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>
32 #include <urcu/uatomic.h>
35 #include "rcuja-internal.h"
38 enum cds_ja_type_class
{
39 RCU_JA_LINEAR
= 0, /* Type A */
40 /* 32-bit: 1 to 25 children, 8 to 128 bytes */
41 /* 64-bit: 1 to 28 children, 16 to 256 bytes */
42 RCU_JA_POOL
= 1, /* Type B */
43 /* 32-bit: 26 to 100 children, 256 to 512 bytes */
44 /* 64-bit: 29 to 112 children, 512 to 1024 bytes */
45 RCU_JA_PIGEON
= 2, /* Type C */
46 /* 32-bit: 101 to 256 children, 1024 bytes */
47 /* 64-bit: 113 to 256 children, 2048 bytes */
48 /* Leaf nodes are implicit from their height in the tree */
51 RCU_JA_NULL
, /* not an encoded type, but keeps code regular */
55 enum cds_ja_type_class type_class
;
56 uint16_t min_child
; /* minimum number of children: 1 to 256 */
57 uint16_t max_child
; /* maximum number of children: 1 to 256 */
58 uint16_t max_linear_child
; /* per-pool max nr. children: 1 to 256 */
59 uint16_t order
; /* node size is (1 << order), in bytes */
60 uint16_t nr_pool_order
; /* number of pools */
61 uint16_t pool_size_order
; /* pool size */
65 * Number of least significant pointer bits reserved to represent the
68 #define JA_TYPE_BITS 3
69 #define JA_TYPE_MAX_NR (1UL << JA_TYPE_BITS)
70 #define JA_TYPE_MASK (JA_TYPE_MAX_NR - 1)
71 #define JA_PTR_MASK (~JA_TYPE_MASK)
73 #define JA_ENTRY_PER_NODE 256UL
74 #define JA_LOG2_BITS_PER_BYTE 3U
75 #define JA_BITS_PER_BYTE (1U << JA_LOG2_BITS_PER_BYTE)
77 #define JA_MAX_DEPTH 9 /* Maximum depth, including leafs */
80 * Entry for NULL node is at index 8 of the table. It is never encoded
83 #define NODE_INDEX_NULL 8
86 * Number of removals needed on a fallback node before we try to shrink
89 #define JA_FALLBACK_REMOVAL_COUNT 8
92 * Iteration on the array to find the right node size for the number of
93 * children stops when it reaches .max_child == 256 (this is the largest
94 * possible node size, which contains 256 children).
95 * The min_child overlaps with the previous max_child to provide an
96 * hysteresis loop to reallocation for patterns of cyclic add/removal
97 * within the same node.
98 * The node the index within the following arrays is represented on 3
99 * bits. It identifies the node type, min/max number of children, and
101 * The max_child values for the RCU_JA_POOL below result from
102 * statistical approximation: over million populations, the max_child
103 * covers between 97% and 99% of the populations generated. Therefore, a
104 * fallback should exist to cover the rare extreme population unbalance
105 * cases, but it will not have a major impact on speed nor space
106 * consumption, since those are rare cases.
109 #if (CAA_BITS_PER_LONG < 64)
110 /* 32-bit pointers */
112 ja_type_0_max_child
= 1,
113 ja_type_1_max_child
= 3,
114 ja_type_2_max_child
= 6,
115 ja_type_3_max_child
= 12,
116 ja_type_4_max_child
= 25,
117 ja_type_5_max_child
= 48,
118 ja_type_6_max_child
= 92,
119 ja_type_7_max_child
= 256,
120 ja_type_8_max_child
= 0, /* NULL */
124 ja_type_0_max_linear_child
= 1,
125 ja_type_1_max_linear_child
= 3,
126 ja_type_2_max_linear_child
= 6,
127 ja_type_3_max_linear_child
= 12,
128 ja_type_4_max_linear_child
= 25,
129 ja_type_5_max_linear_child
= 24,
130 ja_type_6_max_linear_child
= 23,
134 ja_type_5_nr_pool_order
= 1,
135 ja_type_6_nr_pool_order
= 2,
138 const struct cds_ja_type ja_types
[] = {
139 { .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, },
140 { .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, },
141 { .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, },
142 { .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, },
143 { .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, },
145 /* Pools may fill sooner than max_child */
146 { .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, },
147 { .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, },
150 * TODO: Upon node removal below min_child, if child pool is
151 * filled beyond capacity, we need to roll back to pigeon.
