(&htab->pcpu_ma, l->ptr_to_pptr);
 898         bpf_mem_cache_free(&htab->ma, l);
 899 }
 900
 901 static void htab_put_fd_value(struct bpf_htab *htab, struct htab_elem *l)
 902 {
 903         struct bpf_map *map = &htab->map;
 904         void *ptr;
 905
 906         if (map->ops->map_fd_put_ptr) {
 907                 ptr = fd_htab_map_get_ptr(map, l);
 908                 map->ops->map_fd_put_ptr(map, ptr, true);
 909         }
 910 }
 911
 912 static bool is_map_full(struct bpf_htab *htab)
 913 {
 914         if (htab->use_percpu_counter)
 915                 return __percpu_counter_compare(&htab->pcount, htab->map.max_entries,
 916                                                 PERCPU_COUNTER_BATCH) >= 0;
 917         return atomic_read(&htab->count) >= htab->map.max_entries;
 918 }
 919
 920 static void inc_elem_count(struct bpf_htab *htab)
 921 {
 922         bpf_map_inc_elem_count(&htab->map);
 923
 924         if (htab->use_percpu_counter)
 925                 percpu_counter_add_batch(&htab->pcount, 1, PERCPU_COUNTER_BATCH);
 926         else
 927                 atomic_inc(&htab->count);
 928 }
 929
 930 static void dec_elem_count(struct bpf_htab *htab)
 931 {
 932         bpf_map_dec_elem_count(&htab->map);
 933
 934         if (htab->use_percpu_counter)
 935                 percpu_counter_add_batch(&htab->pcount, -1, PERCPU_COUNTER_BATCH);
 936         else
 937                 atomic_dec(&htab->count);
 938 }
 939
 940
 941 static void free_htab_elem(struct bpf_htab *htab, struct htab_elem *l)
 942 {
 943         htab_put_fd_value(htab, l);
 944
 945         if (htab_is_prealloc(htab)) {
 946                 bpf_map_dec_elem_count(&htab->map);
 947                 check_and_free_fields(htab, l);
 948                 __pcpu_freelist_push(&htab->freelist, &l->fnode);
 949         } else {
 950                 dec_elem_count(htab);
 951                 htab_elem_free(htab, l);
 952         }
 953 }
 954
 955 static void pcpu_copy_value(struct bpf_htab *htab, void __percpu *pptr,
 956                             void *value, bool onallcpus)
 957 {
 958         if (!onallcpus) {
 959                 /* copy true value_size bytes */
 960                 copy_map_value(&htab->map, this_cpu_ptr(pptr), value);
 961         } else {
 962                 u32 size = round_up(htab->map.value_size, 8);
 963                 int off = 0, cpu;
 964
 965                 for_each_possible_cpu(cpu) {
 966                         copy_map_value_long(&htab->map, per_cpu_ptr(pptr, cpu), value + off);
 967                         off += size;
 968                 }
 969         }
 970 }
 971
 972 static void pcpu_init_value(struct bpf_htab *htab, void __percpu *pptr,
 973                             void *value, bool onallcpus)
 974 {
 975         /* When not setting the initial value on all cpus, zero-fill element
 976          * values for other cpus. Otherwise, bpf program has no way to ensure
 977          * known initial values for cpus other than current one
 978          * (onallcpus=false always when coming from bpf prog).
 979          */
 980         if (!onallcpus) {
 981                 int current_cpu = raw_smp_processor_id();
 982                 int cpu;
 983
 984                 for_each_possible_cpu(cpu) {
 985                         if (cpu == current_cpu)
 986                                 copy_map_value_long(&htab->map, per_cpu_ptr(pptr, cpu), value);
 987                         else /* Since elem is preallocated, we cannot touch special fields */
 988                                 zero_map_value(&htab->map, per_cpu_ptr(pptr, cpu));
 989                 }
 990         } else {
 991                 pcpu_copy_value(htab, pptr, value, onallcpus);
 992         }
 993 }
 994
 995 static bool fd_htab_map_needs_adjust(const struct bpf_htab *htab)
 996 {
 997         return htab->map.map_type == BPF_MAP_TYPE_HASH_OF_MAPS &&
 998                BITS_PER_LONG == 64;
 999 }
1000
1001 static struct htab_elem *alloc_htab_elem(struct bpf_htab *htab, void *key,
1002                                          void *value, u32 key_size, u32 hash,
1003                                          bool percpu, bool onallcpus,
1004                                          struct htab_elem *old_elem)
1005 {
1006         u32 size = htab->map.value_size;
1007         bool prealloc = htab_is_prealloc(htab);
1008         struct htab_elem *l_new, **pl_new;
1009         void __percpu *pptr;
1010
1011         if (prealloc) {
1012                 if (old_elem) {
1013                         /* if we're updating the existing element,
1014                          * use per-cpu extra elems to avoid freelist_pop/push
1015                          */
1016                         pl_new = this_cpu_ptr(htab->extra_elems);
1017                         l_new = *pl_new;
1018                         htab_put_fd_value(htab, old_elem);
1019                         *pl_new = old_elem;
1020                 } else {
1021                         struct pcpu_freelist_node *l;
1022
1023                         l = __pcpu_freelist_pop(&htab->freelist);
1024                         if (!l)
1025                                 return ERR_PTR(-E2BIG);
1026                         l_new = container_of(l, struct htab_elem, fnode);
1027                         bpf_map_inc_elem_count(&htab->map);
1028                 }
1029         } else {
1030                 if (is_map_full(htab))
1031                         if (!old_elem)
1032                                 /* when map is full and update() is replacing
1033                                  * old element, it's ok to allocate, since
1034                                  * old element will be freed immediately.
1035                                  * Otherwise return an error
1036                                  */
1037                                 return ERR_PTR(-E2BIG);
1038                 inc_elem_count(htab);
1039                 l_new = bpf_mem_cache_alloc(&htab->ma);
1040                 if (!l_new) {
1041                         l_new = ERR_PTR(-ENOMEM);
1042                         goto dec_count;
1043                 }
1044         }
1045
1046         memcpy(l_new->key, key, key_size);
1047         if (percpu) {
1048                 if (prealloc) {
1049                         pptr = htab_elem_get_ptr(l_new, key_size);
1050                 } else {
1051                         /* alloc_percpu zero-fills */
1052                         pptr = bpf_mem_cache_alloc(&htab->pcpu_ma);
1053                         if (!pptr) {
1054                                 bpf_mem_cache_free(&htab->ma, l_new);
1055                                 l_new = ERR_PTR(-ENOMEM);
1056                                 goto dec_count;
1057                         }
1058                         l_new->ptr_to_pptr = pptr;
1059                         pptr = *(void **)pptr;
1060                 }
1061
1062                 pcpu_init_value(htab, pptr, value, onallcpus);
1063
1064                 if (!prealloc)
1065                         htab_elem_set_ptr(l_new, key_size, pptr);
1066         } else if (fd_htab_map_needs_adjust(htab)) {
1067                 size = round_up(size, 8);
1068                 memcpy(l_new->key + round_up(key_size, 8), value, size);
1069         } else {
1070                 copy_map_value(&htab->map,
1071                                l_new->key + round_up(key_size, 8),
1072                                value);
1073         }
1074
1075         l_new->hash = hash;
1076         return l_new;
1077 dec_count:
1078         dec_elem_count(htab);
1079         return l_new;
1080 }
1081
1082 static int check_flags(struct bpf_htab *htab, struct htab_elem *l_old,
1083                        u64 map_flags)
1084 {
1085         if (l_old && (map_flags & ~BPF_F_LOCK) == BPF_NOEXIST)
1086                 /* elem already exists */
1087                 return -EEXIST;
1088
1089         if (!l_old && (map_flags & ~BPF_F_LOCK) == BPF_EXIST)
1090                 /* elem doesn't exist, cannot update it */
1091                 return -ENOENT;
1092
1093         return 0;
1094 }
1095
1096 /* Called from syscall or from eBPF program */
1097 static long htab_map_update_elem(struct bpf_map *map, void *key, void *value,
1098                                  u64 map_flags)
1099 {
1100         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1101         struct htab_elem *l_new = NULL, *l_old;
1102         struct hlist_nulls_head *head;
1103         unsigned long flags;
1104         struct bucket *b;
1105         u32 key_size, hash;
1106         int ret;
1107
1108         if (unlikely((map_flags & ~BPF_F_LOCK) > BPF_EXIST))
1109                 /* unknown flags */
1110                 return -EINVAL;
1111
1112         WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
1113                      !rcu_read_lock_bh_held());
1114
1115         key_size = map->key_size;
1116
1117         hash = htab_map_hash(key, key_size, htab->hashrnd);
1118
1119         b = __select_bucket(htab, hash);
1120         head = &b->head;
1121
1122         if (unlikely(map_flags & BPF_F_LOCK)) {
1123                 if (unlikely(!btf_record_has_field(map->record, BPF_SPIN_LOCK)))
1124                         return -EINVAL;
1125                 /* find an element without taking the bucket lock */
1126                 l_old = lookup_nulls_elem_raw(head, hash, key, key_size,
1127                                               htab->n_buckets);
1128                 ret = check_flags(htab, l_old, map_flags);
1129                 if (ret)
1130                         return ret;
1131                 if (l_old) {
1132                         /* grab the element lock and update value in place */
1133                         copy_map_value_locked(map,
1134                                               l_old->key + round_up(key_size, 8),
1135                                               value, false);
1136                         return 0;
1137                 }
1138                 /* fall through, grab the bucket lock and lookup again.
