kernel/bpf/hashtab.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /* Copyright (c) 2011-2014 PLUMgrid, http://plumgrid.com
   3  * Copyright (c) 2016 Facebook
   4  */
   5 #include <linux/bpf.h>
   6 #include <linux/btf.h>
   7 #include <linux/jhash.h>
   8 #include <linux/filter.h>
   9 #include <linux/rculist_nulls.h>
  10 #include <linux/random.h>
  11 #include <uapi/linux/btf.h>
  12 #include <linux/rcupdate_trace.h>
  13 #include <linux/btf_ids.h>
  14 #include "percpu_freelist.h"
  15 #include "bpf_lru_list.h"
  16 #include "map_in_map.h"
  17 #include <linux/bpf_mem_alloc.h>
  18
  19 #define HTAB_CREATE_FLAG_MASK                                           \
  20         (BPF_F_NO_PREALLOC | BPF_F_NO_COMMON_LRU | BPF_F_NUMA_NODE |    \
  21          BPF_F_ACCESS_MASK | BPF_F_ZERO_SEED)
  22
  23 #define BATCH_OPS(_name)                        \
  24         .map_lookup_batch =                     \
  25         _name##_map_lookup_batch,               \
  26         .map_lookup_and_delete_batch =          \
  27         _name##_map_lookup_and_delete_batch,    \
  28         .map_update_batch =                     \
  29         generic_map_update_batch,               \
  30         .map_delete_batch =                     \
  31         generic_map_delete_batch
  32
  33 /*
  34  * The bucket lock has two protection scopes:
  35  *
  36  * 1) Serializing concurrent operations from BPF programs on different
  37  *    CPUs
  38  *
  39  * 2) Serializing concurrent operations from BPF programs and sys_bpf()
  40  *
  41  * BPF programs can execute in any context including perf, kprobes and
  42  * tracing. As there are almost no limits where perf, kprobes and tracing
  43  * can be invoked from the lock operations need to be protected against
  44  * deadlocks. Deadlocks can be caused by recursion and by an invocation in
  45  * the lock held section when functions which acquire this lock are invoked
  46  * from sys_bpf(). BPF recursion is prevented by incrementing the per CPU
  47  * variable bpf_prog_active, which prevents BPF programs attached to perf
  48  * events, kprobes and tracing to be invoked before the prior invocation
  49  * from one of these contexts completed. sys_bpf() uses the same mechanism
  50  * by pinning the task to the current CPU and incrementing the recursion
  51  * protection across the map operation.
  52  *
  53  * This has subtle implications on PREEMPT_RT. PREEMPT_RT forbids certain
  54  * operations like memory allocations (even with GFP_ATOMIC) from atomic
  55  * contexts. This is required because even with GFP_ATOMIC the memory
  56  * allocator calls into code paths which acquire locks with long held lock
  57  * sections. To ensure the deterministic behaviour these locks are regular
  58  * spinlocks, which are converted to 'sleepable' spinlocks on RT. The only
  59  * true atomic contexts on an RT kernel are the low level hardware
  60  * handling, scheduling, low level interrupt handling, NMIs etc. None of
  61  * these contexts should ever do memory allocations.
  62  *
  63  * As regular device interrupt handlers and soft interrupts are forced into
  64  * thread context, the existing code which does
  65  *   spin_lock*(); alloc(GFP_ATOMIC); spin_unlock*();
  66  * just works.
  67  *
  68  * In theory the BPF locks could be converted to regular spinlocks as well,
  69  * but the bucket locks and percpu_freelist locks can be taken from
  70  * arbitrary contexts (perf, kprobes, tracepoints) which are required to be
  71  * atomic contexts even on RT. Before the introduction of bpf_mem_alloc,
  72  * it is only safe to use raw spinlock for preallocated hash map on a RT kernel,
  73  * because there is no memory allocation within the lock held sections. However
  74  * after hash map was fully converted to use bpf_mem_alloc, there will be
  75  * non-synchronous memory allocation for non-preallocated hash map, so it is
  76  * safe to always use raw spinlock for bucket lock.
  77  */
  78 struct bucket {
  79         struct hlist_nulls_head head;
  80         raw_spinlock_t raw_lock;
  81 };
  82
  83 #define HASHTAB_MAP_LOCK_COUNT 8
  84 #define HASHTAB_MAP_LOCK_MASK (HASHTAB_MAP_LOCK_COUNT - 1)
  85
  86 struct bpf_htab {
  87         struct bpf_map map;
  88         struct bpf_mem_alloc ma;
  89         struct bpf_mem_alloc pcpu_ma;
  90         struct bucket *buckets;
  91         void *elems;
  92         union {
  93                 struct pcpu_freelist freelist;
  94                 struct bpf_lru lru;
  95         };
  96         struct htab_elem *__percpu *extra_elems;
  97         /* number of elements in non-preallocated hashtable are kept
  98          * in either pcount or count
  99          */
 100         struct percpu_counter pcount;
 101         atomic_t count;
 102         bool use_percpu_counter;
 103         u32 n_buckets;  /* number of hash buckets */
 104         u32 elem_size;  /* size of each element in bytes */
 105         u32 hashrnd;
 106         struct lock_class_key lockdep_key;
 107         int __percpu *map_locked[HASHTAB_MAP_LOCK_COUNT];
 108 };
 109
 110 /* each htab element is struct htab_elem + key + value */
 111 struct htab_elem {
 112         union {
 113                 struct hlist_nulls_node hash_node;
 114                 struct {
 115                         void *padding;
 116                         union {
 117                                 struct pcpu_freelist_node fnode;
 118                                 struct htab_elem *batch_flink;
 119                         };
 120                 };
 121         };
 122         union {
 123                 /* pointer to per-cpu pointer */
 124                 void *ptr_to_pptr;
 125                 struct bpf_lru_node lru_node;
 126         };
 127         u32 hash;
 128         char key[] __aligned(8);
 129 };
 130
 131 static inline bool htab_is_prealloc(const struct bpf_htab *htab)
 132 {
 133         return !(htab->map.map_flags & BPF_F_NO_PREALLOC);
 134 }
 135
 136 static void htab_init_buckets(struct bpf_htab *htab)
 137 {
 138         unsigned int i;
 139
 140         for (i = 0; i < htab->n_buckets; i++) {
 141                 INIT_HLIST_NULLS_HEAD(&htab->buckets[i].head, i);
 142                 raw_spin_lock_init(&htab->buckets[i].raw_lock);
 143                 lockdep_set_class(&htab->buckets[i].raw_lock,
 144                                           &htab->lockdep_key);
 145                 cond_resched();
 146         }
 147 }
 148
 149 static inline int htab_lock_bucket(const struct bpf_htab *htab,
 150                                    struct bucket *b, u32 hash,
 151                                    unsigned long *pflags)
 152 {
 153         unsigned long flags;
 154
 155         hash = hash & min_t(u32, HASHTAB_MAP_LOCK_MASK, htab->n_buckets - 1);
 156
 157         preempt_disable();
 158         local_irq_save(flags);
 159         if (unlikely(__this_cpu_inc_return(*(htab->map_locked[hash])) != 1)) {
 160                 __this_cpu_dec(*(htab->map_locked[hash]));
 161                 local_irq_restore(flags);
 162                 preempt_enable();
 163                 return -EBUSY;
 164         }
 165
 166         raw_spin_lock(&b->raw_lock);
 167         *pflags = flags;
 168
 169         return 0;
 170 }
 171
 172 static inline void htab_unlock_bucket(const struct bpf_htab *htab,
 173                                       struct bucket *b, u32 hash,
 174                                       unsigned long flags)
 175 {
 176         hash = hash & min_t(u32, HASHTAB_MAP_LOCK_MASK, htab->n_buckets - 1);
 177         raw_spin_unlock(&b->raw_lock);
 178         __this_cpu_dec(*(htab->map_locked[hash]));
 179         local_irq_restore(flags);
 180         preempt_enable();
 181 }
 182
 183 static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node);
 184
 185 static bool htab_is_lru(const struct bpf_htab *htab)
 186 {
 187         return htab->map.map_type == BPF_MAP_TYPE_LRU_HASH ||
 188                 htab->map.map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH;
 189 }
 190
 191 static bool htab_is_percpu(const struct bpf_htab *htab)
 192 {
 193         return htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH ||
 194                 htab->map.map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH;
 195 }
 196
 197 static inline void htab_elem_set_ptr(struct htab_elem *l, u32 key_size,
 198                                      void __percpu *pptr)
 199 {
 200         *(void __percpu **)(l->key + key_size) = pptr;
 201 }
 202
 203 static inline void __percpu *htab_elem_get_ptr(struct htab_elem *l, u32 key_size)
 204 {
 205         return *(void __percpu **)(l->key + key_size);
 206 }
 207
 208 static void *fd_htab_map_get_ptr(const struct bpf_map *map, struct htab_elem *l)
 209 {
 210         return *(void **)(l->key + roundup(map->key_size, 8));
 211 }
 212
 213 static struct htab_elem *get_htab_elem(struct bpf_htab *htab, int i)
 214 {
 215         return (struct htab_elem *) (htab->elems + i * (u64)htab->elem_size);
 216 }
 217
 218 static bool htab_has_extra_elems(struct bpf_htab *htab)
 219 {
 220         return !htab_is_percpu(htab) && !htab_is_lru(htab);
 221 }
 222
 223 static void htab_free_prealloced_timers(struct bpf_htab *htab)
 224 {
 225         u32 num_entries = htab->map.max_entries;
 226         int i;
 227
 228         if (!btf_record_has_field(htab->map.record, BPF_TIMER))
 229                 return;
 230         if (htab_has_extra_elems(htab))
 231                 num_entries += num_possible_cpus();
 232
 233         for (i = 0; i < num_entries; i++) {
 234                 struct htab_elem *elem;
 235
 236                 elem = get_htab_elem(htab, i);
 237                 bpf_obj_free_timer(htab->map.record, elem->key + round_up(htab->map.key_size, 8));
 238                 cond_resched();
 239         }
 240 }
 241
 242 static void htab_free_prealloced_fields(struct bpf_htab *htab)
 243 {
 244         u32 num_entries = htab->map.max_entries;
 245         int i;
 246
 247         if (IS_ERR_OR_NULL(htab->map.record))
 248                 return;
 249         if (htab_has_extra_elems(htab))
 250                 num_entries += num_possible_cpus();
 251         for (i = 0; i < num_entries; i++) {
 252                 struct htab_elem *elem;
 253
 254                 elem = get_htab_elem(htab, i);
 255                 if (htab_is_percpu(htab)) {
 256                         void __percpu *pptr = htab_elem_get_ptr(elem, htab->map.key_size);
 257                         int cpu;
 258
 259                         for_each_possible_cpu(cpu) {
 260                                 bpf_obj_free_fields(htab->map.record, per_cpu_ptr(pptr, cpu));
 261                                 cond_resched();
 262                         }
 263                 } else {
 264                         bpf_obj_free_fields(htab->map.record, elem->key + round_up(htab->map.key_size, 8));
 265                         cond_resched();
 266                 }
 267                 cond_resched();
 268         }
 269 }
 270
 271 static void htab_free_elems(struct bpf_htab *htab)
 272 {
 273         int i;
 274
 275         if (!htab_is_percpu(htab))
 276                 goto free_elems;
 277
 278         for (i = 0; i < htab->map.max_entries; i++) {
 279                 void __percpu *pptr;
 280
 281                 pptr = htab_elem_get_ptr(get_htab_elem(htab, i),
 282                                          htab->map.key_size);
 283                 free_percpu(pptr);
 284                 cond_resched();
 285         }
 286 free_elems:
 287         bpf_map_area_free(htab->elems);
 288 }
 289
 290 /* The LRU list has a lock (lru_lock). Each htab bucket has a lock
 291  * (bucket_lock). If both locks need to be acquired together, the lock
 292  * order is always lru_lock -> bucket_lock and this only happens in
 293  * bpf_lru_list.c logic. For example, certain code path of
 294  * bpf_lru_pop_free(), which is called by function prealloc_lru_pop(),
 295  * will acquire lru_lock first followed by acquiring bucket_lock.
 296  *
 297  * In hashtab.c, to avoid deadlock, lock acquisition of
 298  * bucket_lock followed by lru_lock is not allowed. In such cases,
 299  * bucket_lock needs to be released first before acquiring lru_lock.
