kernel/bpf/hashtab.c

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