libcfs/include/libcfs/libcfs_hash.h

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  */
  30 /*
  31  * This file is part of Lustre, http://www.lustre.org/
  32  * Lustre is a trademark of Sun Microsystems, Inc.
  33  *
  34  * libcfs/include/libcfs/libcfs_hash.h
  35  *
  36  * Hashing routines
  37  *
  38  */
  39
  40 #ifndef __LIBCFS_HASH_H__
  41 #define __LIBCFS_HASH_H__
  42 /*
  43  * Knuth recommends primes in approximately golden ratio to the maximum
  44  * integer representable by a machine word for multiplicative hashing.
  45  * Chuck Lever verified the effectiveness of this technique:
  46  * http://www.citi.umich.edu/techreports/reports/citi-tr-00-1.pdf
  47  *
  48  * These primes are chosen to be bit-sparse, that is operations on
  49  * them can use shifts and additions instead of multiplications for
  50  * machines where multiplications are slow.
  51  */
  52 /* 2^31 + 2^29 - 2^25 + 2^22 - 2^19 - 2^16 + 1 */
  53 #define CFS_GOLDEN_RATIO_PRIME_32 0x9e370001UL
  54 /*  2^63 + 2^61 - 2^57 + 2^54 - 2^51 - 2^18 + 1 */
  55 #define CFS_GOLDEN_RATIO_PRIME_64 0x9e37fffffffc0001ULL
  56
  57 /*
  58  * Ideally we would use HAVE_HASH_LONG for this, but on linux we configure
  59  * the linux kernel and user space at the same time, so we need to differentiate
  60  * between them explicitely. If this is not needed on other architectures, then
  61  * we'll need to move the functions to archi specific headers.
  62  */
  63
  64 #if (defined __linux__ && defined __KERNEL__)
  65 #include <linux/hash.h>
  66
  67 #define cfs_hash_long(val, bits)    hash_long(val, bits)
  68 #else
  69 /* Fast hashing routine for a long.
  70    (C) 2002 William Lee Irwin III, IBM */
  71
  72 #if BITS_PER_LONG == 32
  73 /* 2^31 + 2^29 - 2^25 + 2^22 - 2^19 - 2^16 + 1 */
  74 #define CFS_GOLDEN_RATIO_PRIME          CFS_GOLDEN_RATIO_PRIME_32
  75 #elif BITS_PER_LONG == 64
  76 /*  2^63 + 2^61 - 2^57 + 2^54 - 2^51 - 2^18 + 1 */
  77 #define CFS_GOLDEN_RATIO_PRIME          CFS_GOLDEN_RATIO_PRIME_64
  78 #else
  79 #error Define CFS_GOLDEN_RATIO_PRIME for your wordsize.
  80 #endif
  81
  82 static inline unsigned long cfs_hash_long(unsigned long val, unsigned int bits)
  83 {
  84         unsigned long hash = val;
  85
  86 #if BITS_PER_LONG == 64
  87         /*  Sigh, gcc can't optimise this alone like it does for 32 bits. */
  88         unsigned long n = hash;
  89         n <<= 18;
  90         hash -= n;
  91         n <<= 33;
  92         hash -= n;
  93         n <<= 3;
  94         hash += n;
  95         n <<= 3;
  96         hash -= n;
  97         n <<= 4;
  98         hash += n;
  99         n <<= 2;
 100         hash += n;
 101 #else
 102         /* On some cpus multiply is faster, on others gcc will do shifts */
 103         hash *= CFS_GOLDEN_RATIO_PRIME;
 104 #endif
 105
 106         /* High bits are more random, so use them. */
 107         return hash >> (BITS_PER_LONG - bits);
 108 }
 109 #if 0
 110 static inline unsigned long hash_ptr(void *ptr, unsigned int bits)
 111 {
 112         return cfs_hash_long((unsigned long)ptr, bits);
 113 }
 114 #endif
 115
 116 /* !(__linux__ && __KERNEL__) */
 117 #endif
 118
 119 /** disable debug */
 120 #define CFS_HASH_DEBUG_NONE         0
 121 /** record hash depth and output to console when it's too deep,
 122  *  computing overhead is low but consume more memory */
 123 #define CFS_HASH_DEBUG_1            1
 124 /** expensive, check key validation */
 125 #define CFS_HASH_DEBUG_2            2
 126
 127 #define CFS_HASH_DEBUG_LEVEL        CFS_HASH_DEBUG_NONE
 128
 129 struct cfs_hash_ops;
 130 struct cfs_hash_lock_ops;
 131 struct cfs_hash_hlist_ops;
 132
 133 typedef union {
 134         rwlock_t                rw;             /**< rwlock */
 135         spinlock_t              spin;           /**< spinlock */
 136 } cfs_hash_lock_t;
 137
 138 /**
 139  * cfs_hash_bucket is a container of:
 140  * - lock, couter ...
