git.sesse.net Git - bcachefs-tools-debian/blob - include/linux/rhashtable.h

   1 /*
   2  * Resizable, Scalable, Concurrent Hash Table
   3  *
   4  * Copyright (c) 2015 Herbert Xu <herbert@gondor.apana.org.au>
   5  * Copyright (c) 2014-2015 Thomas Graf <tgraf@suug.ch>
   6  * Copyright (c) 2008-2014 Patrick McHardy <kaber@trash.net>
   7  *
   8  * Code partially derived from nft_hash
   9  * Rewritten with rehash code from br_multicast plus single list
  10  * pointer as suggested by Josh Triplett
  11  *
  12  * This program is free software; you can redistribute it and/or modify
  13  * it under the terms of the GNU General Public License version 2 as
  14  * published by the Free Software Foundation.
  15  */
  16
  17 #ifndef _LINUX_RHASHTABLE_H
  18 #define _LINUX_RHASHTABLE_H
  19
  20 #include <linux/atomic.h>
  21 #include <linux/cache.h>
  22 #include <linux/compiler.h>
  23 #include <linux/cpumask.h>
  24 #include <linux/err.h>
  25 #include <linux/errno.h>
  26 #include <linux/jhash.h>
  27 #include <linux/list_nulls.h>
  28 #include <linux/workqueue.h>
  29 #include <linux/mutex.h>
  30 #include <linux/spinlock.h>
  31 #include <linux/rcupdate.h>
  32
  33 #define RHT_BASE_BITS           4
  34 #define RHT_HASH_BITS           27
  35 #define RHT_BASE_SHIFT          RHT_HASH_BITS
  36 #define RHT_HASH_RESERVED_SPACE (RHT_BASE_BITS + 1)
  37
  38 struct rhash_head {
  39         struct rhash_head __rcu         *next;
  40 };
  41
  42 struct bucket_table {
  43         unsigned int            size;
  44         unsigned int            rehash;
  45         u32                     hash_rnd;
  46         unsigned int            locks_mask;
  47         spinlock_t              *locks;
  48         struct list_head        walkers;
  49         struct rcu_head         rcu;
  50
  51         struct bucket_table __rcu *future_tbl;
  52
  53         struct rhash_head __rcu *buckets[] ____cacheline_aligned_in_smp;
  54 };
  55
  56 struct rhashtable_compare_arg {
  57         struct rhashtable *ht;
  58         const void *key;
  59 };
  60
  61 typedef u32 (*rht_hashfn_t)(const void *data, u32 len, u32 seed);
  62 typedef u32 (*rht_obj_hashfn_t)(const void *data, u32 len, u32 seed);
  63 typedef int (*rht_obj_cmpfn_t)(struct rhashtable_compare_arg *arg,
  64                                const void *obj);
  65
  66 struct rhashtable_params {
  67         size_t                  nelem_hint;
  68         size_t                  key_len;
  69         size_t                  key_offset;
  70         size_t                  head_offset;
  71         unsigned int            insecure_max_entries;
  72         unsigned int            max_size;
  73         unsigned int            min_size;
  74         u32                     nulls_base;
  75         bool                    insecure_elasticity;
  76         bool                    automatic_shrinking;
  77         size_t                  locks_mul;
  78         rht_hashfn_t            hashfn;
  79         rht_obj_hashfn_t        obj_hashfn;
  80         rht_obj_cmpfn_t         obj_cmpfn;
  81 };
  82
  83 struct rhashtable {
  84         struct bucket_table __rcu       *tbl;
  85         atomic_t                        nelems;
  86         unsigned int                    key_len;
  87         unsigned int                    elasticity;
  88         struct rhashtable_params        p;
  89         struct work_struct              run_work;
  90         struct mutex                    mutex;
  91         spinlock_t                      lock;
  92 };
  93
  94 struct rhashtable_walker {
  95         struct list_head list;
  96         struct bucket_table *tbl;
  97 };
  98
  99 static inline unsigned long rht_marker(const struct rhashtable *ht, u32 hash)
 100 {
 101         return NULLS_MARKER(ht->p.nulls_base + hash);
 102 }
 103
 104 #define INIT_RHT_NULLS_HEAD(ptr, ht, hash) \
 105         ((ptr) = (typeof(ptr)) rht_marker(ht, hash))
 106
 107 static inline bool rht_is_a_nulls(const struct rhash_head *ptr)
 108 {
 109         return ((unsigned long) ptr & 1);
 110 }
 111
 112 static inline unsigned long rht_get_nulls_value(const struct rhash_head *ptr)
 113 {
 114         return ((unsigned long) ptr) >> 1;
 115 }
 116
 117 static inline void *rht_obj(const struct rhashtable *ht,