1139                  * 99.9% chance that the element won't be found,
1140                  * but second lookup under lock has to be done.
1141                  */
1142         }
1143
1144         ret = htab_lock_bucket(htab, b, hash, &flags);
1145         if (ret)
1146                 return ret;
1147
1148         l_old = lookup_elem_raw(head, hash, key, key_size);
1149
1150         ret = check_flags(htab, l_old, map_flags);
1151         if (ret)
1152                 goto err;
1153
1154         if (unlikely(l_old && (map_flags & BPF_F_LOCK))) {
1155                 /* first lookup without the bucket lock didn't find the element,
1156                  * but second lookup with the bucket lock found it.
1157                  * This case is highly unlikely, but has to be dealt with:
1158                  * grab the element lock in addition to the bucket lock
1159                  * and update element in place
1160                  */
1161                 copy_map_value_locked(map,
1162                                       l_old->key + round_up(key_size, 8),
1163                                       value, false);
1164                 ret = 0;
1165                 goto err;
1166         }
1167
1168         l_new = alloc_htab_elem(htab, key, value, key_size, hash, false, false,
1169                                 l_old);
1170         if (IS_ERR(l_new)) {
1171                 /* all pre-allocated elements are in use or memory exhausted */
1172                 ret = PTR_ERR(l_new);
1173                 goto err;
1174         }
1175
1176         /* add new element to the head of the list, so that
1177          * concurrent search will find it before old elem
1178          */
1179         hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1180         if (l_old) {
1181                 hlist_nulls_del_rcu(&l_old->hash_node);
1182                 if (!htab_is_prealloc(htab))
1183                         free_htab_elem(htab, l_old);
1184                 else
1185                         check_and_free_fields(htab, l_old);
1186         }
1187         ret = 0;
1188 err:
1189         htab_unlock_bucket(htab, b, hash, flags);
1190         return ret;
1191 }
1192
1193 static void htab_lru_push_free(struct bpf_htab *htab, struct htab_elem *elem)
1194 {
1195         check_and_free_fields(htab, elem);
1196         bpf_map_dec_elem_count(&htab->map);
1197         bpf_lru_push_free(&htab->lru, &elem->lru_node);
1198 }
1199
1200 static long htab_lru_map_update_elem(struct bpf_map *map, void *key, void *value,
1201                                      u64 map_flags)
1202 {
1203         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1204         struct htab_elem *l_new, *l_old = NULL;
1205         struct hlist_nulls_head *head;
1206         unsigned long flags;
1207         struct bucket *b;
1208         u32 key_size, hash;
1209         int ret;
1210
1211         if (unlikely(map_flags > BPF_EXIST))
1212                 /* unknown flags */
1213                 return -EINVAL;
1214
1215         WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
1216                      !rcu_read_lock_bh_held());
1217
1218         key_size = map->key_size;
1219
1220         hash = htab_map_hash(key, key_size, htab->hashrnd);
1221
1222         b = __select_bucket(htab, hash);
1223         head = &b->head;
1224
1225         /* For LRU, we need to alloc before taking bucket's
1226          * spinlock because getting free nodes from LRU may need
1227          * to remove older elements from htab and this removal
1228          * operation will need a bucket lock.
1229          */
1230         l_new = prealloc_lru_pop(htab, key, hash);
1231         if (!l_new)
1232                 return -ENOMEM;
1233         copy_map_value(&htab->map,
1234                        l_new->key + round_up(map->key_size, 8), value);
1235
1236         ret = htab_lock_bucket(htab, b, hash, &flags);
1237         if (ret)
1238                 goto err_lock_bucket;
1239
1240         l_old = lookup_elem_raw(head, hash, key, key_size);
1241
1242         ret = check_flags(htab, l_old, map_flags);
1243         if (ret)
1244                 goto err;
1245
1246         /* add new element to the head of the list, so that
1247          * concurrent search will find it before old elem
1248          */
1249         hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1250         if (l_old) {
1251                 bpf_lru_node_set_ref(&l_new->lru_node);
1252                 hlist_nulls_del_rcu(&l_old->hash_node);
1253         }
1254         ret = 0;
1255
1256 err:
1257         htab_unlock_bucket(htab, b, hash, flags);
1258
1259 err_lock_bucket:
1260         if (ret)
1261                 htab_lru_push_free(htab, l_new);
1262         else if (l_old)
1263                 htab_lru_push_free(htab, l_old);
1264
1265         return ret;
1266 }
1267
1268 static long __htab_percpu_map_update_elem(struct bpf_map *map, void *key,
1269                                           void *value, u64 map_flags,
1270                                           bool onallcpus)
1271 {
1272         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1273         struct htab_elem *l_new = NULL, *l_old;
1274         struct hlist_nulls_head *head;
1275         unsigned long flags;
1276         struct bucket *b;
1277         u32 key_size, hash;
1278         int ret;
1279
1280         if (unlikely(map_flags > BPF_EXIST))
1281                 /* unknown flags */
1282                 return -EINVAL;
1283
1284         WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
1285                      !rcu_read_lock_bh_held());
1286
1287         key_size = map->key_size;
1288
1289         hash = htab_map_hash(key, key_size, htab->hashrnd);
1290
1291         b = __select_bucket(htab, hash);
1292         head = &b->head;
1293
1294         ret = htab_lock_bucket(htab, b, hash, &flags);
1295         if (ret)
1296                 return ret;
1297
1298         l_old = lookup_elem_raw(head, hash, key, key_size);
1299
1300         ret = check_flags(htab, l_old, map_flags);
1301         if (ret)
1302                 goto err;
1303
1304         if (l_old) {
1305                 /* per-cpu hash map can update value in-place */
1306                 pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
1307                                 value, onallcpus);
1308         } else {
1309                 l_new = alloc_htab_elem(htab, key, value, key_size,
1310                                         hash, true, onallcpus, NULL);
1311                 if (IS_ERR(l_new)) {
1312                         ret = PTR_ERR(l_new);
1313                         goto err;
1314                 }
1315                 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1316         }
1317         ret = 0;
1318 err:
1319         htab_unlock_bucket(htab, b, hash, flags);
1320         return ret;
1321 }
1322
1323 static long __htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
1324                                               void *value, u64 map_flags,
1325                                               bool onallcpus)
1326 {
1327         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1328         struct htab_elem *l_new = NULL, *l_old;
1329         struct hlist_nulls_head *head;
1330         unsigned long flags;
1331         struct bucket *b;
1332         u32 key_size, hash;
1333         int ret;
1334
1335         if (unlikely(map_flags > BPF_EXIST))
1336                 /* unknown flags */
1337                 return -EINVAL;
1338
1339         WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
1340                      !rcu_read_lock_bh_held());
1341
1342         key_size = map->key_size;
1343
1344         hash = htab_map_hash(key, key_size, htab->hashrnd);
1345
1346         b = __select_bucket(htab, hash);
1347         head = &b->head;
1348
1349         /* For LRU, we need to alloc before taking bucket's
1350          * spinlock because LRU's elem alloc may need
1351          * to remove older elem from htab and this removal
1352          * operation will need a bucket lock.