 141  * - array of hash-head starting from hsb_head[0], hash-head can be one of
 142  *   . cfs_hash_head_t
 143  *   . cfs_hash_head_dep_t
 144  *   . cfs_hash_dhead_t
 145  *   . cfs_hash_dhead_dep_t
 146  *   which depends on requirement of user
 147  * - some extra bytes (caller can require it while creating hash)
 148  */
 149 typedef struct cfs_hash_bucket {
 150         cfs_hash_lock_t             hsb_lock;       /**< bucket lock */
 151         __u32                       hsb_count;      /**< current entries */
 152         __u32                       hsb_version;    /**< change version */
 153         unsigned int                hsb_index;      /**< index of bucket */
 154         int                         hsb_depmax;     /**< max depth on bucket */
 155         char                        hsb_head[0];    /**< hash-head array */
 156 } cfs_hash_bucket_t;
 157
 158 /**
 159  * cfs_hash bucket descriptor, it's normally in stack of caller
 160  */
 161 typedef struct cfs_hash_bd {
 162         cfs_hash_bucket_t          *bd_bucket;      /**< address of bucket */
 163         unsigned int                bd_offset;      /**< offset in bucket */
 164 } cfs_hash_bd_t;
 165
 166 #define CFS_HASH_NAME_LEN           16      /**< default name length */
 167 #define CFS_HASH_BIGNAME_LEN        64      /**< bigname for param tree */
 168
 169 #define CFS_HASH_BKT_BITS           3       /**< default bits of bucket */
 170 #define CFS_HASH_BITS_MAX           30      /**< max bits of bucket */
 171 #define CFS_HASH_BITS_MIN           CFS_HASH_BKT_BITS
 172
 173 /**
 174  * common hash attributes.
 175  */
 176 enum cfs_hash_tag {
 177         /**
 178          * don't need any lock, caller will protect operations with it's
 179          * own lock. With this flag:
 180          *  . CFS_HASH_NO_BKTLOCK, CFS_HASH_RW_BKTLOCK, CFS_HASH_SPIN_BKTLOCK
 181          *    will be ignored.
 182          *  . Some functions will be disabled with this flag, i.e:
 183          *    cfs_hash_for_each_empty, cfs_hash_rehash
 184          */
 185         CFS_HASH_NO_LOCK        = 1 << 0,
 186         /** no bucket lock, use one spinlock to protect the whole hash */
 187         CFS_HASH_NO_BKTLOCK     = 1 << 1,
 188         /** rwlock to protect bucket */
 189         CFS_HASH_RW_BKTLOCK     = 1 << 2,
 190         /** spinlcok to protect bucket */
 191         CFS_HASH_SPIN_BKTLOCK   = 1 << 3,
 192         /** always add new item to tail */
 193         CFS_HASH_ADD_TAIL       = 1 << 4,
 194         /** hash-table doesn't have refcount on item */
 195         CFS_HASH_NO_ITEMREF     = 1 << 5,
 196         /** big name for param-tree */
 197         CFS_HASH_BIGNAME        = 1 << 6,
 198         /** track global count */
 199         CFS_HASH_COUNTER        = 1 << 7,
 200         /** rehash item by new key */
 201         CFS_HASH_REHASH_KEY     = 1 << 8,
 202         /** Enable dynamic hash resizing */
 203         CFS_HASH_REHASH         = 1 << 9,
 204         /** can shrink hash-size */
 205         CFS_HASH_SHRINK         = 1 << 10,
 206         /** assert hash is empty on exit */
 207         CFS_HASH_ASSERT_EMPTY   = 1 << 11,
 208         /** record hlist depth */
 209         CFS_HASH_DEPTH          = 1 << 12,
 210         /**
 211          * rehash is always scheduled in a different thread, so current
 212          * change on hash table is non-blocking
 213          */
 214         CFS_HASH_NBLK_CHANGE    = 1 << 13,
 215         /** NB, we typed hs_flags as  __u16, please change it
 216          * if you need to extend >=16 flags */
 217 };
 218
 219 /** most used attributes */
 220 #define CFS_HASH_DEFAULT       (CFS_HASH_RW_BKTLOCK | \
 221                                 CFS_HASH_COUNTER | CFS_HASH_REHASH)
 222
 223 /**
 224  * cfs_hash is a hash-table implementation for general purpose, it can support:
 225  *    . two refcount modes
 226  *      hash-table with & without refcount
 227  *    . four lock modes
 228  *      nolock, one-spinlock, rw-bucket-lock, spin-bucket-lock
 229  *    . general operations
 230  *      lookup, add(add_tail or add_head), delete
 231  *    . rehash
 232  *      grows or shrink
 233  *    . iteration
 234  *      locked iteration and unlocked iteration
 235  *    . bigname
 236  *      support long name hash
 237  *    . debug
 238  *      trace max searching depth
 239  *
 240  * Rehash:
 241  * When the htable grows or shrinks, a separate task (cfs_hash_rehash_worker)
 242  * is spawned to handle the rehash in the background, it's possible that other
 243  * processes can concurrently perform additions, deletions, and lookups
 244  * without being blocked on rehash completion, because rehash will release
 245  * the global wrlock for each bucket.
 246  *
 247  * rehash and iteration can't run at the same time because it's too tricky
 248  * to keep both of them safe and correct.
