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
27 #include <urcu/rcuja.h>
28 #include <urcu/compiler.h>
29 #include <urcu/arch.h>
31 #include <urcu-pointer.h>
33 #include "rcuja-internal.h"
36 enum rcu_ja_type_class
{
37 RCU_JA_LINEAR
= 0, /* Type A */
38 /* 32-bit: 1 to 25 children, 8 to 128 bytes */
39 /* 64-bit: 1 to 28 children, 16 to 256 bytes */
40 RCU_JA_POOL
= 1, /* Type B */
41 /* 32-bit: 26 to 100 children, 256 to 512 bytes */
42 /* 64-bit: 29 to 112 children, 512 to 1024 bytes */
43 RCU_JA_PIGEON
= 2, /* Type C */
44 /* 32-bit: 101 to 256 children, 1024 bytes */
45 /* 64-bit: 113 to 256 children, 2048 bytes */
46 /* Leaf nodes are implicit from their height in the tree */
49 RCU_JA_NULL
, /* not an encoded type, but keeps code regular */
53 enum rcu_ja_type_class type_class
;
54 uint16_t min_child
; /* minimum number of children: 1 to 256 */
55 uint16_t max_child
; /* maximum number of children: 1 to 256 */
56 uint16_t max_linear_child
; /* per-pool max nr. children: 1 to 256 */
57 uint16_t order
; /* node size is (1 << order), in bytes */
58 uint16_t nr_pool_order
; /* number of pools */
59 uint16_t pool_size_order
; /* pool size */
63 * Number of least significant pointer bits reserved to represent the
66 #define JA_TYPE_BITS 3
67 #define JA_TYPE_MAX_NR (1U << JA_TYPE_BITS)
68 #define JA_TYPE_MASK (JA_TYPE_MAX_NR - 1)
69 #define JA_PTR_MASK (~JA_TYPE_MASK)
71 #define JA_ENTRY_PER_NODE 256UL
74 * Entry for NULL node is at index 8 of the table. It is never encoded
77 #define NODE_INDEX_NULL 8
80 * Iteration on the array to find the right node size for the number of
81 * children stops when it reaches .max_child == 256 (this is the largest
82 * possible node size, which contains 256 children).
83 * The min_child overlaps with the previous max_child to provide an
84 * hysteresis loop to reallocation for patterns of cyclic add/removal
85 * within the same node.
86 * The node the index within the following arrays is represented on 3
87 * bits. It identifies the node type, min/max number of children, and
89 * The max_child values for the RCU_JA_POOL below result from
90 * statistical approximation: over million populations, the max_child
91 * covers between 97% and 99% of the populations generated. Therefore, a
92 * fallback should exist to cover the rare extreme population unbalance
93 * cases, but it will not have a major impact on speed nor space
94 * consumption, since those are rare cases.
97 #if (CAA_BITS_PER_LONG < 64)
100 ja_type_0_max_child
= 1,
101 ja_type_1_max_child
= 3,
102 ja_type_2_max_child
= 6,
103 ja_type_3_max_child
= 12,
104 ja_type_4_max_child
= 25,
105 ja_type_5_max_child
= 48,
106 ja_type_6_max_child
= 92,
107 ja_type_7_max_child
= 256,
108 ja_type_8_max_child
= 0, /* NULL */
112 ja_type_0_max_linear_child
= 1,
113 ja_type_1_max_linear_child
= 3,
114 ja_type_2_max_linear_child
= 6,
115 ja_type_3_max_linear_child
= 12,
116 ja_type_4_max_linear_child
= 25,
117 ja_type_5_max_linear_child
= 24,
118 ja_type_6_max_linear_child
= 23,
122 ja_type_5_nr_pool_order
= 1,
123 ja_type_6_nr_pool_order
= 2,
126 const struct rcu_ja_type ja_types
[] = {
127 { .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, },
128 { .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, },
129 { .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, },
130 { .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, },
131 { .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, },
133 /* Pools may fill sooner than max_child */
134 { .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, },
135 { .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, },
138 * TODO: Upon node removal below min_child, if child pool is
139 * filled beyond capacity, we need to roll back to pigeon.
