-/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
- * vim:expandtab:shiftwidth=8:tabstop=8:
- *
+/*
* GPL HEADER START
*
* DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
* GPL HEADER END
*/
/*
- * Copyright 2008 Sun Microsystems, Inc. All rights reserved
+ * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
* Use is subject to license terms.
+ *
+ * Copyright (c) 2012, 2013, Intel Corporation.
*/
/*
* This file is part of Lustre, http://www.lustre.org/
* we'll need to move the functions to archi specific headers.
*/
-#if (defined __linux__ && defined __KERNEL__)
-#include <linux/hash.h>
-#else
+#ifdef __KERNEL__
+# include <linux/hash.h>
+#else /* __KERNEL__ */
/* Fast hashing routine for a long.
(C) 2002 William Lee Irwin III, IBM */
-#if BITS_PER_LONG == 32
+# if BITS_PER_LONG == 32
/* 2^31 + 2^29 - 2^25 + 2^22 - 2^19 - 2^16 + 1 */
-#define CFS_GOLDEN_RATIO_PRIME CFS_GOLDEN_RATIO_PRIME_32
-#elif BITS_PER_LONG == 64
+# define CFS_GOLDEN_RATIO_PRIME CFS_GOLDEN_RATIO_PRIME_32
+# elif BITS_PER_LONG == 64
/* 2^63 + 2^61 - 2^57 + 2^54 - 2^51 - 2^18 + 1 */
-#define CFS_GOLDEN_RATIO_PRIME CFS_GOLDEN_RATIO_PRIME_64
-#else
-#error Define CFS_GOLDEN_RATIO_PRIME for your wordsize.
-#endif
+# define CFS_GOLDEN_RATIO_PRIME CFS_GOLDEN_RATIO_PRIME_64
+# else
+# error Define CFS_GOLDEN_RATIO_PRIME for your wordsize.
+# endif /* BITS_PER_LONG == 64 */
static inline unsigned long hash_long(unsigned long val, unsigned int bits)
{
unsigned long hash = val;
-#if BITS_PER_LONG == 64
+# if BITS_PER_LONG == 64
/* Sigh, gcc can't optimise this alone like it does for 32 bits. */
unsigned long n = hash;
n <<= 18;
hash += n;
n <<= 2;
hash += n;
-#else
+# else /* BITS_PER_LONG == 64 */
/* On some cpus multiply is faster, on others gcc will do shifts */
hash *= CFS_GOLDEN_RATIO_PRIME;
-#endif
+# endif /* BITS_PER_LONG != 64 */
/* High bits are more random, so use them. */
return hash >> (BITS_PER_LONG - bits);
}
-#if 0
-static inline unsigned long hash_ptr(void *ptr, unsigned int bits)
-{
- return hash_long((unsigned long)ptr, bits);
-}
-#endif
+#endif /* !__KERNEL__ */
-/* !(__linux__ && __KERNEL__) */
-#endif
+/** disable debug */
+#define CFS_HASH_DEBUG_NONE 0
+/** record hash depth and output to console when it's too deep,
+ * computing overhead is low but consume more memory */
+#define CFS_HASH_DEBUG_1 1
+/** expensive, check key validation */
+#define CFS_HASH_DEBUG_2 2
-struct cfs_hash_ops;
+#define CFS_HASH_DEBUG_LEVEL CFS_HASH_DEBUG_NONE
+struct cfs_hash_ops;
+struct cfs_hash_lock_ops;
+struct cfs_hash_hlist_ops;
+
+typedef union {
+ rwlock_t rw; /**< rwlock */
+ spinlock_t spin; /**< spinlock */
+} cfs_hash_lock_t;
+
+/**
+ * cfs_hash_bucket is a container of:
+ * - lock, couter ...
+ * - array of hash-head starting from hsb_head[0], hash-head can be one of
+ * . cfs_hash_head_t
+ * . cfs_hash_head_dep_t
+ * . cfs_hash_dhead_t
+ * . cfs_hash_dhead_dep_t
+ * which depends on requirement of user
+ * - some extra bytes (caller can require it while creating hash)
+ */
typedef struct cfs_hash_bucket {
- struct hlist_head hsb_head; /* entries list */
- atomic_t hsb_count; /* current entries */
- rwlock_t hsb_rwlock; /* cfs_hash_bucket */
+ cfs_hash_lock_t hsb_lock; /**< bucket lock */
+ __u32 hsb_count; /**< current entries */
+ __u32 hsb_version; /**< change version */
+ unsigned int hsb_index; /**< index of bucket */
+ int hsb_depmax; /**< max depth on bucket */
+ long hsb_head[0]; /**< hash-head array */
} cfs_hash_bucket_t;
-#define CFS_MAX_HASH_NAME 16
+/**
+ * cfs_hash bucket descriptor, it's normally in stack of caller
+ */
+typedef struct cfs_hash_bd {
+ cfs_hash_bucket_t *bd_bucket; /**< address of bucket */
+ unsigned int bd_offset; /**< offset in bucket */
+} cfs_hash_bd_t;
+
+#define CFS_HASH_NAME_LEN 16 /**< default name length */
+#define CFS_HASH_BIGNAME_LEN 64 /**< bigname for param tree */
+
+#define CFS_HASH_BKT_BITS 3 /**< default bits of bucket */
+#define CFS_HASH_BITS_MAX 30 /**< max bits of bucket */
+#define CFS_HASH_BITS_MIN CFS_HASH_BKT_BITS
+
+/**
+ * common hash attributes.