 249  * As they are relatively rare operations, so:
 250  *   . if iteration is in progress while we try to launch rehash, then
 251  *     it just giveup, iterator will launch rehash at the end.
 252  *   . if rehash is in progress while we try to iterate the hash table,
 253  *     then we just wait (shouldn't be very long time), anyway, nobody
 254  *     should expect iteration of whole hash-table to be non-blocking.
 255  *
 256  * During rehashing, a (key,object) pair may be in one of two buckets,
 257  * depending on whether the worker task has yet to transfer the object
 258  * to its new location in the table. Lookups and deletions need to search both
 259  * locations; additions must take care to only insert into the new bucket.
 260  */
 261
 262 typedef struct cfs_hash {
 263         /** serialize with rehash, or serialize all operations if
 264          * the hash-table has CFS_HASH_NO_BKTLOCK */
 265         cfs_hash_lock_t             hs_lock;
 266         /** hash operations */
 267         struct cfs_hash_ops        *hs_ops;
 268         /** hash lock operations */
 269         struct cfs_hash_lock_ops   *hs_lops;
 270         /** hash list operations */
 271         struct cfs_hash_hlist_ops  *hs_hops;
 272         /** hash buckets-table */
 273         cfs_hash_bucket_t         **hs_buckets;
 274         /** total number of items on this hash-table */
 275         cfs_atomic_t                hs_count;
 276         /** hash flags, see cfs_hash_tag for detail */
 277         __u16                       hs_flags;
 278         /** # of extra-bytes for bucket, for user saving extended attributes */
 279         __u16                       hs_extra_bytes;
 280         /** wants to iterate */
 281         __u8                        hs_iterating;
 282         /** hash-table is dying */
 283         __u8                        hs_exiting;
 284         /** current hash bits */
 285         __u8                        hs_cur_bits;
 286         /** min hash bits */
 287         __u8                        hs_min_bits;
 288         /** max hash bits */
 289         __u8                        hs_max_bits;
 290         /** bits for rehash */
 291         __u8                        hs_rehash_bits;
 292         /** bits for each bucket */
 293         __u8                        hs_bkt_bits;
 294         /** resize min threshold */
 295         __u16                       hs_min_theta;
 296         /** resize max threshold */
 297         __u16                       hs_max_theta;
 298         /** resize count */
 299         __u32                       hs_rehash_count;
 300         /** # of iterators (caller of cfs_hash_for_each_*) */
 301         __u32                       hs_iterators;
 302         /** rehash workitem */
 303         cfs_workitem_t              hs_rehash_wi;
 304         /** refcount on this hash table */
 305         cfs_atomic_t                hs_refcount;
 306         /** rehash buckets-table */
 307         cfs_hash_bucket_t         **hs_rehash_buckets;
 308 #if CFS_HASH_DEBUG_LEVEL >= CFS_HASH_DEBUG_1
 309         /** serialize debug members */
 310         spinlock_t                      hs_dep_lock;
 311         /** max depth */
 312         unsigned int                hs_dep_max;
 313         /** id of the deepest bucket */
 314         unsigned int                hs_dep_bkt;
 315         /** offset in the deepest bucket */
 316         unsigned int                hs_dep_off;
 317         /** bits when we found the max depth */
 318         unsigned int                hs_dep_bits;
 319         /** workitem to output max depth */
 320         cfs_workitem_t              hs_dep_wi;
 321 #endif
 322         /** name of htable */
 323         char                        hs_name[0];
 324 } cfs_hash_t;
 325
 326 typedef struct cfs_hash_lock_ops {
 327         /** lock the hash table */
 328         void    (*hs_lock)(cfs_hash_lock_t *lock, int exclusive);
 329         /** unlock the hash table */
 330         void    (*hs_unlock)(cfs_hash_lock_t *lock, int exclusive);
 331         /** lock the hash bucket */
 332         void    (*hs_bkt_lock)(cfs_hash_lock_t *lock, int exclusive);
