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
,
280 struct cds_ja_inode_flag
***node_flag_ptr
,
285 struct cds_ja_inode_flag
**pointers
;
286 struct cds_ja_inode_flag
*ptr
;
289 assert(type
->type_class
== RCU_JA_LINEAR
|| type
->type_class
== RCU_JA_POOL
);
291 nr_child
= ja_linear_node_get_nr_child(type
, node
);
292 cmm_smp_rmb(); /* read nr_child before values and pointers */
293 assert(nr_child
<= type
->max_linear_child
);
294 assert(type
->type_class
!= RCU_JA_LINEAR
|| nr_child
>= type
->min_child
);
296 values
= &node
->u
.data
[1];
297 for (i
= 0; i
< nr_child
; i
++) {
298 if (CMM_LOAD_SHARED(values
[i
]) == n
)
302 if (caa_unlikely(node_flag_ptr
))
303 *node_flag_ptr
= NULL
;
306 pointers
= (struct cds_ja_inode_flag
**) align_ptr_size(&values
[type
->max_linear_child
]);
307 ptr
= rcu_dereference(pointers
[i
]);
308 if (caa_unlikely(child_node_flag_ptr
) && ptr
)
309 *child_node_flag_ptr
= &pointers
[i
];
310 if (caa_unlikely(node_flag_ptr
))
311 *node_flag_ptr
= &pointers
[i
];
316 void ja_linear_node_get_ith_pos(const struct cds_ja_type
*type
,
317 struct cds_ja_inode
*node
,
320 struct cds_ja_inode_flag
**iter
)
323 struct cds_ja_inode_flag
**pointers
;
325 assert(type
->type_class
== RCU_JA_LINEAR
|| type
->type_class
== RCU_JA_POOL
);
326 assert(i
< ja_linear_node_get_nr_child(type
, node
));
328 values
= &node
->u
.data
[1];
330 pointers
= (struct cds_ja_inode_flag
**) align_ptr_size(&values
[type
->max_linear_child
]);
335 struct cds_ja_inode_flag
*ja_pool_node_get_nth(const struct cds_ja_type
*type
,
336 struct cds_ja_inode
*node
,
337 struct cds_ja_inode_flag
***child_node_flag_ptr
,
338 struct cds_ja_inode_flag
***node_flag_ptr
,
341 struct cds_ja_inode
*linear
;
343 assert(type
->type_class
== RCU_JA_POOL
);
345 * TODO: currently, we select the pool by highest bits. We
346 * should support various encodings.
348 linear
= (struct cds_ja_inode
*)
349 &node
->u
.data
[((unsigned long) n
>> (CHAR_BIT
- type
->nr_pool_order
)) << type
->pool_size_order
];
350 return ja_linear_node_get_nth(type
, linear
, child_node_flag_ptr
,
355 struct cds_ja_inode
*ja_pool_node_get_ith_pool(const struct cds_ja_type
*type
,
356 struct cds_ja_inode
*node
,
359 assert(type
->type_class
== RCU_JA_POOL
);
360 return (struct cds_ja_inode
*)
361 &node
->u
.data
[(unsigned int) i
<< type
->pool_size_order
];
365 struct cds_ja_inode_flag
*ja_pigeon_node_get_nth(const struct cds_ja_type
*type
,
366 struct cds_ja_inode
*node
,
367 struct cds_ja_inode_flag
***child_node_flag_ptr
,
368 struct cds_ja_inode_flag
***node_flag_ptr
,
371 struct cds_ja_inode_flag
**child_node_flag
;
373 assert(type
->type_class
== RCU_JA_PIGEON
);
374 child_node_flag
= &((struct cds_ja_inode_flag
**) node
->u
.data
)[n
];
375 dbg_printf("ja_pigeon_node_get_nth child_node_flag_ptr %p\n",
377 if (caa_unlikely(child_node_flag_ptr
) && *child_node_flag
)
378 *child_node_flag_ptr
= child_node_flag
;
379 if (caa_unlikely(node_flag_ptr
))
380 *node_flag_ptr
= child_node_flag
;
381 return rcu_dereference(*child_node_flag
);
385 struct cds_ja_inode_flag
*ja_pigeon_node_get_ith_pos(const struct cds_ja_type
*type
,
386 struct cds_ja_inode
*node
,
389 return ja_pigeon_node_get_nth(type
, node
, NULL
, NULL
, i
);
393 * ja_node_get_nth: get nth item from a node.
394 * node_flag is already rcu_dereference'd.