141 { .type_class
= RCU_JA_PIGEON
, .min_child
= 89, .max_child
= ja_type_7_max_child
, .order
= 10, },
143 { .type_class
= RCU_JA_NULL
, .min_child
= 0, .max_child
= ja_type_8_max_child
, },
145 #else /* !(CAA_BITS_PER_LONG < 64) */
146 /* 64-bit pointers */
148 ja_type_0_max_child
= 1,
149 ja_type_1_max_child
= 3,
150 ja_type_2_max_child
= 7,
151 ja_type_3_max_child
= 14,
152 ja_type_4_max_child
= 28,
153 ja_type_5_max_child
= 54,
154 ja_type_6_max_child
= 104,
155 ja_type_7_max_child
= 256,
156 ja_type_8_max_child
= 256,
160 ja_type_0_max_linear_child
= 1,
161 ja_type_1_max_linear_child
= 3,
162 ja_type_2_max_linear_child
= 7,
163 ja_type_3_max_linear_child
= 14,
164 ja_type_4_max_linear_child
= 28,
165 ja_type_5_max_linear_child
= 27,
166 ja_type_6_max_linear_child
= 26,
170 ja_type_5_nr_pool_order
= 1,
171 ja_type_6_nr_pool_order
= 2,
174 const struct rcu_ja_type ja_types
[] = {
175 { .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, },
176 { .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, },
177 { .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, },
178 { .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, },
179 { .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, },
181 /* Pools may fill sooner than max_child. */
182 { .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, },
183 { .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, },
186 * TODO: Upon node removal below min_child, if child pool is
187 * filled beyond capacity, we need to roll back to pigeon.
189 { .type_class
= RCU_JA_PIGEON
, .min_child
= 101, .max_child
= ja_type_7_max_child
, .order
= 11, },
191 { .type_class
= RCU_JA_NULL
, .min_child
= 0, .max_child
= ja_type_8_max_child
, },
193 #endif /* !(BITS_PER_LONG < 64) */
195 static inline __attribute__((unused
))
196 void static_array_size_check(void)
198 CAA_BUILD_BUG_ON(CAA_ARRAY_SIZE(ja_types
) < JA_TYPE_MAX_NR
);
202 * The rcu_ja_node contains the compressed node data needed for
203 * read-side. For linear and pool node configurations, it starts with a
204 * byte counting the number of children in the node. Then, the
205 * node-specific data is placed.
206 * The node mutex, if any is needed, protecting concurrent updated of
207 * each node is placed in a separate hash table indexed by node address.
208 * For the pigeon configuration, the number of children is also kept in
209 * a separate hash table, indexed by node address, because it is only
210 * required for updates.
