4 * Userspace RCU library - RCU Judy Array
6 * Copyright 2012 - 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
26 #include <urcu/rcuja.h>
27 #include <urcu/compiler.h>
28 #include <urcu/arch.h>
30 #include <urcu-pointer.h>
32 #include "rcuja-internal.h"
35 enum rcu_ja_type_class
{
36 RCU_JA_LINEAR
= 0, /* Type A */
37 /* 32-bit: 1 to 25 children, 8 to 128 bytes */
38 /* 64-bit: 1 to 28 children, 16 to 256 bytes */
39 RCU_JA_POOL
= 1, /* Type B */
40 /* 32-bit: 26 to 100 children, 256 to 512 bytes */
41 /* 64-bit: 29 to 112 children, 512 to 1024 bytes */
42 RCU_JA_PIGEON
= 2, /* Type C */
43 /* 32-bit: 101 to 256 children, 1024 bytes */
44 /* 64-bit: 113 to 256 children, 2048 bytes */
45 /* Leaf nodes are implicit from their height in the tree */
50 enum rcu_ja_type_class type_class
;
51 uint16_t min_child
; /* minimum number of children: 1 to 256 */
52 uint16_t max_child
; /* maximum number of children: 1 to 256 */
53 uint16_t max_linear_child
; /* per-pool max nr. children: 1 to 256 */
54 uint16_t order
; /* node size is (1 << order), in bytes */
55 uint16_t nr_pool_order
; /* number of pools */
56 uint16_t pool_size_order
; /* pool size */
60 * Number of least significant pointer bits reserved to represent the
63 #define JA_TYPE_BITS 3
64 #define JA_TYPE_MAX_NR (1U << JA_TYPE_BITS)
65 #define JA_TYPE_MASK (JA_TYPE_MAX_NR - 1)
66 #define JA_PTR_MASK (~JA_TYPE_MASK)
68 #define JA_ENTRY_PER_NODE 256UL
71 * Iteration on the array to find the right node size for the number of
72 * children stops when it reaches .max_child == 256 (this is the largest
73 * possible node size, which contains 256 children).
74 * The min_child overlaps with the previous max_child to provide an
75 * hysteresis loop to reallocation for patterns of cyclic add/removal
76 * within the same node.
77 * The node the index within the following arrays is represented on 3
78 * bits. It identifies the node type, min/max number of children, and
80 * The max_child values for the RCU_JA_POOL below result from
81 * statistical approximation: over million populations, the max_child
82 * covers between 97% and 99% of the populations generated. Therefore, a
83 * fallback should exist to cover the rare extreme population unbalance
84 * cases, but it will not have a major impact on speed nor space
85 * consumption, since those are rare cases.
88 #if (CAA_BITS_PER_LONG < 64)
91 ja_type_0_max_child
= 1,
92 ja_type_1_max_child
= 3,
93 ja_type_2_max_child
= 6,
94 ja_type_3_max_child
= 12,
95 ja_type_4_max_child
= 25,
96 ja_type_5_max_child
= 48,
97 ja_type_6_max_child
= 92,
98 ja_type_7_max_child
= 256,
102 ja_type_0_max_linear_child
= 1,
103 ja_type_1_max_linear_child
= 3,
104 ja_type_2_max_linear_child
= 6,
105 ja_type_3_max_linear_child
= 12,
106 ja_type_4_max_linear_child
= 25,
107 ja_type_5_max_linear_child
= 24,
108 ja_type_6_max_linear_child
= 23,
112 ja_type_5_nr_pool_order
= 1,
113 ja_type_6_nr_pool_order
= 2,
116 const struct rcu_ja_type ja_types
[] = {
117 { .type_class
= RCU_JA_LINEAR
, .min_child
= 1, .max_child
= ja_type_0_max_child
, .max_linear_child
= ja_type_0_max_linear_child
, .order
= 3, },
118 { .type_class
= RCU_JA_LINEAR
, .min_child
= 1, .max_child
= ja_type_1_max_child
, .max_linear_child
