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 * Iteration on the array to find the right node size for the number of
66 * children stops when it reaches .max_child == 256 (this is the largest
67 * possible node size, which contains 256 children).
68 * The min_child overlaps with the previous max_child to provide an
69 * hysteresis loop to reallocation for patterns of cyclic add/removal
70 * within the same node.
71 * The node the index within the following arrays is represented on 3
72 * bits. It identifies the node type, min/max number of children, and
74 * The max_child values for the RCU_JA_POOL below result from
75 * statistical approximation: over million populations, the max_child
76 * covers between 97% and 99% of the populations generated. Therefore, a
77 * fallback should exist to cover the rare extreme population unbalance
78 * cases, but it will not have a major impact on speed nor space
79 * consumption, since those are rare cases.
82 #if (CAA_BITS_PER_LONG < 64)
85 ja_type_0_max_child
= 1,
86 ja_type_1_max_child
= 3,
87 ja_type_2_max_child
= 6,
88 ja_type_3_max_child
= 12,
89 ja_type_4_max_child
= 25,
90 ja_type_5_max_child
= 48,
91 ja_type_6_max_child
= 92,
92 ja_type_7_max_child
= 256,
93 ja_type_8_max_child
= 0, /* NULL */
97 ja_type_0_max_linear_child
= 1,
98 ja_type_1_max_linear_child
= 3,
99 ja_type_2_max_linear_child
= 6,
100 ja_type_3_max_linear_child
= 12,
101 ja_type_4_max_linear_child
= 25,
102 ja_type_5_max_linear_child
= 24,
103 ja_type_6_max_linear_child
= 23,
107 ja_type_5_nr_pool_order
= 1,
108 ja_type_6_nr_pool_order
= 2,
111 const struct cds_ja_type ja_types
[] = {
112 { .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, },
113 { .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, },
114 { .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, },
115 { .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, },
116 { .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, },
118 /* Pools may fill sooner than max_child */
119 { .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, },
120 { .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, },
123 * TODO: Upon node removal below min_child, if child pool is
124 * filled beyond capacity, we need to roll back to pigeon.
126 { .type_class
= RCU_JA_PIGEON
, .min_child
= 89, .max_child
= ja_type_7_max_child
, .order
= 10, },
128 { .type_class
= RCU_JA_NULL
, .min_child
= 0, .max_child
= ja_type_8_max_child
, },
130 #else /* !(CAA_BITS_PER_LONG < 64) */
131 /* 64-bit pointers */
133 ja_type_0_max_child
= 1,
134 ja_type_1_max_child
= 3,
135 ja_type_2_max_child
= 7,
136 ja_type_3_max_child
= 14,
137 ja_type_4_max_child
= 28,
138 ja_type_5_max_child
= 54,
139 ja_type_6_max_child
= 104,
140 ja_type_7_max_child
= 256,
141 ja_type_8_max_child
= 256,
145 ja_type_0_max_linear_child
= 1,
146 ja_type_1_max_linear_child
= 3,
147 ja_type_2_max_linear_child
= 7,
148 ja_type_3_max_linear_child
= 14,
149 ja_type_4_max_linear_child
= 28,
150 ja_type_5_max_linear_child
= 27,
151 ja_type_6_max_linear_child
= 26,
155 ja_type_5_nr_pool_order
= 1,
156 ja_type_6_nr_pool_order
= 2,
159 const struct cds_ja_type ja_types
[] = {
160 { .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, },
161 { .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, },
162 { .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, },
163 { .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, },
164 { .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, },
166 /* Pools may fill sooner than max_child. */
167 { .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, },
168 { .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, },
171 * TODO: Upon node removal below min_child, if child pool is
172 * filled beyond capacity, we need to roll back to pigeon.
174 { .type_class
= RCU_JA_PIGEON
, .min_child
= 101, .max_child
= ja_type_7_max_child
, .order
= 11, },
176 { .type_class
= RCU_JA_NULL
, .min_child
= 0, .max_child
= ja_type_8_max_child
, },
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
);
187 * The cds_ja_node contains the compressed node data needed for
188 * read-side. For linear and pool node configurations, it starts with a
189 * byte counting the number of children in the node. Then, the
190 * node-specific data is placed.
191 * The node mutex, if any is needed, protecting concurrent updated of
192 * each node is placed in a separate hash table indexed by node address.
193 * For the pigeon configuration, the number of children is also kept in
194 * a separate hash table, indexed by node address, because it is only
195 * required for updates.
198 #define DECLARE_LINEAR_NODE(index) \
201 uint8_t child_value[ja_type_## index ##_max_linear_child]; \
202 struct cds_ja_inode_flag *child_ptr[ja_type_## index ##_max_linear_child]; \
205 #define DECLARE_POOL_NODE(index) \
209 uint8_t child_value[ja_type_## index ##_max_linear_child]; \
210 struct cds_ja_inode_flag *child_ptr[ja_type_## index ##_max_linear_child]; \
211 } linear[1U << ja_type_## index ##_nr_pool_order]; \
214 struct cds_ja_inode
{
216 /* Linear configuration */
217 DECLARE_LINEAR_NODE(0) conf_0
;
218 DECLARE_LINEAR_NODE(1) conf_1
;
219 DECLARE_LINEAR_NODE(2) conf_2
;
220 DECLARE_LINEAR_NODE(3) conf_3
;
221 DECLARE_LINEAR_NODE(4) conf_4
;
223 /* Pool configuration */
224 DECLARE_POOL_NODE(5) conf_5
;
225 DECLARE_POOL_NODE(6) conf_6
;
227 /* Pigeon configuration */
229 struct cds_ja_inode_flag
*child
[ja_type_7_max_child
];
231 /* data aliasing nodes for computed accesses */
232 uint8_t data
[sizeof(struct cds_ja_inode_flag
*) * ja_type_7_max_child
];
242 struct cds_ja_inode
*alloc_cds_ja_node(const struct cds_ja_type
*ja_type
)
244 return calloc(1U << ja_type
->order
, sizeof(char));
247 void free_cds_ja_node(struct cds_ja_inode
*node
)
252 #define __JA_ALIGN_MASK(v, mask) (((v) + (mask)) & ~(mask))
253 #define JA_ALIGN(v, align) __JA_ALIGN_MASK(v, (typeof(v)) (align) - 1)
254 #define __JA_FLOOR_MASK(v, mask) ((v) & ~(mask))
255 #define JA_FLOOR(v, align) __JA_FLOOR_MASK(v, (typeof(v)) (align) - 1)
258 uint8_t *align_ptr_size(uint8_t *ptr
)
260 return (uint8_t *) JA_ALIGN((unsigned long) ptr
, sizeof(void *));
264 uint8_t ja_linear_node_get_nr_child(const struct cds_ja_type
*type
,
265 struct cds_ja_inode
*node
)
267 assert(type
->type_class
== RCU_JA_LINEAR
|| type
->type_class
== RCU_JA_POOL
);
268 return rcu_dereference(node
->u
.data
[0]);
272 * The order in which values and pointers are does does not matter: if
273 * a value is missing, we return NULL. If a value is there, but its
274 * associated pointers is still NULL, we return NULL too.