153 { .type_class
= RCU_JA_PIGEON
, .min_child
= 89, .max_child
= ja_type_7_max_child
, .order
= 10, },
155 { .type_class
= RCU_JA_NULL
, .min_child
= 0, .max_child
= ja_type_8_max_child
, },
157 #else /* !(CAA_BITS_PER_LONG < 64) */
158 /* 64-bit pointers */
160 ja_type_0_max_child
= 1,
161 ja_type_1_max_child
= 3,
162 ja_type_2_max_child
= 7,
163 ja_type_3_max_child
= 14,
164 ja_type_4_max_child
= 28,
165 ja_type_5_max_child
= 54,
166 ja_type_6_max_child
= 104,
167 ja_type_7_max_child
= 256,
168 ja_type_8_max_child
= 256,
172 ja_type_0_max_linear_child
= 1,
173 ja_type_1_max_linear_child
= 3,
174 ja_type_2_max_linear_child
= 7,
175 ja_type_3_max_linear_child
= 14,
176 ja_type_4_max_linear_child
= 28,
177 ja_type_5_max_linear_child
= 27,
178 ja_type_6_max_linear_child
= 26,
182 ja_type_5_nr_pool_order
= 1,
183 ja_type_6_nr_pool_order
= 2,
186 const struct cds_ja_type ja_types
[] = {
187 { .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, },
188 { .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, },
189 { .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, },
190 { .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, },
191 { .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, },
193 /* Pools may fill sooner than max_child. */
194 { .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, },
195 { .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, },
198 * TODO: Upon node removal below min_child, if child pool is
199 * filled beyond capacity, we need to roll back to pigeon.
201 { .type_class
= RCU_JA_PIGEON
, .min_child
= 101, .max_child
= ja_type_7_max_child
, .order
= 11, },
203 { .type_class
= RCU_JA_NULL
, .min_child
= 0, .max_child
= ja_type_8_max_child
, },
205 #endif /* !(BITS_PER_LONG < 64) */
207 static inline __attribute__((unused
))
208 void static_array_size_check(void)
210 CAA_BUILD_BUG_ON(CAA_ARRAY_SIZE(ja_types
) < JA_TYPE_MAX_NR
);
214 * The cds_ja_node contains the compressed node data needed for
215 * read-side. For linear and pool node configurations, it starts with a
216 * byte counting the number of children in the node. Then, the
217 * node-specific data is placed.
218 * The node mutex, if any is needed, protecting concurrent updated of
219 * each node is placed in a separate hash table indexed by node address.
220 * For the pigeon configuration, the number of children is also kept in
221 * a separate hash table, indexed by node address, because it is only
222 * required for updates.
225 #define DECLARE_LINEAR_NODE(index) \
228 uint8_t child_value[ja_type_## index ##_max_linear_child]; \
229 struct cds_ja_inode_flag *child_ptr[ja_type_## index ##_max_linear_child]; \
232 #define DECLARE_POOL_NODE(index) \
236 uint8_t child_value[ja_type_## index ##_max_linear_child]; \
237 struct cds_ja_inode_flag *child_ptr[ja_type_## index ##_max_linear_child]; \
238 } linear[1U << ja_type_## index ##_nr_pool_order]; \
241 struct cds_ja_inode
{
243 /* Linear configuration */
244 DECLARE_LINEAR_NODE(0) conf_0
;
245 DECLARE_LINEAR_NODE(1) conf_1
;
246 DECLARE_LINEAR_NODE(2) conf_2
;
247 DECLARE_LINEAR_NODE(3) conf_3
;
248 DECLARE_LINEAR_NODE(4) conf_4
;
250 /* Pool configuration */
251 DECLARE_POOL_NODE(5) conf_5
;
252 DECLARE_POOL_NODE(6) conf_6
;
254 /* Pigeon configuration */
256 struct cds_ja_inode_flag
*child
[ja_type_7_max_child
];
258 /* data aliasing nodes for computed accesses */
259 uint8_t data
[sizeof(struct cds_ja_inode_flag