 333         /** unlock the hash bucket */
 334         void    (*hs_bkt_unlock)(cfs_hash_lock_t *lock, int exclusive);
 335 } cfs_hash_lock_ops_t;
 336
 337 typedef struct cfs_hash_hlist_ops {
 338         /** return hlist_head of hash-head of @bd */
 339         cfs_hlist_head_t *(*hop_hhead)(cfs_hash_t *hs, cfs_hash_bd_t *bd);
 340         /** return hash-head size */
 341         int (*hop_hhead_size)(cfs_hash_t *hs);
 342         /** add @hnode to hash-head of @bd */
 343         int (*hop_hnode_add)(cfs_hash_t *hs,
 344                              cfs_hash_bd_t *bd, cfs_hlist_node_t *hnode);
 345         /** remove @hnode from hash-head of @bd */
 346         int (*hop_hnode_del)(cfs_hash_t *hs,
 347                              cfs_hash_bd_t *bd, cfs_hlist_node_t *hnode);
 348 } cfs_hash_hlist_ops_t;
 349
 350 typedef struct cfs_hash_ops {
 351         /** return hashed value from @key */
 352         unsigned (*hs_hash)(cfs_hash_t *hs, const void *key, unsigned mask);
 353         /** return key address of @hnode */
 354         void *   (*hs_key)(cfs_hlist_node_t *hnode);
 355         /** copy key from @hnode to @key */
 356         void     (*hs_keycpy)(cfs_hlist_node_t *hnode, void *key);
 357         /**
 358          *  compare @key with key of @hnode
 359          *  returns 1 on a match
 360          */
 361         int      (*hs_keycmp)(const void *key, cfs_hlist_node_t *hnode);
 362         /** return object address of @hnode, i.e: container_of(...hnode) */
 363         void *   (*hs_object)(cfs_hlist_node_t *hnode);
 364         /** get refcount of item, always called with holding bucket-lock */
 365         void     (*hs_get)(cfs_hash_t *hs, cfs_hlist_node_t *hnode);
 366         /** release refcount of item */
 367         void     (*hs_put)(cfs_hash_t *hs, cfs_hlist_node_t *hnode);
 368         /** release refcount of item, always called with holding bucket-lock */
 369         void     (*hs_put_locked)(cfs_hash_t *hs, cfs_hlist_node_t *hnode);
 370         /** it's called before removing of @hnode */
 371         void     (*hs_exit)(cfs_hash_t *hs, cfs_hlist_node_t *hnode);
 372 } cfs_hash_ops_t;
 373
 374 /** total number of buckets in @hs */
 375 #define CFS_HASH_NBKT(hs)       \
 376         (1U << ((hs)->hs_cur_bits - (hs)->hs_bkt_bits))
 377
 378 /** total number of buckets in @hs while rehashing */
 379 #define CFS_HASH_RH_NBKT(hs)    \
 380         (1U << ((hs)->hs_rehash_bits - (hs)->hs_bkt_bits))
 381
 382 /** number of hlist for in bucket */
 383 #define CFS_HASH_BKT_NHLIST(hs) (1U << (hs)->hs_bkt_bits)
 384
 385 /** total number of hlist in @hs */
 386 #define CFS_HASH_NHLIST(hs)     (1U << (hs)->hs_cur_bits)
 387
 388 /** total number of hlist in @hs while rehashing */
 389 #define CFS_HASH_RH_NHLIST(hs)  (1U << (hs)->hs_rehash_bits)
 390
 391 static inline int
 392 cfs_hash_with_no_lock(cfs_hash_t *hs)
 393 {
 394         /* caller will serialize all operations for this hash-table */
 395         return (hs->hs_flags & CFS_HASH_NO_LOCK) != 0;
 396 }
 397
 398 static inline int
 399 cfs_hash_with_no_bktlock(cfs_hash_t *hs)
 400 {
 401         /* no bucket lock, one single lock to protect the hash-table */
 402         return (hs->hs_flags & CFS_HASH_NO_BKTLOCK) != 0;
 403 }
 404
 405 static inline int
 406 cfs_hash_with_rw_bktlock(cfs_hash_t *hs)
 407 {
 408         /* rwlock to protect hash bucket */
 409         return (hs->hs_flags & CFS_HASH_RW_BKTLOCK) != 0;
 410 }
 411
 412 static inline int
 413 cfs_hash_with_spin_bktlock(cfs_hash_t *hs)
 414 {
 415         /* spinlock to protect hash bucket */
 416         return (hs->hs_flags & CFS_HASH_SPIN_BKTLOCK) != 0;
 417 }
 418
 419 static inline int
 420 cfs_hash_with_add_tail(cfs_hash_t *hs)
 421 {
 422         return (hs->hs_flags & CFS_HASH_ADD_TAIL) != 0;
 423 }
 424
 425 static inline int
 426 cfs_hash_with_no_itemref(cfs_hash_t *hs)
 427 {
 428         /* hash-table doesn't keep refcount on item,
 429          * item can't be removed from hash unless it's
 430          * ZERO refcount */
 431         return (hs->hs_flags & CFS_HASH_NO_ITEMREF) != 0;
 432 }
 433
 434 static inline int