277 struct cds_ja_inode_flag
*ja_linear_node_get_nth(const struct cds_ja_type
*type
,
278 struct cds_ja_inode
*node
,
279 struct cds_ja_inode_flag
***child_node_flag_ptr
,
284 struct cds_ja_inode_flag
**pointers
;
285 struct cds_ja_inode_flag
*ptr
;
288 assert(type
->type_class
== RCU_JA_LINEAR
|| type
->type_class
== RCU_JA_POOL
);
290 nr_child
= ja_linear_node_get_nr_child(type
, node
);
291 cmm_smp_rmb(); /* read nr_child before values and pointers */
292 assert(nr_child
<= type
->max_linear_child
);
293 assert(type
->type_class
!= RCU_JA_LINEAR
|| nr_child
>= type
->min_child
);
295 values
= &node
->u
.data
[1];
296 for (i
= 0; i
< nr_child
; i
++) {
297 if (CMM_LOAD_SHARED(values
[i
]) == n
)
302 pointers
= (struct cds_ja_inode_flag
**) align_ptr_size(&values
[type
->max_linear_child
]);
303 ptr
= rcu_dereference(pointers
[i
]);
304 if (caa_unlikely(child_node_flag_ptr
) && ptr
)
305 *child_node_flag_ptr
= &pointers
[i
];
310 void ja_linear_node_get_ith_pos(const struct cds_ja_type
*type
,
311 struct cds_ja_inode
*node
,
314 struct cds_ja_inode_flag
**iter
)
317 struct cds_ja_inode_flag
**pointers
;
319 assert(type
->type_class
== RCU_JA_LINEAR
|| type
->type_class
== RCU_JA_POOL
);
320 assert(i
< ja_linear_node_get_nr_child(type
, node
));
322 values
= &node
->u
.data
[1];
324 pointers
= (struct cds_ja_inode_flag
**) align_ptr_size(&values
[type
->max_linear_child
]);
329 struct cds_ja_inode_flag
*ja_pool_node_get_nth(const struct cds_ja_type
*type
,
330 struct cds_ja_inode
*node
,
331 struct cds_ja_inode_flag
***child_node_flag_ptr
,
334 struct cds_ja_inode
*linear
;
336 assert(type
->type_class
== RCU_JA_POOL
);
338 * TODO: currently, we select the pool by highest bits. We
339 * should support various encodings.
341 linear
= (struct cds_ja_inode
*)
342 &node
->u
.data
[((unsigned long) n
>> (CHAR_BIT
- type
->nr_pool_order
)) << type
->pool_size_order
];
343 return ja_linear_node_get_nth(type
, linear
, child_node_flag_ptr
, n
);
347 struct cds_ja_inode
*ja_pool_node_get_ith_pool(const struct cds_ja_type
*type
,
348 struct cds_ja_inode
*node
,
351 assert(type
->type_class
== RCU_JA_POOL
);
352 return (struct cds_ja_inode
*)
353 &node
->u
.data
[(unsigned int) i
<< type
->pool_size_order
];
357 struct cds_ja_inode_flag
*ja_pigeon_node_get_nth(const struct cds_ja_type
*type
,
358 struct cds_ja_inode
*node
,
359 struct cds_ja_inode_flag
***child_node_flag_ptr
,
362 struct cds_ja_inode_flag
**child_node_flag
;
364 assert(type
->type_class
== RCU_JA_PIGEON
);
365 child_node_flag
= &((struct cds_ja_inode_flag
**) node
->u
.data
)[n
];
366 dbg_printf("ja_pigeon_node_get_nth child_node_flag_ptr %p\n",
368 if (caa_unlikely(child_node_flag_ptr
) && *child_node_flag
)
369 *child_node_flag_ptr
= child_node_flag
;
370 return rcu_dereference(*child_node_flag
);
374 struct cds_ja_inode_flag
*ja_pigeon_node_get_ith_pos(const struct cds_ja_type
*type
,
375 struct cds_ja_inode
*node
,
378 return ja_pigeon_node_get_nth(type
, node
, NULL
, i
);
382 * ja_node_get_nth: get nth item from a node.
383 * node_flag is already rcu_dereference'd.