*) * ja_type_7_max_child
];
264 struct cds_ja_inode_flag
*ja_node_flag(struct cds_ja_inode
*node
,
267 assert(type
< (1UL << JA_TYPE_BITS
));
268 return (struct cds_ja_inode_flag
*) (((unsigned long) node
) | type
);
272 struct cds_ja_inode
*ja_node_ptr(struct cds_ja_inode_flag
*node
)
274 return (struct cds_ja_inode
*) (((unsigned long) node
) & JA_PTR_MASK
);
278 unsigned long ja_node_type(struct cds_ja_inode_flag
*node
)
282 if (ja_node_ptr(node
) == NULL
) {
283 return NODE_INDEX_NULL
;
285 type
= (unsigned int) ((unsigned long) node
& JA_TYPE_MASK
);
286 assert(type
< (1UL << JA_TYPE_BITS
));
290 struct cds_ja_inode
*alloc_cds_ja_node(const struct cds_ja_type
*ja_type
)
292 return calloc(1U << ja_type
->order
, sizeof(char));
295 void free_cds_ja_node(struct cds_ja_inode
*node
)
300 #define __JA_ALIGN_MASK(v, mask) (((v) + (mask)) & ~(mask))
301 #define JA_ALIGN(v, align) __JA_ALIGN_MASK(v, (typeof(v)) (align) - 1)
302 #define __JA_FLOOR_MASK(v, mask) ((v) & ~(mask))
303 #define JA_FLOOR(v, align) __JA_FLOOR_MASK(v, (typeof(v)) (align) - 1)
306 uint8_t *align_ptr_size(uint8_t *ptr
)
308 return (uint8_t *) JA_ALIGN((unsigned long) ptr
, sizeof(void *));
312 uint8_t ja_linear_node_get_nr_child(const struct cds_ja_type
*type
,
313 struct cds_ja_inode
*node
)
315 assert(type
->type_class
== RCU_JA_LINEAR
|| type
->type_class
== RCU_JA_POOL
);
316 return CMM_LOAD_SHARED(node
->u
.data
[0]);
320 * The order in which values and pointers are does does not matter: if
321 * a value is missing, we return NULL. If a value is there, but its
322 * associated pointers is still NULL, we return NULL too.
325 struct cds_ja_inode_flag
*ja_linear_node_get_nth(const struct cds_ja_type
*type
,
326 struct cds_ja_inode
*node
,
327 struct cds_ja_inode_flag
***child_node_flag_ptr
,
332 struct cds_ja_inode_flag
**pointers
;
333 struct cds_ja_inode_flag
*ptr
;
336 assert(type
->type_class
== RCU_JA_LINEAR
|| type
->type_class
== RCU_JA_POOL
);
338 nr_child
= ja_linear_node_get_nr_child(type
, node
);
339 cmm_smp_rmb(); /* read nr_child before values and pointers */
340 assert(nr_child
<= type
->max_linear_child
);
341 assert(type
->type_class
!= RCU_JA_LINEAR
|| nr_child
>= type
->min_child
);
343 values
= &node
->u
.data
[1];
344 for (i
= 0; i
< nr_child
; i
++) {
345 if (CMM_LOAD_SHARED(values
[i
]) == n
)
350 pointers
= (struct cds_ja_inode_flag
**) align_ptr_size(&values
[type
->max_linear_child
]);
351 if (caa_unlikely(child_node_flag_ptr
))
352 *child_node_flag_ptr
= &pointers
[i
];
353 ptr
= rcu_dereference(pointers
[i
]);
354 assert(ja_node_ptr(ptr
) != NULL
);
359 void ja_linear_node_get_ith_pos(const struct cds_ja_type
*type
,
360 struct cds_ja_inode
*node
,
363 struct cds_ja_inode_flag
**iter
)
366 struct cds_ja_inode_flag
**pointers
;
368 assert(type
->type_class
== RCU_JA_LINEAR
|| type
->type_class
== RCU_JA_POOL
);
369 assert(i
< ja_linear_node_get_nr_child(type
, node
));
371 values
= &node
->u
.data
[1];
373 pointers
= (struct cds_ja_inode_flag
**) align_ptr_size(&values
[type
->max_linear_child
]);
378 struct cds_ja_inode_flag
*ja_pool_node_get_nth(const struct cds_ja_type
*type
,
379 struct cds_ja_inode
*node
,
380 struct cds_ja_inode_flag
***child_node_flag_ptr
,
383 struct cds_ja_inode
*linear
;
385 assert(type
->type_class
== RCU_JA_POOL
);
387 * TODO: currently, we select the pool by highest bits. We
388 * should support various encodings.