+ */
+enum cfs_hash_tag {
+ /**
+ * don't need any lock, caller will protect operations with it's
+ * own lock. With this flag:
+ * . CFS_HASH_NO_BKTLOCK, CFS_HASH_RW_BKTLOCK, CFS_HASH_SPIN_BKTLOCK
+ * will be ignored.
+ * . Some functions will be disabled with this flag, i.e:
+ * cfs_hash_for_each_empty, cfs_hash_rehash
+ */
+ CFS_HASH_NO_LOCK = 1 << 0,
+ /** no bucket lock, use one spinlock to protect the whole hash */
+ CFS_HASH_NO_BKTLOCK = 1 << 1,
+ /** rwlock to protect bucket */
+ CFS_HASH_RW_BKTLOCK = 1 << 2,
+ /** spinlcok to protect bucket */
+ CFS_HASH_SPIN_BKTLOCK = 1 << 3,
+ /** always add new item to tail */
+ CFS_HASH_ADD_TAIL = 1 << 4,
+ /** hash-table doesn't have refcount on item */
+ CFS_HASH_NO_ITEMREF = 1 << 5,
+ /** big name for param-tree */
+ CFS_HASH_BIGNAME = 1 << 6,
+ /** track global count */
+ CFS_HASH_COUNTER = 1 << 7,
+ /** rehash item by new key */
+ CFS_HASH_REHASH_KEY = 1 << 8,
+ /** Enable dynamic hash resizing */
+ CFS_HASH_REHASH = 1 << 9,
+ /** can shrink hash-size */
+ CFS_HASH_SHRINK = 1 << 10,
+ /** assert hash is empty on exit */
+ CFS_HASH_ASSERT_EMPTY = 1 << 11,
+ /** record hlist depth */
+ CFS_HASH_DEPTH = 1 << 12,
+ /**
+ * rehash is always scheduled in a different thread, so current
+ * change on hash table is non-blocking
+ */
+ CFS_HASH_NBLK_CHANGE = 1 << 13,
+ /** NB, we typed hs_flags as __u16, please change it
+ * if you need to extend >=16 flags */
+};
+
+/** most used attributes */
+#define CFS_HASH_DEFAULT (CFS_HASH_RW_BKTLOCK | \
+ CFS_HASH_COUNTER | CFS_HASH_REHASH)
+
+/**
+ * cfs_hash is a hash-table implementation for general purpose, it can support:
+ * . two refcount modes
+ * hash-table with & without refcount
+ * . four lock modes
+ * nolock, one-spinlock, rw-bucket-lock, spin-bucket-lock
+ * . general operations
+ * lookup, add(add_tail or add_head), delete
+ * . rehash
+ * grows or shrink
+ * . iteration
+ * locked iteration and unlocked iteration
+ * . bigname
+ * support long name hash
+ * . debug
+ * trace max searching depth
+ *
+ * Rehash:
+ * When the htable grows or shrinks, a separate task (cfs_hash_rehash_worker)
+ * is spawned to handle the rehash in the background, it's possible that other
+ * processes can concurrently perform additions, deletions, and lookups
+ * without being blocked on rehash completion, because rehash will release
+ * the global wrlock for each bucket.
+ *
+ * rehash and iteration can't run at the same time because it's too tricky
+ * to keep both of them safe and correct.
+ * As they are relatively rare operations, so:
+ * . if iteration is in progress while we try to launch rehash, then
+ * it just giveup, iterator will launch rehash at the end.
+ * . if rehash is in progress while we try to iterate the hash table,
+ * then we just wait (shouldn't be very long time), anyway, nobody
+ * should expect iteration of whole hash-table to be non-blocking.
+ *
+ * During rehashing, a (key,object) pair may be in one of two buckets,
+ * depending on whether the worker task has yet to transfer the object
+ * to its new location in the table. Lookups and deletions need to search both
+ * locations; additions must take care to only insert into the new bucket.