213 #define DECLARE_LINEAR_NODE(index) \
216 uint8_t child_value[ja_type_## index ##_max_linear_child]; \
217 struct rcu_ja_node_flag *child_ptr[ja_type_## index ##_max_linear_child]; \
220 #define DECLARE_POOL_NODE(index) \
224 uint8_t child_value[ja_type_## index ##_max_linear_child]; \
225 struct rcu_ja_node_flag *child_ptr[ja_type_## index ##_max_linear_child]; \
226 } linear[1U << ja_type_## index ##_nr_pool_order]; \
231 /* Linear configuration */
232 DECLARE_LINEAR_NODE(0) conf_0
;
233 DECLARE_LINEAR_NODE(1) conf_1
;
234 DECLARE_LINEAR_NODE(2) conf_2
;
235 DECLARE_LINEAR_NODE(3) conf_3
;
236 DECLARE_LINEAR_NODE(4) conf_4
;
238 /* Pool configuration */
239 DECLARE_POOL_NODE(5) conf_5
;
240 DECLARE_POOL_NODE(6) conf_6
;
242 /* Pigeon configuration */
244 struct rcu_ja_node_flag
*child
[ja_type_7_max_child
];
246 /* data aliasing nodes for computed accesses */
247 uint8_t data
[sizeof(struct rcu_ja_node_flag
*) * ja_type_7_max_child
];
252 struct rcu_ja_node_flag
*ja_node_flag(struct rcu_ja_node
*node
,
255 assert(type
< RCU_JA_NR_TYPES
);
256 return (struct rcu_ja_node_flag
*) (((unsigned long) node
) | type
);
260 struct rcu_ja_node
*ja_node_ptr(struct rcu_ja_node_flag
*node
)
262 return (struct rcu_ja_node
*) (((unsigned long) node
) | JA_PTR_MASK
);
266 unsigned int ja_node_type(struct rcu_ja_node_flag
*node
)
270 if (ja_node_ptr(node
) == NULL
) {
271 return NODE_INDEX_NULL
;
273 type
= (unsigned int) ((unsigned long) node
& JA_TYPE_MASK
);
274 assert(type
< RCU_JA_NR_TYPES
);
278 struct rcu_ja_node
*alloc_rcu_ja_node(const struct rcu_ja_type
*ja_type
)
280 return calloc(1U << ja_type
->order
, sizeof(char));
283 void free_rcu_ja_node(struct rcu_ja_node
*node
)
288 #define __JA_ALIGN_MASK(v, mask) (((v) + (mask)) & ~(mask))
289 #define JA_ALIGN(v, align) __JA_ALIGN_MASK(v, (typeof(v)) (align) - 1)
290 #define __JA_FLOOR_MASK(v, mask) ((v) & ~(mask))
291 #define JA_FLOOR(v, align) __JA_FLOOR_MASK(v, (typeof(v)) (align) - 1)
294 uint8_t *align_ptr_size(uint8_t *ptr
)
296 return (uint8_t *) JA_ALIGN((unsigned long) ptr
, sizeof(void *));
300 * The order in which values and pointers are does does not matter: if
301 * a value is missing, we return NULL. If a value is there, but its
302 * associated pointers is still NULL, we return NULL too.
305 struct rcu_ja_node_flag
*ja_linear_node_get_nth(const struct rcu_ja_type
*type
,
306 struct rcu_ja_node
*node
,
311 struct rcu_ja_node_flag
**pointers
;
312 struct rcu_ja_node_flag
*ptr
;
315 assert(type
->type_class
== RCU_JA_LINEAR
|| type
->type_class
== RCU_JA_POOL
);
317 nr_child
= CMM_LOAD_SHARED(node
->u
.data
[0]);
318 cmm_smp_rmb(); /* read nr_child before values and pointers */
319 assert(nr_child
<= type
->max_linear_child
);
320 assert(type
->type_class
!= RCU_JA_LINEAR
|| nr_child
>= type
->min_child
);
322 values
= &node
->u
.data
[1];
323 for (i
= 0; i
< nr_child
; i
++) {
324 if (CMM_LOAD_SHARED(values
[i
]) == n
)
329 pointers
= (struct rcu_ja_node_flag
**) align_ptr_size(&values
[type
->max_linear_child
]);
330 ptr
= rcu_dereference(pointers
[i
]);
331 assert(ja_node_ptr(ptr
) != NULL
);
336 struct rcu_ja_node_flag
*ja_pool_node_get_nth(const struct rcu_ja_type
*type
,
337 struct rcu_ja_node
*node
,
340 struct rcu_ja_node
*linear
;
342 assert(type
->type_class
== RCU_JA_POOL
);
344 * TODO: currently, we select the pool by highest bits. We
345 * should support various encodings.