 118                             const struct rhash_head *he)
 119 {
 120         return (char *)he - ht->p.head_offset;
 121 }
 122
 123 static inline unsigned int rht_bucket_index(const struct bucket_table *tbl,
 124                                             unsigned int hash)
 125 {
 126         return (hash >> RHT_HASH_RESERVED_SPACE) & (tbl->size - 1);
 127 }
 128
 129 static inline unsigned int rht_key_hashfn(
 130         struct rhashtable *ht, const struct bucket_table *tbl,
 131         const void *key, const struct rhashtable_params params)
 132 {
 133         unsigned int hash;
 134
 135         /* params must be equal to ht->p if it isn't constant. */
 136         if (!__builtin_constant_p(params.key_len))
 137                 hash = ht->p.hashfn(key, ht->key_len, tbl->hash_rnd);
 138         else if (params.key_len) {
 139                 unsigned int key_len = params.key_len;
 140
 141                 if (params.hashfn)
 142                         hash = params.hashfn(key, key_len, tbl->hash_rnd);
 143                 else if (key_len & (sizeof(u32) - 1))
 144                         hash = jhash(key, key_len, tbl->hash_rnd);
 145                 else
 146                         hash = jhash2(key, key_len / sizeof(u32),
 147                                       tbl->hash_rnd);
 148         } else {
 149                 unsigned int key_len = ht->p.key_len;
 150
 151                 if (params.hashfn)
 152                         hash = params.hashfn(key, key_len, tbl->hash_rnd);
 153                 else
 154                         hash = jhash(key, key_len, tbl->hash_rnd);
 155         }
 156
 157         return rht_bucket_index(tbl, hash);
 158 }
 159
 160 static inline unsigned int rht_head_hashfn(
 161         struct rhashtable *ht, const struct bucket_table *tbl,
 162         const struct rhash_head *he, const struct rhashtable_params params)
 163 {
 164         const char *ptr = rht_obj(ht, he);
 165
 166         return likely(params.obj_hashfn) ?
 167                rht_bucket_index(tbl, params.obj_hashfn(ptr, params.key_len ?:
 168                                                             ht->p.key_len,
 169                                                        tbl->hash_rnd)) :
 170                rht_key_hashfn(ht, tbl, ptr + params.key_offset, params);
 171 }
 172
 173 static inline bool rht_grow_above_75(const struct rhashtable *ht,
 174                                      const struct bucket_table *tbl)
 175 {
 176         /* Expand table when exceeding 75% load */
 177         return atomic_read(&ht->nelems) > (tbl->size / 4 * 3) &&
 178                (!ht->p.max_size || tbl->size < ht->p.max_size);
 179 }
 180
 181 static inline bool rht_shrink_below_30(const struct rhashtable *ht,
 182                                        const struct bucket_table *tbl)
 183 {
 184         /* Shrink table beneath 30% load */
 185         return atomic_read(&ht->nelems) < (tbl->size * 3 / 10) &&
 186                tbl->size > ht->p.min_size;
 187 }
 188
 189 static inline bool rht_grow_above_100(const struct rhashtable *ht,
 190                                       const struct bucket_table *tbl)
 191 {
 192         return atomic_read(&ht->nelems) > tbl->size &&
 193                 (!ht->p.max_size || tbl->size < ht->p.max_size);
 194 }
 195
 196 static inline bool rht_grow_above_max(const struct rhashtable *ht,
 197                                       const struct bucket_table *tbl)
 198 {
 199         return ht->p.insecure_max_entries &&
 200                atomic_read(&ht->nelems) >= ht->p.insecure_max_entries;
 201 }
 202
 203 static inline spinlock_t *rht_bucket_lock(const struct bucket_table *tbl,
 204                                           unsigned int hash)
 205 {
 206         return &tbl->locks[hash & tbl->locks_mask];
 207 }
 208
 209 int rhashtable_insert_rehash(struct rhashtable *, struct bucket_table *);
 210 struct bucket_table *rhashtable_insert_slow(struct rhashtable *,
 211                                             const void *,
 212                                             struct rhash_head *,
 213                                             struct bucket_table *);
 214
 215 int rhashtable_init(struct rhashtable *, const struct rhashtable_params *);