386 struct cds_ja_inode_flag
* ja_node_get_nth(struct cds_ja_inode_flag
*node_flag
,
387 struct cds_ja_inode_flag
***child_node_flag_ptr
,
390 unsigned int type_index
;
391 struct cds_ja_inode
*node
;
392 const struct cds_ja_type
*type
;
394 node
= ja_node_ptr(node_flag
);
395 assert(node
!= NULL
);
396 type_index
= ja_node_type(node_flag
);
397 type
= &ja_types
[type_index
];
399 switch (type
->type_class
) {
401 return ja_linear_node_get_nth(type
, node
,
402 child_node_flag_ptr
, n
);
404 return ja_pool_node_get_nth(type
, node
,
405 child_node_flag_ptr
, n
);
407 return ja_pigeon_node_get_nth(type
, node
,
408 child_node_flag_ptr
, n
);
411 return (void *) -1UL;
416 int ja_linear_node_set_nth(const struct cds_ja_type
*type
,
417 struct cds_ja_inode
*node
,
418 struct cds_ja_shadow_node
*shadow_node
,
420 struct cds_ja_inode_flag
*child_node_flag
)
423 uint8_t *values
, *nr_child_ptr
;
424 struct cds_ja_inode_flag
**pointers
;
425 unsigned int i
, unused
= 0;
427 assert(type
->type_class
== RCU_JA_LINEAR
|| type
->type_class
== RCU_JA_POOL
);
429 nr_child_ptr
= &node
->u
.data
[0];
430 dbg_printf("linear set nth: nr_child_ptr %p\n", nr_child_ptr
);
431 nr_child
= *nr_child_ptr
;
432 assert(nr_child
<= type
->max_linear_child
);
434 values
= &node
->u
.data
[1];
435 pointers
= (struct cds_ja_inode_flag
**) align_ptr_size(&values
[type
->max_linear_child
]);
436 /* Check if node value is already populated */
437 for (i
= 0; i
< nr_child
; i
++) {
438 if (values
[i
] == n
) {
448 if (i
== nr_child
&& nr_child
>= type
->max_linear_child
) {
450 return -ERANGE
; /* recompact node */
452 return -ENOSPC
; /* No space left in this node type */
455 assert(pointers
[i
] == NULL
);
456 rcu_assign_pointer(pointers
[i
], child_node_flag
);
457 /* If we expanded the nr_child, increment it */
459 CMM_STORE_SHARED(values
[nr_child
], n
);
460 /* write pointer and value before nr_child */
462 CMM_STORE_SHARED(*nr_child_ptr
, nr_child
+ 1);
464 shadow_node
->nr_child
++;
465 dbg_printf("linear set nth: %u child, shadow: %u child, for node %p shadow %p\n",
466 (unsigned int) CMM_LOAD_SHARED(*nr_child_ptr
),
467 (unsigned int) shadow_node
->nr_child
,
474 int ja_pool_node_set_nth(const struct cds_ja_type
*type
,
475 struct cds_ja_inode
*node
,
476 struct cds_ja_shadow_node
*shadow_node
,
478 struct cds_ja_inode_flag
*child_node_flag
)
480 struct cds_ja_inode
*linear
;
482 assert(type
->type_class
== RCU_JA_POOL
);
483 linear
= (struct cds_ja_inode
*)
484 &node
->u
.data
[((unsigned long) n
>> (CHAR_BIT
- type
->nr_pool_order
)) << type
->pool_size_order
];
485 return ja_linear_node_set_nth(type
, linear
, shadow_node
,
490 int ja_pigeon_node_set_nth(const struct cds_ja_type
*type
,
491 struct cds_ja_inode
*node
,
492 struct cds_ja_shadow_node
*shadow_node
,
494 struct cds_ja_inode_flag
*child_node_flag
)
496 struct cds_ja_inode_flag
**ptr
;
498 assert(type
->type_class
== RCU_JA_PIGEON
);
499 ptr
= &((struct cds_ja_inode_flag
**) node
->u
.data
)[n
];
502 rcu_assign_pointer(*ptr
, child_node_flag
);
503 shadow_node
->nr_child
++;
508 * _ja_node_set_nth: set nth item within a node. Return an error
509 * (negative error value) if it is already there.
512 int _ja_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 switch (type
->type_class
) {
520 return ja_linear_node_set_nth(type
, node
, shadow_node
, n
,
523 return ja_pool_node_set_nth(type
, node
, shadow_node
, n
,
526 return ja_pigeon_node_set_nth(type
, node
, shadow_node
, n
,
539 int ja_linear_node_clear_ptr(const struct cds_ja_type
*type
,
540 struct cds_ja_inode
*node
,
541 struct cds_ja_shadow_node
*shadow_node
,
542 struct cds_ja_inode_flag
**node_flag_ptr
)
545 uint8_t *nr_child_ptr
;
547 assert(type
->type_class
== RCU_JA_LINEAR
|| type
->type_class
== RCU_JA_POOL
);
549 nr_child_ptr
= &node
->u
.data
[0];
550 dbg_printf("linear clear ptr: nr_child_ptr %p\n", nr_child_ptr
);
551 nr_child
= *nr_child_ptr
;
552 assert(nr_child
<= type
->max_linear_child
);
554 if (shadow_node
->fallback_removal_count
) {
555 shadow_node
->fallback_removal_count
--;
557 if (shadow_node
->nr_child
<= type
->min_child
) {
558 /* We need to try recompacting the node */
562 assert(*node_flag_ptr
!= NULL
);
563 rcu_assign_pointer(*node_flag_ptr
, NULL
);
565 * Value and nr_child are never changed (would cause ABA issue).
566 * Instead, we leave the pointer to NULL and recompact the node
567 * once in a while. It is allowed to set a NULL pointer to a new
568 * value without recompaction though.
569 * Only update the shadow node accounting.
571 shadow_node
->nr_child
--;
572 dbg_printf("linear clear ptr: %u child, shadow: %u child, for node %p shadow %p\n",
573 (unsigned int) CMM_LOAD_SHARED(*nr_child_ptr
),
574 (unsigned int) shadow_node
->nr_child
,
581 int ja_pool_node_clear_ptr(const struct cds_ja_type
*type
,
582 struct cds_ja_inode
*node
,
583 struct cds_ja_shadow_node
*shadow_node
,
584 struct cds_ja_inode_flag
**node_flag_ptr
,
587 struct cds_ja_inode
*linear
;
589 assert(type
->type_class
== RCU_JA_POOL
);
590 linear
= (struct cds_ja_inode
*)
591 &node
->u
.data
[((unsigned long) n
>> (CHAR_BIT
- type
->nr_pool_order
)) << type
->pool_size_order
];
592 return ja_linear_node_clear_ptr(type
, linear
, shadow_node
, node_flag_ptr
);
596 int ja_pigeon_node_clear_ptr(const struct cds_ja_type
*type
,
597 struct cds_ja_inode
*node
,
598 struct cds_ja_shadow_node
*shadow_node
,
599 struct cds_ja_inode_flag
**node_flag_ptr
)
601 assert(type
->type_class
== RCU_JA_PIGEON
);
602 dbg_printf("ja_pigeon_node_clear_ptr: clearing ptr: %p\n", *node_flag_ptr
);
603 rcu_assign_pointer(*node_flag_ptr
, NULL
);
604 shadow_node
->nr_child
--;
609 * _ja_node_clear_ptr: clear ptr item within a node. Return an error
610 * (negative error value) if it is not found (-ENOENT).