= ja_type_1_max_linear_child
, .order
= 4, },
119 { .type_class
= RCU_JA_LINEAR
, .min_child
= 3, .max_child
= ja_type_2_max_child
, .max_linear_child
= ja_type_2_max_linear_child
, .order
= 5, },
120 { .type_class
= RCU_JA_LINEAR
, .min_child
= 4, .max_child
= ja_type_3_max_child
, .max_linear_child
= ja_type_3_max_linear_child
, .order
= 6, },
121 { .type_class
= RCU_JA_LINEAR
, .min_child
= 10, .max_child
= ja_type_4_max_child
, .max_linear_child
= ja_type_4_max_linear_child
, .order
= 7, },
123 /* Pools may fill sooner than max_child */
124 { .type_class
= RCU_JA_POOL
, .min_child
= 20, .max_child
= ja_type_5_max_child
, .max_linear_child
= ja_type_5_max_linear_child
, .order
= 8, .nr_pool_order
= ja_type_5_nr_pool_order
, .pool_size_order
= 7, },
125 { .type_class
= RCU_JA_POOL
, .min_child
= 45, .max_child
= ja_type_6_max_child
, .max_linear_child
= ja_type_6_max_linear_child
, .order
= 9, .nr_pool_order
= ja_type_6_nr_pool_order
, .pool_size_order
= 7, },
128 * TODO: Upon node removal below min_child, if child pool is
129 * filled beyond capacity, we need to roll back to pigeon.
131 { .type_class
= RCU_JA_PIGEON
, .min_child
= 89, .max_child
= ja_type_7_max_child
, .order
= 10, },
133 #else /* !(CAA_BITS_PER_LONG < 64) */
134 /* 64-bit pointers */
136 ja_type_0_max_child
= 1,
137 ja_type_1_max_child
= 3,
138 ja_type_2_max_child
= 7,
139 ja_type_3_max_child
= 14,
140 ja_type_4_max_child
= 28,
141 ja_type_5_max_child
= 54,
142 ja_type_6_max_child
= 104,
143 ja_type_7_max_child
= 256,
147 ja_type_0_max_linear_child
= 1,
148 ja_type_1_max_linear_child
= 3,
149 ja_type_2_max_linear_child
= 7,
150 ja_type_3_max_linear_child
= 14,
151 ja_type_4_max_linear_child
= 28,
152 ja_type_5_max_linear_child
= 27,
153 ja_type_6_max_linear_child
= 26,
157 ja_type_5_nr_pool_order
= 1,
158 ja_type_6_nr_pool_order
= 2,
161 const struct rcu_ja_type ja_types
[] = {
162 { .type_class
= RCU_JA_LINEAR
, .min_child
= 1, .max_child
= ja_type_0_max_child
, .max_linear_child
= ja_type_0_max_linear_child
, .order
= 4, },
163 { .type_class
= RCU_JA_LINEAR
, .min_child
= 1, .max_child
= ja_type_1_max_child
, .max_linear_child
= ja_type_1_max_linear_child
, .order
= 5, },
164 { .type_class
= RCU_JA_LINEAR
, .min_child
= 3, .max_child
= ja_type_2_max_child
, .max_linear_child
= ja_type_2_max_linear_child
, .order
= 6, },
165 { .type_class
= RCU_JA_LINEAR
, .min_child
= 5, .max_child
= ja_type_3_max_child
, .max_linear_child
= ja_type_3_max_linear_child
, .order
= 7, },
166 { .type_class
= RCU_JA_LINEAR
, .min_child
= 10, .max_child
= ja_type_4_max_child
, .max_linear_child
= ja_type_4_max_linear_child
, .order
= 8, },
168 /* Pools may fill sooner than max_child. */
169 { .type_class
= RCU_JA_POOL
, .min_child
= 22, .max_child
= ja_type_5_max_child
, .max_linear_child
= ja_type_5_max_linear_child
, .order
= 9, .nr_pool_order
= ja_type_5_nr_pool_order
, .pool_size_order
= 8, },
170 { .type_class
= RCU_JA_POOL
, .min_child
= 51, .max_child
= ja_type_6_max_child
, .max_linear_child
= ja_type_6_max_linear_child
, .order
= 10, .nr_pool_order
= ja_type_6_nr_pool_order
, .pool_size_order
= 8, },
173 * TODO: Upon node removal below min_child, if child pool is
174 * filled beyond capacity, we need to roll back to pigeon.
