libcfs/include/libcfs/libcfs_hash.h

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