1353          */
1354         if (map_flags != BPF_EXIST) {
1355                 l_new = prealloc_lru_pop(htab, key, hash);
1356                 if (!l_new)
1357                         return -ENOMEM;
1358         }
1359
1360         ret = htab_lock_bucket(htab, b, hash, &flags);
1361         if (ret)
1362                 goto err_lock_bucket;
1363
1364         l_old = lookup_elem_raw(head, hash, key, key_size);
1365
1366         ret = check_flags(htab, l_old, map_flags);
1367         if (ret)
1368                 goto err;
1369
1370         if (l_old) {
1371                 bpf_lru_node_set_ref(&l_old->lru_node);
1372
1373                 /* per-cpu hash map can update value in-place */
1374                 pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
1375                                 value, onallcpus);
1376         } else {
1377                 pcpu_init_value(htab, htab_elem_get_ptr(l_new, key_size),
1378                                 value, onallcpus);
1379                 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1380                 l_new = NULL;
1381         }
1382         ret = 0;
1383 err:
1384         htab_unlock_bucket(htab, b, hash, flags);
1385 err_lock_bucket:
1386         if (l_new) {
1387                 bpf_map_dec_elem_count(&htab->map);
1388                 bpf_lru_push_free(&htab->lru, &l_new->lru_node);
1389         }
1390         return ret;
1391 }
1392
1393 static long htab_percpu_map_update_elem(struct bpf_map *map, void *key,
1394                                         void *value, u64 map_flags)
1395 {
1396         return __htab_percpu_map_update_elem(map, key, value, map_flags, false);
1397 }
1398
1399 static long htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
1400                                             void *value, u64 map_flags)
1401 {
1402         return __htab_lru_percpu_map_update_elem(map, key, value, map_flags,
1403                                                  false);
1404 }
1405
1406 /* Called from syscall or from eBPF program */
1407 static long htab_map_delete_elem(struct bpf_map *map, void *key)
1408 {
1409         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1410         struct hlist_nulls_head *head;
1411         struct bucket *b;
1412         struct htab_elem *l;
1413         unsigned long flags;
1414         u32 hash, key_size;
1415         int ret;
1416
1417         WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
1418                      !rcu_read_lock_bh_held());
1419
1420         key_size = map->key_size;
1421
1422         hash = htab_map_hash(key, key_size, htab->hashrnd);
1423         b = __select_bucket(htab, hash);
1424         head = &b->head;
1425
1426         ret = htab_lock_bucket(htab, b, hash, &flags);
1427         if (ret)
1428                 return ret;
1429
1430         l = lookup_elem_raw(head, hash, key, key_size);
1431
1432         if (l) {
1433                 hlist_nulls_del_rcu(&l->hash_node);
1434                 free_htab_elem(htab, l);
1435         } else {
1436                 ret = -ENOENT;
1437         }
1438
1439         htab_unlock_bucket(htab, b, hash, flags);
1440         return ret;
1441 }
1442
1443 static long htab_lru_map_delete_elem(struct bpf_map *map, void *key)
1444 {
1445         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1446         struct hlist_nulls_head *head;
1447         struct bucket *b;
1448         struct htab_elem *l;
1449         unsigned long flags;
1450         u32 hash, key_size;
1451         int ret;
1452
1453         WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
1454                      !rcu_read_lock_bh_held());
1455
1456         key_size = map->key_size;
1457
1458         hash = htab_map_hash(key, key_size, htab->hashrnd);
1459         b = __select_bucket(htab, hash);
1460         head = &b->head;
1461
1462         ret = htab_lock_bucket(htab, b, hash, &flags);
1463         if (ret)
1464                 return ret;
1465
1466         l = lookup_elem_raw(head, hash, key, key_size);
1467
1468         if (l)
1469                 hlist_nulls_del_rcu(&l->hash_node);
1470         else
1471                 ret = -ENOENT;
1472
1473         htab_unlock_bucket(htab, b, hash, flags);
1474         if (l)
1475                 htab_lru_push_free(htab, l);
1476         return ret;
1477 }
1478
1479 static void delete_all_elements(struct bpf_htab *htab)
1480 {
1481         int i;
1482
1483         /* It's called from a worker thread, so disable migration here,
1484          * since bpf_mem_cache_free() relies on that.
1485          */
1486         migrate_disable();
1487         for (i = 0; i < htab->n_buckets; i++) {
1488                 struct hlist_nulls_head *head = select_bucket(htab, i);
1489                 struct hlist_nulls_node *n;
1490                 struct htab_elem *l;
1491
1492                 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1493                         hlist_nulls_del_rcu(&l->hash_node);
1494                         htab_elem_free(htab, l);
1495                 }
1496                 cond_resched();
1497         }
1498         migrate_enable();
1499 }
1500
1501 static void htab_free_malloced_timers_and_wq(struct bpf_htab *htab)
1502 {
1503         int i;
1504
1505         rcu_read_lock();
1506         for (i = 0; i < htab->n_buckets; i++) {
1507                 struct hlist_nulls_head *head = select_bucket(htab, i);
1508                 struct hlist_nulls_node *n;
1509                 struct htab_elem *l;
1510
1511                 hlist_nulls_for_each_entry(l, n, head, hash_node) {
1512                         /* We only free timer on uref dropping to zero */
1513                         if (btf_record_has_field(htab->map.record, BPF_TIMER))
1514                                 bpf_obj_free_timer(htab->map.record,
1515                                                    l->key + round_up(htab->map.key_size, 8));
1516                         if (btf_record_has_field(htab->map.record, BPF_WORKQUEUE))
1517                                 bpf_obj_free_workqueue(htab->map.record,
1518                                                        l->key + round_up(htab->map.key_size, 8));
1519                 }
1520                 cond_resched_rcu();
1521         }
1522         rcu_read_unlock();
1523 }
1524
1525 static void htab_map_free_timers_and_wq(struct bpf_map *map)
1526 {
1527         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1528
1529         /* We only free timer and workqueue on uref dropping to zero */
1530         if (btf_record_has_field(htab->map.record, BPF_TIMER | BPF_WORKQUEUE)) {
1531                 if (!htab_is_prealloc(htab))
1532                         htab_free_malloced_timers_and_wq(htab);
1533                 else
1534                         htab_free_prealloced_timers_and_wq(htab);
1535         }
1536 }
1537
1538 /* Called when map->refcnt goes to zero, either from workqueue or from syscall */
1539 static void htab_map_free(struct bpf_map *map)
1540 {
1541         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1542         int i;
1543
1544         /* bpf_free_used_maps() or close(map_fd) will trigger this map_free callback.
1545          * bpf_free_used_maps() is called after bpf prog is no longer executing.
1546          * There is no need to synchronize_rcu() here to protect map elements.
1547          */
1548
1549         /* htab no longer uses call_rcu() directly. bpf_mem_alloc does it
1550          * underneath and is responsible for waiting for callbacks to finish
1551          * during bpf_mem_alloc_destroy().