 435 cfs_hash_with_bigname(cfs_hash_t *hs)
 436 {
 437         return (hs->hs_flags & CFS_HASH_BIGNAME) != 0;
 438 }
 439
 440 static inline int
 441 cfs_hash_with_counter(cfs_hash_t *hs)
 442 {
 443         return (hs->hs_flags & CFS_HASH_COUNTER) != 0;
 444 }
 445
 446 static inline int
 447 cfs_hash_with_rehash(cfs_hash_t *hs)
 448 {
 449         return (hs->hs_flags & CFS_HASH_REHASH) != 0;
 450 }
 451
 452 static inline int
 453 cfs_hash_with_rehash_key(cfs_hash_t *hs)
 454 {
 455         return (hs->hs_flags & CFS_HASH_REHASH_KEY) != 0;
 456 }
 457
 458 static inline int
 459 cfs_hash_with_shrink(cfs_hash_t *hs)
 460 {
 461         return (hs->hs_flags & CFS_HASH_SHRINK) != 0;
 462 }
 463
 464 static inline int
 465 cfs_hash_with_assert_empty(cfs_hash_t *hs)
 466 {
 467         return (hs->hs_flags & CFS_HASH_ASSERT_EMPTY) != 0;
 468 }
 469
 470 static inline int
 471 cfs_hash_with_depth(cfs_hash_t *hs)
 472 {
 473         return (hs->hs_flags & CFS_HASH_DEPTH) != 0;
 474 }
 475
 476 static inline int
 477 cfs_hash_with_nblk_change(cfs_hash_t *hs)
 478 {
 479         return (hs->hs_flags & CFS_HASH_NBLK_CHANGE) != 0;
 480 }
 481
 482 static inline int
 483 cfs_hash_is_exiting(cfs_hash_t *hs)
 484 {       /* cfs_hash_destroy is called */
 485         return hs->hs_exiting;
 486 }
 487
 488 static inline int
 489 cfs_hash_is_rehashing(cfs_hash_t *hs)
 490 {       /* rehash is launched */
 491         return hs->hs_rehash_bits != 0;
 492 }
 493
 494 static inline int
 495 cfs_hash_is_iterating(cfs_hash_t *hs)
 496 {       /* someone is calling cfs_hash_for_each_* */
 497         return hs->hs_iterating || hs->hs_iterators != 0;
 498 }
 499
 500 static inline int
 501 cfs_hash_bkt_size(cfs_hash_t *hs)
 502 {
 503         return offsetof(cfs_hash_bucket_t, hsb_head[0]) +
 504                hs->hs_hops->hop_hhead_size(hs) * CFS_HASH_BKT_NHLIST(hs) +
 505                hs->hs_extra_bytes;
 506 }
 507
 508 #define CFS_HOP(hs, op)           (hs)->hs_ops->hs_ ## op
 509
 510 static inline unsigned
 511 cfs_hash_id(cfs_hash_t *hs, const void *key, unsigned mask)
 512 {
 513         return CFS_HOP(hs, hash)(hs, key, mask);
 514 }
 515
 516 static inline void *
 517 cfs_hash_key(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
 518 {
 519         return CFS_HOP(hs, key)(hnode);
 520 }
 521
 522 static inline void
 523 cfs_hash_keycpy(cfs_hash_t *hs, cfs_hlist_node_t *hnode, void *key)
 524 {
 525         if (CFS_HOP(hs, keycpy) != NULL)
 526                 CFS_HOP(hs, keycpy)(hnode, key);
 527 }
 528
 529 /**
 530  * Returns 1 on a match,
 531  */
 532 static inline int
 533 cfs_hash_keycmp(cfs_hash_t *hs, const void *key, cfs_hlist_node_t *hnode)
 534 {
 535         return CFS_HOP(hs, keycmp)(key, hnode);
 536 }
 537
 538 static inline void *
 539 cfs_hash_object(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
 540 {
 541         return CFS_HOP(hs, object)(hnode);
 542 }
 543
 544 static inline void
 545 cfs_hash_get(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
 546 {
 547         return CFS_HOP(hs, get)(hs, hnode);
 548 }
 549
 550 static inline void
 551 cfs_hash_put_locked(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
 552 {
 553         LASSERT(CFS_HOP(hs, put_locked) != NULL);
 554
 555         return CFS_HOP(hs, put_locked)(hs, hnode);
 556 }
 557
 558 static inline void
 559 cfs_hash_put(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
 560 {
 561         LASSERT(CFS_HOP(hs, put) != NULL);
 562
 563         return CFS_HOP(hs, put)(hs, hnode);
 564 }
 565
 566 static inline void
 567 cfs_hash_exit(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
 568 {
 569         if (CFS_HOP(hs, exit))
 570                 CFS_HOP(hs, exit)(hs, hnode);
 571 }
 572
 573 static inline void cfs_hash_lock(cfs_hash_t *hs, int excl)
 574 {
 575         hs->hs_lops->hs_lock(&hs->hs_lock, excl);
 576 }
 577
 578 static inline void cfs_hash_unlock(cfs_hash_t *hs, int excl)
 579 {
 580         hs->hs_lops->hs_unlock(&hs->hs_lock, excl);
 581 }
 582
 583 static inline int cfs_hash_dec_and_lock(cfs_hash_t *hs,
 584                                         cfs_atomic_t *condition)