613 int _ja_node_clear_ptr(const struct cds_ja_type
*type
,
614 struct cds_ja_inode
*node
,
615 struct cds_ja_shadow_node
*shadow_node
,
616 struct cds_ja_inode_flag
**node_flag_ptr
,
619 switch (type
->type_class
) {
621 return ja_linear_node_clear_ptr(type
, node
, shadow_node
, node_flag_ptr
);
623 return ja_pool_node_clear_ptr(type
, node
, shadow_node
, node_flag_ptr
, n
);
625 return ja_pigeon_node_clear_ptr(type
, node
, shadow_node
, node_flag_ptr
);
637 * ja_node_recompact_add: recompact a node, adding a new child.
638 * TODO: for pool type, take selection bit(s) into account.
639 * Return 0 on success, -EAGAIN if need to retry, or other negative
640 * error value otherwise.
643 int ja_node_recompact(enum ja_recompact mode
,
645 unsigned int old_type_index
,
646 const struct cds_ja_type
*old_type
,
647 struct cds_ja_inode
*old_node
,
648 struct cds_ja_shadow_node
*shadow_node
,
649 struct cds_ja_inode_flag
**old_node_flag_ptr
, uint8_t n
,
650 struct cds_ja_inode_flag
*child_node_flag
,
651 struct cds_ja_inode_flag
**nullify_node_flag_ptr
)
653 unsigned int new_type_index
;
654 struct cds_ja_inode
*new_node
;
655 struct cds_ja_shadow_node
*new_shadow_node
= NULL
;
656 const struct cds_ja_type
*new_type
;
657 struct cds_ja_inode_flag
*new_node_flag
, *old_node_flag
;
661 old_node_flag
= *old_node_flag_ptr
;
665 new_type_index
= old_type_index
;
667 case JA_RECOMPACT_ADD
:
668 if (!shadow_node
|| old_type_index
== NODE_INDEX_NULL
) {
671 new_type_index
= old_type_index
+ 1;
674 case JA_RECOMPACT_DEL
:
675 if (old_type_index
== 0) {
676 new_type_index
= NODE_INDEX_NULL
;
678 new_type_index
= old_type_index
- 1;
685 retry
: /* for fallback */
686 dbg_printf("Recompact from type %d to type %d\n",
687 old_type_index
, new_type_index
);
688 new_type
= &ja_types
[new_type_index
];
689 if (new_type_index
!= NODE_INDEX_NULL
) {
690 new_node
= alloc_cds_ja_node(new_type
);
693 new_node_flag
= ja_node_flag(new_node
, new_type_index
);
694 dbg_printf("Recompact inherit lock from %p\n", shadow_node
);
695 new_shadow_node
= rcuja_shadow_set(ja
->ht
, new_node_flag
, shadow_node
, ja
);
696 if (!new_shadow_node
) {
701 new_shadow_node
->fallback_removal_count
=
702 JA_FALLBACK_REMOVAL_COUNT
;
705 new_node_flag
= NULL
;
708 assert(mode
!= JA_RECOMPACT_ADD
|| old_type
->type_class
!= RCU_JA_PIGEON
);
710 if (new_type_index
== NODE_INDEX_NULL
)
713 switch (old_type
->type_class
) {
717 ja_linear_node_get_nr_child(old_type
, old_node
);
720 for (i
= 0; i
< nr_child
; i
++) {
721 struct cds_ja_inode_flag
*iter
;
724 ja_linear_node_get_ith_pos(old_type
, old_node
, i
, &v
, &iter
);
727 if (mode
== JA_RECOMPACT_DEL
&& *nullify_node_flag_ptr
== iter
)
729 ret
= _ja_node_set_nth(new_type
, new_node
,
732 if (new_type
->type_class
== RCU_JA_POOL
&& ret
) {
733 goto fallback_toosmall
;
741 unsigned int pool_nr
;
743 for (pool_nr
= 0; pool_nr
< (1U << old_type
->nr_pool_order
); pool_nr
++) {
744 struct cds_ja_inode
*pool
=
745 ja_pool_node_get_ith_pool(old_type
,
748 ja_linear_node_get_nr_child(old_type
, pool
);
751 for (j
= 0; j
< nr_child
; j
++) {
752 struct cds_ja_inode_flag
*iter
;
755 ja_linear_node_get_ith_pos(old_type
, pool
,
759 if (mode
== JA_RECOMPACT_DEL
&& *nullify_node_flag_ptr
== iter
)
761 ret
= _ja_node_set_nth(new_type
, new_node
,
764 if (new_type
->type_class
== RCU_JA_POOL
766 goto fallback_toosmall
;
774 assert(mode
== JA_RECOMPACT_ADD
);
781 assert(mode
== JA_RECOMPACT_DEL
);
782 nr_child
= shadow_node
->nr_child
;
783 for (i
= 0; i
< nr_child
; i
++) {
784 struct cds_ja_inode_flag
*iter
;
786 iter
= ja_pigeon_node_get_ith_pos(old_type
, old_node
, i
);
789 if (mode
== JA_RECOMPACT_DEL
&& *nullify_node_flag_ptr
== iter
)
791 ret
= _ja_node_set_nth(new_type
, new_node
,
794 if (new_type
->type_class
== RCU_JA_POOL
&& ret
) {
795 goto fallback_toosmall
;
808 if (mode
== JA_RECOMPACT_ADD
) {
810 ret
= _ja_node_set_nth(new_type
, new_node
,
815 /* Return pointer to new recompacted node through old_node_flag_ptr */
816 *old_node_flag_ptr
= new_node_flag
;
820 flags
= RCUJA_SHADOW_CLEAR_FREE_NODE
;
822 * It is OK to free the lock associated with a node
823 * going to NULL, since we are holding the parent lock.
824 * This synchronizes removal with re-add of that node.
826 if (new_type_index
== NODE_INDEX_NULL
)
827 flags
= RCUJA_SHADOW_CLEAR_FREE_LOCK
;
828 ret
= rcuja_shadow_clear(ja
->ht
, old_node_flag
, shadow_node
,
838 /* fallback if next pool is too small */
839 assert(new_shadow_node
);
840 ret
= rcuja_shadow_clear(ja
->ht
, new_node_flag
, new_shadow_node
,
841 RCUJA_SHADOW_CLEAR_FREE_NODE
);
844 /* Choose fallback type: pigeon */
845 new_type_index
= (1UL << JA_TYPE_BITS
) - 1;
846 dbg_printf("Fallback to type %d\n", new_type_index
);
847 uatomic_inc(&ja
->nr_fallback
);
853 * Return 0 on success, -EAGAIN if need to retry, or other negative
854 * error value otherwise.