+ */
typedef struct cfs_hash {
- int hs_cur_bits; /* current hash bits */
- int hs_cur_mask; /* current hash mask */
- int hs_min_bits; /* min hash bits */
- int hs_max_bits; /* max hash bits */
- int hs_min_theta; /* resize min threshold */
- int hs_max_theta; /* resize max threshold */
- int hs_flags; /* hash flags */
- atomic_t hs_count; /* current entries */
- atomic_t hs_rehash_count;/* resize count */
- struct cfs_hash_bucket **hs_buckets; /* hash buckets */
- struct cfs_hash_ops *hs_ops; /* hash operations */
- rwlock_t hs_rwlock; /* cfs_hash */
- char hs_name[CFS_MAX_HASH_NAME];
+ /** serialize with rehash, or serialize all operations if
+ * the hash-table has CFS_HASH_NO_BKTLOCK */
+ cfs_hash_lock_t hs_lock;
+ /** hash operations */
+ struct cfs_hash_ops *hs_ops;
+ /** hash lock operations */
+ struct cfs_hash_lock_ops *hs_lops;
+ /** hash list operations */
+ struct cfs_hash_hlist_ops *hs_hops;
+ /** hash buckets-table */
+ cfs_hash_bucket_t **hs_buckets;
+ /** total number of items on this hash-table */
+ atomic_t hs_count;
+ /** hash flags, see cfs_hash_tag for detail */
+ __u16 hs_flags;
+ /** # of extra-bytes for bucket, for user saving extended attributes */
+ __u16 hs_extra_bytes;
+ /** wants to iterate */
+ __u8 hs_iterating;
+ /** hash-table is dying */
+ __u8 hs_exiting;
+ /** current hash bits */
+ __u8 hs_cur_bits;
+ /** min hash bits */
+ __u8 hs_min_bits;
+ /** max hash bits */
+ __u8 hs_max_bits;
+ /** bits for rehash */
+ __u8 hs_rehash_bits;
+ /** bits for each bucket */
+ __u8 hs_bkt_bits;
+ /** resize min threshold */
+ __u16 hs_min_theta;
+ /** resize max threshold */
+ __u16 hs_max_theta;
+ /** resize count */
+ __u32 hs_rehash_count;
+ /** # of iterators (caller of cfs_hash_for_each_*) */
+ __u32 hs_iterators;
+ /** rehash workitem */
+ cfs_workitem_t hs_rehash_wi;
+ /** refcount on this hash table */
+ atomic_t hs_refcount;
+ /** rehash buckets-table */
+ cfs_hash_bucket_t **hs_rehash_buckets;
+#if CFS_HASH_DEBUG_LEVEL >= CFS_HASH_DEBUG_1
+ /** serialize debug members */
+ spinlock_t hs_dep_lock;
+ /** max depth */
+ unsigned int hs_dep_max;
+ /** id of the deepest bucket */
+ unsigned int hs_dep_bkt;
+ /** offset in the deepest bucket */
+ unsigned int hs_dep_off;
+ /** bits when we found the max depth */
+ unsigned int hs_dep_bits;
+ /** workitem to output max depth */
+ cfs_workitem_t hs_dep_wi;
+#endif
+ /** name of htable */
+ char hs_name[0];
} cfs_hash_t;
+typedef struct cfs_hash_lock_ops {
+ /** lock the hash table */
+ void (*hs_lock)(cfs_hash_lock_t *lock, int exclusive);
+ /** unlock the hash table */
+ void (*hs_unlock)(cfs_hash_lock_t *lock, int exclusive);
+ /** lock the hash bucket */
+ void (*hs_bkt_lock)(cfs_hash_lock_t *lock, int exclusive);
+ /** unlock the hash bucket */
+ void (*hs_bkt_unlock)(cfs_hash_lock_t *lock, int exclusive);
+} cfs_hash_lock_ops_t;
+
+typedef struct cfs_hash_hlist_ops {
+ /** return hlist_head of hash-head of @bd */
+ struct hlist_head *(*hop_hhead)(cfs_hash_t *hs, cfs_hash_bd_t *bd);
+ /** return hash-head size */
+ int (*hop_hhead_size)(cfs_hash_t *hs);
+ /** add @hnode to hash-head of @bd */
+ int (*hop_hnode_add)(cfs_hash_t *hs, cfs_hash_bd_t *bd,
+ struct hlist_node *hnode);
+ /** remove @hnode from hash-head of @bd */
+ int (*hop_hnode_del)(cfs_hash_t *hs, cfs_hash_bd_t *bd,
+ struct hlist_node *hnode);
+} cfs_hash_hlist_ops_t;
+
typedef struct cfs_hash_ops {
- unsigned (*hs_hash)(cfs_hash_t *hs, void *key, unsigned mask);
- void * (*hs_key)(struct hlist_node *hnode);
- int (*hs_compare)(void *key, struct hlist_node *hnode);
- void * (*hs_get)(struct hlist_node *hnode);
- void * (*hs_put)(struct hlist_node *hnode);
- void (*hs_exit)(struct hlist_node *hnode);
+ /** return hashed value from @key */
+ unsigned (*hs_hash)(cfs_hash_t *hs, const void *key, unsigned mask);
+ /** return key address of @hnode */
+ void * (*hs_key)(struct hlist_node *hnode);