 585 {
 586         LASSERT(cfs_hash_with_no_bktlock(hs));
 587         return cfs_atomic_dec_and_lock(condition, &hs->hs_lock.spin);
 588 }
 589
 590 static inline void cfs_hash_bd_lock(cfs_hash_t *hs,
 591                                     cfs_hash_bd_t *bd, int excl)
 592 {
 593         hs->hs_lops->hs_bkt_lock(&bd->bd_bucket->hsb_lock, excl);
 594 }
 595
 596 static inline void cfs_hash_bd_unlock(cfs_hash_t *hs,
 597                                       cfs_hash_bd_t *bd, int excl)
 598 {
 599         hs->hs_lops->hs_bkt_unlock(&bd->bd_bucket->hsb_lock, excl);
 600 }
 601
 602 /**
 603  * operations on cfs_hash bucket (bd: bucket descriptor),
 604  * they are normally for hash-table without rehash
 605  */
 606 void cfs_hash_bd_get(cfs_hash_t *hs, const void *key, cfs_hash_bd_t *bd);
 607
 608 static inline void cfs_hash_bd_get_and_lock(cfs_hash_t *hs, const void *key,
 609                                             cfs_hash_bd_t *bd, int excl)
 610 {
 611         cfs_hash_bd_get(hs, key, bd);
 612         cfs_hash_bd_lock(hs, bd, excl);
 613 }
 614
 615 static inline unsigned cfs_hash_bd_index_get(cfs_hash_t *hs, cfs_hash_bd_t *bd)
 616 {
 617         return bd->bd_offset | (bd->bd_bucket->hsb_index << hs->hs_bkt_bits);
 618 }
 619
 620 static inline void cfs_hash_bd_index_set(cfs_hash_t *hs,
 621                                          unsigned index, cfs_hash_bd_t *bd)
 622 {
 623         bd->bd_bucket = hs->hs_buckets[index >> hs->hs_bkt_bits];
 624         bd->bd_offset = index & (CFS_HASH_BKT_NHLIST(hs) - 1U);
 625 }
 626
 627 static inline void *
 628 cfs_hash_bd_extra_get(cfs_hash_t *hs, cfs_hash_bd_t *bd)
 629 {
 630         return (void *)bd->bd_bucket +
 631                cfs_hash_bkt_size(hs) - hs->hs_extra_bytes;
 632 }
 633
 634 static inline __u32
 635 cfs_hash_bd_version_get(cfs_hash_bd_t *bd)
 636 {
 637         /* need hold cfs_hash_bd_lock */
 638         return bd->bd_bucket->hsb_version;
 639 }
 640
 641 static inline __u32
 642 cfs_hash_bd_count_get(cfs_hash_bd_t *bd)
 643 {
 644         /* need hold cfs_hash_bd_lock */
 645         return bd->bd_bucket->hsb_count;
 646 }
 647
 648 static inline int
 649 cfs_hash_bd_depmax_get(cfs_hash_bd_t *bd)
 650 {
 651         return bd->bd_bucket->hsb_depmax;
 652 }
 653
 654 static inline int
 655 cfs_hash_bd_compare(cfs_hash_bd_t *bd1, cfs_hash_bd_t *bd2)
 656 {
 657         if (bd1->bd_bucket->hsb_index != bd2->bd_bucket->hsb_index)
 658                 return bd1->bd_bucket->hsb_index - bd2->bd_bucket->hsb_index;
 659
 660         if (bd1->bd_offset != bd2->bd_offset)
 661                 return bd1->bd_offset - bd2->bd_offset;
 662
 663         return 0;
 664 }
 665
 666 void cfs_hash_bd_add_locked(cfs_hash_t *hs, cfs_hash_bd_t *bd,
 667                             cfs_hlist_node_t *hnode);
 668 void cfs_hash_bd_del_locked(cfs_hash_t *hs, cfs_hash_bd_t *bd,
 669                             cfs_hlist_node_t *hnode);
 670 void cfs_hash_bd_move_locked(cfs_hash_t *hs, cfs_hash_bd_t *bd_old,
 671                              cfs_hash_bd_t *bd_new, cfs_hlist_node_t *hnode);
 672
 673 static inline int cfs_hash_bd_dec_and_lock(cfs_hash_t *hs, cfs_hash_bd_t *bd,
 674                                            cfs_atomic_t *condition)
 675 {
 676         LASSERT(cfs_hash_with_spin_bktlock(hs));
 677         return cfs_atomic_dec_and_lock(condition,
 678                                        &bd->bd_bucket->hsb_lock.spin);
 679 }
 680
 681 static inline cfs_hlist_head_t *cfs_hash_bd_hhead(cfs_hash_t *hs,
 682                                                   cfs_hash_bd_t *bd)
 683 {
 684         return hs->hs_hops->hop_hhead(hs, bd);
 685 }
 686
 687 cfs_hlist_node_t *cfs_hash_bd_lookup_locked(cfs_hash_t *hs,
 688                                             cfs_hash_bd_t *bd, const void *key);
 689 cfs_hlist_node_t *cfs_hash_bd_findadd_locked(cfs_hash_t *hs,
 690                                              cfs_hash_bd_t *bd, const void *key,
 691                                              cfs_hlist_node_t *hnode,
 692                                              int insist_add);