176 { .type_class
= RCU_JA_PIGEON
, .min_child
= 101, .max_child
= ja_type_7_max_child
, .order
= 11, },
178 #endif /* !(BITS_PER_LONG < 64) */
180 static inline __attribute__((unused
))
181 void static_array_size_check(void)
183 CAA_BUILD_BUG_ON(CAA_ARRAY_SIZE(ja_types
) > JA_TYPE_MAX_NR
);
186 /* Never declared. Opaque type used to store flagged node pointers. */
187 struct rcu_ja_node_flag
;
190 * The rcu_ja_node contains the compressed node data needed for
191 * read-side. For linear and pool node configurations, it starts with a
192 * byte counting the number of children in the node. Then, the
193 * node-specific data is placed.
194 * The node mutex, if any is needed, protecting concurrent updated of
195 * each node is placed in a separate hash table indexed by node address.
196 * For the pigeon configuration, the number of children is also kept in
197 * a separate hash table, indexed by node address, because it is only
198 * required for updates.
201 #define DECLARE_LINEAR_NODE(index) \
204 uint8_t child_value[ja_type_## index ##_max_linear_child]; \
205 struct rcu_ja_node_flag *child_ptr[ja_type_## index ##_max_linear_child]; \
208 #define DECLARE_POOL_NODE(index) \
212 uint8_t child_value[ja_type_## index ##_max_linear_child]; \
213 struct rcu_ja_node_flag *child_ptr[ja_type_## index ##_max_linear_child]; \
214 } linear[1U << ja_type_## index ##_nr_pool_order]; \
219 /* Linear configuration */
220 DECLARE_LINEAR_NODE(0) conf_0
;
221 DECLARE_LINEAR_NODE(1) conf_1
;
222 DECLARE_LINEAR_NODE(2) conf_2
;
223 DECLARE_LINEAR_NODE(3) conf_3
;
224 DECLARE_LINEAR_NODE(4) conf_4
;
226 /* Pool configuration */
227 DECLARE_POOL_NODE(5) conf_5
;
228 DECLARE_POOL_NODE(6) conf_6
;
230 /* Pigeon configuration */
232 struct rcu_ja_node_flag
*child
[ja_type_7_max_child
];
234 /* data aliasing nodes for computed accesses */
235 uint8_t data
[sizeof(struct rcu_ja_node_flag
*) * ja_type_7_max_child
];
240 struct rcu_ja_node_flag
*ja_node_flag(struct rcu_ja_node
*node
,
243 assert(type
< RCU_JA_NR_TYPES
);
244 return (struct rcu_ja_node_flag
*) (((unsigned long) node
) | type
);
248 unsigned int ja_node_type(struct rcu_ja_node_flag
*node
)
252 type
= (unsigned int) ((unsigned long) node
& JA_TYPE_MASK
);
253 assert(type
< RCU_JA_NR_TYPES
);
258 struct rcu_ja_node
*ja_node_ptr(struct rcu_ja_node_flag
*node
)
260 return (struct rcu_ja_node
*) (((unsigned long) node
) | JA_PTR_MASK
);
263 struct rcu_ja_node
*alloc_rcu_ja_node(const struct rcu_ja_type
*ja_type
)
265 return calloc(1U << ja_type
->order
, sizeof(char));
268 void free_rcu_ja_node(struct rcu_ja_node
*node
)
273 #define __JA_ALIGN_MASK(v, mask) (((v) + (mask)) & ~(mask))
274 #define JA_ALIGN(v, align) __JA_ALIGN_MASK(v, (typeof(v)) (align) - 1)
275 #define __JA_FLOOR_MASK(v, mask) ((v) & ~(mask))
276 #define JA_FLOOR(v, align) __JA_FLOOR_MASK(v, (typeof(v)) (align) - 1)
279 uint8_t *align_ptr_size(uint8_t *ptr
)
281 return (uint8_t *) JA_ALIGN((unsigned long) ptr
, sizeof(void *));
285 * The order in which values and pointers are does does not matter: if
286 * a value is missing, we return NULL. If a value is there, but its
287 * associated pointers is still NULL, we return NULL too.