 300  */
 301 static struct htab_elem *prealloc_lru_pop(struct bpf_htab *htab, void *key,
 302                                           u32 hash)
 303 {
 304         struct bpf_lru_node *node = bpf_lru_pop_free(&htab->lru, hash);
 305         struct htab_elem *l;
 306
 307         if (node) {
 308                 bpf_map_inc_elem_count(&htab->map);
 309                 l = container_of(node, struct htab_elem, lru_node);
 310                 memcpy(l->key, key, htab->map.key_size);
 311                 return l;
 312         }
 313
 314         return NULL;
 315 }
 316
 317 static int prealloc_init(struct bpf_htab *htab)
 318 {
 319         u32 num_entries = htab->map.max_entries;
 320         int err = -ENOMEM, i;
 321
 322         if (htab_has_extra_elems(htab))
 323                 num_entries += num_possible_cpus();
 324
 325         htab->elems = bpf_map_area_alloc((u64)htab->elem_size * num_entries,
 326                                          htab->map.numa_node);
 327         if (!htab->elems)
 328                 return -ENOMEM;
 329
 330         if (!htab_is_percpu(htab))
 331                 goto skip_percpu_elems;
 332
 333         for (i = 0; i < num_entries; i++) {
 334                 u32 size = round_up(htab->map.value_size, 8);
 335                 void __percpu *pptr;
 336
 337                 pptr = bpf_map_alloc_percpu(&htab->map, size, 8,
 338                                             GFP_USER | __GFP_NOWARN);
 339                 if (!pptr)
 340                         goto free_elems;
 341                 htab_elem_set_ptr(get_htab_elem(htab, i), htab->map.key_size,
 342                                   pptr);
 343                 cond_resched();
 344         }
 345
 346 skip_percpu_elems:
 347         if (htab_is_lru(htab))
 348                 err = bpf_lru_init(&htab->lru,
 349                                    htab->map.map_flags & BPF_F_NO_COMMON_LRU,
 350                                    offsetof(struct htab_elem, hash) -
 351                                    offsetof(struct htab_elem, lru_node),
 352                                    htab_lru_map_delete_node,
 353                                    htab);
 354         else
 355                 err = pcpu_freelist_init(&htab->freelist);
 356
 357         if (err)
 358                 goto free_elems;
 359
 360         if (htab_is_lru(htab))
 361                 bpf_lru_populate(&htab->lru, htab->elems,
 362                                  offsetof(struct htab_elem, lru_node),
 363                                  htab->elem_size, num_entries);
 364         else
 365                 pcpu_freelist_populate(&htab->freelist,
 366                                        htab->elems + offsetof(struct htab_elem, fnode),
 367                                        htab->elem_size, num_entries);
 368
 369         return 0;
 370
 371 free_elems:
 372         htab_free_elems(htab);
 373         return err;
 374 }
 375
 376 static void prealloc_destroy(struct bpf_htab *htab)
 377 {
 378         htab_free_elems(htab);
 379
 380         if (htab_is_lru(htab))
 381                 bpf_lru_destroy(&htab->lru);
 382         else
 383                 pcpu_freelist_destroy(&htab->freelist);
 384 }
 385
 386 static int alloc_extra_elems(struct bpf_htab *htab)
 387 {
 388         struct htab_elem *__percpu *pptr, *l_new;
 389         struct pcpu_freelist_node *l;
 390         int cpu;
 391
 392         pptr = bpf_map_alloc_percpu(&htab->map, sizeof(struct htab_elem *), 8,
 393                                     GFP_USER | __GFP_NOWARN);
 394         if (!pptr)
 395                 return -ENOMEM;
 396
 397         for_each_possible_cpu(cpu) {
 398                 l = pcpu_freelist_pop(&htab->freelist);
 399                 /* pop will succeed, since prealloc_init()
 400                  * preallocated extra num_possible_cpus elements
 401                  */
 402                 l_new = container_of(l, struct htab_elem, fnode);
 403                 *per_cpu_ptr(pptr, cpu) = l_new;
 404         }
 405         htab->extra_elems = pptr;
 406         return 0;
 407 }
 408
 409 /* Called from syscall */
 410 static int htab_map_alloc_check(union bpf_attr *attr)
 411 {
 412         bool percpu = (attr->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
 413                        attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
 414         bool lru = (attr->map_type == BPF_MAP_TYPE_LRU_HASH ||
 415                     attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
 416         /* percpu_lru means each cpu has its own LRU list.
 417          * it is different from BPF_MAP_TYPE_PERCPU_HASH where
 418          * the map's value itself is percpu.  percpu_lru has
 419          * nothing to do with the map's value.
 420          */
 421         bool percpu_lru = (attr->map_flags & BPF_F_NO_COMMON_LRU);
 422         bool prealloc = !(attr->map_flags & BPF_F_NO_PREALLOC);
 423         bool zero_seed = (attr->map_flags & BPF_F_ZERO_SEED);
 424         int numa_node = bpf_map_attr_numa_node(attr);
 425
 426         BUILD_BUG_ON(offsetof(struct htab_elem, fnode.next) !=
 427                      offsetof(struct htab_elem, hash_node.pprev));
 428
 429         if (zero_seed && !capable(CAP_SYS_ADMIN))
 430                 /* Guard against local DoS, and discourage production use. */
 431                 return -EPERM;
 432
 433         if (attr->map_flags & ~HTAB_CREATE_FLAG_MASK ||
 434             !bpf_map_flags_access_ok(attr->map_flags))
 435                 return -EINVAL;
 436
 437         if (!lru && percpu_lru)
 438                 return -EINVAL;
 439
 440         if (lru && !prealloc)
 441                 return -ENOTSUPP;
 442
 443         if (numa_node != NUMA_NO_NODE && (percpu || percpu_lru))
 444                 return -EINVAL;
 445
 446         /* check sanity of attributes.
 447          * value_size == 0 may be allowed in the future to use map as a set
 448          */
 449         if (attr->max_entries == 0 || attr->key_size == 0 ||
 450             attr->value_size == 0)
 451                 return -EINVAL;
 452
 453         if ((u64)attr->key_size + attr->value_size >= KMALLOC_MAX_SIZE -
 454            sizeof(struct htab_elem))
 455                 /* if key_size + value_size is bigger, the user space won't be
 456                  * able to access the elements via bpf syscall. This check
 457                  * also makes sure that the elem_size doesn't overflow and it's
 458                  * kmalloc-able later in htab_map_update_elem()
 459                  */
 460                 return -E2BIG;
 461
 462         return 0;
 463 }
 464
 465 static struct bpf_map *htab_map_alloc(union bpf_attr *attr)
 466 {
 467         bool percpu = (attr->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
 468                        attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
 469         bool lru = (attr->map_type == BPF_MAP_TYPE_LRU_HASH ||
 470                     attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
 471         /* percpu_lru means each cpu has its own LRU list.
 472          * it is different from BPF_MAP_TYPE_PERCPU_HASH where
 473          * the map's value itself is percpu.  percpu_lru has
 474          * nothing to do with the map's value.
 475          */
 476         bool percpu_lru = (attr->map_flags & BPF_F_NO_COMMON_LRU);
 477         bool prealloc = !(attr->map_flags & BPF_F_NO_PREALLOC);
 478         struct bpf_htab *htab;
 479         int err, i;
 480
 481         htab = bpf_map_area_alloc(sizeof(*htab), NUMA_NO_NODE);
 482         if (!htab)
 483                 return ERR_PTR(-ENOMEM);
 484
 485         lockdep_register_key(&htab->lockdep_key);
 486
 487         bpf_map_init_from_attr(&htab->map, attr);
 488
 489         if (percpu_lru) {
 490                 /* ensure each CPU's lru list has >=1 elements.
 491                  * since we are at it, make each lru list has the same
 492                  * number of elements.
 493                  */
 494                 htab->map.max_entries = roundup(attr->max_entries,
 495                                                 num_possible_cpus());
 496                 if (htab->map.max_entries < attr->max_entries)
 497                         htab->map.max_entries = rounddown(attr->max_entries,
 498                                                           num_possible_cpus());
 499         }
 500
 501         /* hash table size must be power of 2 */
 502         htab->n_buckets = roundup_pow_of_two(htab->map.max_entries);
 503
 504         htab->elem_size = sizeof(struct htab_elem) +
 505                           round_up(htab->map.key_size, 8);
 506         if (percpu)
 507                 htab->elem_size += sizeof(void *);
 508         else
 509                 htab->elem_size += round_up(htab->map.value_size, 8);
 510
 511         err = -E2BIG;
 512         /* prevent zero size kmalloc and check for u32 overflow */
 513         if (htab->n_buckets == 0 ||
 514             htab->n_buckets > U32_MAX / sizeof(struct bucket))
 515                 goto free_htab;
 516
 517         err = bpf_map_init_elem_count(&htab->map);
 518         if (err)
 519                 goto free_htab;
 520
 521         err = -ENOMEM;
 522         htab->buckets = bpf_map_area_alloc(htab->n_buckets *
 523                                            sizeof(struct bucket),
 524                                            htab->map.numa_node);
 525         if (!htab->buckets)
 526                 goto free_elem_count;
 527
 528         for (i = 0; i < HASHTAB_MAP_LOCK_COUNT; i++) {
 529                 htab->map_locked[i] = bpf_map_alloc_percpu(&htab->map,
 530                                                            sizeof(int),
 531                                                            sizeof(int),
 532                                                            GFP_USER);
 533                 if (!htab->map_locked[i])
 534                         goto free_map_locked;
 535         }
 536
 537         if (htab->map.map_flags & BPF_F_ZERO_SEED)
 538                 htab->hashrnd = 0;
 539         else
 540                 htab->hashrnd = get_random_u32();
 541
 542         htab_init_buckets(htab);
 543
 544 /* compute_batch_value() computes batch value as num_online_cpus() * 2
 545  * and __percpu_counter_compare() needs
 546  * htab->max_entries - cur_number_of_elems to be more than batch * num_online_cpus()
 547  * for percpu_counter to be faster than atomic_t. In practice the average bpf
 548  * hash map size is 10k, which means that a system with 64 cpus will fill
 549  * hashmap to 20% of 10k before percpu_counter becomes ineffective. Therefore
 550  * define our own batch count as 32 then 10k hash map can be filled up to 80%:
 551  * 10k - 8k > 32 _batch_ * 64 _cpus_
 552  * and __percpu_counter_compare() will still be fast. At that point hash map
 553  * collisions will dominate its performance anyway. Assume that hash map filled
 554  * to 50+% isn't going to be O(1) and use the following formula to choose
 555  * between percpu_counter and atomic_t.
 556  */
 557 #define PERCPU_COUNTER_BATCH 32
 558         if (attr->max_entries / 2 > num_online_cpus() * PERCPU_COUNTER_BATCH)
 559                 htab->use_percpu_counter = true;
 560
 561         if (htab->use_percpu_counter) {
 562                 err = percpu_counter_init(&htab->pcount, 0, GFP_KERNEL);
 563                 if (err)
 564                         goto free_map_locked;
 565         }
 566
 567         if (prealloc) {
 568                 err = prealloc_init(htab);
 569                 if (err)
 570                         goto free_map_locked;
 571
 572                 if (!percpu && !lru) {
 573                         /* lru itself can remove the least used element, so
 574                          * there is no need for an extra elem during map_update.