1552          */
1553         if (!htab_is_prealloc(htab)) {
1554                 delete_all_elements(htab);
1555         } else {
1556                 htab_free_prealloced_fields(htab);
1557                 prealloc_destroy(htab);
1558         }
1559
1560         bpf_map_free_elem_count(map);
1561         free_percpu(htab->extra_elems);
1562         bpf_map_area_free(htab->buckets);
1563         bpf_mem_alloc_destroy(&htab->pcpu_ma);
1564         bpf_mem_alloc_destroy(&htab->ma);
1565         if (htab->use_percpu_counter)
1566                 percpu_counter_destroy(&htab->pcount);
1567         for (i = 0; i < HASHTAB_MAP_LOCK_COUNT; i++)
1568                 free_percpu(htab->map_locked[i]);
1569         lockdep_unregister_key(&htab->lockdep_key);
1570         bpf_map_area_free(htab);
1571 }
1572
1573 static void htab_map_seq_show_elem(struct bpf_map *map, void *key,
1574                                    struct seq_file *m)
1575 {
1576         void *value;
1577
1578         rcu_read_lock();
1579
1580         value = htab_map_lookup_elem(map, key);
1581         if (!value) {
1582                 rcu_read_unlock();
1583                 return;
1584         }
1585
1586         btf_type_seq_show(map->btf, map->btf_key_type_id, key, m);
1587         seq_puts(m, ": ");
1588         btf_type_seq_show(map->btf, map->btf_value_type_id, value, m);
1589         seq_puts(m, "\n");
1590
1591         rcu_read_unlock();
1592 }
1593
1594 static int __htab_map_lookup_and_delete_elem(struct bpf_map *map, void *key,
1595                                              void *value, bool is_lru_map,
1596                                              bool is_percpu, u64 flags)
1597 {
1598         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1599         struct hlist_nulls_head *head;
1600         unsigned long bflags;
1601         struct htab_elem *l;
1602         u32 hash, key_size;
1603         struct bucket *b;
1604         int ret;
1605
1606         key_size = map->key_size;
1607
1608         hash = htab_map_hash(key, key_size, htab->hashrnd);
1609         b = __select_bucket(htab, hash);
1610         head = &b->head;
1611
1612         ret = htab_lock_bucket(htab, b, hash, &bflags);
1613         if (ret)
1614                 return ret;
1615
1616         l = lookup_elem_raw(head, hash, key, key_size);
1617         if (!l) {
1618                 ret = -ENOENT;
1619         } else {
1620                 if (is_percpu) {
1621                         u32 roundup_value_size = round_up(map->value_size, 8);
1622                         void __percpu *pptr;
1623                         int off = 0, cpu;
1624
1625                         pptr = htab_elem_get_ptr(l, key_size);
1626                         for_each_possible_cpu(cpu) {
1627                                 copy_map_value_long(&htab->map, value + off, per_cpu_ptr(pptr, cpu));
1628                                 check_and_init_map_value(&htab->map, value + off);
1629                                 off += roundup_value_size;
1630                         }
1631                 } else {
1632                         u32 roundup_key_size = round_up(map->key_size, 8);
1633
1634                         if (flags & BPF_F_LOCK)
1635                                 copy_map_value_locked(map, value, l->key +
1636                                                       roundup_key_size,
1637                                                       true);
1638                         else
1639                                 copy_map_value(map, value, l->key +
1640                                                roundup_key_size);
1641                         /* Zeroing special fields in the temp buffer */
1642                         check_and_init_map_value(map, value);
1643                 }
1644
1645                 hlist_nulls_del_rcu(&l->hash_node);
1646                 if (!is_lru_map)
1647                         free_htab_elem(htab, l);
1648         }
1649
1650         htab_unlock_bucket(htab, b, hash, bflags);
1651
1652         if (is_lru_map && l)
1653                 htab_lru_push_free(htab, l);
1654
1655         return ret;
1656 }
1657
1658 static int htab_map_lookup_and_delete_elem(struct bpf_map *map, void *key,
1659                                            void *value, u64 flags)
1660 {
1661         return __htab_map_lookup_and_delete_elem(map, key, value, false, false,
1662                                                  flags);
1663 }
1664
1665 static int htab_percpu_map_lookup_and_delete_elem(struct bpf_map *map,
1666                                                   void *key, void *value,
1667                                                   u64 flags)
1668 {
1669         return __htab_map_lookup_and_delete_elem(map, key, value, false, true,
1670                                                  flags);
1671 }
1672
1673 static int htab_lru_map_lookup_and_delete_elem(struct bpf_map *map, void *key,
1674                                                void *value, u64 flags)
1675 {
1676         return __htab_map_lookup_and_delete_elem(map, key, value, true, false,
1677                                                  flags);
1678 }
1679
1680 static int htab_lru_percpu_map_lookup_and_delete_elem(struct bpf_map *map,
1681                                                       void *key, void *value,
1682                                                       u64 flags)
1683 {
1684         return __htab_map_lookup_and_delete_elem(map, key, value, true, true,
1685                                                  flags);
1686 }
1687
1688 static int
1689 __htab_map_lookup_and_delete_batch(struct bpf_map *map,
1690                                    const union bpf_attr *attr,
1691                                    union bpf_attr __user *uattr,
1692                                    bool do_delete, bool is_lru_map,
1693                                    bool is_percpu)
1694 {
1695         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1696         u32 bucket_cnt, total, key_size, value_size, roundup_key_size;
1697         void *keys = NULL, *values = NULL, *value, *dst_key, *dst_val;
1698         void __user *uvalues = u64_to_user_ptr(attr->batch.values);
1699         void __user *ukeys = u64_to_user_ptr(attr->batch.keys);
1700         void __user *ubatch = u64_to_user_ptr(attr->batch.in_batch);
1701         u32 batch, max_count, size, bucket_size, map_id;
1702         struct htab_elem *node_to_free = NULL;
1703         u64 elem_map_flags, map_flags;
1704         struct hlist_nulls_head *head;
1705         struct hlist_nulls_node *n;
1706         unsigned long flags = 0;
1707         bool locked = false;
1708         struct htab_elem *l;
1709         struct bucket *b;
1710         int ret = 0;
1711
1712         elem_map_flags = attr->batch.elem_flags;
1713         if ((elem_map_flags & ~BPF_F_LOCK) ||
1714             ((elem_map_flags & BPF_F_LOCK) && !btf_record_has_field(map->record, BPF_SPIN_LOCK)))
1715                 return -EINVAL;
1716
1717         map_flags = attr->batch.flags;
1718         if (map_flags)
1719                 return -EINVAL;
1720
1721         max_count = attr->batch.count;
1722         if (!max_count)
1723                 return 0;
1724
1725         if (put_user(0, &uattr->batch.count))
1726                 return -EFAULT;
1727
1728         batch = 0;
1729         if (ubatch && copy_from_user(&batch, ubatch, sizeof(batch)))
1730                 return -EFAULT;
1731
1732         if (batch >= htab->n_buckets)
1733                 return -ENOENT;
1734
1735         key_size = htab->map.key_size;
1736         roundup_key_size = round_up(htab->map.key_size, 8);
1737         value_size = htab->map.value_size;
1738         size = round_up(value_size, 8);
1739         if (is_percpu)
1740                 value_size = size * num_possible_cpus();
1741         total = 0;
1742         /* while experimenting with hash tables with sizes ranging from 10 to
1743          * 1000, it was observed that a bucket can have up to 5 entries.
1744          */
1745         bucket_size = 5;
1746
1747 alloc:
1748         /* We cannot do copy_from_user or copy_to_user inside
1749          * the rcu_read_lock. Allocate enough space here.
1750          */
1751         keys = kvmalloc_array(key_size, bucket_size, GFP_USER | __GFP_NOWARN);
1752         values = kvmalloc_array(value_size, bucket_size, GFP_USER | __GFP_NOWARN);
1753         if (!keys || !values) {
1754                 ret = -ENOMEM;
1755                 goto after_loop;
1756         }
1757
1758 again:
1759         bpf_disable_instrumentation();
1760         rcu_read_lock();
1761 again_nocopy:
1762         dst_key = keys;
1763         dst_val = values;
1764         b = &htab->buckets[batch];
1765         head = &b->head;
1766         /* do not grab the lock unless need it (bucket_cnt > 0). */
1767         if (locked) {
1768                 ret = htab_lock_bucket(htab, b, batch, &flags);
1769                 if (ret) {
1770                         rcu_read_unlock();
1771                         bpf_enable_instrumentation();
1772                         goto after_loop;
1773                 }
1774         }
1775
1776         bucket_cnt = 0;
1777         hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
1778                 bucket_cnt++;
1779
1780         if (bucket_cnt && !locked) {
1781                 locked = true;
1782                 goto again_nocopy;
1783         }
1784
1785         if (bucket_cnt > (max_count - total)) {
1786                 if (total == 0)
1787                         ret = -ENOSPC;
1788                 /* Note that since bucket_cnt > 0 here, it is implicit
1789                  * that the locked was grabbed, so release it.
1790                  */
1791                 htab_unlock_bucket(htab, b, batch, flags);
1792                 rcu_read_unlock();
1793                 bpf_enable_instrumentation();
1794                 goto after_loop;
1795         }
1796
1797         if (bucket_cnt > bucket_size) {
1798                 bucket_size = bucket_cnt;
1799                 /* Note that since bucket_cnt > 0 here, it is implicit
1800                  * that the locked was grabbed, so release it.