857 int ja_node_set_nth(struct cds_ja
*ja
,
858 struct cds_ja_inode_flag
**node_flag
, uint8_t n
,
859 struct cds_ja_inode_flag
*child_node_flag
,
860 struct cds_ja_shadow_node
*shadow_node
)
863 unsigned int type_index
;
864 const struct cds_ja_type
*type
;
865 struct cds_ja_inode
*node
;
867 dbg_printf("ja_node_set_nth for n=%u, node %p, shadow %p\n",
868 (unsigned int) n
, ja_node_ptr(*node_flag
), shadow_node
);
870 node
= ja_node_ptr(*node_flag
);
871 type_index
= ja_node_type(*node_flag
);
872 type
= &ja_types
[type_index
];
873 ret
= _ja_node_set_nth(type
, node
, shadow_node
,
877 /* Not enough space in node, need to recompact. */
878 ret
= ja_node_recompact(JA_RECOMPACT_ADD
, ja
, type_index
, type
, node
,
879 shadow_node
, node_flag
, n
, child_node_flag
, NULL
);
882 /* Node needs to be recompacted. */
883 ret
= ja_node_recompact(JA_RECOMPACT
, ja
, type_index
, type
, node
,
884 shadow_node
, node_flag
, n
, child_node_flag
, NULL
);
891 * Return 0 on success, -EAGAIN if need to retry, or other negative
892 * error value otherwise.
895 int ja_node_clear_ptr(struct cds_ja
*ja
,
896 struct cds_ja_inode_flag
**node_flag_ptr
, /* Pointer to location to nullify */
897 struct cds_ja_inode_flag
**parent_node_flag_ptr
, /* Address of parent ptr in its parent */
898 struct cds_ja_shadow_node
*shadow_node
, /* of parent */
902 unsigned int type_index
;
903 const struct cds_ja_type
*type
;
904 struct cds_ja_inode
*node
;
906 dbg_printf("ja_node_clear_ptr for node %p, shadow %p, target ptr %p\n",
907 ja_node_ptr(*parent_node_flag_ptr
), shadow_node
, node_flag_ptr
);
909 node
= ja_node_ptr(*parent_node_flag_ptr
);
910 type_index
= ja_node_type(*parent_node_flag_ptr
);
911 type
= &ja_types
[type_index
];
912 ret
= _ja_node_clear_ptr(type
, node
, shadow_node
, node_flag_ptr
, n
);
914 /* Should to try recompaction. */
915 ret
= ja_node_recompact(JA_RECOMPACT_DEL
, ja
, type_index
, type
, node
,
916 shadow_node
, parent_node_flag_ptr
, n
, NULL
,
922 struct cds_hlist_head
cds_ja_lookup(struct cds_ja
*ja
, uint64_t key
)
924 unsigned int tree_depth
, i
;
925 struct cds_ja_inode_flag
*node_flag
;
926 struct cds_hlist_head head
= { NULL
};
928 if (caa_unlikely(key
> ja
->key_max
))
930 tree_depth
= ja
->tree_depth
;
931 node_flag
= rcu_dereference(ja
->root
);
933 /* level 0: root node */
934 if (!ja_node_ptr(node_flag
))
937 for (i
= 1; i
< tree_depth
; i
++) {
940 iter_key
= (uint8_t) (key
>> (JA_BITS_PER_BYTE
* (tree_depth
- i
- 1)));
941 node_flag
= ja_node_get_nth(node_flag
, NULL
,
943 dbg_printf("cds_ja_lookup iter key lookup %u finds node_flag %p\n",
944 (unsigned int) iter_key
, node_flag
);
945 if (!ja_node_ptr(node_flag
))
949 /* Last level lookup succeded. We got an actual match. */
950 head
.next
= (struct cds_hlist_node
*) node_flag
;
955 * We reached an unpopulated node. Create it and the children we need,
956 * and then attach the entire branch to the current node. This may
957 * trigger recompaction of the current node. Locks needed: node lock
958 * (for add), and, possibly, parent node lock (to update pointer due to
959 * node recompaction).
961 * First take node lock, check if recompaction is needed, then take
962 * parent lock (if needed). Then we can proceed to create the new
963 * branch. Publish the new branch, and release locks.
964 * TODO: we currently always take the parent lock even when not needed.
967 int ja_attach_node(struct cds_ja
*ja
,
968 struct cds_ja_inode_flag
**node_flag_ptr
,
969 struct cds_ja_inode_flag
*node_flag
,
970 struct cds_ja_inode_flag
*parent_node_flag
,
973 struct cds_ja_node
*child_node
)
975 struct cds_ja_shadow_node
*shadow_node
= NULL
,
976 *parent_shadow_node
= NULL
;
977 struct cds_ja_inode
*node
= ja_node_ptr(node_flag
);
978 struct cds_ja_inode
*parent_node
= ja_node_ptr(parent_node_flag
);
979 struct cds_hlist_head head
;
980 struct cds_ja_inode_flag
*iter_node_flag
, *iter_dest_node_flag
;
982 struct cds_ja_inode_flag
*created_nodes
[JA_MAX_DEPTH
];
983 int nr_created_nodes
= 0;
985 dbg_printf("Attach node at level %u (node %p, node_flag %p)\n",
986 level
, node
, node_flag
);
989 shadow_node
= rcuja_shadow_lookup_lock(ja
->ht
, node_flag
);
995 parent_shadow_node
= rcuja_shadow_lookup_lock(ja
->ht
,
997 if (!parent_shadow_node
) {
1003 if (*node_flag_ptr
!= NULL
) {
1005 * Attach point is non-NULL: it has been updated between
1006 * RCU lookup and lock acquisition. We need to re-try
1007 * lookup and attach.