290 struct rcu_ja_node_flag
*ja_linear_node_get_nth(const struct rcu_ja_type
*type
,
291 struct rcu_ja_node
*node
,
296 struct rcu_ja_node_flag
**pointers
;
297 struct rcu_ja_node_flag
*ptr
;
300 assert(type
->type_class
== RCU_JA_LINEAR
|| type
->type_class
== RCU_JA_POOL
);
302 nr_child
= CMM_LOAD_SHARED(node
->u
.data
[0]);
303 cmm_smp_rmb(); /* read nr_child before values and pointers */
304 assert(nr_child
<= type
->max_linear_child
);
305 assert(type
->type_class
!= RCU_JA_LINEAR
|| nr_child
>= type
->min_child
);
307 values
= &node
->u
.data
[1];
308 for (i
= 0; i
< nr_child
; i
++) {
309 if (CMM_LOAD_SHARED(values
[i
]) == n
)
314 pointers
= (struct rcu_ja_node_flag
**) align_ptr_size(&values
[type
->max_linear_child
]);
315 ptr
= rcu_dereference(pointers
[i
]);
316 assert(ja_node_ptr(ptr
) != NULL
);
321 struct rcu_ja_node_flag
*ja_pool_node_get_nth(const struct rcu_ja_type
*type
,
322 struct rcu_ja_node
*node
,
325 struct rcu_ja_node
*linear
;
327 assert(type
->type_class
== RCU_JA_POOL
);
328 linear
= (struct rcu_ja_node
*)
329 &node
->u
.data
[((unsigned long) n
>> (CHAR_BIT
- type
->nr_pool_order
)) << type
->pool_size_order
];
330 return ja_linear_node_get_nth(type
, linear
, n
);
334 struct rcu_ja_node_flag
*ja_pigeon_node_get_nth(const struct rcu_ja_type
*type
,
335 struct rcu_ja_node
*node
,
338 assert(type
->type_class
== RCU_JA_PIGEON
);
339 return rcu_dereference(((struct rcu_ja_node_flag
**) node
->u
.data
)[n
]);
343 * ja_node_get_nth: get nth item from a node.
344 * node_flag is already rcu_dereference'd.
347 struct rcu_ja_node_flag
*ja_node_get_nth(struct rcu_ja_node_flag
*node_flag
,
350 unsigned int type_index
;
351 struct rcu_ja_node
*node
;
352 const struct rcu_ja_type
*type
;
354 node
= ja_node_ptr(node_flag
);
355 assert(node
!= NULL
);
356 type_index
= ja_node_type(node_flag
);
357 type
= &ja_types
[type_index
];
359 switch (type
->type_class
) {
361 return ja_linear_node_get_nth(type
, node
, n
);
363 return ja_pool_node_get_nth(type
, node
, n
);
365 return ja_pigeon_node_get_nth(type
, node
, n
);
368 return (void *) -1UL;
373 int ja_linear_node_set_nth(const struct rcu_ja_type
*type
,
374 struct rcu_ja_node
*node
,
376 struct rcu_ja_node_flag
*child_node_flag
)
379 uint8_t *values
, *nr_child_ptr
;
380 struct rcu_ja_node_flag
**pointers
;
383 assert(type
->type_class
== RCU_JA_LINEAR
|| type
->type_class
== RCU_JA_POOL
);
385 nr_child_ptr
= &node
->u
.data
[0];
386 nr_child
= *nr_child_ptr
;
387 assert(nr_child
<= type
->max_linear_child
);
388 assert(type
->type_class
!= RCU_JA_LINEAR
|| nr_child
>= type
->min_child
);
390 values
= &node
->u
.data
[1];
391 for (i
= 0; i
< nr_child
; i
++) {
395 if (nr_child
>= type
->max_linear_child
) {
396 /* No space left in this node type */
399 pointers
= (struct rcu_ja_node_flag
**) align_ptr_size(&values
[type
->max_linear_child
]);
400 assert(pointers
[nr_child
] == NULL
);
401 rcu_assign_pointer(pointers
[nr_child
], child_node_flag
);
402 CMM_STORE_SHARED(values
[nr_child
], n
);
403 cmm_smp_wmb(); /* write value and pointer before nr_child */