+ /** copy key from @hnode to @key */
+ void (*hs_keycpy)(struct hlist_node *hnode, void *key);
+ /**
+ * compare @key with key of @hnode
+ * returns 1 on a match
+ */
+ int (*hs_keycmp)(const void *key, struct hlist_node *hnode);
+ /** return object address of @hnode, i.e: container_of(...hnode) */
+ void * (*hs_object)(struct hlist_node *hnode);
+ /** get refcount of item, always called with holding bucket-lock */
+ void (*hs_get)(cfs_hash_t *hs, struct hlist_node *hnode);
+ /** release refcount of item */
+ void (*hs_put)(cfs_hash_t *hs, struct hlist_node *hnode);
+ /** release refcount of item, always called with holding bucket-lock */
+ void (*hs_put_locked)(cfs_hash_t *hs, struct hlist_node *hnode);
+ /** it's called before removing of @hnode */
+ void (*hs_exit)(cfs_hash_t *hs, struct hlist_node *hnode);
} cfs_hash_ops_t;
-#define CFS_HASH_DEBUG 0x0001 /* Enable expensive debug checks */
-#define CFS_HASH_REHASH 0x0002 /* Enable dynamic hash resizing */
+/** total number of buckets in @hs */
+#define CFS_HASH_NBKT(hs) \
+ (1U << ((hs)->hs_cur_bits - (hs)->hs_bkt_bits))
-#define CFS_HO(hs) (hs)->hs_ops
-#define CFS_HOP(hs, op) (hs)->hs_ops->hs_ ## op
+/** total number of buckets in @hs while rehashing */
+#define CFS_HASH_RH_NBKT(hs) \
+ (1U << ((hs)->hs_rehash_bits - (hs)->hs_bkt_bits))
-static inline unsigned
-cfs_hash_id(cfs_hash_t *hs, void *key, unsigned mask)
+/** number of hlist for in bucket */
+#define CFS_HASH_BKT_NHLIST(hs) (1U << (hs)->hs_bkt_bits)
+
+/** total number of hlist in @hs */
+#define CFS_HASH_NHLIST(hs) (1U << (hs)->hs_cur_bits)
+
+/** total number of hlist in @hs while rehashing */
+#define CFS_HASH_RH_NHLIST(hs) (1U << (hs)->hs_rehash_bits)
+
+static inline int
+cfs_hash_with_no_lock(cfs_hash_t *hs)
{
- LASSERT(hs);
- LASSERT(CFS_HO(hs));
- LASSERT(CFS_HOP(hs, hash));
+ /* caller will serialize all operations for this hash-table */
+ return (hs->hs_flags & CFS_HASH_NO_LOCK) != 0;
+}
- return CFS_HOP(hs, hash)(hs, key, mask);
+static inline int
+cfs_hash_with_no_bktlock(cfs_hash_t *hs)
+{
+ /* no bucket lock, one single lock to protect the hash-table */
+ return (hs->hs_flags & CFS_HASH_NO_BKTLOCK) != 0;
}
-static inline void *
-cfs_hash_key(cfs_hash_t *hs, struct hlist_node *hnode)
+static inline int
+cfs_hash_with_rw_bktlock(cfs_hash_t *hs)
{
- LASSERT(hs);
- LASSERT(hnode);
- LASSERT(CFS_HO(hs));
+ /* rwlock to protect hash bucket */
+ return (hs->hs_flags & CFS_HASH_RW_BKTLOCK) != 0;
+}
- if (CFS_HOP(hs, key))
- return CFS_HOP(hs, key)(hnode);
+static inline int
+cfs_hash_with_spin_bktlock(cfs_hash_t *hs)
+{
+ /* spinlock to protect hash bucket */
+ return (hs->hs_flags & CFS_HASH_SPIN_BKTLOCK) != 0;
+}
- return NULL;
+static inline int
+cfs_hash_with_add_tail(cfs_hash_t *hs)
+{
+ return (hs->hs_flags & CFS_HASH_ADD_TAIL) != 0;
}
-/* Returns 1 on a match,
- * XXX: This would be better if it returned, -1, 0, or 1 for
- * <, =, > respectivly. It could then be used to implement
- * a CFS_HASH_SORT feature flags which could keep each hash
- * bucket in order. This would increase insertion times
- * but could reduce lookup times for deep chains. Ideally,
- * the rehash should keep chain depth short but if that
- * ends up not being the case this would be a nice feature.
- */
static inline int
-cfs_hash_compare(cfs_hash_t *hs, void *key, struct hlist_node *hnode)
+cfs_hash_with_no_itemref(cfs_hash_t *hs)
{
- LASSERT(hs);
- LASSERT(hnode);
- LASSERT(CFS_HO(hs));
+ /* hash-table doesn't keep refcount on item,
+ * item can't be removed from hash unless it's
+ * ZERO refcount */
+ return (hs->hs_flags & CFS_HASH_NO_ITEMREF) != 0;
+}
- if (CFS_HOP(hs, compare))
- return CFS_HOP(hs, compare)(key, hnode);
+static inline int
+cfs_hash_with_bigname(cfs_hash_t *hs)
+{
+ return (hs->hs_flags & CFS_HASH_BIGNAME) != 0;
+}
- return -EOPNOTSUPP;
+static inline int
+cfs_hash_with_counter(cfs_hash_t *hs)
+{
+ return (hs->hs_flags & CFS_HASH_COUNTER) != 0;