 575                          */
 576                         err = alloc_extra_elems(htab);
 577                         if (err)
 578                                 goto free_prealloc;
 579                 }
 580         } else {
 581                 err = bpf_mem_alloc_init(&htab->ma, htab->elem_size, false);
 582                 if (err)
 583                         goto free_map_locked;
 584                 if (percpu) {
 585                         err = bpf_mem_alloc_init(&htab->pcpu_ma,
 586                                                  round_up(htab->map.value_size, 8), true);
 587                         if (err)
 588                                 goto free_map_locked;
 589                 }
 590         }
 591
 592         return &htab->map;
 593
 594 free_prealloc:
 595         prealloc_destroy(htab);
 596 free_map_locked:
 597         if (htab->use_percpu_counter)
 598                 percpu_counter_destroy(&htab->pcount);
 599         for (i = 0; i < HASHTAB_MAP_LOCK_COUNT; i++)
 600                 free_percpu(htab->map_locked[i]);
 601         bpf_map_area_free(htab->buckets);
 602         bpf_mem_alloc_destroy(&htab->pcpu_ma);
 603         bpf_mem_alloc_destroy(&htab->ma);
 604 free_elem_count:
 605         bpf_map_free_elem_count(&htab->map);
 606 free_htab:
 607         lockdep_unregister_key(&htab->lockdep_key);
 608         bpf_map_area_free(htab);
 609         return ERR_PTR(err);
 610 }
 611
 612 static inline u32 htab_map_hash(const void *key, u32 key_len, u32 hashrnd)
 613 {
 614         if (likely(key_len % 4 == 0))
 615                 return jhash2(key, key_len / 4, hashrnd);
 616         return jhash(key, key_len, hashrnd);
 617 }
 618
 619 static inline struct bucket *__select_bucket(struct bpf_htab *htab, u32 hash)
 620 {
 621         return &htab->buckets[hash & (htab->n_buckets - 1)];
 622 }
 623
 624 static inline struct hlist_nulls_head *select_bucket(struct bpf_htab *htab, u32 hash)
 625 {
 626         return &__select_bucket(htab, hash)->head;
 627 }
 628
 629 /* this lookup function can only be called with bucket lock taken */
 630 static struct htab_elem *lookup_elem_raw(struct hlist_nulls_head *head, u32 hash,
 631                                          void *key, u32 key_size)
 632 {
 633         struct hlist_nulls_node *n;
 634         struct htab_elem *l;
 635
 636         hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
 637                 if (l->hash == hash && !memcmp(&l->key, key, key_size))
 638                         return l;
 639
 640         return NULL;
 641 }
 642
 643 /* can be called without bucket lock. it will repeat the loop in
 644  * the unlikely event when elements moved from one bucket into another
 645  * while link list is being walked
 646  */
 647 static struct htab_elem *lookup_nulls_elem_raw(struct hlist_nulls_head *head,
 648                                                u32 hash, void *key,
 649                                                u32 key_size, u32 n_buckets)
 650 {
 651         struct hlist_nulls_node *n;
 652         struct htab_elem *l;
 653
 654 again:
 655         hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
 656                 if (l->hash == hash && !memcmp(&l->key, key, key_size))
 657                         return l;
 658
 659         if (unlikely(get_nulls_value(n) != (hash & (n_buckets - 1))))
 660                 goto again;
 661
 662         return NULL;
 663 }
 664
 665 /* Called from syscall or from eBPF program directly, so
 666  * arguments have to match bpf_map_lookup_elem() exactly.
 667  * The return value is adjusted by BPF instructions
 668  * in htab_map_gen_lookup().
 669  */
 670 static void *__htab_map_lookup_elem(struct bpf_map *map, void *key)
 671 {
 672         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 673         struct hlist_nulls_head *head;
 674         struct htab_elem *l;
 675         u32 hash, key_size;
 676
 677         WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
 678                      !rcu_read_lock_bh_held());
 679
 680         key_size = map->key_size;
 681
 682         hash = htab_map_hash(key, key_size, htab->hashrnd);
 683
 684         head = select_bucket(htab, hash);
 685
 686         l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
 687
 688         return l;
 689 }
 690
 691 static void *htab_map_lookup_elem(struct bpf_map *map, void *key)
 692 {
 693         struct htab_elem *l = __htab_map_lookup_elem(map, key);
 694
 695         if (l)
 696                 return l->key + round_up(map->key_size, 8);
 697
 698         return NULL;
 699 }
 700
 701 /* inline bpf_map_lookup_elem() call.
 702  * Instead of:
 703  * bpf_prog
 704  *   bpf_map_lookup_elem
 705  *     map->ops->map_lookup_elem
 706  *       htab_map_lookup_elem
 707  *         __htab_map_lookup_elem
 708  * do:
 709  * bpf_prog
 710  *   __htab_map_lookup_elem
 711  */
 712 static int htab_map_gen_lookup(struct bpf_map *map, struct bpf_insn *insn_buf)
 713 {
 714         struct bpf_insn *insn = insn_buf;
 715         const int ret = BPF_REG_0;
 716
 717         BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
 718                      (void *(*)(struct bpf_map *map, void *key))NULL));
 719         *insn++ = BPF_EMIT_CALL(__htab_map_lookup_elem);
 720         *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 1);
 721         *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
 722                                 offsetof(struct htab_elem, key) +
 723                                 round_up(map->key_size, 8));
 724         return insn - insn_buf;
 725 }
 726
 727 static __always_inline void *__htab_lru_map_lookup_elem(struct bpf_map *map,
 728                                                         void *key, const bool mark)
 729 {
 730         struct htab_elem *l = __htab_map_lookup_elem(map, key);
 731
 732         if (l) {
 733                 if (mark)
 734                         bpf_lru_node_set_ref(&l->lru_node);
 735                 return l->key + round_up(map->key_size, 8);
 736         }
 737
 738         return NULL;
 739 }
 740
 741 static void *htab_lru_map_lookup_elem(struct bpf_map *map, void *key)
 742 {
 743         return __htab_lru_map_lookup_elem(map, key, true);
 744 }
 745
 746 static void *htab_lru_map_lookup_elem_sys(struct bpf_map *map, void *key)
 747 {
 748         return __htab_lru_map_lookup_elem(map, key, false);
 749 }
 750
 751 static int htab_lru_map_gen_lookup(struct bpf_map *map,
 752                                    struct bpf_insn *insn_buf)
 753 {
 754         struct bpf_insn *insn = insn_buf;
 755         const int ret = BPF_REG_0;
 756         const int ref_reg = BPF_REG_1;
 757
 758         BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
 759                      (void *(*)(struct bpf_map *map, void *key))NULL));
 760         *insn++ = BPF_EMIT_CALL(__htab_map_lookup_elem);
 761         *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 4);
 762         *insn++ = BPF_LDX_MEM(BPF_B, ref_reg, ret,
 763                               offsetof(struct htab_elem, lru_node) +
 764                               offsetof(struct bpf_lru_node, ref));
 765         *insn++ = BPF_JMP_IMM(BPF_JNE, ref_reg, 0, 1);
 766         *insn++ = BPF_ST_MEM(BPF_B, ret,
 767                              offsetof(struct htab_elem, lru_node) +
 768                              offsetof(struct bpf_lru_node, ref),
 769                              1);
 770         *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
 771                                 offsetof(struct htab_elem, key) +
 772                                 round_up(map->key_size, 8));
 773         return insn - insn_buf;
 774 }
 775
 776 static void check_and_free_fields(struct bpf_htab *htab,
 777                                   struct htab_elem *elem)
 778 {
 779         if (htab_is_percpu(htab)) {
 780                 void __percpu *pptr = htab_elem_get_ptr(elem, htab->map.key_size);
 781                 int cpu;
 782
 783                 for_each_possible_cpu(cpu)
 784                         bpf_obj_free_fields(htab->map.record, per_cpu_ptr(pptr, cpu));
 785         } else {
 786                 void *map_value = elem->key + round_up(htab->map.key_size, 8);
 787
 788                 bpf_obj_free_fields(htab->map.record, map_value);
 789         }
 790 }
 791
 792 /* It is called from the bpf_lru_list when the LRU needs to delete
 793  * older elements from the htab.
 794  */
 795 static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node)
 796 {
 797         struct bpf_htab *htab = arg;
 798         struct htab_elem *l = NULL, *tgt_l;
 799         struct hlist_nulls_head *head;
 800         struct hlist_nulls_node *n;
 801         unsigned long flags;
 802         struct bucket *b;
 803         int ret;
 804
 805         tgt_l = container_of(node, struct htab_elem, lru_node);
 806         b = __select_bucket(htab, tgt_l->hash);
 807         head = &b->head;
 808
 809         ret = htab_lock_bucket(htab, b, tgt_l->hash, &flags);
 810         if (ret)
 811                 return false;
 812
 813         hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
 814                 if (l == tgt_l) {
 815                         hlist_nulls_del_rcu(&l->hash_node);
 816                         check_and_free_fields(htab, l);
 817                         bpf_map_dec_elem_count(&htab->map);
 818                         break;
 819                 }
 820
 821         htab_unlock_bucket(htab, b, tgt_l->hash, flags);
 822
 823         return l == tgt_l;
 824 }
 825
 826 /* Called from syscall */
 827 static int htab_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
 828 {
 829         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 830         struct hlist_nulls_head *head;
 831         struct htab_elem *l, *next_l;
 832         u32 hash, key_size;
 833         int i = 0;
 834
 835         WARN_ON_ONCE(!rcu_read_lock_held());
 836
 837         key_size = map->key_size;
 838
 839         if (!key)
 840                 goto find_first_elem;
 841
 842         hash = htab_map_hash(key, key_size, htab->hashrnd);
 843
 844         head = select_bucket(htab, hash);
 845
 846         /* lookup the key */
 847         l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
 848
 849         if (!l)
 850                 goto find_first_elem;
 851
 852         /* key was found, get next key in the same bucket */
 853         next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_next_rcu(&l->hash_node)),
 854                                   struct htab_elem, hash_node);
 855
 856         if (next_l) {
 857                 /* if next elem in this hash list is non-zero, just return it */
 858                 memcpy(next_key, next_l->key, key_size);
 859                 return 0;
 860         }
 861
 862         /* no more elements in this hash list, go to the next bucket */
 863         i = hash & (htab->n_buckets - 1);
 864         i++;
 865
 866 find_first_elem:
 867         /* iterate over buckets */
 868         for (; i < htab->n_buckets; i++) {
 869                 head = select_bucket(htab, i);
 870
 871                 /* pick first element in the bucket */
 872                 next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_first_rcu(head)),
 873                                           struct htab_elem, hash_node);
 874                 if (next_l) {
 875                         /* if it's not empty, just return it */
 876                         memcpy(next_key, next_l->key, key_size);
 877                         return 0;
 878                 }
 879         }
 880
 881         /* iterated over all buckets and all elements */
 882         return -ENOENT;
 883 }
 884
 885 static void htab_elem_free(struct bpf_htab *htab, struct htab_elem *l)
 886 {
 887         check_and_free_fields(htab, l);
 888         if (htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH)
 889                 bpf_mem_cache_free(&htab->pcpu_ma, l->ptr_to_pptr);
 890         bpf_mem_cache_free(&htab->ma, l);
 891 }
 892
 893 static void htab_put_fd_value(struct bpf_htab *htab, struct htab_elem *l)
 894 {
 895         struct bpf_map *map = &htab->map;
 896         void *ptr;
 897
 898         if (map->ops->map_fd_put_ptr) {
 899                 ptr = fd_htab_map_get_ptr(map, l);
 900                 map->ops->map_fd_put_ptr(ptr);
 901         }
 902 }
 903
 904 static bool is_map_full(struct bpf_htab *htab)
 905 {
 906         if (htab->use_percpu_counter)
 907                 return __percpu_counter_compare(&htab->pcount, htab->map.max_entries,
 908                                                 PERCPU_COUNTER_BATCH) >= 0;
 909         return atomic_read(&htab->count) >= htab->map.max_entries;
 910 }
 911
 912 static void inc_elem_count(struct bpf_htab *htab)
 913 {
 914         bpf_map_inc_elem_count(&htab->map);
 915
 916         if (htab->use_percpu_counter)
 917                 percpu_counter_add_batch(&htab->pcount, 1, PERCPU_COUNTER_BATCH);
 918         else
 919                 atomic_inc(&htab->count);
 920 }
 921
 922 static void dec_elem_count(struct bpf_htab *htab)
 923 {
 924         bpf_map_dec_elem_count(&htab->map);
 925
 926         if (htab->use_percpu_counter)
 927                 percpu_counter_add_batch(&htab->pcount, -1, PERCPU_COUNTER_BATCH);
 928         else
 929                 atomic_dec(&htab->count);
 930 }
 931
 932
 933 static void free_htab_elem(struct bpf_htab *htab, struct htab_elem *l)
 934 {
 935         htab_put_fd_value(htab, l);
 936
 937         if (htab_is_prealloc(htab)) {
 938                 bpf_map_dec_elem_count(&htab->map);
 939                 check_and_free_fields(htab, l);
 940                 __pcpu_freelist_push(&htab->freelist, &l->fnode);
 941         } else {
 942                 dec_elem_count(htab);
 943                 htab_elem_free(htab, l);
 944         }
 945 }
 946
 947 static void pcpu_copy_value(struct bpf_htab *htab, void __percpu *pptr,
 948                             void *value, bool onallcpus)
 949 {
 950         if (!onallcpus) {
 951                 /* copy true value_size bytes */
 952                 copy_map_value(&htab->map, this_cpu_ptr(pptr), value);
 953         } else {
 954                 u32 size = round_up(htab->map.value_size, 8);
 955                 int off = 0, cpu;
 956
 957                 for_each_possible_cpu(cpu) {
 958                         copy_map_value_long(&htab->map, per_cpu_ptr(pptr, cpu), value + off);
 959                         off += size;
 960                 }
 961         }
 962 }
 963
 964 static void pcpu_init_value(struct bpf_htab *htab, void __percpu *pptr,
 965                             void *value, bool onallcpus)
 966 {
 967         /* When not setting the initial value on all cpus, zero-fill element
 968          * values for other cpus. Otherwise, bpf program has no way to ensure
 969          * known initial values for cpus other than current one
 970          * (onallcpus=false always when coming from bpf prog).