1013 /* Create new branch, starting from bottom */
1014 CDS_INIT_HLIST_HEAD(&head
);
1015 cds_hlist_add_head_rcu(&child_node
->list
, &head
);
1016 iter_node_flag
= (struct cds_ja_inode_flag
*) head
.next
;
1018 for (i
= ja
->tree_depth
; i
> (int) level
; i
--) {
1021 iter_key
= (uint8_t) (key
>> (JA_BITS_PER_BYTE
* (ja
->tree_depth
- i
)));
1022 dbg_printf("branch creation level %d, key %u\n",
1023 i
- 1, (unsigned int) iter_key
);
1024 iter_dest_node_flag
= NULL
;
1025 ret
= ja_node_set_nth(ja
, &iter_dest_node_flag
,
1031 created_nodes
[nr_created_nodes
++] = iter_dest_node_flag
;
1032 iter_node_flag
= iter_dest_node_flag
;
1038 iter_key
= (uint8_t) (key
>> (JA_BITS_PER_BYTE
* (ja
->tree_depth
- level
)));
1039 /* We need to use set_nth on the previous level. */
1040 iter_dest_node_flag
= node_flag
;
1041 ret
= ja_node_set_nth(ja
, &iter_dest_node_flag
,
1047 created_nodes
[nr_created_nodes
++] = iter_dest_node_flag
;
1048 iter_node_flag
= iter_dest_node_flag
;
1051 /* Publish new branch */
1052 dbg_printf("Publish branch %p, replacing %p\n",
1053 iter_node_flag
, *node_flag_ptr
);
1054 rcu_assign_pointer(*node_flag_ptr
, iter_node_flag
);
1061 for (i
= 0; i
< nr_created_nodes
; i
++) {
1065 flags
= RCUJA_SHADOW_CLEAR_FREE_LOCK
;
1067 flags
|= RCUJA_SHADOW_CLEAR_FREE_NODE
;
1068 tmpret
= rcuja_shadow_clear(ja
->ht
,
1076 if (parent_shadow_node
)
1077 rcuja_shadow_unlock(parent_shadow_node
);
1080 rcuja_shadow_unlock(shadow_node
);
1086 * Lock the parent containing the hlist head pointer, and add node to list of
1087 * duplicates. Failure can happen if concurrent update changes the
1088 * parent before we get the lock. We return -EAGAIN in that case.
1089 * Return 0 on success, negative error value on failure.
1092 int ja_chain_node(struct cds_ja
*ja
,
1093 struct cds_ja_inode_flag
*parent_node_flag
,
1094 struct cds_hlist_head
*head
,
1095 struct cds_ja_node
*node
)
1097 struct cds_ja_shadow_node
*shadow_node
;
1099 shadow_node
= rcuja_shadow_lookup_lock(ja
->ht
, parent_node_flag
);
1101 dbg_printf("AGAIN3\n");
1104 cds_hlist_add_head_rcu(&node
->list
, head
);
1105 rcuja_shadow_unlock(shadow_node
);
1109 int cds_ja_add(struct cds_ja
*ja
, uint64_t key
,
1110 struct cds_ja_node
*new_node
)
1112 unsigned int tree_depth
, i
;
1113 struct cds_ja_inode_flag
**node_flag_ptr
; /* in parent */
1114 struct cds_ja_inode_flag
*node_flag
,
1119 if (caa_unlikely(key
> ja
->key_max
)) {
1122 tree_depth
= ja
->tree_depth
;
1125 dbg_printf("cds_ja_add attempt: key %" PRIu64
", node %p\n",
1127 parent2_node_flag
= NULL
;
1129 (struct cds_ja_inode_flag
*) &ja
->root
; /* Use root ptr address as key for mutex */
1130 node_flag_ptr
= &ja
->root
;
1131 node_flag
= rcu_dereference(ja
->root
);
1133 /* Iterate on all internal levels */
1134 for (i
= 1; i
< tree_depth
; i
++) {
1137 dbg_printf("cds_ja_add iter node_flag_ptr %p node_flag %p\n",
1138 *node_flag_ptr
, node_flag
);
1139 if (!ja_node_ptr(node_flag
)) {
1140 ret
= ja_attach_node(ja
, node_flag_ptr
,
1141 parent_node_flag
, parent2_node_flag
,
1143 if (ret
== -EAGAIN
|| ret
== -EEXIST
)
1148 iter_key
= (uint8_t) (key
>> (JA_BITS_PER_BYTE
* (tree_depth
- i
- 1)));
1149 parent2_node_flag
= parent_node_flag
;
1150 parent_node_flag
= node_flag
;
1151 node_flag
= ja_node_get_nth(node_flag
,
1154 dbg_printf("cds_ja_add iter key lookup %u finds node_flag %p node_flag_ptr %p\n",
1155 (unsigned int) iter_key
, node_flag
, *node_flag_ptr
);
1159 * We reached bottom of tree, simply add node to last internal
1160 * level, or chain it if key is already present.
1162 if (!ja_node_ptr(node_flag
)) {
1163 dbg_printf("cds_ja_add last node_flag_ptr %p node_flag %p\n",
1164 *node_flag_ptr
, node_flag
);
1165 ret
= ja_attach_node(ja
, node_flag_ptr
, parent_node_flag
,
1166 parent2_node_flag
, key
, i
, new_node
);
1168 ret
= ja_chain_node(ja
,
1170 (struct cds_hlist_head
*) node_flag_ptr
,
1173 if (ret
== -EAGAIN
|| ret
== -EEXIST
)
1180 * Note: there is no need to lookup the pointer address associated with
1181 * each node's nth item after taking the lock: it's already been done by
1182 * cds_ja_del while holding the rcu read-side lock, and our node rules
1183 * ensure that when a match value -> pointer is found in a node, it is
1184 * _NEVER_ changed for that node without recompaction, and recompaction
1185 * reallocates the node.
1186 * However, when a child is removed from "linear" nodes, its pointer
1187 * is set to NULL. We therefore check, while holding the locks, if this
1188 * pointer is NULL, and return -ENOENT to the caller if it is the case.
1191 int ja_detach_node(struct cds_ja
*ja
,
1192 struct cds_ja_inode_flag
**snapshot
,
1193 struct cds_ja_inode_flag
***snapshot_ptr
,
1194 uint8_t *snapshot_n
,
1197 struct cds_ja_node
*node
)
1199 struct cds_ja_shadow_node
*shadow_nodes
[JA_MAX_DEPTH
];
1200 struct cds_ja_inode_flag
**node_flag_ptr
= NULL
,
1201 *parent_node_flag
= NULL
,
1202 **parent_node_flag_ptr
= NULL
;
1203 struct cds_ja_inode_flag
*iter_node_flag
;
1204 int ret
, i
, nr_shadow
= 0, nr_clear
= 0, nr_branch
= 0;
1207 assert(nr_snapshot
== ja
->tree_depth
+ 1);
1210 * From the last internal level node going up, get the node
1211 * lock, check if the node has only one child left. If it is the
1212 * case, we continue iterating upward. When we reach a node
1213 * which has more that one child left, we lock the parent, and
1214 * proceed to the node deletion (removing its children too).