}
-static inline void *
-cfs_hash_get(cfs_hash_t *hs, struct hlist_node *hnode)
+static inline int
+cfs_hash_with_rehash(cfs_hash_t *hs)
+{
+ return (hs->hs_flags & CFS_HASH_REHASH) != 0;
+}
+
+static inline int
+cfs_hash_with_rehash_key(cfs_hash_t *hs)
+{
+ return (hs->hs_flags & CFS_HASH_REHASH_KEY) != 0;
+}
+
+static inline int
+cfs_hash_with_shrink(cfs_hash_t *hs)
{
- LASSERT(hs);
- LASSERT(hnode);
- LASSERT(CFS_HO(hs));
+ return (hs->hs_flags & CFS_HASH_SHRINK) != 0;
+}
- if (CFS_HOP(hs, get))
- return CFS_HOP(hs, get)(hnode);
+static inline int
+cfs_hash_with_assert_empty(cfs_hash_t *hs)
+{
+ return (hs->hs_flags & CFS_HASH_ASSERT_EMPTY) != 0;
+}
- return NULL;
+static inline int
+cfs_hash_with_depth(cfs_hash_t *hs)
+{
+ return (hs->hs_flags & CFS_HASH_DEPTH) != 0;
+}
+
+static inline int
+cfs_hash_with_nblk_change(cfs_hash_t *hs)
+{
+ return (hs->hs_flags & CFS_HASH_NBLK_CHANGE) != 0;
+}
+
+static inline int
+cfs_hash_is_exiting(cfs_hash_t *hs)
+{ /* cfs_hash_destroy is called */
+ return hs->hs_exiting;
+}
+
+static inline int
+cfs_hash_is_rehashing(cfs_hash_t *hs)
+{ /* rehash is launched */
+ return hs->hs_rehash_bits != 0;
+}
+
+static inline int
+cfs_hash_is_iterating(cfs_hash_t *hs)
+{ /* someone is calling cfs_hash_for_each_* */
+ return hs->hs_iterating || hs->hs_iterators != 0;
+}
+
+static inline int
+cfs_hash_bkt_size(cfs_hash_t *hs)
+{
+ return offsetof(cfs_hash_bucket_t, hsb_head[0]) +
+ hs->hs_hops->hop_hhead_size(hs) * CFS_HASH_BKT_NHLIST(hs) +
+ hs->hs_extra_bytes;
+}
+
+#define CFS_HOP(hs, op) (hs)->hs_ops->hs_ ## op
+
+static inline unsigned
+cfs_hash_id(cfs_hash_t *hs, const void *key, unsigned mask)
+{
+ return CFS_HOP(hs, hash)(hs, key, mask);
}
static inline void *
-cfs_hash_put(cfs_hash_t *hs, struct hlist_node *hnode)
+cfs_hash_key(cfs_hash_t *hs, struct hlist_node *hnode)
+{
+ return CFS_HOP(hs, key)(hnode);
+}
+
+static inline void
+cfs_hash_keycpy(cfs_hash_t *hs, struct hlist_node *hnode, void *key)
{
- LASSERT(hs);
- LASSERT(hnode);
- LASSERT(CFS_HO(hs));
+ if (CFS_HOP(hs, keycpy) != NULL)
+ CFS_HOP(hs, keycpy)(hnode, key);
+}
+
+/**
+ * Returns 1 on a match,
+ */
+static inline int
+cfs_hash_keycmp(cfs_hash_t *hs, const void *key, struct hlist_node *hnode)
+{
+ return CFS_HOP(hs, keycmp)(key, hnode);
+}
- if (CFS_HOP(hs, put))
- return CFS_HOP(hs, put)(hnode);
+static inline void *
+cfs_hash_object(cfs_hash_t *hs, struct hlist_node *hnode)
+{
+ return CFS_HOP(hs, object)(hnode);
+}
- return NULL;
+static inline void
+cfs_hash_get(cfs_hash_t *hs, struct hlist_node *hnode)
+{
+ return CFS_HOP(hs, get)(hs, hnode);
}
static inline void
-cfs_hash_exit(cfs_hash_t *hs, struct hlist_node *hnode)
+cfs_hash_put_locked(cfs_hash_t *hs, struct hlist_node *hnode)
{
- LASSERT(hs);
- LASSERT(hnode);
- LASSERT(CFS_HO(hs));
+ LASSERT(CFS_HOP(hs, put_locked) != NULL);
- if (CFS_HOP(hs, exit))
- return CFS_HOP(hs, exit)(hnode);
+ return CFS_HOP(hs, put_locked)(hs, hnode);
}
-/* Validate hnode references the correct key */
static inline void
-__cfs_hash_key_validate(cfs_hash_t *hs, void *key,
- struct hlist_node *hnode)
+cfs_hash_put(cfs_hash_t *hs, struct hlist_node *hnode)
{
- if (unlikely(hs->hs_flags & CFS_HASH_DEBUG))
- LASSERT(cfs_hash_compare(hs, key, hnode) > 0);
+ LASSERT(CFS_HOP(hs, put) != NULL);
+
+ return CFS_HOP(hs, put)(hs, hnode);
}
-/* Validate hnode is in the correct bucket */
static inline void
-__cfs_hash_bucket_validate(cfs_hash_t *hs, cfs_hash_bucket_t *hsb,
- struct hlist_node *hnode)
+cfs_hash_exit(cfs_hash_t *hs, struct hlist_node *hnode)
{
- unsigned i;
+ if (CFS_HOP(hs, exit))
+ CFS_HOP(hs, exit)(hs, hnode);
+}
- if (unlikely(hs->hs_flags & CFS_HASH_DEBUG)) {
- i = cfs_hash_id(hs, cfs_hash_key(hs, hnode), hs->hs_cur_mask);
- LASSERT(hs->hs_buckets[i] == hsb);
- }
+static inline void cfs_hash_lock(cfs_hash_t *hs, int excl)
+{
+ hs->hs_lops->hs_lock(&hs->hs_lock, excl);
}
-static inline struct hlist_node *