 971          */
 972         if (!onallcpus) {
 973                 int current_cpu = raw_smp_processor_id();
 974                 int cpu;
 975
 976                 for_each_possible_cpu(cpu) {
 977                         if (cpu == current_cpu)
 978                                 copy_map_value_long(&htab->map, per_cpu_ptr(pptr, cpu), value);
 979                         else /* Since elem is preallocated, we cannot touch special fields */
 980                                 zero_map_value(&htab->map, per_cpu_ptr(pptr, cpu));
 981                 }
 982         } else {
 983                 pcpu_copy_value(htab, pptr, value, onallcpus);
 984         }
 985 }
 986
 987 static bool fd_htab_map_needs_adjust(const struct bpf_htab *htab)
 988 {
 989         return htab->map.map_type == BPF_MAP_TYPE_HASH_OF_MAPS &&
 990                BITS_PER_LONG == 64;
 991 }
 992
 993 static struct htab_elem *alloc_htab_elem(struct bpf_htab *htab, void *key,
 994                                          void *value, u32 key_size, u32 hash,
 995                                          bool percpu, bool onallcpus,
 996                                          struct htab_elem *old_elem)
 997 {
 998         u32 size = htab->map.value_size;
 999         bool prealloc = htab_is_prealloc(htab);
1000         struct htab_elem *l_new, **pl_new;
1001         void __percpu *pptr;
1002
1003         if (prealloc) {
1004                 if (old_elem) {
1005                         /* if we're updating the existing element,
1006                          * use per-cpu extra elems to avoid freelist_pop/push
1007                          */
1008                         pl_new = this_cpu_ptr(htab->extra_elems);
1009                         l_new = *pl_new;
1010                         htab_put_fd_value(htab, old_elem);
1011                         *pl_new = old_elem;
1012                 } else {
1013                         struct pcpu_freelist_node *l;
1014
1015                         l = __pcpu_freelist_pop(&htab->freelist);
1016                         if (!l)
1017                                 return ERR_PTR(-E2BIG);
1018                         l_new = container_of(l, struct htab_elem, fnode);
1019                         bpf_map_inc_elem_count(&htab->map);
1020                 }
1021         } else {
1022                 if (is_map_full(htab))
1023                         if (!old_elem)
1024                                 /* when map is full and update() is replacing
1025                                  * old element, it's ok to allocate, since
1026                                  * old element will be freed immediately.
1027                                  * Otherwise return an error
1028                                  */
1029                                 return ERR_PTR(-E2BIG);
1030                 inc_elem_count(htab);
1031                 l_new = bpf_mem_cache_alloc(&htab->ma);
1032                 if (!l_new) {
1033                         l_new = ERR_PTR(-ENOMEM);
1034                         goto dec_count;
1035                 }
1036         }
1037
1038         memcpy(l_new->key, key, key_size);
1039         if (percpu) {
1040                 if (prealloc) {
1041                         pptr = htab_elem_get_ptr(l_new, key_size);
1042                 } else {
1043                         /* alloc_percpu zero-fills */
1044                         pptr = bpf_mem_cache_alloc(&htab->pcpu_ma);
1045                         if (!pptr) {
1046                                 bpf_mem_cache_free(&htab->ma, l_new);
1047                                 l_new = ERR_PTR(-ENOMEM);
1048                                 goto dec_count;
1049                         }
1050                         l_new->ptr_to_pptr = pptr;
1051                         pptr = *(void **)pptr;
1052                 }
1053
1054                 pcpu_init_value(htab, pptr, value, onallcpus);
1055
1056                 if (!prealloc)
1057                         htab_elem_set_ptr(l_new, key_size, pptr);
1058         } else if (fd_htab_map_needs_adjust(htab)) {
1059                 size = round_up(size, 8);
1060                 memcpy(l_new->key + round_up(key_size, 8), value, size);
1061         } else {
1062                 copy_map_value(&htab->map,
1063                                l_new->key + round_up(key_size, 8),
1064                                value);
1065         }
1066
1067         l_new->hash = hash;
1068         return l_new;
1069 dec_count:
1070         dec_elem_count(htab);
1071         return l_new;
1072 }
1073
1074 static int check_flags(struct bpf_htab *htab, struct htab_elem *l_old,
1075                        u64 map_flags)
1076 {
1077         if (l_old && (map_flags & ~BPF_F_LOCK) == BPF_NOEXIST)
1078                 /* elem already exists */
1079                 return -EEXIST;
1080
1081         if (!l_old && (map_flags & ~BPF_F_LOCK) == BPF_EXIST)
1082                 /* elem doesn't exist, cannot update it */
1083                 return -ENOENT;
1084
1085         return 0;
1086 }
1087
1088 /* Called from syscall or from eBPF program */
1089 static long htab_map_update_elem(struct bpf_map *map, void *key, void *value,
1090                                  u64 map_flags)
1091 {
1092         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1093         struct htab_elem *l_new = NULL, *l_old;
1094         struct hlist_nulls_head *head;
1095         unsigned long flags;
1096         struct bucket *b;
1097         u32 key_size, hash;
1098         int ret;
1099
1100         if (unlikely((map_flags & ~BPF_F_LOCK) > BPF_EXIST))
1101                 /* unknown flags */
1102                 return -EINVAL;
1103
1104         WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
1105                      !rcu_read_lock_bh_held());
1106
1107         key_size = map->key_size;
1108
1109         hash = htab_map_hash(key, key_size, htab->hashrnd);
1110
1111         b = __select_bucket(htab, hash);
1112         head = &b->head;
1113
1114         if (unlikely(map_flags & BPF_F_LOCK)) {
1115                 if (unlikely(!btf_record_has_field(map->record, BPF_SPIN_LOCK)))
1116                         return -EINVAL;
1117                 /* find an element without taking the bucket lock */
1118                 l_old = lookup_nulls_elem_raw(head, hash, key, key_size,
1119                                               htab->n_buckets);
1120                 ret = check_flags(htab, l_old, map_flags);
1121                 if (ret)
1122                         return ret;
1123                 if (l_old) {
1124                         /* grab the element lock and update value in place */
1125                         copy_map_value_locked(map,
1126                                               l_old->key + round_up(key_size, 8),
1127                                               value, false);
1128                         return 0;
1129                 }
1130                 /* fall through, grab the bucket lock and lookup again.
1131                  * 99.9% chance that the element won't be found,
1132                  * but second lookup under lock has to be done.
1133                  */
1134         }
1135
1136         ret = htab_lock_bucket(htab, b, hash, &flags);
1137         if (ret)
1138                 return ret;
1139
1140         l_old = lookup_elem_raw(head, hash, key, key_size);
1141
1142         ret = check_flags(htab, l_old, map_flags);
1143         if (ret)
1144                 goto err;
1145
1146         if (unlikely(l_old && (map_flags & BPF_F_LOCK))) {
1147                 /* first lookup without the bucket lock didn't find the element,
1148                  * but second lookup with the bucket lock found it.
1149                  * This case is highly unlikely, but has to be dealt with:
1150                  * grab the element lock in addition to the bucket lock
1151                  * and update element in place
1152                  */
1153                 copy_map_value_locked(map,
1154                                       l_old->key + round_up(key_size, 8),
1155                                       value, false);
1156                 ret = 0;
1157                 goto err;
1158         }
1159
1160         l_new = alloc_htab_elem(htab, key, value, key_size, hash, false, false,
1161                                 l_old);
1162         if (IS_ERR(l_new)) {
1163                 /* all pre-allocated elements are in use or memory exhausted */
1164                 ret = PTR_ERR(l_new);
1165                 goto err;
1166         }
1167
1168         /* add new element to the head of the list, so that
1169          * concurrent search will find it before old elem
1170          */
1171         hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1172         if (l_old) {
1173                 hlist_nulls_del_rcu(&l_old->hash_node);
1174                 if (!htab_is_prealloc(htab))
1175                         free_htab_elem(htab, l_old);
1176                 else
1177                         check_and_free_fields(htab, l_old);
1178         }
1179         ret = 0;
1180 err:
1181         htab_unlock_bucket(htab, b, hash, flags);
1182         return ret;
1183 }
1184
1185 static void htab_lru_push_free(struct bpf_htab *htab, struct htab_elem *elem)
1186 {
1187         check_and_free_fields(htab, elem);
1188         bpf_map_dec_elem_count(&htab->map);
1189         bpf_lru_push_free(&htab->lru, &elem->lru_node);
1190 }
1191
1192 static long htab_lru_map_update_elem(struct bpf_map *map, void *key, void *value,
1193                                      u64 map_flags)
1194 {
1195         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1196         struct htab_elem *l_new, *l_old = NULL;
1197         struct hlist_nulls_head *head;
1198         unsigned long flags;
1199         struct bucket *b;
1200         u32 key_size, hash;
1201         int ret;
1202
1203         if (unlikely(map_flags > BPF_EXIST))
1204                 /* unknown flags */
1205                 return -EINVAL;
1206
1207         WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
1208                      !rcu_read_lock_bh_held());
1209
1210         key_size = map->key_size;
1211
1212         hash = htab_map_hash(key, key_size, htab->hashrnd);
1213
1214         b = __select_bucket(htab, hash);
1215         head = &b->head;
1216
1217         /* For LRU, we need to alloc before taking bucket's
1218          * spinlock because getting free nodes from LRU may need
1219          * to remove older elements from htab and this removal
1220          * operation will need a bucket lock.