1801                  */
1802                 htab_unlock_bucket(htab, b, batch, flags);
1803                 rcu_read_unlock();
1804                 bpf_enable_instrumentation();
1805                 kvfree(keys);
1806                 kvfree(values);
1807                 goto alloc;
1808         }
1809
1810         /* Next block is only safe to run if you have grabbed the lock */
1811         if (!locked)
1812                 goto next_batch;
1813
1814         hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1815                 memcpy(dst_key, l->key, key_size);
1816
1817                 if (is_percpu) {
1818                         int off = 0, cpu;
1819                         void __percpu *pptr;
1820
1821                         pptr = htab_elem_get_ptr(l, map->key_size);
1822                         for_each_possible_cpu(cpu) {
1823                                 copy_map_value_long(&htab->map, dst_val + off, per_cpu_ptr(pptr, cpu));
1824                                 check_and_init_map_value(&htab->map, dst_val + off);
1825                                 off += size;
1826                         }
1827                 } else {
1828                         value = l->key + roundup_key_size;
1829                         if (map->map_type == BPF_MAP_TYPE_HASH_OF_MAPS) {
1830                                 struct bpf_map **inner_map = value;
1831
1832                                  /* Actual value is the id of the inner map */
1833                                 map_id = map->ops->map_fd_sys_lookup_elem(*inner_map);
1834                                 value = &map_id;
1835                         }
1836
1837                         if (elem_map_flags & BPF_F_LOCK)
1838                                 copy_map_value_locked(map, dst_val, value,
1839                                                       true);
1840                         else
1841                                 copy_map_value(map, dst_val, value);
1842                         /* Zeroing special fields in the temp buffer */
1843                         check_and_init_map_value(map, dst_val);
1844                 }
1845                 if (do_delete) {
1846                         hlist_nulls_del_rcu(&l->hash_node);
1847
1848                         /* bpf_lru_push_free() will acquire lru_lock, which
1849                          * may cause deadlock. See comments in function
1850                          * prealloc_lru_pop(). Let us do bpf_lru_push_free()
1851                          * after releasing the bucket lock.
1852                          */
1853                         if (is_lru_map) {
1854                                 l->batch_flink = node_to_free;
1855                                 node_to_free = l;
1856                         } else {
1857                                 free_htab_elem(htab, l);
1858                         }
1859                 }
1860                 dst_key += key_size;
1861                 dst_val += value_size;
1862         }
1863
1864         htab_unlock_bucket(htab, b, batch, flags);
1865         locked = false;
1866
1867         while (node_to_free) {
1868                 l = node_to_free;
1869                 node_to_free = node_to_free->batch_flink;
1870                 htab_lru_push_free(htab, l);
1871         }
1872
1873 next_batch:
1874         /* If we are not copying data, we can go to next bucket and avoid
1875          * unlocking the rcu.
1876          */
1877         if (!bucket_cnt && (batch + 1 < htab->n_buckets)) {
1878                 batch++;
1879                 goto again_nocopy;
1880         }
1881
1882         rcu_read_unlock();
1883         bpf_enable_instrumentation();
1884         if (bucket_cnt && (copy_to_user(ukeys + total * key_size, keys,
1885             key_size * bucket_cnt) ||
1886             copy_to_user(uvalues + total * value_size, values,
1887             value_size * bucket_cnt))) {
1888                 ret = -EFAULT;
1889                 goto after_loop;
1890         }
1891
1892         total += bucket_cnt;
1893         batch++;
1894         if (batch >= htab->n_buckets) {
1895                 ret = -ENOENT;
1896                 goto after_loop;
1897         }
1898         goto again;
1899
1900 after_loop:
1901         if (ret == -EFAULT)
1902                 goto out;
1903
1904         /* copy # of entries and next batch */
1905         ubatch = u64_to_user_ptr(attr->batch.out_batch);
1906         if (copy_to_user(ubatch, &batch, sizeof(batch)) ||
1907             put_user(total, &uattr->batch.count))
1908                 ret = -EFAULT;
1909
1910 out:
1911         kvfree(keys);
1912         kvfree(values);
1913         return ret;
1914 }
1915
1916 static int
1917 htab_percpu_map_lookup_batch(struct bpf_map *map, const union bpf_attr *attr,
1918                              union bpf_attr __user *uattr)
1919 {
1920         return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1921                                                   false, true);
1922 }
1923
1924 static int
1925 htab_percpu_map_lookup_and_delete_batch(struct bpf_map *map,
1926                                         const union bpf_attr *attr,
1927                                         union bpf_attr __user *uattr)
1928 {
1929         return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1930                                                   false, true);
1931 }
1932
1933 static int
1934 htab_map_lookup_batch(struct bpf_map *map, const union bpf_attr *attr,
1935                       union bpf_attr __user *uattr)
1936 {
1937         return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1938                                                   false, false);
1939 }
1940
1941 static int
1942 htab_map_lookup_and_delete_batch(struct bpf_map *map,
1943                                  const union bpf_attr *attr,
1944                                  union bpf_attr __user *uattr)
1945 {
1946         return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1947                                                   false, false);
1948 }
1949
1950 static int
1951 htab_lru_percpu_map_lookup_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, false,
1956                                                   true, true);
1957 }
1958
1959 static int
1960 htab_lru_percpu_map_lookup_and_delete_batch(struct bpf_map *map,
1961                                             const union bpf_attr *attr,
1962                                             union bpf_attr __user *uattr)
1963 {
1964         return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1965                                                   true, true);
1966 }
1967
1968 static int
1969 htab_lru_map_lookup_batch(struct bpf_map *map, const union bpf_attr *attr,
1970                           union bpf_attr __user *uattr)
1971 {
1972         return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1973                                                   true, false);
1974 }
1975
1976 static int
1977 htab_lru_map_lookup_and_delete_batch(struct bpf_map *map,
1978                                      const union bpf_attr *attr,
1979                                      union bpf_attr __user *uattr)
1980 {
1981         return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1982                                                   true, false);
1983 }
1984
1985 struct bpf_iter_seq_hash_map_info {
1986         struct bpf_map *map;
1987         struct bpf_htab *htab;
1988         void *percpu_value_buf; // non-zero means percpu hash
1989         u32 bucket_id;
1990         u32 skip_elems;
1991 };
1992
1993 static struct htab_elem *
1994 bpf_hash_map_seq_find_next(struct bpf_iter_seq_hash_map_info *info,
1995                            struct htab_elem *prev_elem)
1996 {
1997         const struct bpf_htab *htab = info->htab;
1998         u32 skip_elems = info->skip_elems;
1999         u32 bucket_id = info->bucket_id;
2000         struct hlist_nulls_head *head;
2001         struct hlist_nulls_node *n;
2002         struct htab_elem *elem;
2003         struct bucket *b;
2004         u32 i, count;
2005
2006         if (bucket_id >= htab->n_buckets)
2007                 return NULL;
2008
2009         /* try to find next elem in the same bucket */
2010         if (prev_elem) {
2011                 /* no update/deletion on this bucket, prev_elem should be still valid
2012                  * and we won't skip elements.