 216 void rhashtable_destroy(struct rhashtable *);
 217
 218 #define rht_dereference(p, ht)                  rcu_dereference(p)
 219 #define rht_dereference_rcu(p, ht)              rcu_dereference(p)
 220 #define rht_dereference_bucket(p, tbl, hash)    rcu_dereference(p)
 221 #define rht_dereference_bucket_rcu(p, tbl, hash) rcu_dereference(p)
 222
 223 #define rht_entry(tpos, pos, member) \
 224         ({ tpos = container_of(pos, typeof(*tpos), member); 1; })
 225
 226 #define rht_for_each_continue(pos, head, tbl, hash) \
 227         for (pos = rht_dereference_bucket(head, tbl, hash); \
 228              !rht_is_a_nulls(pos); \
 229              pos = rht_dereference_bucket((pos)->next, tbl, hash))
 230
 231 #define rht_for_each(pos, tbl, hash) \
 232         rht_for_each_continue(pos, (tbl)->buckets[hash], tbl, hash)
 233
 234 #define rht_for_each_rcu_continue(pos, head, tbl, hash)                 \
 235         for (({barrier(); }),                                           \
 236              pos = rht_dereference_bucket_rcu(head, tbl, hash);         \
 237              !rht_is_a_nulls(pos);                                      \
 238              pos = rcu_dereference_raw(pos->next))
 239
 240 #define rht_for_each_rcu(pos, tbl, hash)                                \
 241         rht_for_each_rcu_continue(pos, (tbl)->buckets[hash], tbl, hash)
 242
 243 #define rht_for_each_entry_rcu_continue(tpos, pos, head, tbl, hash, member) \
 244         for (({barrier(); }),                                               \
 245              pos = rht_dereference_bucket_rcu(head, tbl, hash);             \
 246              (!rht_is_a_nulls(pos)) && rht_entry(tpos, pos, member);        \
 247              pos = rht_dereference_bucket_rcu(pos->next, tbl, hash))
 248
 249 #define rht_for_each_entry_rcu(tpos, pos, tbl, hash, member)            \
 250         rht_for_each_entry_rcu_continue(tpos, pos, (tbl)->buckets[hash],\
 251                                         tbl, hash, member)
 252
 253 static inline int rhashtable_compare(struct rhashtable_compare_arg *arg,
 254                                      const void *obj)
 255 {
 256         struct rhashtable *ht = arg->ht;
 257         const char *ptr = obj;
 258
 259         return memcmp(ptr + ht->p.key_offset, arg->key, ht->p.key_len);
 260 }
 261
 262 static inline void *rhashtable_lookup_fast(
 263         struct rhashtable *ht, const void *key,
 264         const struct rhashtable_params params)
 265 {
 266         struct rhashtable_compare_arg arg = {
 267                 .ht = ht,
 268                 .key = key,
 269         };
 270         const struct bucket_table *tbl;
 271         struct rhash_head *he;
 272         unsigned int hash;
 273
 274         rcu_read_lock();
 275
 276         tbl = rht_dereference_rcu(ht->tbl, ht);
 277 restart:
 278         hash = rht_key_hashfn(ht, tbl, key, params);
 279         rht_for_each_rcu(he, tbl, hash) {
 280                 if (params.obj_cmpfn ?
 281                     params.obj_cmpfn(&arg, rht_obj(ht, he)) :
 282                     rhashtable_compare(&arg, rht_obj(ht, he)))
 283                         continue;
 284                 rcu_read_unlock();
 285                 return rht_obj(ht, he);
 286         }
 287
 288         /* Ensure we see any new tables. */
 289         smp_rmb();
 290
 291         tbl = rht_dereference_rcu(tbl->future_tbl, ht);
 292         if (unlikely(tbl))
 293                 goto restart;
 294         rcu_read_unlock();
 295
 296         return NULL;
 297 }
 298
 299 static inline int __rhashtable_insert_fast(
 300         struct rhashtable *ht, const void *key, struct rhash_head *obj,
 301         const struct rhashtable_params params)
 302 {
 303         struct rhashtable_compare_arg arg = {
 304                 .ht = ht,
 305                 .key = key,
 306         };
 307         struct bucket_table *tbl, *new_tbl;
 308         struct rhash_head *head;
 309         spinlock_t *lock;
 310         unsigned int elasticity;
 311         unsigned int hash;
 312         int err;
 313
 314 restart:
 315         rcu_read_lock();
 316
 317         tbl = rht_dereference_rcu(ht->tbl, ht);
 318
 319         /* All insertions must grab the oldest table containing
 320          * the hashed bucket that is yet to be rehashed.