390 linear
= (struct cds_ja_inode
*)
391 &node
->u
.data
[((unsigned long) n
>> (CHAR_BIT
- type
->nr_pool_order
)) << type
->pool_size_order
];
392 return ja_linear_node_get_nth(type
, linear
, child_node_flag_ptr
, n
);
396 struct cds_ja_inode
*ja_pool_node_get_ith_pool(const struct cds_ja_type
*type
,
397 struct cds_ja_inode
*node
,
400 assert(type
->type_class
== RCU_JA_POOL
);
401 return (struct cds_ja_inode
*)
402 &node
->u
.data
[(unsigned int) i
<< type
->pool_size_order
];
406 struct cds_ja_inode_flag
*ja_pigeon_node_get_nth(const struct cds_ja_type
*type
,
407 struct cds_ja_inode
*node
,
408 struct cds_ja_inode_flag
***child_node_flag_ptr
,
411 struct cds_ja_inode_flag
**child_node_flag
;
413 assert(type
->type_class
== RCU_JA_PIGEON
);
414 child_node_flag
= &((struct cds_ja_inode_flag
**) node
->u
.data
)[n
];
415 dbg_printf("ja_pigeon_node_get_nth child_node_flag_ptr %p\n",
417 if (caa_unlikely(child_node_flag_ptr
) && *child_node_flag
)
418 *child_node_flag_ptr
= child_node_flag
;
419 return rcu_dereference(*child_node_flag
);
423 * ja_node_get_nth: get nth item from a node.
424 * node_flag is already rcu_dereference'd.
427 struct cds_ja_inode_flag
* ja_node_get_nth(struct cds_ja_inode_flag
*node_flag
,
428 struct cds_ja_inode_flag
***child_node_flag_ptr
,
431 unsigned int type_index
;
432 struct cds_ja_inode
*node
;
433 const struct cds_ja_type
*type
;
435 node
= ja_node_ptr(node_flag
);
436 assert(node
!= NULL
);
437 type_index
= ja_node_type(node_flag
);
438 type
= &ja_types
[type_index
];
440 switch (type
->type_class
) {
442 return ja_linear_node_get_nth(type
, node
,
443 child_node_flag_ptr
, n
);
445 return ja_pool_node_get_nth(type
, node
,
446 child_node_flag_ptr
, n
);
448 return ja_pigeon_node_get_nth(type
, node
,
449 child_node_flag_ptr
, n
);
452 return (void *) -1UL;
457 * TODO: use ja_get_nr_child to monitor limits triggering shrink
459 * Also use ja_get_nr_child to make the difference between resize and
460 * pool change of compaction bit(s).
463 unsigned int ja_get_nr_child(struct cds_ja_shadow_node
*shadow_node
)
465 return shadow_node
->nr_child
;
469 int ja_linear_node_set_nth(const struct cds_ja_type
*type
,
470 struct cds_ja_inode
*node
,
471 struct cds_ja_shadow_node
*shadow_node
,
473 struct cds_ja_inode_flag
*child_node_flag
)
476 uint8_t *values
, *nr_child_ptr
;
477 struct cds_ja_inode_flag
**pointers
;
480 assert(type
->type_class
== RCU_JA_LINEAR
|| type
->type_class
== RCU_JA_POOL
);
482 nr_child_ptr
= &node
->u
.data
[0];
483 dbg_printf("linear set nth: nr_child_ptr %p\n", nr_child_ptr
);
484 nr_child
= *nr_child_ptr
;
485 assert(nr_child
<= type
->max_linear_child
);
487 values
= &node
->u
.data
[1];
488 for (i
= 0; i
< nr_child
; i
++) {
492 if (nr_child
>= type
->max_linear_child
) {
493 /* No space left in this node type */
496 pointers
= (struct cds_ja_inode_flag
**) align_ptr_size(&values
[type
->max_linear_child
]);
497 assert(pointers
[nr_child
] == NULL
);
498 rcu_assign_pointer(pointers
[nr_child
], child_node_flag
);
499 CMM_STORE_SHARED(values
[nr_child
], n
);
500 cmm_smp_wmb(); /* write value and pointer before nr_child */
501 CMM_STORE_SHARED(*nr_child_ptr
, nr_child
+ 1);
502 shadow_node
->nr_child
++;
503 dbg_printf("linear set nth: %u child, shadow: %u child, for node %p shadow %p\n",
504 (unsigned int) CMM_LOAD_SHARED(*nr_child_ptr