2013                  */
2014                 n = rcu_dereference_raw(hlist_nulls_next_rcu(&prev_elem->hash_node));
2015                 elem = hlist_nulls_entry_safe(n, struct htab_elem, hash_node);
2016                 if (elem)
2017                         return elem;
2018
2019                 /* not found, unlock and go to the next bucket */
2020                 b = &htab->buckets[bucket_id++];
2021                 rcu_read_unlock();
2022                 skip_elems = 0;
2023         }
2024
2025         for (i = bucket_id; i < htab->n_buckets; i++) {
2026                 b = &htab->buckets[i];
2027                 rcu_read_lock();
2028
2029                 count = 0;
2030                 head = &b->head;
2031                 hlist_nulls_for_each_entry_rcu(elem, n, head, hash_node) {
2032                         if (count >= skip_elems) {
2033                                 info->bucket_id = i;
2034                                 info->skip_elems = count;
2035                                 return elem;
2036                         }
2037                         count++;
2038                 }
2039
2040                 rcu_read_unlock();
2041                 skip_elems = 0;
2042         }
2043
2044         info->bucket_id = i;
2045         info->skip_elems = 0;
2046         return NULL;
2047 }
2048
2049 static void *bpf_hash_map_seq_start(struct seq_file *seq, loff_t *pos)
2050 {
2051         struct bpf_iter_seq_hash_map_info *info = seq->private;
2052         struct htab_elem *elem;
2053
2054         elem = bpf_hash_map_seq_find_next(info, NULL);
2055         if (!elem)
2056                 return NULL;
2057
2058         if (*pos == 0)
2059                 ++*pos;
2060         return elem;
2061 }
2062
2063 static void *bpf_hash_map_seq_next(struct seq_file *seq, void *v, loff_t *pos)
2064 {
2065         struct bpf_iter_seq_hash_map_info *info = seq->private;
2066
2067         ++*pos;
2068         ++info->skip_elems;
2069         return bpf_hash_map_seq_find_next(info, v);
2070 }
2071
2072 static int __bpf_hash_map_seq_show(struct seq_file *seq, struct htab_elem *elem)
2073 {
2074         struct bpf_iter_seq_hash_map_info *info = seq->private;
2075         u32 roundup_key_size, roundup_value_size;
2076         struct bpf_iter__bpf_map_elem ctx = {};
2077         struct bpf_map *map = info->map;
2078         struct bpf_iter_meta meta;
2079         int ret = 0, off = 0, cpu;
2080         struct bpf_prog *prog;
2081         void __percpu *pptr;
2082
2083         meta.seq = seq;
2084         prog = bpf_iter_get_info(&meta, elem == NULL);
2085         if (prog) {
2086                 ctx.meta = &meta;
2087                 ctx.map = info->map;
2088                 if (elem) {
2089                         roundup_key_size = round_up(map->key_size, 8);
2090                         ctx.key = elem->key;
2091                         if (!info->percpu_value_buf) {
2092                                 ctx.value = elem->key + roundup_key_size;
2093                         } else {
2094                                 roundup_value_size = round_up(map->value_size, 8);
2095                                 pptr = htab_elem_get_ptr(elem, map->key_size);
2096                                 for_each_possible_cpu(cpu) {
2097                                         copy_map_value_long(map, info->percpu_value_buf + off,
2098                                                             per_cpu_ptr(pptr, cpu));
2099                                         check_and_init_map_value(map, info->percpu_value_buf + off);
2100                                         off += roundup_value_size;
2101                                 }
2102                                 ctx.value = info->percpu_value_buf;
2103                         }
2104                 }
2105                 ret = bpf_iter_run_prog(prog, &ctx);
2106         }
2107
2108         return ret;
2109 }
2110
2111 static int bpf_hash_map_seq_show(struct seq_file *seq, void *v)
2112 {
2113         return __bpf_hash_map_seq_show(seq, v);
2114 }
2115
2116 static void bpf_hash_map_seq_stop(struct seq_file *seq, void *v)
2117 {
2118         if (!v)
2119                 (void)__bpf_hash_map_seq_show(seq, NULL);
2120         else
2121                 rcu_read_unlock();
2122 }
2123
2124 static int bpf_iter_init_hash_map(void *priv_data,
2125                                   struct bpf_iter_aux_info *aux)
2126 {
2127         struct bpf_iter_seq_hash_map_info *seq_info = priv_data;
2128         struct bpf_map *map = aux->map;
2129         void *value_buf;
2130         u32 buf_size;
2131
2132         if (map->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
2133             map->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH) {
2134                 buf_size = round_up(map->value_size, 8) * num_possible_cpus();
2135                 value_buf = kmalloc(buf_size, GFP_USER | __GFP_NOWARN);
2136                 if (!value_buf)
2137                         return -ENOMEM;
2138
2139                 seq_info->percpu_value_buf = value_buf;
2140         }
2141
2142         bpf_map_inc_with_uref(map);
2143         seq_info->map = map;
2144         seq_info->htab = container_of(map, struct bpf_htab, map);
2145         return 0;
2146 }
2147
2148 static void bpf_iter_fini_hash_map(void *priv_data)
2149 {
2150         struct bpf_iter_seq_hash_map_info *seq_info = priv_data;
2151
2152         bpf_map_put_with_uref(seq_info->map);
2153         kfree(seq_info->percpu_value_buf);
2154 }
2155
2156 static const struct seq_operations bpf_hash_map_seq_ops = {
2157         .start  = bpf_hash_map_seq_start,
2158         .next   = bpf_hash_map_seq_next,
2159         .stop   = bpf_hash_map_seq_stop,
2160         .show   = bpf_hash_map_seq_show,
2161 };
2162
2163 static const struct bpf_iter_seq_info iter_seq_info = {
2164         .seq_ops                = &bpf_hash_map_seq_ops,
2165         .init_seq_private       = bpf_iter_init_hash_map,
2166         .fini_seq_private       = bpf_iter_fini_hash_map,
2167         .seq_priv_size          = sizeof(struct bpf_iter_seq_hash_map_info),
2168 };
2169
2170 static long bpf_for_each_hash_elem(struct bpf_map *map, bpf_callback_t callback_fn,
2171                                    void *callback_ctx, u64 flags)
2172 {
2173         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
2174         struct hlist_nulls_head *head;
2175         struct hlist_nulls_node *n;
2176         struct htab_elem *elem;
2177         u32 roundup_key_size;
2178         int i, num_elems = 0;
2179         void __percpu *pptr;
2180         struct bucket *b;
2181         void *key, *val;
2182         bool is_percpu;
2183         u64 ret = 0;
2184
2185         if (flags != 0)
2186                 return -EINVAL;
2187
2188         is_percpu = htab_is_percpu(htab);
2189
2190         roundup_key_size = round_up(map->key_size, 8);
2191         /* disable migration so percpu value prepared here will be the
2192          * same as the one seen by the bpf program with bpf_map_lookup_elem().
2193          */
2194         if (is_percpu)
2195                 migrate_disable();
2196         for (i = 0; i < htab->n_buckets; i++) {
2197                 b = &htab->buckets[i];
2198                 rcu_read_lock();
2199                 head = &b->head;
2200                 hlist_nulls_for_each_entry_rcu(elem, n, head, hash_node) {
2201                         key = elem->key;
2202                         if (is_percpu) {
2203                                 /* current cpu value for percpu map */
2204                                 pptr = htab_elem_get_ptr(elem, map->key_size);
2205                                 val = this_cpu_ptr(pptr);
2206                         } else {
2207                                 val = elem->key + roundup_key_size;
2208                         }
2209                         num_elems++;
2210                         ret = callback_fn((u64)(long)map, (u64)(long)key,
2211                                           (u64)(long)val, (u64)(long)callback_ctx, 0);
2212                         /* return value: 0 - continue, 1 - stop and return */
2213                         if (ret) {
2214                                 rcu_read_unlock();
2215                                 goto out;
2216                         }
2217                 }
2218                 rcu_read_unlock();
2219         }
2220 out:
2221         if (is_percpu)
2222                 migrate_enable();
2223         return num_elems;
2224 }
2225
2226 static u64 htab_map_mem_usage(const struct bpf_map *map)
2227 {
2228         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
2229         u32 value_size = round_up(htab->map.value_size, 8);
2230         bool prealloc = htab_is_prealloc(htab);
2231         bool percpu = htab_is_percpu(htab);
2232         bool lru = htab_is_lru(htab);
2233         u64 num_entries;
2234         u64 usage = sizeof(struct bpf_htab);
2235
2236         usage += sizeof(struct bucket) * htab->n_buckets;
2237         usage += sizeof(int) * num_possible_cpus() * HASHTAB_MAP_LOCK_COUNT;
2238         if (prealloc) {
2239                 num_entries = map->max_entries;
2240                 if (htab_has_extra_elems(htab))
2241                         num_entries += num_possible_cpus();
2242
2243                 usage += htab->elem_size * num_entries;
2244
2245                 if (percpu)
2246                         usage += value_size * num_possible_cpus() * num_entries;
2247                 else if (!lru)
2248                         usage += sizeof(struct htab_elem *) * num_possible_cpus();
2249         } else {
2250 #define LLIST_NODE_SZ sizeof(struct llist_node)
2251
2252                 num_entries = htab->use_percpu_counter ?