 321          */
 322         for (;;) {
 323                 hash = rht_head_hashfn(ht, tbl, obj, params);
 324                 lock = rht_bucket_lock(tbl, hash);
 325                 spin_lock_bh(lock);
 326
 327                 if (tbl->rehash <= hash)
 328                         break;
 329
 330                 spin_unlock_bh(lock);
 331                 tbl = rht_dereference_rcu(tbl->future_tbl, ht);
 332         }
 333
 334         new_tbl = rht_dereference_rcu(tbl->future_tbl, ht);
 335         if (unlikely(new_tbl)) {
 336                 tbl = rhashtable_insert_slow(ht, key, obj, new_tbl);
 337                 if (!IS_ERR_OR_NULL(tbl))
 338                         goto slow_path;
 339
 340                 err = PTR_ERR(tbl);
 341                 goto out;
 342         }
 343
 344         err = -E2BIG;
 345         if (unlikely(rht_grow_above_max(ht, tbl)))
 346                 goto out;
 347
 348         if (unlikely(rht_grow_above_100(ht, tbl))) {
 349 slow_path:
 350                 spin_unlock_bh(lock);
 351                 err = rhashtable_insert_rehash(ht, tbl);
 352                 rcu_read_unlock();
 353                 if (err)
 354                         return err;
 355
 356                 goto restart;
 357         }
 358
 359         err = -EEXIST;
 360         elasticity = ht->elasticity;
 361         rht_for_each(head, tbl, hash) {
 362                 if (key &&
 363                     unlikely(!(params.obj_cmpfn ?
 364                                params.obj_cmpfn(&arg, rht_obj(ht, head)) :
 365                                rhashtable_compare(&arg, rht_obj(ht, head)))))
 366                         goto out;
 367                 if (!--elasticity)
 368                         goto slow_path;
 369         }
 370
 371         err = 0;
 372
 373         head = rht_dereference_bucket(tbl->buckets[hash], tbl, hash);
 374
 375         RCU_INIT_POINTER(obj->next, head);
 376
 377         rcu_assign_pointer(tbl->buckets[hash], obj);
 378
 379         atomic_inc(&ht->nelems);
 380         if (rht_grow_above_75(ht, tbl))
 381                 schedule_work(&ht->run_work);
 382
 383 out:
 384         spin_unlock_bh(lock);
 385         rcu_read_unlock();
 386
 387         return err;
 388 }
 389
 390 static inline int rhashtable_lookup_insert_fast(
 391         struct rhashtable *ht, struct rhash_head *obj,
 392         const struct rhashtable_params params)
 393 {
 394         const char *key = rht_obj(ht, obj);
 395
 396         BUG_ON(ht->p.obj_hashfn);
 397
 398         return __rhashtable_insert_fast(ht, key + ht->p.key_offset, obj,
 399                                         params);
 400 }
 401
 402 static inline int __rhashtable_remove_fast(
 403         struct rhashtable *ht, struct bucket_table *tbl,
 404         struct rhash_head *obj, const struct rhashtable_params params)
 405 {
 406         struct rhash_head __rcu **pprev;
 407         struct rhash_head *he;
 408         spinlock_t * lock;
 409         unsigned int hash;
 410         int err = -ENOENT;
 411
 412         hash = rht_head_hashfn(ht, tbl, obj, params);
 413         lock = rht_bucket_lock(tbl, hash);
 414
 415         spin_lock_bh(lock);
 416
 417         pprev = &tbl->buckets[hash];
 418         rht_for_each(he, tbl, hash) {
 419                 if (he != obj) {
 420                         pprev = &he->next;
 421                         continue;
 422                 }
 423
 424                 rcu_assign_pointer(*pprev, obj->next);
 425                 err = 0;
 426                 break;
 427         }
 428
 429         spin_unlock_bh(lock);
 430
 431         return err;
 432 }
 433
 434 static inline int rhashtable_remove_fast(
 435         struct rhashtable *ht, struct rhash_head *obj,
 436         const struct rhashtable_params params)
 437 {
 438         struct bucket_table *tbl;
 439         int err;
 440
 441         rcu_read_lock();
 442
 443         tbl = rht_dereference_rcu(ht->tbl, ht);
 444
 445         /* Because we have already taken (and released) the bucket
 446          * lock in old_tbl, if we find that future_tbl is not yet
 447          * visible then that guarantees the entry to still be in
 448          * the old tbl if it exists.
 449          */
 450         while ((err = __rhashtable_remove_fast(ht, tbl, obj, params)) &&
 451                (tbl = rht_dereference_rcu(tbl->future_tbl, ht)))
 452                 ;
 453
 454         if (err)
 455                 goto out;
 456
 457         atomic_dec(&ht->nelems);
 458         if (unlikely(ht->p.automatic_shrinking &&
 459                      rht_shrink_below_30(ht, tbl)))
 460                 schedule_work(&ht->run_work);
 461
 462 out:
 463         rcu_read_unlock();
 464
 465         return err;
 466 }
 467
 468 #endif /* _LINUX_RHASHTABLE_H */