-__cfs_hash_bucket_lookup(cfs_hash_t *hs,
- cfs_hash_bucket_t *hsb, void *key)
+static inline void cfs_hash_unlock(cfs_hash_t *hs, int excl)
{
- struct hlist_node *hnode;
+ hs->hs_lops->hs_unlock(&hs->hs_lock, excl);
+}
- hlist_for_each(hnode, &hsb->hsb_head)
- if (cfs_hash_compare(hs, key, hnode) > 0)
- return hnode;
+static inline int cfs_hash_dec_and_lock(cfs_hash_t *hs,
+ atomic_t *condition)
+{
+ LASSERT(cfs_hash_with_no_bktlock(hs));
+ return atomic_dec_and_lock(condition, &hs->hs_lock.spin);
+}
- return NULL;
+static inline void cfs_hash_bd_lock(cfs_hash_t *hs,
+ cfs_hash_bd_t *bd, int excl)
+{
+ hs->hs_lops->hs_bkt_lock(&bd->bd_bucket->hsb_lock, excl);
}
-static inline void *
-__cfs_hash_bucket_add(cfs_hash_t *hs,
- cfs_hash_bucket_t *hsb,
- struct hlist_node *hnode)
+static inline void cfs_hash_bd_unlock(cfs_hash_t *hs,
+ cfs_hash_bd_t *bd, int excl)
{
- hlist_add_head(hnode, &(hsb->hsb_head));
- atomic_inc(&hsb->hsb_count);
- atomic_inc(&hs->hs_count);
+ hs->hs_lops->hs_bkt_unlock(&bd->bd_bucket->hsb_lock, excl);
+}
+
+/**
+ * operations on cfs_hash bucket (bd: bucket descriptor),
+ * they are normally for hash-table without rehash
+ */
+void cfs_hash_bd_get(cfs_hash_t *hs, const void *key, cfs_hash_bd_t *bd);
- return cfs_hash_get(hs, hnode);
+static inline void cfs_hash_bd_get_and_lock(cfs_hash_t *hs, const void *key,
+ cfs_hash_bd_t *bd, int excl)
+{
+ cfs_hash_bd_get(hs, key, bd);
+ cfs_hash_bd_lock(hs, bd, excl);
+}
+
+static inline unsigned cfs_hash_bd_index_get(cfs_hash_t *hs, cfs_hash_bd_t *bd)
+{
+ return bd->bd_offset | (bd->bd_bucket->hsb_index << hs->hs_bkt_bits);
+}
+
+static inline void cfs_hash_bd_index_set(cfs_hash_t *hs,
+ unsigned index, cfs_hash_bd_t *bd)
+{
+ bd->bd_bucket = hs->hs_buckets[index >> hs->hs_bkt_bits];
+ bd->bd_offset = index & (CFS_HASH_BKT_NHLIST(hs) - 1U);
}
static inline void *
-__cfs_hash_bucket_del(cfs_hash_t *hs,
- cfs_hash_bucket_t *hsb,
- struct hlist_node *hnode)
+cfs_hash_bd_extra_get(cfs_hash_t *hs, cfs_hash_bd_t *bd)
+{
+ return (void *)bd->bd_bucket +
+ cfs_hash_bkt_size(hs) - hs->hs_extra_bytes;
+}
+
+static inline __u32
+cfs_hash_bd_version_get(cfs_hash_bd_t *bd)
{
- hlist_del_init(hnode);
- LASSERT(atomic_read(&hsb->hsb_count) > 0);
- atomic_dec(&hsb->hsb_count);
- LASSERT(atomic_read(&hs->hs_count) > 0);
- atomic_dec(&hs->hs_count);
+ /* need hold cfs_hash_bd_lock */
+ return bd->bd_bucket->hsb_version;
+}
+
+static inline __u32
+cfs_hash_bd_count_get(cfs_hash_bd_t *bd)
+{
+ /* need hold cfs_hash_bd_lock */
+ return bd->bd_bucket->hsb_count;
+}
+
+static inline int
+cfs_hash_bd_depmax_get(cfs_hash_bd_t *bd)
+{
+ return bd->bd_bucket->hsb_depmax;
+}
+
+static inline int
+cfs_hash_bd_compare(cfs_hash_bd_t *bd1, cfs_hash_bd_t *bd2)
+{
+ if (bd1->bd_bucket->hsb_index != bd2->bd_bucket->hsb_index)
+ return bd1->bd_bucket->hsb_index - bd2->bd_bucket->hsb_index;
- return cfs_hash_put(hs, hnode);
+ if (bd1->bd_offset != bd2->bd_offset)
+ return bd1->bd_offset - bd2->bd_offset;
+
+ return 0;
+}
+
+void cfs_hash_bd_add_locked(cfs_hash_t *hs, cfs_hash_bd_t *bd,
+ struct hlist_node *hnode);
+void cfs_hash_bd_del_locked(cfs_hash_t *hs, cfs_hash_bd_t *bd,
+ struct hlist_node *hnode);
+void cfs_hash_bd_move_locked(cfs_hash_t *hs, cfs_hash_bd_t *bd_old,
+ cfs_hash_bd_t *bd_new, struct hlist_node *hnode);
+
+static inline int cfs_hash_bd_dec_and_lock(cfs_hash_t *hs, cfs_hash_bd_t *bd,
+ atomic_t *condition)
+{
+ LASSERT(cfs_hash_with_spin_bktlock(hs));
+ return atomic_dec_and_lock(condition, &bd->bd_bucket->hsb_lock.spin);
}
+static inline struct hlist_head *cfs_hash_bd_hhead(cfs_hash_t *hs,
+ cfs_hash_bd_t *bd)
+{
+ return hs->hs_hops->hop_hhead(hs, bd);
+}
+
+struct hlist_node *cfs_hash_bd_lookup_locked(cfs_hash_t *hs, cfs_hash_bd_t *bd,
+ const void *key);
+struct hlist_node *cfs_hash_bd_peek_locked(cfs_hash_t *hs, cfs_hash_bd_t *bd,
+ const void *key);