2253                                           percpu_counter_sum(&htab->pcount) :
2254                                           atomic_read(&htab->count);
2255                 usage += (htab->elem_size + LLIST_NODE_SZ) * num_entries;
2256                 if (percpu) {
2257                         usage += (LLIST_NODE_SZ + sizeof(void *)) * num_entries;
2258                         usage += value_size * num_possible_cpus() * num_entries;
2259                 }
2260         }
2261         return usage;
2262 }
2263
2264 BTF_ID_LIST_SINGLE(htab_map_btf_ids, struct, bpf_htab)
2265 const struct bpf_map_ops htab_map_ops = {
2266         .map_meta_equal = bpf_map_meta_equal,
2267         .map_alloc_check = htab_map_alloc_check,
2268         .map_alloc = htab_map_alloc,
2269         .map_free = htab_map_free,
2270         .map_get_next_key = htab_map_get_next_key,
2271         .map_release_uref = htab_map_free_timers_and_wq,
2272         .map_lookup_elem = htab_map_lookup_elem,
2273         .map_lookup_and_delete_elem = htab_map_lookup_and_delete_elem,
2274         .map_update_elem = htab_map_update_elem,
2275         .map_delete_elem = htab_map_delete_elem,
2276         .map_gen_lookup = htab_map_gen_lookup,
2277         .map_seq_show_elem = htab_map_seq_show_elem,
2278         .map_set_for_each_callback_args = map_set_for_each_callback_args,
2279         .map_for_each_callback = bpf_for_each_hash_elem,
2280         .map_mem_usage = htab_map_mem_usage,
2281         BATCH_OPS(htab),
2282         .map_btf_id = &htab_map_btf_ids[0],
2283         .iter_seq_info = &iter_seq_info,
2284 };
2285
2286 const struct bpf_map_ops htab_lru_map_ops = {
2287         .map_meta_equal = bpf_map_meta_equal,
2288         .map_alloc_check = htab_map_alloc_check,
2289         .map_alloc = htab_map_alloc,
2290         .map_free = htab_map_free,
2291         .map_get_next_key = htab_map_get_next_key,
2292         .map_release_uref = htab_map_free_timers_and_wq,
2293         .map_lookup_elem = htab_lru_map_lookup_elem,
2294         .map_lookup_and_delete_elem = htab_lru_map_lookup_and_delete_elem,
2295         .map_lookup_elem_sys_only = htab_lru_map_lookup_elem_sys,
2296         .map_update_elem = htab_lru_map_update_elem,
2297         .map_delete_elem = htab_lru_map_delete_elem,
2298         .map_gen_lookup = htab_lru_map_gen_lookup,
2299         .map_seq_show_elem = htab_map_seq_show_elem,
2300         .map_set_for_each_callback_args = map_set_for_each_callback_args,
2301         .map_for_each_callback = bpf_for_each_hash_elem,
2302         .map_mem_usage = htab_map_mem_usage,
2303         BATCH_OPS(htab_lru),
2304         .map_btf_id = &htab_map_btf_ids[0],
2305         .iter_seq_info = &iter_seq_info,
2306 };
2307
2308 /* Called from eBPF program */
2309 static void *htab_percpu_map_lookup_elem(struct bpf_map *map, void *key)
2310 {
2311         struct htab_elem *l = __htab_map_lookup_elem(map, key);
2312
2313         if (l)
2314                 return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
2315         else
2316                 return NULL;
2317 }
2318
2319 /* inline bpf_map_lookup_elem() call for per-CPU hashmap */
2320 static int htab_percpu_map_gen_lookup(struct bpf_map *map, struct bpf_insn *insn_buf)
2321 {
2322         struct bpf_insn *insn = insn_buf;
2323
2324         if (!bpf_jit_supports_percpu_insn())
2325                 return -EOPNOTSUPP;
2326
2327         BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
2328                      (void *(*)(struct bpf_map *map, void *key))NULL));
2329         *insn++ = BPF_EMIT_CALL(__htab_map_lookup_elem);
2330         *insn++ = BPF_JMP_IMM(BPF_JEQ, BPF_REG_0, 0, 3);
2331         *insn++ = BPF_ALU64_IMM(BPF_ADD, BPF_REG_0,
2332                                 offsetof(struct htab_elem, key) + map->key_size);
2333         *insn++ = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_0, 0);
2334         *insn++ = BPF_MOV64_PERCPU_REG(BPF_REG_0, BPF_REG_0);
2335
2336         return insn - insn_buf;
2337 }
2338
2339 static void *htab_percpu_map_lookup_percpu_elem(struct bpf_map *map, void *key, u32 cpu)
2340 {
2341         struct htab_elem *l;
2342
2343         if (cpu >= nr_cpu_ids)
2344                 return NULL;
2345
2346         l = __htab_map_lookup_elem(map, key);
2347         if (l)
2348                 return per_cpu_ptr(htab_elem_get_ptr(l, map->key_size), cpu);
2349         else
2350                 return NULL;
2351 }
2352
2353 static void *htab_lru_percpu_map_lookup_elem(struct bpf_map *map, void *key)
2354 {
2355         struct htab_elem *l = __htab_map_lookup_elem(map, key);
2356
2357         if (l) {
2358                 bpf_lru_node_set_ref(&l->lru_node);
2359                 return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
2360         }
2361
2362         return NULL;
2363 }
2364
2365 static void *htab_lru_percpu_map_lookup_percpu_elem(struct bpf_map *map, void *key, u32 cpu)
2366 {
2367         struct htab_elem *l;
2368
2369         if (cpu >= nr_cpu_ids)
2370                 return NULL;
2371
2372         l = __htab_map_lookup_elem(map, key);
2373         if (l) {
2374                 bpf_lru_node_set_ref(&l->lru_node);
2375                 return per_cpu_ptr(htab_elem_get_ptr(l, map->key_size), cpu);
2376         }
2377
2378         return NULL;
2379 }
2380
2381 int bpf_percpu_hash_copy(struct bpf_map *map, void *key, void *value)
2382 {
2383         struct htab_elem *l;
2384         void __percpu *pptr;
2385         int ret = -ENOENT;
2386         int cpu, off = 0;
2387         u32 size;
2388
2389         /* per_cpu areas are zero-filled and bpf programs can only
2390          * access 'value_size' of them, so copying rounded areas
2391          * will not leak any kernel data
2392          */
2393         size = round_up(map->value_size, 8);
2394         rcu_read_lock();
2395         l = __htab_map_lookup_elem(map, key);
2396         if (!l)
2397                 goto out;
2398         /* We do not mark LRU map element here in order to not mess up
2399          * eviction heuristics when user space does a map walk.
