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