+struct hlist_node *cfs_hash_bd_findadd_locked(cfs_hash_t *hs, cfs_hash_bd_t *bd,
+ const void *key,
+ struct hlist_node *hnode,
+ int insist_add);
+struct hlist_node *cfs_hash_bd_finddel_locked(cfs_hash_t *hs, cfs_hash_bd_t *bd,
+ const void *key,
+ struct hlist_node *hnode);
+
+/**
+ * operations on cfs_hash bucket (bd: bucket descriptor),
+ * they are safe for hash-table with rehash
+ */
+void cfs_hash_dual_bd_get(cfs_hash_t *hs, const void *key, cfs_hash_bd_t *bds);
+void cfs_hash_dual_bd_lock(cfs_hash_t *hs, cfs_hash_bd_t *bds, int excl);
+void cfs_hash_dual_bd_unlock(cfs_hash_t *hs, cfs_hash_bd_t *bds, int excl);
+
+static inline void cfs_hash_dual_bd_get_and_lock(cfs_hash_t *hs, const void *key,
+ cfs_hash_bd_t *bds, int excl)
+{
+ cfs_hash_dual_bd_get(hs, key, bds);
+ cfs_hash_dual_bd_lock(hs, bds, excl);
+}
+
+struct hlist_node *
+cfs_hash_dual_bd_lookup_locked(cfs_hash_t *hs, cfs_hash_bd_t *bds,
+ const void *key);
+struct hlist_node *
+cfs_hash_dual_bd_findadd_locked(cfs_hash_t *hs, cfs_hash_bd_t *bds,
+ const void *key, struct hlist_node *hnode,
+ int insist_add);
+struct hlist_node *
+cfs_hash_dual_bd_finddel_locked(cfs_hash_t *hs, cfs_hash_bd_t *bds,
+ const void *key, struct hlist_node *hnode);
+
/* Hash init/cleanup functions */
-cfs_hash_t *cfs_hash_create(char *name, unsigned int cur_bits,
- unsigned int max_bits,
- cfs_hash_ops_t *ops, int flags);
-void cfs_hash_destroy(cfs_hash_t *hs);
+cfs_hash_t *cfs_hash_create(char *name, unsigned cur_bits, unsigned max_bits,
+ unsigned bkt_bits, unsigned extra_bytes,
+ unsigned min_theta, unsigned max_theta,
+ cfs_hash_ops_t *ops, unsigned flags);
+
+cfs_hash_t *cfs_hash_getref(cfs_hash_t *hs);
+void cfs_hash_putref(cfs_hash_t *hs);
/* Hash addition functions */
-void cfs_hash_add(cfs_hash_t *hs, void *key,
- struct hlist_node *hnode);
-int cfs_hash_add_unique(cfs_hash_t *hs, void *key,
- struct hlist_node *hnode);
-void *cfs_hash_findadd_unique(cfs_hash_t *hs, void *key,
- struct hlist_node *hnode);
+void cfs_hash_add(cfs_hash_t *hs, const void *key,
+ struct hlist_node *hnode);
+int cfs_hash_add_unique(cfs_hash_t *hs, const void *key,
+ struct hlist_node *hnode);
+void *cfs_hash_findadd_unique(cfs_hash_t *hs, const void *key,
+ struct hlist_node *hnode);
/* Hash deletion functions */
-void *cfs_hash_del(cfs_hash_t *hs, void *key, struct hlist_node *hnode);
-void *cfs_hash_del_key(cfs_hash_t *hs, void *key);
+void *cfs_hash_del(cfs_hash_t *hs, const void *key, struct hlist_node *hnode);
+void *cfs_hash_del_key(cfs_hash_t *hs, const void *key);
/* Hash lookup/for_each functions */
-void *cfs_hash_lookup(cfs_hash_t *hs, void *key);
-typedef void (*cfs_hash_for_each_cb_t)(void *obj, void *data);
+#define CFS_HASH_LOOP_HOG 1024
+
+typedef int (*cfs_hash_for_each_cb_t)(cfs_hash_t *hs, cfs_hash_bd_t *bd,
+ struct hlist_node *node, void *data);
+void *cfs_hash_lookup(cfs_hash_t *hs, const void *key);
void cfs_hash_for_each(cfs_hash_t *hs, cfs_hash_for_each_cb_t, void *data);
void cfs_hash_for_each_safe(cfs_hash_t *hs, cfs_hash_for_each_cb_t, void *data);
-void cfs_hash_for_each_empty(cfs_hash_t *hs, cfs_hash_for_each_cb_t, void *data);
-void cfs_hash_for_each_key(cfs_hash_t *hs, void *key,
- cfs_hash_for_each_cb_t, void *data);
+int cfs_hash_for_each_nolock(cfs_hash_t *hs, cfs_hash_for_each_cb_t,
+ void *data);
+int cfs_hash_for_each_empty(cfs_hash_t *hs, cfs_hash_for_each_cb_t,
+ void *data);
+void cfs_hash_for_each_key(cfs_hash_t *hs, const void *key,
+ cfs_hash_for_each_cb_t, void *data);
+typedef int (*cfs_hash_cond_opt_cb_t)(void *obj, void *data);
+void cfs_hash_cond_del(cfs_hash_t *hs, cfs_hash_cond_opt_cb_t, void *data);
+
+void cfs_hash_hlist_for_each(cfs_hash_t *hs, unsigned hindex,
+ cfs_hash_for_each_cb_t, void *data);
+int cfs_hash_is_empty(cfs_hash_t *hs);
+__u64 cfs_hash_size_get(cfs_hash_t *hs);
/*
* Rehash - Theta is calculated to be the average chained
* hash depth assuming a perfectly uniform hash funcion.