),
505 (unsigned int) shadow_node
->nr_child
,
512 int ja_pool_node_set_nth(const struct cds_ja_type
*type
,
513 struct cds_ja_inode
*node
,
514 struct cds_ja_shadow_node
*shadow_node
,
516 struct cds_ja_inode_flag
*child_node_flag
)
518 struct cds_ja_inode
*linear
;
520 assert(type
->type_class
== RCU_JA_POOL
);
521 linear
= (struct cds_ja_inode
*)
522 &node
->u
.data
[((unsigned long) n
>> (CHAR_BIT
- type
->nr_pool_order
)) << type
->pool_size_order
];
523 return ja_linear_node_set_nth(type
, linear
, shadow_node
,
528 int ja_pigeon_node_set_nth(const struct cds_ja_type
*type
,
529 struct cds_ja_inode
*node
,
530 struct cds_ja_shadow_node
*shadow_node
,
532 struct cds_ja_inode_flag
*child_node_flag
)
534 struct cds_ja_inode_flag
**ptr
;
536 assert(type
->type_class
== RCU_JA_PIGEON
);
537 ptr
= &((struct cds_ja_inode_flag
**) node
->u
.data
)[n
];
540 rcu_assign_pointer(*ptr
, child_node_flag
);
541 shadow_node
->nr_child
++;
546 * _ja_node_set_nth: set nth item within a node. Return an error
547 * (negative error value) if it is already there.
548 * TODO: exclusive access on node.
551 int _ja_node_set_nth(const struct cds_ja_type
*type
,
552 struct cds_ja_inode
*node
,
553 struct cds_ja_shadow_node
*shadow_node
,
555 struct cds_ja_inode_flag
*child_node_flag
)
557 switch (type
->type_class
) {
559 return ja_linear_node_set_nth(type
, node
, shadow_node
, n
,
562 return ja_pool_node_set_nth(type
, node
, shadow_node
, n
,
565 return ja_pigeon_node_set_nth(type
, node
, shadow_node
, n
,
578 * ja_node_recompact_add: recompact a node, adding a new child.
579 * TODO: for pool type, take selection bit(s) into account.
580 * Return 0 on success, -ENOENT if need to retry, or other negative
581 * error value otherwise.
584 int ja_node_recompact_add(struct cds_ja
*ja
,
585 unsigned int old_type_index
,
586 const struct cds_ja_type
*old_type
,
587 struct cds_ja_inode
*old_node
,
588 struct cds_ja_shadow_node
*shadow_node
,
589 struct cds_ja_inode_flag
**old_node_flag
, uint8_t n
,
590 struct cds_ja_inode_flag
*child_node_flag
)
592 unsigned int new_type_index
;
593 struct cds_ja_inode
*new_node
;
594 struct cds_ja_shadow_node
*new_shadow_node
;
595 const struct cds_ja_type
*new_type
;
596 struct cds_ja_inode_flag
*new_node_flag
;
600 if (!shadow_node
|| old_type_index
== NODE_INDEX_NULL
) {
603 new_type_index
= old_type_index
+ 1;
606 retry
: /* for fallback */
607 dbg_printf("Recompact from type %d to type %d\n",
608 old_type_index
, new_type_index
);
609 new_type
= &ja_types
[new_type_index
];
610 new_node
= alloc_cds_ja_node(new_type
);
613 new_node_flag
= ja_node_flag(new_node
, new_type_index
);
615 dbg_printf("Recompact inherit lock from %p\n", shadow_node
);
616 new_shadow_node
= rcuja_shadow_set(ja
->ht
, new_node
, shadow_node
);
617 if (!new_shadow_node
) {
622 new_shadow_node
->fallback_removal_count
=
623 JA_FALLBACK_REMOVAL_COUNT
;
625 assert(old_type
->type_class
!= RCU_JA_PIGEON
);
626 switch (old_type
->type_class
) {
630 ja_linear_node_get_nr_child(old_type
, old_node
);
633 for (i
= 0; i
< nr_child
; i
++) {
634 struct cds_ja_inode_flag
*iter
;
637 ja_linear_node_get_ith_pos(old_type
, old_node
, i
, &v
, &iter
);
640 ret
= _ja_node_set_nth(new_type
, new_node
,
643 if (new_type
->type_class
== RCU_JA_POOL
&& ret
) {
644 goto fallback_toosmall