 693 cfs_hlist_node_t *cfs_hash_bd_finddel_locked(cfs_hash_t *hs,
 694                                              cfs_hash_bd_t *bd, const void *key,
 695                                              cfs_hlist_node_t *hnode);
 696
 697 /**
 698  * operations on cfs_hash bucket (bd: bucket descriptor),
 699  * they are safe for hash-table with rehash
 700  */
 701 void cfs_hash_dual_bd_get(cfs_hash_t *hs, const void *key, cfs_hash_bd_t *bds);
 702 void cfs_hash_dual_bd_lock(cfs_hash_t *hs, cfs_hash_bd_t *bds, int excl);
 703 void cfs_hash_dual_bd_unlock(cfs_hash_t *hs, cfs_hash_bd_t *bds, int excl);
 704
 705 static inline void cfs_hash_dual_bd_get_and_lock(cfs_hash_t *hs, const void *key,
 706                                                  cfs_hash_bd_t *bds, int excl)
 707 {
 708         cfs_hash_dual_bd_get(hs, key, bds);
 709         cfs_hash_dual_bd_lock(hs, bds, excl);
 710 }
 711
 712 cfs_hlist_node_t *cfs_hash_dual_bd_lookup_locked(cfs_hash_t *hs,
 713                                                  cfs_hash_bd_t *bds,
 714                                                  const void *key);
 715 cfs_hlist_node_t *cfs_hash_dual_bd_findadd_locked(cfs_hash_t *hs,
 716                                                   cfs_hash_bd_t *bds,
 717                                                   const void *key,
 718                                                   cfs_hlist_node_t *hnode,
 719                                                   int insist_add);
 720 cfs_hlist_node_t *cfs_hash_dual_bd_finddel_locked(cfs_hash_t *hs,
 721                                                   cfs_hash_bd_t *bds,
 722                                                   const void *key,
 723                                                   cfs_hlist_node_t *hnode);
 724
 725 /* Hash init/cleanup functions */
 726 cfs_hash_t *cfs_hash_create(char *name, unsigned cur_bits, unsigned max_bits,
 727                             unsigned bkt_bits, unsigned extra_bytes,
 728                             unsigned min_theta, unsigned max_theta,
 729                             cfs_hash_ops_t *ops, unsigned flags);
 730
 731 cfs_hash_t *cfs_hash_getref(cfs_hash_t *hs);
 732 void cfs_hash_putref(cfs_hash_t *hs);
 733
 734 /* Hash addition functions */
 735 void cfs_hash_add(cfs_hash_t *hs, const void *key,
 736                   cfs_hlist_node_t *hnode);
 737 int cfs_hash_add_unique(cfs_hash_t *hs, const void *key,
 738                         cfs_hlist_node_t *hnode);
 739 void *cfs_hash_findadd_unique(cfs_hash_t *hs, const void *key,
 740                               cfs_hlist_node_t *hnode);
 741
 742 /* Hash deletion functions */
 743 void *cfs_hash_del(cfs_hash_t *hs, const void *key, cfs_hlist_node_t *hnode);
 744 void *cfs_hash_del_key(cfs_hash_t *hs, const void *key);
 745
 746 /* Hash lookup/for_each functions */
 747 #define CFS_HASH_LOOP_HOG       1024
 748
 749 typedef int (*cfs_hash_for_each_cb_t)(cfs_hash_t *hs, cfs_hash_bd_t *bd,
 750                                       cfs_hlist_node_t *node, void *data);
 751 void *cfs_hash_lookup(cfs_hash_t *hs, const void *key);
 752 void cfs_hash_for_each(cfs_hash_t *hs, cfs_hash_for_each_cb_t, void *data);
 753 void cfs_hash_for_each_safe(cfs_hash_t *hs, cfs_hash_for_each_cb_t, void *data);
 754 int  cfs_hash_for_each_nolock(cfs_hash_t *hs,
 755                               cfs_hash_for_each_cb_t, void *data);
 756 int  cfs_hash_for_each_empty(cfs_hash_t *hs,
 757                              cfs_hash_for_each_cb_t, void *data);
 758 void cfs_hash_for_each_key(cfs_hash_t *hs, const void *key,
 759                            cfs_hash_for_each_cb_t, void *data);
 760 typedef int (*cfs_hash_cond_opt_cb_t)(void *obj, void *data);
 761 void cfs_hash_cond_del(cfs_hash_t *hs, cfs_hash_cond_opt_cb_t, void *data);
 762
 763 void cfs_hash_hlist_for_each(cfs_hash_t *hs, unsigned hindex,
 764                              cfs_hash_for_each_cb_t, void *data);
 765 int  cfs_hash_is_empty(cfs_hash_t *hs);
 766 __u64 cfs_hash_size_get(cfs_hash_t *hs);
 767
 768 /*
 769  * Rehash - Theta is calculated to be the average chained
 770  * hash depth assuming a perfectly uniform hash funcion.
