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