;
652 unsigned int pool_nr
;
654 for (pool_nr
= 0; pool_nr
< (1U << old_type
->nr_pool_order
); pool_nr
++) {
655 struct cds_ja_inode
*pool
=
656 ja_pool_node_get_ith_pool(old_type
,
659 ja_linear_node_get_nr_child(old_type
, pool
);
662 for (j
= 0; j
< nr_child
; j
++) {
663 struct cds_ja_inode_flag
*iter
;
666 ja_linear_node_get_ith_pos(old_type
, pool
,
670 ret
= _ja_node_set_nth(new_type
, new_node
,
673 if (new_type
->type_class
== RCU_JA_POOL
675 goto fallback_toosmall
;
683 /* Nothing to copy */
693 ret
= _ja_node_set_nth(new_type
, new_node
,
697 /* Return pointer to new recompacted new through old_node_flag */
698 *old_node_flag
= new_node_flag
;
700 ret
= rcuja_shadow_clear(ja
->ht
, old_node
, shadow_node
,
701 RCUJA_SHADOW_CLEAR_FREE_NODE
);
710 /* fallback if next pool is too small */
711 ret
= rcuja_shadow_clear(ja
->ht
, new_node
, new_shadow_node
,
712 RCUJA_SHADOW_CLEAR_FREE_NODE
);
715 /* Last type: pigeon */
716 new_type_index
= (1UL << JA_TYPE_BITS
) - 1;
717 dbg_printf("Fallback to type %d\n", new_type_index
);
718 uatomic_inc(&ja
->nr_fallback
);
724 * Return 0 on success, -ENOENT if need to retry, or other negative
725 * error value otherwise.
728 int ja_node_set_nth(struct cds_ja
*ja
,
729 struct cds_ja_inode_flag
**node_flag
, uint8_t n
,
730 struct cds_ja_inode_flag
*child_node_flag
,
731 struct cds_ja_shadow_node
*shadow_node
)
734 unsigned int type_index
;
735 const struct cds_ja_type
*type
;
736 struct cds_ja_inode
*node
;
738 dbg_printf("ja_node_set_nth for n=%u, node %p, shadow %p\n",
739 (unsigned int) n
, ja_node_ptr(*node_flag
), shadow_node
);
741 node
= ja_node_ptr(*node_flag
);
742 type_index
= ja_node_type(*node_flag
);
743 type
= &ja_types
[type_index
];
744 ret
= _ja_node_set_nth(type
, node
, shadow_node
,
746 if (ret
== -ENOSPC
) {
747 /* Not enough space in node, need to recompact. */
748 ret
= ja_node_recompact_add(ja
, type_index
, type
, node
,
749 shadow_node
, node_flag
, n
, child_node_flag
);
754 struct cds_hlist_head
*cds_ja_lookup(struct cds_ja
*ja
, uint64_t key
)
756 unsigned int tree_depth
, i
;
757 struct cds_ja_inode_flag
*node_flag
;
759 if (caa_unlikely(key
> ja
->key_max
))
761 tree_depth
= ja
->tree_depth
;
762 node_flag
= rcu_dereference(ja
->root
);
764 /* level 0: root node */
765 if (!ja_node_ptr(node_flag
))
768 for (i
= 1; i
< tree_depth
; i
++) {
771 iter_key
= (uint8_t) (key
>> (JA_BITS_PER_BYTE
* (tree_depth
- i
- 1)));
772 node_flag
= ja_node_get_nth(node_flag
, NULL
,
774 dbg_printf("cds_ja_lookup iter key lookup %u finds node_flag %p\n",
775 (unsigned int) iter_key
, node_flag
);
776 if (!ja_node_ptr(node_flag
))
780 /* Last level lookup succeded. We got an actual match. */
781 return (struct cds_hlist_head
*) node_flag
;
785 * We reached an unpopulated node. Create it and the children we need,
786 * and then attach the entire branch to the current node. This may
787 * trigger recompaction of the current node. Locks needed: node lock
788 * (for add), and, possibly, parent node lock (to update pointer due to
789 * node recompaction).
791 * First take node lock, check if recompaction is needed, then take
792 * parent lock (if needed). Then we can proceed to create the new
793 * branch. Publish the new branch, and release locks.
794 * TODO: we currently always take the parent lock even when not needed.