1221          */
1222         l_new = prealloc_lru_pop(htab, key, hash);
1223         if (!l_new)
1224                 return -ENOMEM;
1225         copy_map_value(&htab->map,
1226                        l_new->key + round_up(map->key_size, 8), value);
1227
1228         ret = htab_lock_bucket(htab, b, hash, &flags);
1229         if (ret)
1230                 goto err_lock_bucket;
1231
1232         l_old = lookup_elem_raw(head, hash, key, key_size);
1233
1234         ret = check_flags(htab, l_old, map_flags);
1235         if (ret)
1236                 goto err;
1237
1238         /* add new element to the head of the list, so that
1239          * concurrent search will find it before old elem
1240          */
1241         hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1242         if (l_old) {
1243                 bpf_lru_node_set_ref(&l_new->lru_node);
1244                 hlist_nulls_del_rcu(&l_old->hash_node);
1245         }
1246         ret = 0;
1247
1248 err:
1249         htab_unlock_bucket(htab, b, hash, flags);
1250
1251 err_lock_bucket:
1252         if (ret)
1253                 htab_lru_push_free(htab, l_new);
1254         else if (l_old)
1255                 htab_lru_push_free(htab, l_old);
1256
1257         return ret;
1258 }
1259
1260 static long __htab_percpu_map_update_elem(struct bpf_map *map, void *key,
1261                                           void *value, u64 map_flags,
1262                                           bool onallcpus)
1263 {
1264         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1265         struct htab_elem *l_new = NULL, *l_old;
1266         struct hlist_nulls_head *head;
1267         unsigned long flags;
1268         struct bucket *b;
1269         u32 key_size, hash;
1270         int ret;
1271
1272         if (unlikely(map_flags > BPF_EXIST))
1273                 /* unknown flags */
1274                 return -EINVAL;
1275
1276         WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
1277                      !rcu_read_lock_bh_held());
1278
1279         key_size = map->key_size;
1280
1281         hash = htab_map_hash(key, key_size, htab->hashrnd);
1282
1283         b = __select_bucket(htab, hash);
1284         head = &b->head;
1285
1286         ret = htab_lock_bucket(htab, b, hash, &flags);
1287         if (ret)
1288                 return ret;
1289
1290         l_old = lookup_elem_raw(head, hash, key, key_size);
1291
1292         ret = check_flags(htab, l_old, map_flags);
1293         if (ret)
1294                 goto err;
1295
1296         if (l_old) {
1297                 /* per-cpu hash map can update value in-place */
1298                 pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
1299                                 value, onallcpus);
1300         } else {
1301                 l_new = alloc_htab_elem(htab, key, value, key_size,
1302                                         hash, true, onallcpus, NULL);
1303                 if (IS_ERR(l_new)) {
1304                         ret = PTR_ERR(l_new);
1305                         goto err;
1306                 }
1307                 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1308         }
1309         ret = 0;
1310 err:
1311         htab_unlock_bucket(htab, b, hash, flags);
1312         return ret;
1313 }
1314
1315 static long __htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
1316                                               void *value, u64 map_flags,
1317                                               bool onallcpus)
1318 {
1319         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1320         struct htab_elem *l_new = NULL, *l_old;
1321         struct hlist_nulls_head *head;
1322         unsigned long flags;
1323         struct bucket *b;
1324         u32 key_size, hash;
1325         int ret;
1326
1327         if (unlikely(map_flags > BPF_EXIST))
1328                 /* unknown flags */
1329                 return -EINVAL;
1330
1331         WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
1332                      !rcu_read_lock_bh_held());
1333
1334         key_size = map->key_size;
1335
1336         hash = htab_map_hash(key, key_size, htab->hashrnd);
1337
1338         b = __select_bucket(htab, hash);
1339         head = &b->head;
1340
1341         /* For LRU, we need to alloc before taking bucket's
1342          * spinlock because LRU's elem alloc may need
1343          * to remove older elem from htab and this removal
1344          * operation will need a bucket lock.
1345          */
1346         if (map_flags != BPF_EXIST) {
1347                 l_new = prealloc_lru_pop(htab, key, hash);
1348                 if (!l_new)
1349                         return -ENOMEM;
1350         }
1351
1352         ret = htab_lock_bucket(htab, b, hash, &flags);
1353         if (ret)
1354                 goto err_lock_bucket;
1355
1356         l_old = lookup_elem_raw(head, hash, key, key_size);
1357
1358         ret = check_flags(htab, l_old, map_flags);
1359         if (ret)
1360                 goto err;
1361
1362         if (l_old) {
1363                 bpf_lru_node_set_ref(&l_old->lru_node);
1364
1365                 /* per-cpu hash map can update value in-place */
1366                 pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
1367                                 value, onallcpus);
1368         } else {
1369                 pcpu_init_value(htab, htab_elem_get_ptr(l_new, key_size),
1370                                 value, onallcpus);
1371                 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1372                 l_new = NULL;
1373         }
1374         ret = 0;
1375 err:
1376         htab_unlock_bucket(htab, b, hash, flags);
1377 err_lock_bucket:
1378         if (l_new) {
1379                 bpf_map_dec_elem_count(&htab->map);
1380                 bpf_lru_push_free(&htab->lru, &l_new->lru_node);
1381         }
1382         return ret;
1383 }
1384
1385 static long htab_percpu_map_update_elem(struct bpf_map *map, void *key,
1386                                         void *value, u64 map_flags)
1387 {
1388         return __htab_percpu_map_update_elem(map, key, value, map_flags, false);
1389 }
1390
1391 static long htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
1392                                             void *value, u64 map_flags)
1393 {
1394         return __htab_lru_percpu_map_update_elem(map, key, value, map_flags,
1395                                                  false);
1396 }
1397
1398 /* Called from syscall or from eBPF program */
1399 static long htab_map_delete_elem(struct bpf_map *map, void *key)
1400 {
1401         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1402         struct hlist_nulls_head *head;
1403         struct bucket *b;
1404         struct htab_elem *l;
1405         unsigned long flags;
1406         u32 hash, key_size;
1407         int ret;
1408
1409         WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
1410                      !rcu_read_lock_bh_held());
1411
1412         key_size = map->key_size;
1413
1414         hash = htab_map_hash(key, key_size, htab->hashrnd);
1415         b = __select_bucket(htab, hash);
1416         head = &b->head;
1417
1418         ret = htab_lock_bucket(htab, b, hash, &flags);
1419         if (ret)
1420                 return ret;
1421
1422         l = lookup_elem_raw(head, hash, key, key_size);
1423
1424         if (l) {
1425                 hlist_nulls_del_rcu(&l->hash_node);
1426                 free_htab_elem(htab, l);
1427         } else {
1428                 ret = -ENOENT;
1429         }
1430
1431         htab_unlock_bucket(htab, b, hash, flags);
1432         return ret;
1433 }
1434
1435 static long htab_lru_map_delete_elem(struct bpf_map *map, void *key)
1436 {
1437         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1438         struct hlist_nulls_head *head;
1439         struct bucket *b;
1440         struct htab_elem *l;
1441         unsigned long flags;
1442         u32 hash, key_size;
1443         int ret;
1444
1445         WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
1446                      !rcu_read_lock_bh_held());
1447
1448         key_size = map->key_size;
1449
1450         hash = htab_map_hash(key, key_size, htab->hashrnd);
1451         b = __select_bucket(htab, hash);
1452         head = &b->head;
1453
1454         ret = htab_lock_bucket(htab, b, hash, &flags);
1455         if (ret)
1456                 return ret;
1457
1458         l = lookup_elem_raw(head, hash, key, key_size);
1459
1460         if (l)
1461                 hlist_nulls_del_rcu(&l->hash_node);
1462         else
1463                 ret = -ENOENT;
1464
1465         htab_unlock_bucket(htab, b, hash, flags);
1466         if (l)
1467                 htab_lru_push_free(htab, l);
1468         return ret;
1469 }
1470
1471 static void delete_all_elements(struct bpf_htab *htab)
1472 {
1473         int i;
1474
1475         /* It's called from a worker thread, so disable migration here,
1476          * since bpf_mem_cache_free() relies on that.
1477          */
1478         migrate_disable();
1479         for (i = 0; i < htab->n_buckets; i++) {
1480                 struct hlist_nulls_head *head = select_bucket(htab, i);
1481                 struct hlist_nulls_node *n;
1482                 struct htab_elem *l;
1483
1484                 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1485                         hlist_nulls_del_rcu(&l->hash_node);
1486                         htab_elem_free(htab, l);
1487                 }
1488         }
1489         migrate_enable();
1490 }
1491
1492 static void htab_free_malloced_timers(struct bpf_htab *htab)
1493 {
1494         int i;
1495
1496         rcu_read_lock();
1497         for (i = 0; i < htab->n_buckets; i++) {
1498                 struct hlist_nulls_head *head = select_bucket(htab, i);
1499                 struct hlist_nulls_node *n;
1500                 struct htab_elem *l;
1501
1502                 hlist_nulls_for_each_entry(l, n, head, hash_node) {
1503                         /* We only free timer on uref dropping to zero */
1504                         bpf_obj_free_timer(htab->map.record, l->key + round_up(htab->map.key_size, 8));
1505                 }
1506                 cond_resched_rcu();
1507         }
1508         rcu_read_unlock();
1509 }
1510
1511 static void htab_map_free_timers(struct bpf_map *map)
1512 {
1513         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1514
1515         /* We only free timer on uref dropping to zero */
1516         if (!btf_record_has_field(htab->map.record, BPF_TIMER))
1517                 return;
1518         if (!htab_is_prealloc(htab))
1519                 htab_free_malloced_timers(htab);
1520         else
1521                 htab_free_prealloced_timers(htab);
1522 }
1523
1524 /* Called when map->refcnt goes to zero, either from workqueue or from syscall */
1525 static void htab_map_free(struct bpf_map *map)
1526 {
1527         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1528         int i;
1529
1530         /* bpf_free_used_maps() or close(map_fd) will trigger this map_free callback.
1531          * bpf_free_used_maps() is called after bpf prog is no longer executing.
1532          * There is no need to synchronize_rcu() here to protect map elements.
1533          */
1534
1535         /* htab no longer uses call_rcu() directly. bpf_mem_alloc does it
1536          * underneath and is reponsible for waiting for callbacks to finish
1537          * during bpf_mem_alloc_destroy().
1538          */
1539         if (!htab_is_prealloc(htab)) {
1540                 delete_all_elements(htab);
1541         } else {
1542                 htab_free_prealloced_fields(htab);
1543                 prealloc_destroy(htab);
1544         }
1545
1546         bpf_map_free_elem_count(map);
1547         free_percpu(htab->extra_elems);
1548         bpf_map_area_free(htab->buckets);
1549         bpf_mem_alloc_destroy(&htab->pcpu_ma);
1550         bpf_mem_alloc_destroy(&htab->ma);
1551         if (htab->use_percpu_counter)
1552                 percpu_counter_destroy(&htab->pcount);
1553         for (i = 0; i < HASHTAB_MAP_LOCK_COUNT; i++)
1554                 free_percpu(htab->map_locked[i]);
1555         lockdep_unregister_key(&htab->lockdep_key);
1556         bpf_map_area_free(htab);
1557 }
1558
1559 static void htab_map_seq_show_elem(struct bpf_map *map, void *key,
1560                                    struct seq_file *m)
1561 {
1562         void *value;
1563
1564         rcu_read_lock();
1565
1566         value = htab_map_lookup_elem(map, key);
1567         if (!value) {
1568                 rcu_read_unlock();
1569                 return;
1570         }
1571
1572         btf_type_seq_show(map->btf, map->btf_key_type_id, key, m);
1573         seq_puts(m, ": ");
1574         btf_type_seq_show(map->btf, map->btf_value_type_id, value, m);
1575         seq_puts(m, "\n");
1576
1577         rcu_read_unlock();
1578 }
1579
1580 static int __htab_map_lookup_and_delete_elem(struct bpf_map *map, void *key,
1581                                              void *value, bool is_lru_map,
1582                                              bool is_percpu, u64 flags)
1583 {
1584         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1585         struct hlist_nulls_head *head;
1586         unsigned long bflags;
1587         struct htab_elem *l;
1588         u32 hash, key_size;
1589         struct bucket *b;
1590         int ret;
1591
1592         key_size = map->key_size;
1593
1594         hash = htab_map_hash(key, key_size, htab->hashrnd);
1595         b = __select_bucket(htab, hash);
1596         head = &b->head;
1597
1598         ret = htab_lock_bucket(htab, b, hash, &bflags);
1599         if (ret)
1600                 return ret;
1601
1602         l = lookup_elem_raw(head, hash, key, key_size);
1603         if (!l) {
1604                 ret = -ENOENT;
1605         } else {
1606                 if (is_percpu) {
1607                         u32 roundup_value_size = round_up(map->value_size, 8);
1608                         void __percpu *pptr;
1609                         int off = 0, cpu;
1610
1611                         pptr = htab_elem_get_ptr(l, key_size);
1612                         for_each_possible_cpu(cpu) {
1613                                 copy_map_value_long(&htab->map, value + off, per_cpu_ptr(pptr, cpu));
1614                                 check_and_init_map_value(&htab->map, value + off);
1615                                 off += roundup_value_size;
1616                         }
1617                 } else {
1618                         u32 roundup_key_size = round_up(map->key_size, 8);
1619
1620                         if (flags & BPF_F_LOCK)
1621                                 copy_map_value_locked(map, value, l->key +
1622                                                       roundup_key_size,
1623                                                       true);
1624                         else
1625                                 copy_map_value(map, value, l->key +
1626                                                roundup_key_size);
1627                         /* Zeroing special fields in the temp buffer */
1628                         check_and_init_map_value(map, value);
1629                 }
1630
1631                 hlist_nulls_del_rcu(&l->hash_node);
1632                 if (!is_lru_map)
1633                         free_htab_elem(htab, l);
1634         }
1635
1636         htab_unlock_bucket(htab, b, hash, bflags);
1637
1638         if (is_lru_map && l)
1639                 htab_lru_push_free(htab, l);
1640
1641         return ret;
1642 }
1643
1644 static int htab_map_lookup_and_delete_elem(struct bpf_map *map, void *key,
1645                                            void *value, u64 flags)
1646 {
1647         return __htab_map_lookup_and_delete_elem(map, key, value, false, false,
1648                                                  flags);
1649 }
1650
1651 static int htab_percpu_map_lookup_and_delete_elem(struct bpf_map *map,
1652                                                   void *key, void *value,
1653                                                   u64 flags)
1654 {
1655         return __htab_map_lookup_and_delete_elem(map, key, value, false, true,
1656                                                  flags);
1657 }
1658
1659 static int htab_lru_map_lookup_and_delete_elem(struct bpf_map *map, void *key,
1660                                                void *value, u64 flags)
1661 {
1662         return __htab_map_lookup_and_delete_elem(map, key, value, true, false,
1663                                                  flags);
1664 }
1665
1666 static int htab_lru_percpu_map_lookup_and_delete_elem(struct bpf_map *map,
1667                                                       void *key, void *value,
1668                                                       u64 flags)
1669 {
1670         return __htab_map_lookup_and_delete_elem(map, key, value, true, true,
1671                                                  flags);
1672 }
1673
1674 static int
1675 __htab_map_lookup_and_delete_batch(struct bpf_map *map,
1676                                    const union bpf_attr *attr,
1677                                    union bpf_attr __user *uattr,
1678                                    bool do_delete, bool is_lru_map,
1679                                    bool is_percpu)
1680 {
1681         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1682         u32 bucket_cnt, total, key_size, value_size, roundup_key_size;
1683         void *keys = NULL, *values = NULL, *value, *dst_key, *dst_val;
1684         void __user *uvalues = u64_to_user_ptr(attr->batch.values);
1685         void __user *ukeys = u64_to_user_ptr(attr->batch.keys);
1686         void __user *ubatch = u64_to_user_ptr(attr->batch.in_batch);
1687         u32 batch, max_count, size, bucket_size, map_id;
1688         struct htab_elem *node_to_free = NULL;
1689         u64 elem_map_flags, map_flags;
1690         struct hlist_nulls_head *head;
1691         struct hlist_nulls_node *n;
1692         unsigned long flags = 0;
1693         bool locked = false;
1694         struct htab_elem *l;
1695         struct bucket *b;
1696         int ret = 0;
1697
1698         elem_map_flags = attr->batch.elem_flags;
1699         if ((elem_map_flags & ~BPF_F_LOCK) ||
1700             ((elem_map_flags & BPF_F_LOCK) && !btf_record_has_field(map->record, BPF_SPIN_LOCK)))
1701                 return -EINVAL;
1702
1703         map_flags = attr->batch.flags;
1704         if (map_flags)
1705                 return -EINVAL;
1706
1707         max_count = attr->batch.count;
1708         if (!max_count)
1709                 return 0;
1710
1711         if (put_user(0, &uattr->batch.count))
1712                 return -EFAULT;
1713
1714         batch = 0;
1715         if (ubatch && copy_from_user(&batch, ubatch, sizeof(batch)))
1716                 return -EFAULT;
1717
1718         if (batch >= htab->n_buckets)
1719                 return -ENOENT;
1720
1721         key_size = htab->map.key_size;
1722         roundup_key_size = round_up(htab->map.key_size, 8);
1723         value_size = htab->map.value_size;
1724         size = round_up(value_size, 8);
1725         if (is_percpu)
1726                 value_size = size * num_possible_cpus();
1727         total = 0;
1728         /* while experimenting with hash tables with sizes ranging from 10 to
1729          * 1000, it was observed that a bucket can have up to 5 entries.
