4 * Userspace RCU library - Lock-Free Expandable RCU Hash Table
6 * Copyright 2010-2011 - Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
8 * This library is free software; you can redistribute it and/or
9 * modify it under the terms of the GNU Lesser General Public
10 * License as published by the Free Software Foundation; either
11 * version 2.1 of the License, or (at your option) any later version.
13 * This library is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 * Lesser General Public License for more details.
18 * You should have received a copy of the GNU Lesser General Public
19 * License along with this library; if not, write to the Free Software
20 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
32 #include <urcu-call-rcu.h>
33 #include <urcu/arch.h>
34 #include <urcu/uatomic.h>
35 #include <urcu/jhash.h>
36 #include <urcu/compiler.h>
37 #include <urcu/rculfhash.h>
41 #define BUCKET_SIZE_RESIZE_THRESHOLD 32
42 #define MAX_NR_BUCKETS 1048576 /* 1M buckets */
45 #define max(a, b) ((a) > (b) ? (a) : (b))
49 unsigned long size
; /* always a power of 2 */
51 struct rcu_ht_node
*tbl
[0];
55 struct rcu_table
*t
; /* shared */
58 pthread_mutex_t resize_mutex
; /* resize mutex: add/del mutex */
59 unsigned long target_size
;
60 void (*ht_call_rcu
)(struct rcu_head
*head
,
61 void (*func
)(struct rcu_head
*head
));
64 struct rcu_resize_work
{
70 void ht_resize_lazy(struct rcu_ht
*ht
, struct rcu_table
*t
, int growth
);
73 void check_resize(struct rcu_ht
*ht
, struct rcu_table
*t
,
74 unsigned long chain_len
)
76 //printf("check resize chain len %lu\n", chain_len);
77 if (chain_len
>= BUCKET_SIZE_RESIZE_THRESHOLD
)
78 ht_resize_lazy(ht
, t
, chain_len
/ BUCKET_SIZE_RESIZE_THRESHOLD
);
82 * Algorithm to reverse bits in a word by lookup table, extended to
85 * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable
88 static const uint8_t BitReverseTable256
[256] =
90 #define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64
91 #define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16)
92 #define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 )
93 R6(0), R6(2), R6(1), R6(3)
100 uint8_t bit_reverse_u8(uint8_t v
)
102 return BitReverseTable256
[v
];
105 static __attribute__((unused
))
106 uint32_t bit_reverse_u32(uint32_t v
)
108 return ((uint32_t) bit_reverse_u8(v
) << 24) |
109 ((uint32_t) bit_reverse_u8(v
>> 8) << 16) |
110 ((uint32_t) bit_reverse_u8(v
>> 16) << 8) |
111 ((uint32_t) bit_reverse_u8(v
>> 24));
114 static __attribute__((unused
))
115 uint64_t bit_reverse_u64(uint64_t v
)
117 return ((uint64_t) bit_reverse_u8(v
) << 56) |
118 ((uint64_t) bit_reverse_u8(v
>> 8) << 48) |
119 ((uint64_t) bit_reverse_u8(v
>> 16) << 40) |
120 ((uint64_t) bit_reverse_u8(v
>> 24) << 32) |
121 ((uint64_t) bit_reverse_u8(v
>> 32) << 24) |
122 ((uint64_t) bit_reverse_u8(v
>> 40) << 16) |
123 ((uint64_t) bit_reverse_u8(v
>> 48) << 8) |
124 ((uint64_t) bit_reverse_u8(v
>> 56));
128 unsigned long bit_reverse_ulong(unsigned long v
)
130 #if (CAA_BITS_PER_LONG == 32)
131 return bit_reverse_u32(v
);
133 return bit_reverse_u64(v
);
138 struct rcu_ht_node
*clear_flag(struct rcu_ht_node
*node
)
140 return (struct rcu_ht_node
*) (((unsigned long) node
) & ~0x1);