797 int ja_attach_node(struct cds_ja
*ja
,
798 struct cds_ja_inode_flag
**node_flag_ptr
,
799 struct cds_ja_inode_flag
*node_flag
,
800 struct cds_ja_inode_flag
*parent_node_flag
,
803 struct cds_ja_node
*child_node
)
805 struct cds_ja_shadow_node
*shadow_node
= NULL
,
806 *parent_shadow_node
= NULL
,
808 struct cds_ja_inode
*node
= ja_node_ptr(node_flag
);
809 struct cds_ja_inode
*parent_node
= ja_node_ptr(parent_node_flag
);
810 struct cds_hlist_head head
;
811 struct cds_ja_inode_flag
*iter_node_flag
, *iter_dest_node_flag
;
813 struct cds_ja_inode_flag
*created_nodes
[JA_MAX_DEPTH
];
814 int nr_created_nodes
= 0;
816 dbg_printf("Attach node at level %u (node %p, node_flag %p)\n",
817 level
, node
, node_flag
);
820 shadow_node
= rcuja_shadow_lookup_lock(ja
->ht
, node
);
826 parent_shadow_node
= rcuja_shadow_lookup_lock(ja
->ht
,
828 if (!parent_shadow_node
) {
834 /* Create new branch, starting from bottom */
835 CDS_INIT_HLIST_HEAD(&head
);
836 cds_hlist_add_head_rcu(&child_node
->list
, &head
);
837 iter_node_flag
= (struct cds_ja_inode_flag
*) head
.next
;
839 /* Create shadow node for the leaf node */
840 dbg_printf("leaf shadow node creation\n");
841 iter_shadow_node
= rcuja_shadow_set(ja
->ht
,
842 ja_node_ptr(iter_node_flag
), NULL
);
843 if (!iter_shadow_node
) {
847 created_nodes
[nr_created_nodes
++] = iter_node_flag
;
849 for (i
= ja
->tree_depth
; i
> (int) level
; i
--) {
852 iter_key
= (uint8_t) (key
>> (JA_BITS_PER_BYTE
* (ja
->tree_depth
- i
)));
853 dbg_printf("branch creation level %d, key %u\n",
854 i
- 1, (unsigned int) iter_key
);
855 iter_dest_node_flag
= NULL
;
856 ret
= ja_node_set_nth(ja
, &iter_dest_node_flag
,
862 created_nodes
[nr_created_nodes
++] = iter_dest_node_flag
;
863 iter_node_flag
= iter_dest_node_flag
;
869 iter_key
= (uint8_t) (key
>> (JA_BITS_PER_BYTE
* (ja
->tree_depth
- level
)));
870 /* We need to use set_nth on the previous level. */
871 iter_dest_node_flag
= node_flag
;
872 ret
= ja_node_set_nth(ja
, &iter_dest_node_flag
,
878 created_nodes
[nr_created_nodes
++] = iter_dest_node_flag
;
879 iter_node_flag
= iter_dest_node_flag
;
882 /* Publish new branch */
883 dbg_printf("Publish branch %p, replacing %p\n",
884 iter_node_flag
, *node_flag_ptr
);
885 rcu_assign_pointer(*node_flag_ptr
, iter_node_flag
);
892 for (i
= 0; i
< nr_created_nodes
; i
++) {
896 flags
= RCUJA_SHADOW_CLEAR_FREE_LOCK
;
898 flags
|= RCUJA_SHADOW_CLEAR_FREE_NODE
;
899 tmpret
= rcuja_shadow_clear(ja
->ht
,
900 ja_node_ptr(created_nodes
[i
]),
906 if (parent_shadow_node
)
907 rcuja_shadow_unlock(parent_shadow_node
);
910 rcuja_shadow_unlock(shadow_node
);
916 * Lock the hlist head shadow node mutex, and add node to list of
917 * duplicates. Failure can happen if concurrent removal removes the last
918 * node with same key before we get the lock.
919 * Return 0 on success, negative error value on failure.