1730          */
1731         bucket_size = 5;
1732
1733 alloc:
1734         /* We cannot do copy_from_user or copy_to_user inside
1735          * the rcu_read_lock. Allocate enough space here.
1736          */
1737         keys = kvmalloc_array(key_size, bucket_size, GFP_USER | __GFP_NOWARN);
1738         values = kvmalloc_array(value_size, bucket_size, GFP_USER | __GFP_NOWARN);
1739         if (!keys || !values) {
1740                 ret = -ENOMEM;
1741                 goto after_loop;
1742         }
1743
1744 again:
1745         bpf_disable_instrumentation();
1746         rcu_read_lock();
1747 again_nocopy:
1748         dst_key = keys;
1749         dst_val = values;
1750         b = &htab->buckets[batch];
1751         head = &b->head;
1752         /* do not grab the lock unless need it (bucket_cnt > 0). */
1753         if (locked) {
1754                 ret = htab_lock_bucket(htab, b, batch, &flags);
1755                 if (ret) {
1756                         rcu_read_unlock();
1757                         bpf_enable_instrumentation();
1758                         goto after_loop;
1759                 }
1760         }
1761
1762         bucket_cnt = 0;
1763         hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
1764                 bucket_cnt++;
1765
1766         if (bucket_cnt && !locked) {
1767                 locked = true;
1768                 goto again_nocopy;
1769         }
1770
1771         if (bucket_cnt > (max_count - total)) {
1772                 if (total == 0)
1773                         ret = -ENOSPC;
1774                 /* Note that since bucket_cnt > 0 here, it is implicit
1775                  * that the locked was grabbed, so release it.
1776                  */
1777                 htab_unlock_bucket(htab, b, batch, flags);
1778                 rcu_read_unlock();
1779                 bpf_enable_instrumentation();
1780                 goto after_loop;
1781         }
1782
1783         if (bucket_cnt > bucket_size) {
1784                 bucket_size = bucket_cnt;
1785                 /* Note that since bucket_cnt > 0 here, it is implicit
1786                  * that the locked was grabbed, so release it.
1787                  */
1788                 htab_unlock_bucket(htab, b, batch, flags);
1789                 rcu_read_unlock();
1790                 bpf_enable_instrumentation();
1791                 kvfree(keys);
1792                 kvfree(values);
1793                 goto alloc;
1794         }
1795
1796         /* Next block is only safe to run if you have grabbed the lock */
1797         if (!locked)
1798                 goto next_batch;
1799
1800         hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1801                 memcpy(dst_key, l->key, key_size);
1802
1803                 if (is_percpu) {
1804                         int off = 0, cpu;
1805                         void __percpu *pptr;
1806
1807                         pptr = htab_elem_get_ptr(l, map->key_size);
1808                         for_each_possible_cpu(cpu) {
1809                                 copy_map_value_long(&htab->map, dst_val + off, per_cpu_ptr(pptr, cpu));
1810                                 check_and_init_map_value(&htab->map, dst_val + off);
1811                                 off += size;
1812                         }
1813                 } else {
1814                         value = l->key + roundup_key_size;
1815                         if (map->map_type == BPF_MAP_TYPE_HASH_OF_MAPS) {
1816                                 struct bpf_map **inner_map = value;
1817
1818                                  /* Actual value is the id of the inner map */
1819                                 map_id = map->ops->map_fd_sys_lookup_elem(*inner_map);
1820                                 value = &map_id;
1821                         }
1822
1823                         if (elem_map_flags & BPF_F_LOCK)
1824                                 copy_map_value_locked(map, dst_val, value,
1825                                                       true);
1826                         else
1827                                 copy_map_value(map, dst_val, value);
1828                         /* Zeroing special fields in the temp buffer */
1829                         check_and_init_map_value(map, dst_val);
1830                 }
1831                 if (do_delete) {
1832                         hlist_nulls_del_rcu(&l->hash_node);
1833
1834                         /* bpf_lru_push_free() will acquire lru_lock, which
1835                          * may cause deadlock. See comments in function
1836                          * prealloc_lru_pop(). Let us do bpf_lru_push_free()
1837                          * after releasing the bucket lock.
1838                          */
1839                         if (is_lru_map) {
1840                                 l->batch_flink = node_to_free;
1841                                 node_to_free = l;
1842                         } else {
1843                                 free_htab_elem(htab, l);
1844                         }
1845                 }
1846                 dst_key += key_size;
1847                 dst_val += value_size;
1848         }
1849
1850         htab_unlock_bucket(htab, b, batch, flags);
1851         locked = false;
1852
1853         while (node_to_free) {
1854                 l = node_to_free;
1855                 node_to_free = node_to_free->batch_flink;
1856                 htab_lru_push_free(htab, l);
1857         }
1858
1859 next_batch:
1860         /* If we are not copying data, we can go to next bucket and avoid
1861          * unlocking the rcu.
1862          */
1863         if (!bucket_cnt && (batch + 1 < htab->n_buckets)) {
1864                 batch++;
1865                 goto again_nocopy;
1866         }
1867
1868         rcu_read_unlock();
1869         bpf_enable_instrumentation();
1870         if (bucket_cnt && (copy_to_user(ukeys + total * key_size, keys,
1871             key_size * bucket_cnt) ||
1872             copy_to_user(uvalues + total * value_size, values,
1873             value_size * bucket_cnt))) {
1874                 ret = -EFAULT;
1875                 goto after_loop;
1876         }
1877
1878         total += bucket_cnt;
1879         batch++;
1880         if (batch >= htab->n_buckets) {
1881                 ret = -ENOENT;
1882                 goto after_loop;
1883         }
1884         goto again;
1885
1886 after_loop:
1887         if (ret == -EFAULT)
1888                 goto out;
1889
1890         /* copy # of entries and next batch */
1891         ubatch = u64_to_user_ptr(attr->batch.out_batch);
1892         if (copy_to_user(ubatch, &batch, sizeof(batch)) ||
1893             put_user(total, &uattr->batch.count))
1894                 ret = -EFAULT;
1895
1896 out:
1897         kvfree(keys);
1898         kvfree(values);
1899         return ret;
1900 }
1901
1902 static int
1903 htab_percpu_map_lookup_batch(struct bpf_map *map, const union bpf_attr *attr,
1904                              union bpf_attr __user *uattr)
1905 {
1906         return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1907                                                   false, true);
1908 }
1909
1910 static int
1911 htab_percpu_map_lookup_and_delete_batch(struct bpf_map *map,
1912                                         const union bpf_attr *attr,
1913                                         union bpf_attr __user *uattr)
1914 {
1915         return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1916                                                   false, true);
1917 }
1918
1919 static int
1920 htab_map_lookup_batch(struct bpf_map *map, const union bpf_attr *attr,
1921                       union bpf_attr __user *uattr)
1922 {
1923         return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1924                                                   false, false);
1925 }
1926
1927 static int
1928 htab_map_lookup_and_delete_batch(struct bpf_map *map,
1929                                  const union bpf_attr *attr,
1930                                  union bpf_attr __user *uattr)
1931 {
1932         return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1933                                                   false, false);
1934 }
1935
1936 static int
1937 htab_lru_percpu_map_lookup_batch(struct bpf_map *map,
1938                                  const union bpf_attr *attr,
1939                                  union bpf_attr __user *uattr)
1940 {
1941         return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1942                                                   true, true);
1943 }
1944
1945 static int
1946 htab_lru_percpu_map_lookup_and_delete_batch(struct bpf_map *map,
1947                                             const union bpf_attr *attr,
1948                                             union bpf_attr __user *uattr)
1949 {
1950         return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1951                                                   true, true);
1952 }
1953
1954 static int
1955 htab_lru_map_lookup_batch(struct bpf_map *map, const union bpf_attr *attr,
1956                           union bpf_attr __user *uattr)
1957 {
1958         return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1959                                                   true, false);
1960 }
1961
1962 static int
1963 htab_lru_map_lookup_and_delete_batch(struct bpf_map *map,
1964                                      const union bpf_attr *attr,
1965                                      union bpf_attr __user *uattr)
1966 {
1967         return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1968                                                   true, false);
1969 }
1970
1971 struct bpf_iter_seq_hash_map_info {
1972         struct bpf_map *map;
1973         struct bpf_htab *htab;
1974         void *percpu_value_buf; // non-zero means percpu hash
1975         u32 bucket_id;
1976         u32 skip_elems;
1977 };
1978
1979 static struct htab_elem *
1980 bpf_hash_map_seq_find_next(struct bpf_iter_seq_hash_map_info *info,
1981                            struct htab_elem *prev_elem)
1982 {
1983         const struct bpf_htab *htab = info->htab;
1984         u32 skip_elems = info->skip_elems;
1985         u32 bucket_id = info->bucket_id;
1986         struct hlist_nulls_head *head;
1987         struct hlist_nulls_node *n;
1988         struct htab_elem *elem;
1989         struct bucket *b;
1990         u32 i, count;
1991
1992         if (bucket_id >= htab->n_buckets)
1993                 return NULL;
1994
1995         /* try to find next elem in the same bucket */
1996         if (prev_elem) {
1997                 /* no update/deletion on this bucket, prev_elem should be still valid
1998                  * and we won't skip elements.