347 linear
= (struct rcu_ja_node
*)
348 &node
->u
.data
[((unsigned long) n
>> (CHAR_BIT
- type
->nr_pool_order
)) << type
->pool_size_order
];
349 return ja_linear_node_get_nth(type
, linear
, n
);
353 struct rcu_ja_node_flag
*ja_pigeon_node_get_nth(const struct rcu_ja_type
*type
,
354 struct rcu_ja_node
*node
,
357 assert(type
->type_class
== RCU_JA_PIGEON
);
358 return rcu_dereference(((struct rcu_ja_node_flag
**) node
->u
.data
)[n
]);
362 * ja_node_get_nth: get nth item from a node.
363 * node_flag is already rcu_dereference'd.
366 struct rcu_ja_node_flag
*ja_node_get_nth(struct rcu_ja_node_flag
*node_flag
,
369 unsigned int type_index
;
370 struct rcu_ja_node
*node
;
371 const struct rcu_ja_type
*type
;
373 node
= ja_node_ptr(node_flag
);
374 assert(node
!= NULL
);
375 type_index
= ja_node_type(node_flag
);
376 type
= &ja_types
[type_index
];
378 switch (type
->type_class
) {
380 return ja_linear_node_get_nth(type
, node
, n
);
382 return ja_pool_node_get_nth(type
, node
, n
);
384 return ja_pigeon_node_get_nth(type
, node
, n
);
387 return (void *) -1UL;
392 unsigned int ja_get_nr_child(struct rcu_ja_shadow_node
*shadow_node
)
394 return shadow_node
->nr_child
;
398 int ja_linear_node_set_nth(const struct rcu_ja_type
*type
,
399 struct rcu_ja_node
*node
,
400 struct rcu_ja_shadow_node
*shadow_node
,
402 struct rcu_ja_node_flag
*child_node_flag
)
405 uint8_t *values
, *nr_child_ptr
;
406 struct rcu_ja_node_flag
**pointers
;
409 assert(type
->type_class
== RCU_JA_LINEAR
|| type
->type_class
== RCU_JA_POOL
);
411 nr_child_ptr
= &node
->u
.data
[0];
412 nr_child
= *nr_child_ptr
;
413 assert(nr_child
<= type
->max_linear_child
);
414 assert(type
->type_class
!= RCU_JA_LINEAR
|| nr_child
>= type
->min_child
);
416 values
= &node
->u
.data
[1];
417 for (i
= 0; i
< nr_child
; i
++) {
421 if (nr_child
>= type
->max_linear_child
) {
422 /* No space left in this node type */
425 pointers
= (struct rcu_ja_node_flag
**) align_ptr_size(&values
[type
->max_linear_child
]);
426 assert(pointers
[nr_child
] == NULL
);
427 rcu_assign_pointer(pointers
[nr_child
], child_node_flag
);
428 CMM_STORE_SHARED(values
[nr_child
], n
);
429 cmm_smp_wmb(); /* write value and pointer before nr_child */
430 CMM_STORE_SHARED(*nr_child_ptr
, nr_child
+ 1);
431 shadow_node
->nr_child
++;
436 int ja_pool_node_set_nth(const struct rcu_ja_type
*type
,
437 struct rcu_ja_node
*node
,
438 struct rcu_ja_shadow_node
*shadow_node
,
440 struct rcu_ja_node_flag
*child_node_flag
)
442 struct rcu_ja_node
*linear
;
444 assert(type
->type_class
== RCU_JA_POOL
);
445 linear
= (struct rcu_ja_node
*)
446 &node
->u
.data
[((unsigned long) n
>> (CHAR_BIT
- type
->nr_pool_order
)) << type
->pool_size_order
];
447 return ja_linear_node_set_nth(type
, linear
, shadow_node
,
452 int ja_pigeon_node_set_nth(const struct rcu_ja_type
*type
,
453 struct rcu_ja_node
*node
,
454 struct rcu_ja_shadow_node
*shadow_node
,
456 struct rcu_ja_node_flag
*child_node_flag
)
458 struct rcu_ja_node_flag
**ptr
;
460 assert(type
->type_class
== RCU_JA_PIGEON
);
461 ptr
= &((struct rcu_ja_node_flag
**) node
->u
.data
)[n
];
464 rcu_assign_pointer(*ptr
, child_node_flag
);
465 shadow_node
->nr_child
++;
470 * _ja_node_set_nth: set nth item within a node. Return an error
471 * (negative error value) if it is already there.