922 int ja_chain_node(struct cds_ja
*ja
,
923 struct cds_hlist_head
*head
,
924 struct cds_ja_node
*node
)
926 struct cds_ja_shadow_node
*shadow_node
;
928 shadow_node
= rcuja_shadow_lookup_lock(ja
->ht
,
929 (struct cds_ja_inode
*) head
);
932 cds_hlist_add_head_rcu(&node
->list
, head
);
933 rcuja_shadow_unlock(shadow_node
);
937 int cds_ja_add(struct cds_ja
*ja
, uint64_t key
,
938 struct cds_ja_node
*new_node
)
940 unsigned int tree_depth
, i
;
941 struct cds_ja_inode_flag
**node_flag_ptr
; /* in parent */
942 struct cds_ja_inode_flag
*node_flag
,
947 if (caa_unlikely(key
> ja
->key_max
))
949 tree_depth
= ja
->tree_depth
;
952 dbg_printf("cds_ja_add attempt: key %" PRIu64
", node %p\n",
954 parent2_node_flag
= NULL
;
956 (struct cds_ja_inode_flag
*) &ja
->root
; /* Use root ptr address as key for mutex */
957 node_flag_ptr
= &ja
->root
;
958 node_flag
= rcu_dereference(*node_flag_ptr
);
960 /* Iterate on all internal levels */
961 for (i
= 1; i
< tree_depth
; i
++) {
964 dbg_printf("cds_ja_add iter node_flag_ptr %p node_flag %p\n",
965 *node_flag_ptr
, node_flag
);
966 if (!ja_node_ptr(node_flag
)) {
967 ret
= ja_attach_node(ja
, node_flag_ptr
,
968 parent_node_flag
, parent2_node_flag
,
970 if (ret
== -ENOENT
|| ret
== -EEXIST
)
975 iter_key
= (uint8_t) (key
>> (JA_BITS_PER_BYTE
* (tree_depth
- i
- 1)));
976 parent2_node_flag
= parent_node_flag
;
977 parent_node_flag
= node_flag
;
978 node_flag
= ja_node_get_nth(node_flag
,
981 dbg_printf("cds_ja_add iter key lookup %u finds node_flag %p node_flag_ptr %p\n",
982 (unsigned int) iter_key
, node_flag
, *node_flag_ptr
);
986 * We reached bottom of tree, simply add node to last internal
987 * level, or chain it if key is already present.
989 if (!ja_node_ptr(node_flag
)) {
990 dbg_printf("cds_ja_add last node_flag_ptr %p node_flag %p\n",
991 *node_flag_ptr
, node_flag
);
992 ret
= ja_attach_node(ja
, node_flag_ptr
, parent_node_flag
,
993 parent2_node_flag
, key
, i
, new_node
);
995 ret
= ja_chain_node(ja
,
996 (struct cds_hlist_head
*) ja_node_ptr(node_flag
),
1005 struct cds_ja
*_cds_ja_new(unsigned int key_bits
,
1006 const struct rcu_flavor_struct
*flavor
)
1010 struct cds_ja_shadow_node
*root_shadow_node
;
1012 ja
= calloc(sizeof(*ja
), 1);
1018 ja
->key_max
= UINT8_MAX
;
1021 ja
->key_max
= UINT16_MAX
;
1024 ja
->key_max
= UINT32_MAX
;
1027 ja
->key_max
= UINT64_MAX
;
1033 /* ja->root is NULL */
1034 /* tree_depth 0 is for pointer to root node */
1035 ja
->tree_depth
= (key_bits
>> JA_LOG2_BITS_PER_BYTE
) + 1;
1036 assert(ja
->tree_depth
<= JA_MAX_DEPTH
);
1037 ja
->ht
= rcuja_create_ht(flavor
);
1042 * Note: we should not free this node until judy array destroy.
1044 root_shadow_node
= rcuja_shadow_set(ja
->ht
,
1045 ja_node_ptr((struct cds_ja_inode_flag
*) &ja
->root
),
1047 if (!root_shadow_node
) {
1051 root_shadow_node
->is_root
= 1;
1056 ret
= rcuja_delete_ht(ja
->ht
);
1066 * There should be no more concurrent add to the judy array while it is
1067 * being destroyed (ensured by the caller).
1069 int cds_ja_destroy(struct cds_ja
*ja
)
1073 rcuja_shadow_prune(ja
->ht
,
1074 RCUJA_SHADOW_CLEAR_FREE_NODE
| RCUJA_SHADOW_CLEAR_FREE_LOCK
);
1075 ret
= rcuja_delete_ht(ja
->ht
);
1078 if (uatomic_read(&ja
->nr_fallback
))
1080 "[warning] RCU Judy Array used %lu fallback node(s)\n",
1081 uatomic_read(&ja
->nr_fallback
));