144 int is_removed(struct rcu_ht_node
*node
)
146 return ((unsigned long) node
) & 0x1;
150 struct rcu_ht_node
*flag_removed(struct rcu_ht_node
*node
)
152 return (struct rcu_ht_node
*) (((unsigned long) node
) | 0x1);
156 void _uatomic_max(unsigned long *ptr
, unsigned long v
)
158 unsigned long old1
, old2
;
160 old1
= uatomic_read(ptr
);
165 } while ((old1
= uatomic_cmpxchg(ptr
, old2
, v
)) != old2
);
169 void _ht_add(struct rcu_ht
*ht
, struct rcu_table
*t
, struct rcu_ht_node
*node
)
171 struct rcu_ht_node
*iter_prev
= NULL
, *iter
= NULL
;
176 unsigned long chain_len
= 0;
178 iter_prev
= rcu_dereference(t
->tbl
[node
->hash
& (t
->size
- 1)]);
179 //printf("iter prev %p hash %lu bucket %lu\n", iter_prev,
180 // node->hash, node->hash & (t->size - 1));
182 assert(iter_prev
->reverse_hash
<= node
->reverse_hash
);
184 iter
= clear_flag(rcu_dereference(iter_prev
->next
));
187 if (iter
->reverse_hash
< node
->reverse_hash
)
190 check_resize(ht
, t
, ++chain_len
);
192 /* add in iter_prev->next */
193 if (is_removed(iter
))
195 assert(node
!= iter
);
197 assert(iter_prev
!= node
);
198 if (uatomic_cmpxchg(&iter_prev
->next
, iter
, node
) != iter
)
205 int _ht_remove(struct rcu_ht
*ht
, struct rcu_table
*t
, struct rcu_ht_node
*node
)
207 struct rcu_ht_node
*iter_prev
, *iter
, *next
, *old
;
208 unsigned long chain_len
;
215 iter_prev
= rcu_dereference(t
->tbl
[node
->hash
& (t
->size
- 1)]);
217 assert(iter_prev
->reverse_hash
<= node
->reverse_hash
);
219 iter
= clear_flag(rcu_dereference(iter_prev
->next
));
222 if (iter
->reverse_hash
< node
->reverse_hash
)
234 next
= rcu_dereference(iter
->next
);
236 if (is_removed(next
)) {
240 /* set deletion flag */
241 if ((old
= uatomic_cmpxchg(&iter
->next
, next
,
242 flag_removed(next
))) != next
) {
243 if (old
== flag_removed(next
)) {
253 * Remove the element from the list. Retry if there has been a
254 * concurrent add (there cannot be a concurrent delete, because
255 * we won the deletion flag cmpxchg).
257 if (uatomic_cmpxchg(&iter_prev
->next
, iter
, clear_flag(next
)) != iter
)
264 void init_table(struct rcu_ht
*ht
, struct rcu_table
*t
,
265 unsigned long first
, unsigned long len
)
267 unsigned long i
, end
;
270 for (i
= first
; i
< end
; i
++) {
271 /* Update table size when power of two */
272 if (i
!= 0 && !(i
& (i
- 1)))
274 t
->tbl
[i
] = calloc(1, sizeof(struct rcu_ht_node
));
275 t
->tbl
[i
]->dummy
= 1;
277 t
->tbl
[i
]->reverse_hash
= bit_reverse_ulong(i
);
278 _ht_add(ht
, t
, t
->tbl
[i
]);
283 struct rcu_ht
*ht_new(ht_hash_fct hash_fct
,
285 unsigned long init_size
,
286 void (*ht_call_rcu
)(struct rcu_head
*head
,
287 void (*func
)(struct rcu_head
*head
)))
291 ht
= calloc(1, sizeof(struct rcu_ht
));
292 ht
->hash_fct
= hash_fct
;
293 ht
->hashseed
= hashseed
;
294 ht
->ht_call_rcu
= ht_call_rcu
;
295 /* this mutex should not nest in read-side C.S. */
296 pthread_mutex_init(&ht
->resize_mutex
, NULL
);
297 ht
->t
= calloc(1, sizeof(struct rcu_table
)
298 + (max(init_size
, 1) * sizeof(struct rcu_ht_node
*)));
300 pthread_mutex_lock(&ht
->resize_mutex
);
301 init_table(ht
, ht
->t
, 0, max(init_size
, 1));
302 pthread_mutex_unlock(&ht
->resize_mutex
);
303 ht
->target_size
= ht
->t
->size
;
307 struct rcu_ht_node
*ht_lookup(struct rcu_ht