1999                  */
2000                 n = rcu_dereference_raw(hlist_nulls_next_rcu(&prev_elem->hash_node));
2001                 elem = hlist_nulls_entry_safe(n, struct htab_elem, hash_node);
2002                 if (elem)
2003                         return elem;
2004
2005                 /* not found, unlock and go to the next bucket */
2006                 b = &htab->buckets[bucket_id++];
2007                 rcu_read_unlock();
2008                 skip_elems = 0;
2009         }
2010
2011         for (i = bucket_id; i < htab->n_buckets; i++) {
2012                 b = &htab->buckets[i];
2013                 rcu_read_lock();
2014
2015                 count = 0;
2016                 head = &b->head;
2017                 hlist_nulls_for_each_entry_rcu(elem, n, head, hash_node) {
2018                         if (count >= skip_elems) {
2019                                 info->bucket_id = i;
2020                                 info->skip_elems = count;
2021                                 return elem;
2022                         }
2023                         count++;
2024                 }
2025
2026                 rcu_read_unlock();
2027                 skip_elems = 0;
2028         }
2029
2030         info->bucket_id = i;
2031         info->skip_elems = 0;
2032         return NULL;
2033 }
2034
2035 static void *bpf_hash_map_seq_start(struct seq_file *seq, loff_t *pos)
2036 {
2037         struct bpf_iter_seq_hash_map_info *info = seq->private;
2038         struct htab_elem *elem;
2039
2040         elem = bpf_hash_map_seq_find_next(info, NULL);
2041         if (!elem)
2042                 return NULL;
2043
2044         if (*pos == 0)
2045                 ++*pos;
2046         return elem;
2047 }
2048
2049 static void *bpf_hash_map_seq_next(struct seq_file *seq, void *v, loff_t *pos)
2050 {
2051         struct bpf_iter_seq_hash_map_info *info = seq->private;
2052
2053         ++*pos;
2054         ++info->skip_elems;
2055         return bpf_hash_map_seq_find_next(info, v);
2056 }
2057
2058 static int __bpf_hash_map_seq_show(struct seq_file *seq, struct htab_elem *elem)
2059 {
2060         struct bpf_iter_seq_hash_map_info *info = seq->private;
2061         u32 roundup_key_size, roundup_value_size;
2062         struct bpf_iter__bpf_map_elem ctx = {};
2063         struct bpf_map *map = info->map;
2064         struct bpf_iter_meta meta;
2065         int ret = 0, off = 0, cpu;
2066         struct bpf_prog *prog;
2067         void __percpu *pptr;
2068
2069         meta.seq = seq;
2070         prog = bpf_iter_get_info(&meta, elem == NULL);
2071         if (prog) {
2072                 ctx.meta = &meta;
2073                 ctx.map = info->map;
2074                 if (elem) {
2075                         roundup_key_size = round_up(map->key_size, 8);
2076                         ctx.key = elem->key;
2077                         if (!info->percpu_value_buf) {
2078                                 ctx.value = elem->key + roundup_key_size;
2079                         } else {
2080                                 roundup_value_size = round_up(map->value_size, 8);
2081                                 pptr = htab_elem_get_ptr(elem, map->key_size);
2082                                 for_each_possible_cpu(cpu) {
2083                                         copy_map_value_long(map, info->percpu_value_buf + off,
2084                                                             per_cpu_ptr(pptr, cpu));
2085                                         check_and_init_map_value(map, info->percpu_value_buf + off);
2086                                         off += roundup_value_size;
2087                                 }
2088                                 ctx.value = info->percpu_value_buf;
2089                         }
2090                 }
2091                 ret = bpf_iter_run_prog(prog, &ctx);
2092         }
2093
2094         return ret;
2095 }
2096
2097 static int bpf_hash_map_seq_show(struct seq_file *seq, void *v)
2098 {
2099         return __bpf_hash_map_seq_show(seq, v);
2100 }
2101
2102 static void bpf_hash_map_seq_stop(struct seq_file *seq, void *v)
2103 {
2104         if (!v)
2105                 (void)__bpf_hash_map_seq_show(seq, NULL);
2106         else
2107                 rcu_read_unlock();
2108 }
2109
2110 static int bpf_iter_init_hash_map(void *priv_data,
2111                                   struct bpf_iter_aux_info *aux)
2112 {
2113         struct bpf_iter_seq_hash_map_info *seq_info = priv_data;
2114         struct bpf_map *map = aux->map;
2115         void *value_buf;
2116         u32 buf_size;
2117
2118         if (map->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
2119             map->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH) {
2120                 buf_size = round_up(map->value_size, 8) * num_possible_cpus();
2121                 value_buf = kmalloc(buf_size, GFP_USER | __GFP_NOWARN);
2122                 if (!value_buf)
2123                         return -ENOMEM;
2124
2125                 seq_info->percpu_value_buf = value_buf;
2126         }
2127
2128         bpf_map_inc_with_uref(map);
2129         seq_info->map = map;
2130         seq_info->htab = container_of(map, struct bpf_htab, map);
2131         return 0;
2132 }
2133
2134 static void bpf_iter_fini_hash_map(void *priv_data)
2135 {
2136         struct bpf_iter_seq_hash_map_info *seq_info = priv_data;
2137
2138         bpf_map_put_with_uref(seq_info->map);
2139         kfree(seq_info->percpu_value_buf);
2140 }
2141
2142 static const struct seq_operations bpf_hash_map_seq_ops = {
2143         .start  = bpf_hash_map_seq_start,
2144         .next   = bpf_hash_map_seq_next,
2145         .stop   = bpf_hash_map_seq_stop,
2146         .show   = bpf_hash_map_seq_show,
2147 };
2148
2149 static const struct bpf_iter_seq_info iter_seq_info = {
2150         .seq_ops                = &bpf_hash_map_seq_ops,
2151         .init_seq_private       = bpf_iter_init_hash_map,
2152         .fini_seq_private       = bpf_iter_fini_hash_map,
2153         .seq_priv_size          = sizeof(struct bpf_iter_seq_hash_map_info),
2154 };
2155
2156 static long bpf_for_each_hash_elem(struct bpf_map *map, bpf_callback_t callback_fn,
2157                                    void *callback_ctx, u64 flags)
2158 {
2159         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
2160         struct hlist_nulls_head *head;
2161         struct hlist_nulls_node *n;
2162         struct htab_elem *elem;
2163         u32 roundup_key_size;
2164         int i, num_elems = 0;
2165         void __percpu *pptr;
2166         struct bucket *b;
2167         void *key, *val;
2168         bool is_percpu;
2169         u64 ret = 0;
2170
2171         if (flags != 0)
2172                 return -EINVAL;
2173
2174         is_percpu = htab_is_percpu(htab);
2175
2176         roundup_key_size = round_up(map->key_size, 8);
2177         /* disable migration so percpu value prepared here will be the
2178          * same as the one seen by the bpf program with bpf_map_lookup_elem().
2179          */
2180         if (is_percpu)
2181                 migrate_disable();
2182         for (i = 0; i < htab->n_buckets; i++) {
2183                 b = &htab->buckets[i];
2184                 rcu_read_lock();
2185                 head = &b->head;
2186                 hlist_nulls_for_each_entry_rcu(elem, n, head, hash_node) {
2187                         key = elem->key;
2188                         if (is_percpu) {
2189                                 /* current cpu value for percpu map */
2190                                 pptr = htab_elem_get_ptr(elem, map->key_size);
2191                                 val = this_cpu_ptr(pptr);
2192                         } else {
2193                                 val = elem->key + roundup_key_size;
2194                         }
2195                         num_elems++;
2196                         ret = callback_fn((u64)(long)map, (u64)(long)key,
2197                                           (u64)(long)val, (u64)(long)callback_ctx, 0);
2198                         /* return value: 0 - continue, 1 - stop and return */
2199                         if (ret) {
2200                                 rcu_read_unlock();
2201                                 goto out;
2202                         }
2203                 }
2204                 rcu_read_unlock();
2205         }
2206 out:
2207         if (is_percpu)
2208                 migrate_enable();
2209         return num_elems;
2210 }
2211
2212 static u64 htab_map_mem_usage(const struct bpf_map *map)
2213 {
2214         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
2215         u32 value_size = round_up(htab->map.value_size, 8);
2216         bool prealloc = htab_is_prealloc(htab);
2217         bool percpu = htab_is_percpu(htab);
2218         bool lru = htab_is_lru(htab);
2219         u64 num_entries;
2220         u64 usage = sizeof(struct bpf_htab);
2221
2222         usage += sizeof(struct bucket) * htab->n_buckets;
2223         usage += sizeof(int) * num_possible_cpus() * HASHTAB_MAP_LOCK_COUNT;
2224         if (prealloc) {
2225                 num_entries = map->max_entries;
2226                 if (htab_has_extra_elems(htab))
2227                         num_entries += num_possible_cpus();
2228
2229                 usage += htab->elem_size * num_entries;
2230
2231                 if (percpu)
2232                         usage += value_size * num_possible_cpus() * num_entries;
2233                 else if (!lru)
2234                         usage += sizeof(struct htab_elem *) * num_possible_cpus();
2235         } else {
2236 #define LLIST_NODE_SZ sizeof(struct llist_node)
2237
2238                 num_entries = htab->use_percpu_counter ?