 771  */
 772 void cfs_hash_rehash_cancel_locked(cfs_hash_t *hs);
 773 void cfs_hash_rehash_cancel(cfs_hash_t *hs);
 774 int  cfs_hash_rehash(cfs_hash_t *hs, int do_rehash);
 775 void cfs_hash_rehash_key(cfs_hash_t *hs, const void *old_key,
 776                          void *new_key, cfs_hlist_node_t *hnode);
 777
 778 #if CFS_HASH_DEBUG_LEVEL > CFS_HASH_DEBUG_1
 779 /* Validate hnode references the correct key */
 780 static inline void
 781 cfs_hash_key_validate(cfs_hash_t *hs, const void *key,
 782                       cfs_hlist_node_t *hnode)
 783 {
 784         LASSERT(cfs_hash_keycmp(hs, key, hnode));
 785 }
 786
 787 /* Validate hnode is in the correct bucket */
 788 static inline void
 789 cfs_hash_bucket_validate(cfs_hash_t *hs, cfs_hash_bd_t *bd,
 790                          cfs_hlist_node_t *hnode)
 791 {
 792         cfs_hash_bd_t   bds[2];
 793
 794         cfs_hash_dual_bd_get(hs, cfs_hash_key(hs, hnode), bds);
 795         LASSERT(bds[0].bd_bucket == bd->bd_bucket ||
 796                 bds[1].bd_bucket == bd->bd_bucket);
 797 }
 798
 799 #else /* CFS_HASH_DEBUG_LEVEL > CFS_HASH_DEBUG_1 */
 800
 801 static inline void
 802 cfs_hash_key_validate(cfs_hash_t *hs, const void *key,
 803                       cfs_hlist_node_t *hnode) {}
 804
 805 static inline void
 806 cfs_hash_bucket_validate(cfs_hash_t *hs, cfs_hash_bd_t *bd,
 807                          cfs_hlist_node_t *hnode) {}
 808
 809 #endif /* CFS_HASH_DEBUG_LEVEL */
 810
 811 #define CFS_HASH_THETA_BITS  10
 812 #define CFS_HASH_MIN_THETA  (1U << (CFS_HASH_THETA_BITS - 1))
 813 #define CFS_HASH_MAX_THETA  (1U << (CFS_HASH_THETA_BITS + 1))
 814
 815 /* Return integer component of theta */
 816 static inline int __cfs_hash_theta_int(int theta)
 817 {
 818         return (theta >> CFS_HASH_THETA_BITS);
 819 }
 820
 821 /* Return a fractional value between 0 and 999 */
 822 static inline int __cfs_hash_theta_frac(int theta)
 823 {
 824         return ((theta * 1000) >> CFS_HASH_THETA_BITS) -
 825                (__cfs_hash_theta_int(theta) * 1000);
 826 }
 827
 828 static inline int __cfs_hash_theta(cfs_hash_t *hs)
 829 {
 830         return (cfs_atomic_read(&hs->hs_count) <<
 831                 CFS_HASH_THETA_BITS) >> hs->hs_cur_bits;
 832 }
 833
 834 static inline void __cfs_hash_set_theta(cfs_hash_t *hs, int min, int max)
 835 {
 836         LASSERT(min < max);
 837         hs->hs_min_theta = (__u16)min;
 838         hs->hs_max_theta = (__u16)max;
 839 }
 840
 841 /* Generic debug formatting routines mainly for proc handler */
 842 int cfs_hash_debug_header(char *str, int size);
 843 int cfs_hash_debug_str(cfs_hash_t *hs, char *str, int size);
 844
 845 /*
 846  * Generic djb2 hash algorithm for character arrays.
 847  */
 848 static inline unsigned
 849 cfs_hash_djb2_hash(const void *key, size_t size, unsigned mask)
 850 {
 851         unsigned i, hash = 5381;
 852
 853         LASSERT(key != NULL);
 854
 855         for (i = 0; i < size; i++)
 856                 hash = hash * 33 + ((char *)key)[i];
 857
 858         return (hash & mask);
 859 }
 860
 861 /*
 862  * Generic u32 hash algorithm.
 863  */
 864 static inline unsigned
 865 cfs_hash_u32_hash(const __u32 key, unsigned mask)
 866 {
 867         return ((key * CFS_GOLDEN_RATIO_PRIME_32) & mask);
 868 }
 869
 870 /*
 871  * Generic u64 hash algorithm.
 872  */
 873 static inline unsigned
 874 cfs_hash_u64_hash(const __u64 key, unsigned mask)
 875 {
 876         return ((unsigned)(key * CFS_GOLDEN_RATIO_PRIME_64) & mask);
 877 }
 878
 879 /** iterate over all buckets in @bds (array of cfs_hash_bd_t) */
 880 #define cfs_hash_for_each_bd(bds, n, i) \
 881         for (i = 0; i < n && (bds)[i].bd_bucket != NULL; i++)
 882
 883 /** iterate over all buckets of @hs */
 884 #define cfs_hash_for_each_bucket(hs, bd, pos)                   \
 885         for (pos = 0;                                           \
 886              pos < CFS_HASH_NBKT(hs) &&                         \
 887              ((bd)->bd_bucket = (hs)->hs_buckets[pos]) != NULL; pos++)
 888
 889 /** iterate over all hlist of bucket @bd */
 890 #define cfs_hash_bd_for_each_hlist(hs, bd, hlist)               \
 891         for ((bd)->bd_offset = 0;                               \
 892              (bd)->bd_offset < CFS_HASH_BKT_NHLIST(hs) &&       \
 893              (hlist = cfs_hash_bd_hhead(hs, bd)) != NULL;       \
 894              (bd)->bd_offset++)
 895
 896 /* !__LIBCFS__HASH_H__ */
 897 #endif