*ht
, void *key
)
310 struct rcu_ht_node
*node
;
311 unsigned long hash
, reverse_hash
;
313 hash
= ht
->hash_fct(ht
->hashseed
, key
);
314 reverse_hash
= bit_reverse_ulong(hash
);
316 t
= rcu_dereference(ht
->t
);
317 node
= rcu_dereference(t
->tbl
[hash
& (t
->size
- 1)]);
321 if (node
->reverse_hash
> reverse_hash
) {
325 if (node
->key
== key
) {
326 if (is_removed(rcu_dereference(node
->next
)))
330 node
= clear_flag(rcu_dereference(node
->next
));
335 void ht_add(struct rcu_ht
*ht
, struct rcu_ht_node
*node
)
339 node
->hash
= ht
->hash_fct(ht
->hashseed
, node
->key
);
340 node
->reverse_hash
= bit_reverse_ulong((unsigned long) node
->hash
);
342 t
= rcu_dereference(ht
->t
);
343 _ht_add(ht
, t
, node
);
346 int ht_remove(struct rcu_ht
*ht
, struct rcu_ht_node
*node
)
350 t
= rcu_dereference(ht
->t
);
351 return _ht_remove(ht
, t
, node
);
355 int ht_delete_dummy(struct rcu_ht
*ht
)
358 struct rcu_ht_node
*node
;
362 /* Check that the table is empty */
369 /* Internal sanity check: all nodes left should be dummy */
370 for (i
= 0; i
< t
->size
; i
++) {
371 assert(t
->tbl
[i
]->dummy
);
378 * Should only be called when no more concurrent readers nor writers can
379 * possibly access the table.
381 int ht_destroy(struct rcu_ht
*ht
)
385 ret
= ht_delete_dummy(ht
);
394 void ht_free_table_cb(struct rcu_head
*head
)
396 struct rcu_table
*t
=
397 caa_container_of(head
, struct rcu_table
, head
);
401 /* called with resize mutex held */
403 void _do_ht_resize(struct rcu_ht
*ht
)
405 unsigned long new_size
, old_size
;
406 struct rcu_table
*new_t
, *old_t
;
411 old_size
= old_t
->size
;
413 new_size
= CMM_LOAD_SHARED(ht
->target_size
);
414 if (old_size
== new_size
)
416 new_t
= malloc(sizeof(struct rcu_table
)
417 + (new_size
* sizeof(struct rcu_ht_node
*)));
418 assert(new_size
> old_size
);
419 memcpy(&new_t
->tbl
, &old_t
->tbl
,
420 old_size
* sizeof(struct rcu_ht_node
*));
421 init_table(ht
, new_t
, old_size
, new_size
- old_size
);
422 new_t
->size
= new_size
;
423 /* Changing table and size atomically wrt lookups */
424 rcu_assign_pointer(ht
->t
, new_t
);
425 ht
->ht_call_rcu(&old_t
->head
, ht_free_table_cb
);
429 void resize_target_update(struct rcu_ht
*ht
, struct rcu_table
*t
,
432 unsigned long new_size
= t
->size
<< growth_order
;
434 if (new_size
> MAX_NR_BUCKETS
)
435 new_size
= MAX_NR_BUCKETS
;
436 //printf("resize update prevtarget %lu current %lu order %d\n",
437 // ht->target_size, t->size, growth_order);
438 _uatomic_max(&ht
->target_size
, new_size
);
441 void ht_resize(struct rcu_ht
*ht
, int growth
)
443 resize_target_update(ht
, rcu_dereference(ht
->t
), growth
);
444 pthread_mutex_lock(&ht
->resize_mutex
);
446 pthread_mutex_unlock(&ht
->resize_mutex
);
450 void do_resize_cb(struct rcu_head
*head
)
452 struct rcu_resize_work
*work
=
453 caa_container_of(head
, struct rcu_resize_work
, head
);
454 struct rcu_ht
*ht
= work
->ht
;
456 pthread_mutex_lock(&ht
->resize_mutex
);
458 pthread_mutex_unlock(&ht
->resize_mutex
);
463 void ht_resize_lazy(struct rcu_ht
*ht
, struct rcu_table
*t
, int growth
)
465 struct rcu_resize_work
*work
;
467 work
= malloc(sizeof(*work
));
469 resize_target_update(ht
, t
, growth
);
470 ht
->ht_call_rcu(&work
->head
, do_resize_cb
);