2400          */
2401         pptr = htab_elem_get_ptr(l, map->key_size);
2402         for_each_possible_cpu(cpu) {
2403                 copy_map_value_long(map, value + off, per_cpu_ptr(pptr, cpu));
2404                 check_and_init_map_value(map, value + off);
2405                 off += size;
2406         }
2407         ret = 0;
2408 out:
2409         rcu_read_unlock();
2410         return ret;
2411 }
2412
2413 int bpf_percpu_hash_update(struct bpf_map *map, void *key, void *value,
2414                            u64 map_flags)
2415 {
2416         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
2417         int ret;
2418
2419         rcu_read_lock();
2420         if (htab_is_lru(htab))
2421                 ret = __htab_lru_percpu_map_update_elem(map, key, value,
2422                                                         map_flags, true);
2423         else
2424                 ret = __htab_percpu_map_update_elem(map, key, value, map_flags,
2425                                                     true);
2426         rcu_read_unlock();
2427
2428         return ret;
2429 }
2430
2431 static void htab_percpu_map_seq_show_elem(struct bpf_map *map, void *key,
2432                                           struct seq_file *m)
2433 {
2434         struct htab_elem *l;
2435         void __percpu *pptr;
2436         int cpu;
2437
2438         rcu_read_lock();
2439
2440         l = __htab_map_lookup_elem(map, key);
2441         if (!l) {
2442                 rcu_read_unlock();
2443                 return;
2444         }
2445
2446         btf_type_seq_show(map->btf, map->btf_key_type_id, key, m);
2447         seq_puts(m, ": {\n");
2448         pptr = htab_elem_get_ptr(l, map->key_size);
2449         for_each_possible_cpu(cpu) {
2450                 seq_printf(m, "\tcpu%d: ", cpu);
2451                 btf_type_seq_show(map->btf, map->btf_value_type_id,
2452                                   per_cpu_ptr(pptr, cpu), m);
2453                 seq_puts(m, "\n");
2454         }
2455         seq_puts(m, "}\n");
2456
2457         rcu_read_unlock();
2458 }
2459
2460 const struct bpf_map_ops htab_percpu_map_ops = {
2461         .map_meta_equal = bpf_map_meta_equal,
2462         .map_alloc_check = htab_map_alloc_check,
2463         .map_alloc = htab_map_alloc,
2464         .map_free = htab_map_free,
2465         .map_get_next_key = htab_map_get_next_key,
2466         .map_lookup_elem = htab_percpu_map_lookup_elem,
2467         .map_gen_lookup = htab_percpu_map_gen_lookup,
2468         .map_lookup_and_delete_elem = htab_percpu_map_lookup_and_delete_elem,
2469         .map_update_elem = htab_percpu_map_update_elem,
2470         .map_delete_elem = htab_map_delete_elem,
2471         .map_lookup_percpu_elem = htab_percpu_map_lookup_percpu_elem,
2472         .map_seq_show_elem = htab_percpu_map_seq_show_elem,
2473         .map_set_for_each_callback_args = map_set_for_each_callback_args,
2474         .map_for_each_callback = bpf_for_each_hash_elem,
2475         .map_mem_usage = htab_map_mem_usage,
2476         BATCH_OPS(htab_percpu),
2477         .map_btf_id = &htab_map_btf_ids[0],
2478         .iter_seq_info = &iter_seq_info,
2479 };
2480
2481 const struct bpf_map_ops htab_lru_percpu_map_ops = {
2482         .map_meta_equal = bpf_map_meta_equal,
2483         .map_alloc_check = htab_map_alloc_check,
2484         .map_alloc = htab_map_alloc,
2485         .map_free = htab_map_free,
2486         .map_get_next_key = htab_map_get_next_key,
2487         .map_lookup_elem = htab_lru_percpu_map_lookup_elem,
2488         .map_lookup_and_delete_elem = htab_lru_percpu_map_lookup_and_delete_elem,
2489         .map_update_elem = htab_lru_percpu_map_update_elem,
2490         .map_delete_elem = htab_lru_map_delete_elem,
2491         .map_lookup_percpu_elem = htab_lru_percpu_map_lookup_percpu_elem,
2492         .map_seq_show_elem = htab_percpu_map_seq_show_elem,
2493         .map_set_for_each_callback_args = map_set_for_each_callback_args,
2494         .map_for_each_callback = bpf_for_each_hash_elem,
2495         .map_mem_usage = htab_map_mem_usage,
2496         BATCH_OPS(htab_lru_percpu),
2497         .map_btf_id = &htab_map_btf_ids[0],
2498         .iter_seq_info = &iter_seq_info,
2499 };
2500
2501 static int fd_htab_map_alloc_check(union bpf_attr *attr)
2502 {
2503         if (attr->value_size != sizeof(u32))
2504                 return -EINVAL;
2505         return htab_map_alloc_check(attr);
2506 }
2507
2508 static void fd_htab_map_free(struct bpf_map *map)
2509 {
2510         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
2511         struct hlist_nulls_node *n;
2512         struct hlist_nulls_head *head;
2513         struct htab_elem *l;
2514         int i;
2515
2516         for (i = 0; i < htab->n_buckets; i++) {
2517                 head = select_bucket(htab, i);
2518
2519                 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
2520                         void *ptr = fd_htab_map_get_ptr(map, l);
2521
2522                         map->ops->map_fd_put_ptr(map, ptr, false);
2523                 }
2524         }
2525
2526         htab_map_free(map);
2527 }
2528
2529 /* only called from syscall */
2530 int bpf_fd_htab_map_lookup_elem(struct bpf_map *map, void *key, u32 *value)
2531 {
2532         void **ptr;
2533         int ret = 0;
2534
2535         if (!map->ops->map_fd_sys_lookup_elem)
2536                 return -ENOTSUPP;
2537
2538         rcu_read_lock();
2539         ptr = htab_map_lookup_elem(map, key);
2540         if (ptr)
2541                 *value = map->ops->map_fd_sys_lookup_elem(READ_ONCE(*ptr));
2542         else
2543                 ret = -ENOENT;
2544         rcu_read_unlock();
2545
2546         return ret;
2547 }
2548
2549 /* only called from syscall */
2550 int bpf_fd_htab_map_update_elem(struct bpf_map *map, struct file *map_file,
2551                                 void *key, void *value, u64 map_flags)
2552 {
2553         void *ptr;
2554         int ret;
2555         u32 ufd = *(u32 *)value;
2556
2557         ptr = map->ops->map_fd_get_ptr(map, map_file, ufd);
2558         if (IS_ERR(ptr))
2559                 return PTR_ERR(ptr);
2560
2561         /* The htab bucket lock is always held during update operations in fd
2562          * htab map, and the following rcu_read_lock() is only used to avoid
2563          * the WARN_ON_ONCE in htab_map_update_elem().
2564          */
2565         rcu_read_lock();
2566         ret = htab_map_update_elem(map, key, &ptr, map_flags);
2567         rcu_read_unlock();
2568         if (ret)
2569                 map->ops->map_fd_put_ptr(map, ptr, false);
2570
2571         return ret;
2572 }
2573
2574 static struct bpf_map *htab_of_map_alloc(union bpf_attr *attr)
2575 {
2576         struct bpf_map *map, *inner_map_meta;
2577
2578         inner_map_meta = bpf_map_meta_alloc(attr->inner_map_fd);
2579         if (IS_ERR(inner_map_meta))
2580                 return inner_map_meta;
2581
2582         map = htab_map_alloc(attr);
2583         if (IS_ERR(map)) {
2584                 bpf_map_meta_free(inner_map_meta);
2585                 return map;
2586         }
2587
2588         map->inner_map_meta = inner_map_meta;
2589
2590         return map;
2591 }
2592
2593 static void *htab_of_map_lookup_elem(struct bpf_map *map, void *key)
2594 {
2595         struct bpf_map **inner_map  = htab_map_lookup_elem(map, key);
2596
2597         if (!inner_map)
2598                 return NULL;
2599
2600         return READ_ONCE(*inner_map);
2601 }
2602
2603 static int htab_of_map_gen_lookup(struct bpf_map *map,
2604                                   struct bpf_insn *insn_buf)
2605 {
2606         struct bpf_insn *insn = insn_buf;
2607         const int ret = BPF_REG_0;
2608
2609         BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
2610                      (void *(*)(struct bpf_map *map, void *key))NULL));
2611         *insn++ = BPF_EMIT_CALL(__htab_map_lookup_elem);
2612         *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 2);
2613         *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
2614                                 offsetof(struct htab_elem, key) +
2615                                 round_up(map->key_size, 8));
2616         *insn++ = BPF_LDX_MEM(BPF_DW, ret, ret, 0);
2617
2618         return insn - insn_buf;
2619 }
2620
2621 static void htab_of_map_free(struct bpf_map *map)
2622 {
2623         bpf_map_meta_free(map->inner_map_meta);
2624         fd_htab_map_free(map);
2625 }
2626
2627 const struct bpf_map_ops htab_of_maps_map_ops = {
2628         .map_alloc_check = fd_htab_map_alloc_check,
2629         .map_alloc = htab_of_map_alloc,
2630         .map_free = htab_of_map_free,
2631         .map_get_next_key = htab_map_get_next_key,
2632         .map_lookup_elem = htab_of_map_lookup_elem,
2633         .map_delete_elem = htab_map_delete_elem,
2634         .map_fd_get_ptr = bpf_map_fd_get_ptr,
2635         .map_fd_put_ptr = bpf_map_fd_put_ptr,
2636         .map_fd_sys_lookup_elem = bpf_map_fd_sys_lookup_elem,
2637         .map_gen_lookup = htab_of_map_gen_lookup,
2638         .map_check_btf = map_check_no_btf,
2639         .map_mem_usage = htab_map_mem_usage,
2640         BATCH_OPS(htab),
2641         .map_btf_id = &htab_map_btf_ids[0],
2642 };