*/
-int cfs_hash_rehash(cfs_hash_t *hs, int bits);
-void cfs_hash_rehash_key(cfs_hash_t *hs, void *old_key,
- void *new_key, struct hlist_node *hnode);
+void cfs_hash_rehash_cancel_locked(cfs_hash_t *hs);
+void cfs_hash_rehash_cancel(cfs_hash_t *hs);
+int cfs_hash_rehash(cfs_hash_t *hs, int do_rehash);
+void cfs_hash_rehash_key(cfs_hash_t *hs, const void *old_key,
+ void *new_key, struct hlist_node *hnode);
+
+#if CFS_HASH_DEBUG_LEVEL > CFS_HASH_DEBUG_1
+/* Validate hnode references the correct key */
+static inline void
+cfs_hash_key_validate(cfs_hash_t *hs, const void *key,
+ struct hlist_node *hnode)
+{
+ LASSERT(cfs_hash_keycmp(hs, key, hnode));
+}
+
+/* Validate hnode is in the correct bucket */
+static inline void
+cfs_hash_bucket_validate(cfs_hash_t *hs, cfs_hash_bd_t *bd,
+ struct hlist_node *hnode)
+{
+ cfs_hash_bd_t bds[2];
+
+ cfs_hash_dual_bd_get(hs, cfs_hash_key(hs, hnode), bds);
+ LASSERT(bds[0].bd_bucket == bd->bd_bucket ||
+ bds[1].bd_bucket == bd->bd_bucket);
+}
+
+#else /* CFS_HASH_DEBUG_LEVEL > CFS_HASH_DEBUG_1 */
+
+static inline void
+cfs_hash_key_validate(cfs_hash_t *hs, const void *key,
+ struct hlist_node *hnode) {}
+
+static inline void
+cfs_hash_bucket_validate(cfs_hash_t *hs, cfs_hash_bd_t *bd,
+ struct hlist_node *hnode) {}
+#endif /* CFS_HASH_DEBUG_LEVEL */
#define CFS_HASH_THETA_BITS 10
+#define CFS_HASH_MIN_THETA (1U << (CFS_HASH_THETA_BITS - 1))
+#define CFS_HASH_MAX_THETA (1U << (CFS_HASH_THETA_BITS + 1))
/* Return integer component of theta */
static inline int __cfs_hash_theta_int(int theta)
static inline int __cfs_hash_theta(cfs_hash_t *hs)
{
- return (atomic_read(&hs->hs_count) <<
- CFS_HASH_THETA_BITS) >> hs->hs_cur_bits;
+ return (atomic_read(&hs->hs_count) <<
+ CFS_HASH_THETA_BITS) >> hs->hs_cur_bits;
}
static inline void __cfs_hash_set_theta(cfs_hash_t *hs, int min, int max)
{
LASSERT(min < max);
- hs->hs_min_theta = min;
- hs->hs_max_theta = max;
+ hs->hs_min_theta = (__u16)min;
+ hs->hs_max_theta = (__u16)max;
}
/* Generic debug formatting routines mainly for proc handler */
-int cfs_hash_debug_header(char *str, int size);
-int cfs_hash_debug_str(cfs_hash_t *hs, char *str, int size);
+struct seq_file;
+int cfs_hash_debug_header(struct seq_file *m);
+int cfs_hash_debug_str(cfs_hash_t *hs, struct seq_file *m);
/*
* Generic djb2 hash algorithm for character arrays.
*/
static inline unsigned
-cfs_hash_djb2_hash(void *key, size_t size, unsigned mask)
+cfs_hash_djb2_hash(const void *key, size_t size, unsigned mask)
{
unsigned i, hash = 5381;
* Generic u32 hash algorithm.
*/
static inline unsigned
-cfs_hash_u32_hash(__u32 key, unsigned mask)
+cfs_hash_u32_hash(const __u32 key, unsigned mask)
{
return ((key * CFS_GOLDEN_RATIO_PRIME_32) & mask);
}
* Generic u64 hash algorithm.
*/
static inline unsigned
-cfs_hash_u64_hash(__u64 key, unsigned mask)
+cfs_hash_u64_hash(const __u64 key, unsigned mask)
{
return ((unsigned)(key * CFS_GOLDEN_RATIO_PRIME_64) & mask);
}
-#define cfs_hash_for_each_bucket(hs, hsb, pos) \
- for (pos = 0; \
- pos <= hs->hs_cur_mask && \
- ({ hsb = hs->hs_buckets[i]; 1; }); \
- pos++)
+/** iterate over all buckets in @bds (array of cfs_hash_bd_t) */
+#define cfs_hash_for_each_bd(bds, n, i) \
+ for (i = 0; i < n && (bds)[i].bd_bucket != NULL; i++)
+
+/** iterate over all buckets of @hs */
+#define cfs_hash_for_each_bucket(hs, bd, pos) \
+ for (pos = 0; \
+ pos < CFS_HASH_NBKT(hs) && \
+ ((bd)->bd_bucket = (hs)->hs_buckets[pos]) != NULL; pos++)
+
+/** iterate over all hlist of bucket @bd */
+#define cfs_hash_bd_for_each_hlist(hs, bd, hlist) \
+ for ((bd)->bd_offset = 0; \
+ (bd)->bd_offset < CFS_HASH_BKT_NHLIST(hs) && \
+ (hlist = cfs_hash_bd_hhead(hs, bd)) != NULL; \
+ (bd)->bd_offset++)
/* !__LIBCFS__HASH_H__ */
#endif
git://git.whamcloud.com / fs / lustre-release.git / blobdiff