472 * TODO: exclusive access on node.
475 int _ja_node_set_nth(const struct rcu_ja_type
*type
,
476 struct rcu_ja_node
*node
,
477 struct rcu_ja_shadow_node
*shadow_node
,
479 struct rcu_ja_node_flag
*child_node_flag
)
481 switch (type
->type_class
) {
483 return ja_linear_node_set_nth(type
, node
, shadow_node
, n
,
486 return ja_pool_node_set_nth(type
, node
, shadow_node
, n
,
489 return ja_pigeon_node_set_nth(type
, node
, shadow_node
, n
,
502 * ja_node_recompact_add: recompact a node, adding a new child.
503 * TODO: for pool type, take selection bit(s) into account.
506 int ja_node_recompact_add(struct rcu_ja
*ja
,
507 unsigned int old_type_index
,
508 const struct rcu_ja_type
*old_type
,
509 struct rcu_ja_node
*old_node
,
510 struct rcu_ja_shadow_node
**shadow_node
,
511 struct rcu_ja_node_flag
**old_node_flag
, uint8_t n
,
512 struct rcu_ja_node_flag
*child_node_flag
)
514 unsigned int new_type_index
;
515 struct rcu_ja_node
*new_node
;
516 const struct rcu_ja_type
*new_type
;
517 struct rcu_ja_node_flag
*new_node_flag
;
521 if (*shadow_node
== NULL
) {
524 new_type_index
= old_type_index
+ 1;
526 new_type
= &ja_types
[new_type_index
];
527 new_node
= alloc_rcu_ja_node(new_type
);
530 new_node_flag
= ja_node_flag(new_node
, new_type_index
);
532 ret
= rcuja_shadow_set(ja
->ht
, new_node
, *shadow_node
);
536 if (*shadow_node
== NULL
) {
537 *shadow_node
= rcuja_shadow_lookup_lock(ja
->ht
, new_node
);
538 assert(*shadow_node
);
541 (*shadow_node
)->nr_child
= 0;
542 for (i
= 0; i
< old_type
->max_child
; i
++) {
543 struct rcu_ja_node_flag
*iter
;
545 iter
= ja_node_get_nth(*old_node_flag
, i
);
548 ret
= _ja_node_set_nth(new_type
, new_node
, *shadow_node
,
553 ret
= _ja_node_set_nth(new_type
, new_node
, *shadow_node
,
556 /* Replace the old node with the new recompacted one */
557 rcu_assign_pointer(*old_node_flag
, new_node_flag
);
558 ret
= rcuja_shadow_clear(ja
->ht
, old_node
,
559 RCUJA_SHADOW_CLEAR_FREE_NODE
);
565 int ja_node_set_nth(struct rcu_ja
*ja
,
566 struct rcu_ja_node_flag
**node_flag
, uint8_t n
,
567 struct rcu_ja_node_flag
*child_node_flag
)
570 unsigned int type_index
;
571 const struct rcu_ja_type
*type
;
572 struct rcu_ja_node
*node
;
573 struct rcu_ja_shadow_node
*shadow_node
= NULL
;
575 node
= ja_node_ptr(*node_flag
);
576 type_index
= ja_node_type(*node_flag
);
577 type
= &ja_types
[type_index
];
579 shadow_node
= rcuja_shadow_lookup_lock(ja
->ht
, node
);
582 ret
= _ja_node_set_nth(type
, node
, shadow_node
,
584 if (ret
== -ENOSPC
) {
585 /* Not enough space in node, need to recompact. */
586 ret
= ja_node_recompact_add(ja
, type_index
, type
, node
,
587 &shadow_node
, node_flag
, n
, child_node_flag
);
588 /* recompact always leave shadow_node locked */
590 rcuja_shadow_unlock(shadow_node
);