1216 for (i
= nr_snapshot
- 2; i
>= 1; i
--) {
1217 struct cds_ja_shadow_node
*shadow_node
;
1219 shadow_node
= rcuja_shadow_lookup_lock(ja
->ht
,
1225 assert(shadow_node
->nr_child
> 0);
1226 shadow_nodes
[nr_shadow
++] = shadow_node
;
1227 if (shadow_node
->nr_child
== 1)
1230 if (shadow_node
->nr_child
> 1 || i
== 1) {
1231 /* Lock parent and break */
1232 shadow_node
= rcuja_shadow_lookup_lock(ja
->ht
,
1238 shadow_nodes
[nr_shadow
++] = shadow_node
;
1239 node_flag_ptr
= snapshot_ptr
[i
+ 1];
1240 n
= snapshot_n
[i
+ 1];
1241 parent_node_flag_ptr
= snapshot_ptr
[i
];
1242 parent_node_flag
= snapshot
[i
];
1245 * Lock parent's parent, in case we need
1246 * to recompact parent.
1248 shadow_node
= rcuja_shadow_lookup_lock(ja
->ht
,
1254 shadow_nodes
[nr_shadow
++] = shadow_node
;
1261 * Check if node has been removed between RCU lookup and lock
1264 if (!*node_flag_ptr
) {
1270 * At this point, we want to delete all nodes that are about to
1271 * be removed from shadow_nodes (except the last one, which is
1272 * either the root or the parent of the upmost node with 1
1273 * child). OK to as to free lock here, because RCU read lock is
1274 * held, and free only performed in call_rcu.
1277 for (i
= 0; i
< nr_clear
; i
++) {
1278 ret
= rcuja_shadow_clear(ja
->ht
,
1279 shadow_nodes
[i
]->node_flag
,
1281 RCUJA_SHADOW_CLEAR_FREE_NODE
1282 | RCUJA_SHADOW_CLEAR_FREE_LOCK
);
1286 iter_node_flag
= parent_node_flag
;
1287 /* Remove from parent */
1288 ret
= ja_node_clear_ptr(ja
,
1289 node_flag_ptr
, /* Pointer to location to nullify */
1290 &iter_node_flag
, /* Old new parent ptr in its parent */
1291 shadow_nodes
[nr_branch
- 1], /* of parent */
1296 dbg_printf("ja_detach_node: publish %p instead of %p\n",
1297 iter_node_flag
, *parent_node_flag_ptr
);
1298 /* Update address of parent ptr in its parent */
1299 rcu_assign_pointer(*parent_node_flag_ptr
, iter_node_flag
);
1302 for (i
= 0; i
< nr_shadow
; i
++)
1303 rcuja_shadow_unlock(shadow_nodes
[i
]);
1308 int ja_unchain_node(struct cds_ja
*ja
,
1309 struct cds_ja_inode_flag
*parent_node_flag
,
1310 struct cds_hlist_head
*head
,
1311 struct cds_ja_node
*node
)
1313 struct cds_ja_shadow_node
*shadow_node
;
1314 struct cds_hlist_node
*hlist_node
;
1315 int ret
= 0, count
= 0;
1317 shadow_node
= rcuja_shadow_lookup_lock(ja
->ht
, parent_node_flag
);
1321 * Retry if another thread removed all but one of duplicates
1322 * since check (that was performed without lock).
1324 cds_hlist_for_each_rcu(hlist_node
, head
, list
) {
1332 cds_hlist_del_rcu(&node
->list
);
1334 rcuja_shadow_unlock(shadow_node
);
1339 * Called with RCU read lock held.
1341 int cds_ja_del(struct cds_ja
*ja
, uint64_t key
,
1342 struct cds_ja_node
*node
)
1344 unsigned int tree_depth
, i
;
1345 struct cds_ja_inode_flag
*snapshot
[JA_MAX_DEPTH
];
1346 struct cds_ja_inode_flag
**snapshot_ptr
[JA_MAX_DEPTH
];
1347 uint8_t snapshot_n
[JA_MAX_DEPTH
];
1348 struct cds_ja_inode_flag
*node_flag
;
1349 struct cds_ja_inode_flag
**prev_node_flag_ptr
;
1353 if (caa_unlikely(key
> ja
->key_max
))
1355 tree_depth
= ja
->tree_depth
;
1359 dbg_printf("cds_ja_del attempt: key %" PRIu64
", node %p\n",
1362 /* snapshot for level 0 is only for shadow node lookup */
1365 snapshot_ptr
[nr_snapshot
] = NULL
;
1366 snapshot
[nr_snapshot
++] = (struct cds_ja_inode_flag
*) &ja
->root
;
1367 node_flag
= rcu_dereference(ja
->root
);
1368 prev_node_flag_ptr
= &ja
->root
;
1370 /* Iterate on all internal levels */
1371 for (i
= 1; i
< tree_depth
; i
++) {
1374 dbg_printf("cds_ja_del iter node_flag %p\n",
1376 if (!ja_node_ptr(node_flag
)) {
1379 iter_key
= (uint8_t) (key
>> (JA_BITS_PER_BYTE
* (tree_depth
- i
- 1)));
1380 snapshot_n
[nr_snapshot
+ 1] = iter_key
;
1381 snapshot_ptr
[nr_snapshot
] = prev_node_flag_ptr
;
1382 snapshot
[nr_snapshot
++] = node_flag
;
1383 node_flag
= ja_node_get_nth(node_flag
,
1384 &prev_node_flag_ptr
,
1386 dbg_printf("cds_ja_del iter key lookup %u finds node_flag %p, prev_node_flag_ptr %p\n",
1387 (unsigned int) iter_key
, node_flag
,
1388 prev_node_flag_ptr
);
1392 * We reached bottom of tree, try to find the node we are trying
1393 * to remove. Fail if we cannot find it.