2239                                           percpu_counter_sum(&htab->pcount) :
2240                                           atomic_read(&htab->count);
2241                 usage += (htab->elem_size + LLIST_NODE_SZ) * num_entries;
2242                 if (percpu) {
2243                         usage += (LLIST_NODE_SZ + sizeof(void *)) * num_entries;
2244                         usage += value_size * num_possible_cpus() * num_entries;
2245                 }
2246         }
2247         return usage;
2248 }
2249
2250 BTF_ID_LIST_SINGLE(htab_map_btf_ids, struct, bpf_htab)
2251 const struct bpf_map_ops htab_map_ops = {
2252         .map_meta_equal = bpf_map_meta_equal,
2253         .map_alloc_check = htab_map_alloc_check,
2254         .map_alloc = htab_map_alloc,
2255         .map_free = htab_map_free,
2256         .map_get_next_key = htab_map_get_next_key,
2257         .map_release_uref = htab_map_free_timers,
2258         .map_lookup_elem = htab_map_lookup_elem,
2259         .map_lookup_and_delete_elem = htab_map_lookup_and_delete_elem,
2260         .map_update_elem = htab_map_update_elem,
2261         .map_delete_elem = htab_map_delete_elem,
2262         .map_gen_lookup = htab_map_gen_lookup,
2263         .map_seq_show_elem = htab_map_seq_show_elem,
2264         .map_set_for_each_callback_args = map_set_for_each_callback_args,
2265         .map_for_each_callback = bpf_for_each_hash_elem,
2266         .map_mem_usage = htab_map_mem_usage,
2267         BATCH_OPS(htab),
2268         .map_btf_id = &htab_map_btf_ids[0],
2269         .iter_seq_info = &iter_seq_info,
2270 };
2271
2272 const struct bpf_map_ops htab_lru_map_ops = {
2273         .map_meta_equal = bpf_map_meta_equal,
2274         .map_alloc_check = htab_map_alloc_check,
2275         .map_alloc = htab_map_alloc,
2276         .map_free = htab_map_free,
2277         .map_get_next_key = htab_map_get_next_key,
2278         .map_release_uref = htab_map_free_timers,
2279         .map_lookup_elem = htab_lru_map_lookup_elem,
2280         .map_lookup_and_delete_elem = htab_lru_map_lookup_and_delete_elem,
2281         .map_lookup_elem_sys_only = htab_lru_map_lookup_elem_sys,
2282         .map_update_elem = htab_lru_map_update_elem,
2283         .map_delete_elem = htab_lru_map_delete_elem,
2284         .map_gen_lookup = htab_lru_map_gen_lookup,
2285         .map_seq_show_elem = htab_map_seq_show_elem,
2286         .map_set_for_each_callback_args = map_set_for_each_callback_args,
2287         .map_for_each_callback = bpf_for_each_hash_elem,
2288         .map_mem_usage = htab_map_mem_usage,
2289         BATCH_OPS(htab_lru),
2290         .map_btf_id = &htab_map_btf_ids[0],
2291         .iter_seq_info = &iter_seq_info,
2292 };
2293
2294 /* Called from eBPF program */
2295 static void *htab_percpu_map_lookup_elem(struct bpf_map *map, void *key)
2296 {
2297         struct htab_elem *l = __htab_map_lookup_elem(map, key);
2298
2299         if (l)
2300                 return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
2301         else
2302                 return NULL;
2303 }
2304
2305 static void *htab_percpu_map_lookup_percpu_elem(struct bpf_map *map, void *key, u32 cpu)
2306 {
2307         struct htab_elem *l;
2308
2309         if (cpu >= nr_cpu_ids)
2310                 return NULL;
2311
2312         l = __htab_map_lookup_elem(map, key);
2313         if (l)
2314                 return per_cpu_ptr(htab_elem_get_ptr(l, map->key_size), cpu);
2315         else
2316                 return NULL;
2317 }
2318
2319 static void *htab_lru_percpu_map_lookup_elem(struct bpf_map *map, void *key)
2320 {
2321         struct htab_elem *l = __htab_map_lookup_elem(map, key);
2322
2323         if (l) {
2324                 bpf_lru_node_set_ref(&l->lru_node);
2325                 return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
2326         }
2327
2328         return NULL;
2329 }
2330
2331 static void *htab_lru_percpu_map_lookup_percpu_elem(struct bpf_map *map, void *key, u32 cpu)
2332 {
2333         struct htab_elem *l;
2334
2335         if (cpu >= nr_cpu_ids)
2336                 return NULL;
2337
2338         l = __htab_map_lookup_elem(map, key);
2339         if (l) {
2340                 bpf_lru_node_set_ref(&l->lru_node);
2341                 return per_cpu_ptr(htab_elem_get_ptr(l, map->key_size), cpu);
2342         }
2343
2344         return NULL;
2345 }
2346
2347 int bpf_percpu_hash_copy(struct bpf_map *map, void *key, void *value)
2348 {
2349         struct htab_elem *l;
2350         void __percpu *pptr;
2351         int ret = -ENOENT;
2352         int cpu, off = 0;
2353         u32 size;
2354
2355         /* per_cpu areas are zero-filled and bpf programs can only
2356          * access 'value_size' of them, so copying rounded areas
2357          * will not leak any kernel data
2358          */
2359         size = round_up(map->value_size, 8);
2360         rcu_read_lock();
2361         l = __htab_map_lookup_elem(map, key);
2362         if (!l)
2363                 goto out;
2364         /* We do not mark LRU map element here in order to not mess up
2365          * eviction heuristics when user space does a map walk.
2366          */
2367         pptr = htab_elem_get_ptr(l, map->key_size);
2368         for_each_possible_cpu(cpu) {
2369                 copy_map_value_long(map, value + off, per_cpu_ptr(pptr, cpu));
2370                 check_and_init_map_value(map, value + off);
2371                 off += size;
2372         }
2373         ret = 0;
2374 out:
2375         rcu_read_unlock();
2376         return ret;
2377 }
2378
2379 int bpf_percpu_hash_update(struct bpf_map *map, void *key, void *value,
2380                            u64 map_flags)
2381 {
2382         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
2383         int ret;
2384
2385         rcu_read_lock();
2386         if (htab_is_lru(htab))
2387                 ret = __htab_lru_percpu_map_update_elem(map, key, value,
2388                                                         map_flags, true);
2389         else
2390                 ret = __htab_percpu_map_update_elem(map, key, value, map_flags,
2391                                                     true);
2392         rcu_read_unlock();
2393
2394         return ret;
2395 }
2396
2397 static void htab_percpu_map_seq_show_elem(struct bpf_map *map, void *key,
2398                                           struct seq_file *m)
2399 {
2400         struct htab_elem *l;
2401         void __percpu *pptr;
2402         int cpu;
2403
2404         rcu_read_lock();
2405
2406         l = __htab_map_lookup_elem(map, key);
2407         if (!l) {
2408                 rcu_read_unlock();
2409                 return;
2410         }
2411
2412         btf_type_seq_show(map->btf, map->btf_key_type_id, key, m);
2413         seq_puts(m, ": {\n");
2414         pptr = htab_elem_get_ptr(l, map->key_size);
2415         for_each_possible_cpu(cpu) {
2416                 seq_printf(m, "\tcpu%d: ", cpu);
2417                 btf_type_seq_show(map->btf, map->btf_value_type_id,
2418                                   per_cpu_ptr(pptr, cpu), m);
2419                 seq_puts(m, "\n");
2420         }
2421         seq_puts(m, "}\n");
2422
2423         rcu_read_unlock();
2424 }
2425
2426 const struct bpf_map_ops htab_percpu_map_ops = {
2427         .map_meta_equal = bpf_map_meta_equal,
2428         .map_alloc_check = htab_map_alloc_check,
2429         .map_alloc = htab_map_alloc,
2430         .map_free = htab_map_free,
2431         .map_get_next_key = htab_map_get_next_key,
2432         .map_lookup_elem = htab_percpu_map_lookup_elem,
2433         .map_lookup_and_delete_elem = htab_percpu_map_lookup_and_delete_elem,
2434         .map_update_elem = htab_percpu_map_update_elem,
2435         .map_delete_elem = htab_map_delete_elem,
2436         .map_lookup_percpu_elem = htab_percpu_map_lookup_percpu_elem,
2437         .map_seq_show_elem = htab_percpu_map_seq_show_elem,
2438         .map_set_for_each_callback_args = map_set_for_each_callback_args,
2439         .map_for_each_callback = bpf_for_each_hash_elem,
2440         .map_mem_usage = htab_map_mem_usage,
2441         BATCH_OPS(htab_percpu),
2442         .map_btf_id = &htab_map_btf_ids[0],
2443         .iter_seq_info = &iter_seq_info,
2444 };
2445
2446 const struct bpf_map_ops htab_lru_percpu_map_ops = {
2447         .map_meta_equal = bpf_map_meta_equal,
2448         .map_alloc_check = htab_map_alloc_check,
2449         .map_alloc = htab_map_alloc,
2450         .map_free = htab_map_free,
2451         .map_get_next_key = htab_map_get_next_key,
2452         .map_lookup_elem = htab_lru_percpu_map_lookup_elem,
2453         .map_lookup_and_delete_elem = htab_lru_percpu_map_lookup_and_delete_elem,
2454         .map_update_elem = htab_lru_percpu_map_update_elem,
2455         .map_delete_elem = htab_lru_map_delete_elem,
2456         .map_lookup_percpu_elem = htab_lru_percpu_map_lookup_percpu_elem,
2457         .map_seq_show_elem = htab_percpu_map_seq_show_elem,
2458         .map_set_for_each_callback_args = map_set_for_each_callback_args,
2459         .map_for_each_callback = bpf_for_each_hash_elem,
2460         .map_mem_usage = htab_map_mem_usage,
2461         BATCH_OPS(htab_lru_percpu),
2462         .map_btf_id = &htab_map_btf_ids[0],
2463         .iter_seq_info = &iter_seq_info,
2464 };
2465
2466 static int fd_htab_map_alloc_check(union bpf_attr *attr)
2467 {
2468         if (attr->value_size != sizeof(u32))
2469                 return -EINVAL;
2470         return htab_map_alloc_check(attr);
2471 }
2472
2473 static void fd_htab_map_free(struct bpf_map *map)
2474 {
2475         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
2476         struct hlist_nulls_node *n;
2477         struct hlist_nulls_head *head;
2478         struct htab_elem *l;
2479         int i;
2480
2481         for (i = 0; i < htab->n_buckets; i++) {
2482                 head = select_bucket(htab, i);
2483
2484                 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
2485                         void *ptr = fd_htab_map_get_ptr(map, l);
2486
2487                         map->ops->map_fd_put_ptr(ptr);
2488                 }
2489         }
2490
2491         htab_map_free(map);
2492 }
2493
2494 /* only called from syscall */
2495 int bpf_fd_htab_map_lookup_elem(struct bpf_map *map, void *key, u32 *value)
2496 {
2497         void **ptr;
2498         int ret = 0;
2499
2500         if (!map->ops->map_fd_sys_lookup_elem)
2501                 return -ENOTSUPP;
2502
2503         rcu_read_lock();
2504         ptr = htab_map_lookup_elem(map, key);
2505         if (ptr)
2506                 *value = map->ops->map_fd_sys_lookup_elem(READ_ONCE(*ptr));
2507         else
2508                 ret = -ENOENT;
2509         rcu_read_unlock();
2510
2511         return ret;
2512 }
2513
2514 /* only called from syscall */
2515 int bpf_fd_htab_map_update_elem(struct bpf_map *map, struct file *map_file,
2516                                 void *key, void *value, u64 map_flags)
2517 {
2518         void *ptr;
2519         int ret;
2520         u32 ufd = *(u32 *)value;
2521
2522         ptr = map->ops->map_fd_get_ptr(map, map_file, ufd);
2523         if (IS_ERR(ptr))
2524                 return PTR_ERR(ptr);
2525
2526         ret = htab_map_update_elem(map, key, &ptr, map_flags);
2527         if (ret)
2528                 map->ops->map_fd_put_ptr(ptr);
2529
2530         return ret;
2531 }
2532
2533 static struct bpf_map *htab_of_map_alloc(union bpf_attr *attr)
2534 {
2535         struct bpf_map *map, *inner_map_meta;
2536
2537         inner_map_meta = bpf_map_meta_alloc(attr->inner_map_fd);
2538         if (IS_ERR(inner_map_meta))
2539                 return inner_map_meta;
2540
2541         map = htab_map_alloc(attr);
2542         if (IS_ERR(map)) {
2543                 bpf_map_meta_free(inner_map_meta);
2544                 return map;
2545         }
2546
2547         map->inner_map_meta = inner_map_meta;
2548
2549         return map;
2550 }
2551
2552 static void *htab_of_map_lookup_elem(struct bpf_map *map, void *key)
2553 {
2554         struct bpf_map **inner_map  = htab_map_lookup_elem(map, key);
2555
2556         if (!inner_map)
2557                 return NULL;
2558
2559         return READ_ONCE(*inner_map);
2560 }
2561
2562 static int htab_of_map_gen_lookup(struct bpf_map *map,
2563                                   struct bpf_insn *insn_buf)
2564 {
2565         struct bpf_insn *insn = insn_buf;
2566         const int ret = BPF_REG_0;
2567
2568         BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
2569                      (void *(*)(struct bpf_map *map, void *key))NULL));
2570         *insn++ = BPF_EMIT_CALL(__htab_map_lookup_elem);
2571         *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 2);
2572         *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
2573                                 offsetof(struct htab_elem, key) +
2574                                 round_up(map->key_size, 8));
2575         *insn++ = BPF_LDX_MEM(BPF_DW, ret, ret, 0);
2576
2577         return insn - insn_buf;
2578 }
2579
2580 static void htab_of_map_free(struct bpf_map *map)
2581 {
2582         bpf_map_meta_free(map->inner_map_meta);
2583         fd_htab_map_free(map);
2584 }
2585
2586 const struct bpf_map_ops htab_of_maps_map_ops = {
2587         .map_alloc_check = fd_htab_map_alloc_check,
2588         .map_alloc = htab_of_map_alloc,
2589         .map_free = htab_of_map_free,
2590         .map_get_next_key = htab_map_get_next_key,
2591         .map_lookup_elem = htab_of_map_lookup_elem,
2592         .map_delete_elem = htab_map_delete_elem,
2593         .map_fd_get_ptr = bpf_map_fd_get_ptr,
2594         .map_fd_put_ptr = bpf_map_fd_put_ptr,
2595         .map_fd_sys_lookup_elem = bpf_map_fd_sys_lookup_elem,
2596         .map_gen_lookup = htab_of_map_gen_lookup,
2597         .map_check_btf = map_check_no_btf,
2598         .map_mem_usage = htab_map_mem_usage,
2599         BATCH_OPS(htab),
2600         .map_btf_id = &htab_map_btf_ids[0],
2601 };