404 CMM_STORE_SHARED(*nr_child_ptr
, nr_child
+ 1);
409 int ja_pool_node_set_nth(const struct rcu_ja_type
*type
,
410 struct rcu_ja_node
*node
,
412 struct rcu_ja_node_flag
*child_node_flag
)
414 struct rcu_ja_node
*linear
;
416 assert(type
->type_class
== RCU_JA_POOL
);
417 linear
= (struct rcu_ja_node
*)
418 &node
->u
.data
[((unsigned long) n
>> (CHAR_BIT
- type
->nr_pool_order
)) << type
->pool_size_order
];
419 return ja_linear_node_set_nth(type
, linear
, n
, child_node_flag
);
423 int ja_pigeon_node_set_nth(const struct rcu_ja_type
*type
,
424 struct rcu_ja_node
*node
,
426 struct rcu_ja_node_flag
*child_node_flag
)
428 struct rcu_ja_node_flag
**ptr
;
430 assert(type
->type_class
== RCU_JA_PIGEON
);
431 ptr
= &((struct rcu_ja_node_flag
**) node
->u
.data
)[n
];
434 rcu_assign_pointer(*ptr
, child_node_flag
);
439 * _ja_node_set_nth: set nth item within a node. Return an error
440 * (negative error value) if it is already there.
441 * TODO: exclusive access on node.
444 int _ja_node_set_nth(struct rcu_ja_node_flag
*node_flag
, uint8_t n
,
445 struct rcu_ja_node_flag
*child_node_flag
)
447 unsigned int type_index
;
448 struct rcu_ja_node
*node
;
449 const struct rcu_ja_type
*type
;
451 node
= ja_node_ptr(node_flag
);
452 assert(node
!= NULL
);
453 type_index
= ja_node_type(node_flag
);
454 type
= &ja_types
[type_index
];
456 switch (type
->type_class
) {
458 return ja_linear_node_set_nth(type
, node
, n
,
461 return ja_pool_node_set_nth(type
, node
, n
,
464 return ja_pigeon_node_set_nth(type
, node
, n
,
475 * ja_node_recompact_add: recompact a node, adding a new child.
478 int ja_node_recompact_add(struct rcu_ja_node_flag
**old_node_flag
, uint8_t n
,
479 struct rcu_ja_node_flag
*child_node_flag
)
482 unsigned int old_type_index
, new_type_index
;
483 struct rcu_ja_node
*old_node
, *new_node
;
484 const struct rcu_ja_type
*old_type
, *new_type
;
485 struct rcu_ja_node_flag
*new_node_flag
;
489 old_node
= ja_node_ptr(*old_node_flag
);
490 if (old_node
== NULL
) {
493 old_type_index
= ja_node_type(*old_node_flag
);
494 old_type
= &ja_types
[old_type_index
];
495 new_type_index
= old_type_index
+ 1;
497 new_type
= &ja_types
[new_type_index
];
498 new_node
= alloc_rcu_ja_node(new_type
);
501 new_node_flag
= ja_node_flag(new_node
, new_type_index
);
503 for (i
= 0; i
< old_type
->max_child
; i
++) {
504 struct rcu_ja_node_flag
*iter
;
506 iter
= ja_node_get_nth(*old_node_flag
, i
);
509 ret
= _ja_node_set_nth(new_node_flag
, i
, iter
);
513 ret
= _ja_node_set_nth(new_node_flag
, n
, child_node_flag
);
515 /* Replace the old node with the new recompacted one */
516 rcu_assign_pointer(*old_node_flag
, new_node_flag
);
517 /* TODO: free old_node (call_rcu) */
522 int ja_node_set_nth(struct rcu_ja_node_flag
**node_flag
, uint8_t n
,
523 struct rcu_ja_node_flag
*child_node_flag
)
527 ret
= _ja_node_set_nth(*node_flag
, n
, child_node_flag
);
528 if (ret
== -ENOSPC
) {
529 /* Not enough space in node, need to recompact */
530 ret
= ja_node_recompact_add(node_flag
, n
,