1395 if (!ja_node_ptr(node_flag
)) {
1396 dbg_printf("cds_ja_del: no node found for key %" PRIu64
"\n",
1400 struct cds_hlist_head hlist_head
;
1401 struct cds_hlist_node
*hlist_node
;
1402 struct cds_ja_node
*entry
, *match
= NULL
;
1406 (struct cds_hlist_node
*) ja_node_ptr(node_flag
);
1407 cds_hlist_for_each_entry_rcu(entry
,
1411 dbg_printf("cds_ja_del: compare %p with entry %p\n", node
, entry
);
1417 dbg_printf("cds_ja_del: no node match for node %p key %" PRIu64
"\n", node
, key
);
1423 * Removing last of duplicates. Last snapshot
1424 * does not have a shadow node (external leafs).
1426 snapshot_ptr
[nr_snapshot
] = prev_node_flag_ptr
;
1427 snapshot
[nr_snapshot
++] = node_flag
;
1428 ret
= ja_detach_node(ja
, snapshot
, snapshot_ptr
,
1429 snapshot_n
, nr_snapshot
, key
, node
);
1431 ret
= ja_unchain_node(ja
, snapshot
[nr_snapshot
- 1],
1432 &hlist_head
, match
);
1436 * Explanation of -ENOENT handling: caused by concurrent delete
1437 * between RCU lookup and actual removal. Need to re-do the
1438 * lookup and removal attempt.
1440 if (ret
== -EAGAIN
|| ret
== -ENOENT
)
1445 struct cds_ja
*_cds_ja_new(unsigned int key_bits
,
1446 const struct rcu_flavor_struct
*flavor
)
1450 struct cds_ja_shadow_node
*root_shadow_node
;
1452 ja
= calloc(sizeof(*ja
), 1);
1464 ja
->key_max
= (1ULL << key_bits
) - 1;
1467 ja
->key_max
= UINT64_MAX
;
1473 /* ja->root is NULL */
1474 /* tree_depth 0 is for pointer to root node */
1475 ja
->tree_depth
= (key_bits
>> JA_LOG2_BITS_PER_BYTE
) + 1;
1476 assert(ja
->tree_depth
<= JA_MAX_DEPTH
);
1477 ja
->ht
= rcuja_create_ht(flavor
);
1482 * Note: we should not free this node until judy array destroy.
1484 root_shadow_node
= rcuja_shadow_set(ja
->ht
,
1485 (struct cds_ja_inode_flag
*) &ja
->root
,
1487 if (!root_shadow_node
) {
1491 root_shadow_node
->level
= 0;
1496 ret
= rcuja_delete_ht(ja
->ht
);
1506 * Called from RCU read-side CS.
1508 __attribute__((visibility("protected")))
1509 void rcuja_free_all_children(struct cds_ja_shadow_node
*shadow_node
,
1510 struct cds_ja_inode_flag
*node_flag
,
1511 void (*free_node_cb
)(struct rcu_head
*head
))
1513 const struct rcu_flavor_struct
*flavor
;
1514 unsigned int type_index
;
1515 struct cds_ja_inode
*node
;
1516 const struct cds_ja_type
*type
;
1518 flavor
= cds_lfht_rcu_flavor(shadow_node
->ja
->ht
);
1519 node
= ja_node_ptr(node_flag
);
1520 assert(node
!= NULL
);
1521 type_index
= ja_node_type(node_flag
);
1522 type
= &ja_types
[type_index
];
1524 switch (type
->type_class
) {
1528 ja_linear_node_get_nr_child(type
, node
);
1531 for (i
= 0; i
< nr_child
; i
++) {
1532 struct cds_ja_inode_flag
*iter
;
1533 struct cds_hlist_head head
;
1534 struct cds_ja_node
*entry
;
1535 struct cds_hlist_node
*pos
;
1538 ja_linear_node_get_ith_pos(type
, node
, i
, &v
, &iter
);
1541 head
.next
= (struct cds_hlist_node
*) iter
;
1542 cds_hlist_for_each_entry_rcu(entry
, pos
, &head
, list
) {
1543 flavor
->update_call_rcu(&entry
->head
, free_node_cb
);
1550 unsigned int pool_nr
;
1552 for (pool_nr
= 0; pool_nr
< (1U << type
->nr_pool_order
); pool_nr
++) {
1553 struct cds_ja_inode
*pool
=
1554 ja_pool_node_get_ith_pool(type
, node
, pool_nr
);
1556 ja_linear_node_get_nr_child(type
, pool
);
1559 for (j
= 0; j
< nr_child
; j
++) {
1560 struct cds_ja_inode_flag
*iter
;
1561 struct cds_hlist_head head
;
1562 struct cds_ja_node
*entry
;
1563 struct cds_hlist_node
*pos
;
1566 ja_linear_node_get_ith_pos(type
, node
, j
, &v
, &iter
);
1569 head
.next
= (struct cds_hlist_node
*) iter
;
1570 cds_hlist_for_each_entry_rcu(entry
, pos
, &head
, list
) {
1571 flavor
->update_call_rcu(&entry
->head
, free_node_cb
);
1584 nr_child
= shadow_node
->nr_child
;
1585 for (i
= 0; i
< nr_child
; i
++) {
1586 struct cds_ja_inode_flag
*iter
;
1587 struct cds_hlist_head head
;
1588 struct cds_ja_node
*entry
;
1589 struct cds_hlist_node
*pos
;
1591 iter
= ja_pigeon_node_get_ith_pos(type
, node
, i
);
1594 head
.next
= (struct cds_hlist_node
*) iter
;
1595 cds_hlist_for_each_entry_rcu(entry
, pos
, &head
, list
) {
1596 flavor
->update_call_rcu(&entry
->head
, free_node_cb
);
1607 * There should be no more concurrent add to the judy array while it is
1608 * being destroyed (ensured by the caller).
1610 int cds_ja_destroy(struct cds_ja
*ja
,
1611 void (*free_node_cb
)(struct rcu_head
*head
))
1615 rcuja_shadow_prune(ja
->ht
,
1616 RCUJA_SHADOW_CLEAR_FREE_NODE
| RCUJA_SHADOW_CLEAR_FREE_LOCK
,
1618 ret
= rcuja_delete_ht(ja
->ht
);
1621 if (uatomic_read(&ja
->nr_fallback
))
1623 "[warning] RCU Judy Array used %lu fallback node(s)\n",
1624 uatomic_read(&ja
->nr_fallback
));