397 struct cds_ja_inode_flag
* ja_node_get_nth(struct cds_ja_inode_flag
*node_flag
,
398 struct cds_ja_inode_flag
***child_node_flag_ptr
,
399 struct cds_ja_inode_flag
***node_flag_ptr
,
402 unsigned int type_index
;
403 struct cds_ja_inode
*node
;
404 const struct cds_ja_type
*type
;
406 node
= ja_node_ptr(node_flag
);
407 assert(node
!= NULL
);
408 type_index
= ja_node_type(node_flag
);
409 type
= &ja_types
[type_index
];
411 switch (type
->type_class
) {
413 return ja_linear_node_get_nth(type
, node
,
414 child_node_flag_ptr
, node_flag_ptr
, n
);
416 return ja_pool_node_get_nth(type
, node
,
417 child_node_flag_ptr
, node_flag_ptr
, n
);
419 return ja_pigeon_node_get_nth(type
, node
,
420 child_node_flag_ptr
, node_flag_ptr
, n
);
423 return (void *) -1UL;
428 int ja_linear_node_set_nth(const struct cds_ja_type
*type
,
429 struct cds_ja_inode
*node
,
430 struct cds_ja_shadow_node
*shadow_node
,
432 struct cds_ja_inode_flag
*child_node_flag
)
435 uint8_t *values
, *nr_child_ptr
;
436 struct cds_ja_inode_flag
**pointers
;
437 unsigned int i
, unused
= 0;
439 assert(type
->type_class
== RCU_JA_LINEAR
|| type
->type_class
== RCU_JA_POOL
);
441 nr_child_ptr
= &node
->u
.data
[0];
442 dbg_printf("linear set nth: nr_child_ptr %p\n", nr_child_ptr
);
443 nr_child
= *nr_child_ptr
;
444 assert(nr_child
<= type
->max_linear_child
);
446 values
= &node
->u
.data
[1];
447 pointers
= (struct cds_ja_inode_flag
**) align_ptr_size(&values
[type
->max_linear_child
]);
448 /* Check if node value is already populated */
449 for (i
= 0; i
< nr_child
; i
++) {
450 if (values
[i
] == n
) {
460 if (i
== nr_child
&& nr_child
>= type
->max_linear_child
) {
462 return -ERANGE
; /* recompact node */
464 return -ENOSPC
; /* No space left in this node type */
467 assert(pointers
[i
] == NULL
);
468 rcu_assign_pointer(pointers
[i
], child_node_flag
);
469 /* If we expanded the nr_child, increment it */
471 CMM_STORE_SHARED(values
[nr_child
], n
);
472 /* write pointer and value before nr_child */
474 CMM_STORE_SHARED(*nr_child_ptr
, nr_child
+ 1);
476 shadow_node
->nr_child
++;
477 dbg_printf("linear set nth: %u child, shadow: %u child, for node %p shadow %p\n",
478 (unsigned int) CMM_LOAD_SHARED(*nr_child_ptr
),
479 (unsigned int) shadow_node
->nr_child
,
486 int ja_pool_node_set_nth(const struct cds_ja_type
*type
,
487 struct cds_ja_inode
*node
,
488 struct cds_ja_shadow_node
*shadow_node
,
490 struct cds_ja_inode_flag
*child_node_flag
)
492 struct cds_ja_inode
*linear
;
494 assert(type
->type_class
== RCU_JA_POOL
);
495 linear
= (struct cds_ja_inode
*)
496 &node
->u
.data
[((unsigned long) n
>> (CHAR_BIT
- type
->nr_pool_order
)) << type
->pool_size_order
];
497 return ja_linear_node_set_nth(type
, linear
, shadow_node
,
502 int ja_pigeon_node_set_nth(const struct cds_ja_type
*type
,
503 struct cds_ja_inode
*node
,
504 struct cds_ja_shadow_node
*shadow_node
,
506 struct cds_ja_inode_flag
*child_node_flag
)
508 struct cds_ja_inode_flag
**ptr
;
510 assert(type
->type_class
== RCU_JA_PIGEON
);
511 ptr
= &((struct cds_ja_inode_flag
**) node
->u
.data
)[n
];
514 rcu_assign_pointer(*ptr
, child_node_flag
);
515 shadow_node
->nr_child
++;
520 * _ja_node_set_nth: set nth item within a node. Return an error
521 * (negative error value) if it is already there.
524 int _ja_node_set_nth(const struct cds_ja_type
*type
,
525 struct cds_ja_inode
*node
,
526 struct cds_ja_shadow_node
*shadow_node
,
528 struct cds_ja_inode_flag
*child_node_flag
)
530 switch (type
->type_class
) {
532 return ja_linear_node_set_nth(type
, node
, shadow_node
, n
,
535 return ja_pool_node_set_nth(type
, node
, shadow_node
, n
,
538 return ja_pigeon_node_set_nth(type
, node
, shadow_node
, n
,
551 int ja_linear_node_clear_ptr(const struct cds_ja_type
*type
,
552 struct cds_ja_inode
*node
,
553 struct cds_ja_shadow_node
*shadow_node
,
554 struct cds_ja_inode_flag
**node_flag_ptr
)
557 uint8_t *nr_child_ptr
;
559 assert(type
->type_class
== RCU_JA_LINEAR
|| type
->type_class
== RCU_JA_POOL
);
561 nr_child_ptr
= &node
->u
.data
[0];
562 dbg_printf("linear clear ptr: nr_child_ptr %p\n", nr_child_ptr
);
563 nr_child
= *nr_child_ptr
;
564 assert(nr_child
<= type
->max_linear_child
);
566 if (shadow_node
->fallback_removal_count
) {
567 shadow_node
->fallback_removal_count
--;
569 if (shadow_node
->nr_child
<= type
->min_child
) {
570 /* We need to try recompacting the node */
574 assert(*node_flag_ptr
!= NULL
);
575 rcu_assign_pointer(*node_flag_ptr
, NULL
);
577 * Value and nr_child are never changed (would cause ABA issue).
578 * Instead, we leave the pointer to NULL and recompact the node
579 * once in a while. It is allowed to set a NULL pointer to a new
580 * value without recompaction though.
581 * Only update the shadow node accounting.
583 shadow_node
->nr_child
--;
584 dbg_printf("linear clear ptr: %u child, shadow: %u child, for node %p shadow %p\n",
585 (unsigned int) CMM_LOAD_SHARED(*nr_child_ptr
),
586 (unsigned int) shadow_node
->nr_child
,
593 int ja_pool_node_clear_ptr(const struct cds_ja_type
*type
,
594 struct cds_ja_inode
*node
,
595 struct cds_ja_shadow_node
*shadow_node
,
596 struct cds_ja_inode_flag
**node_flag_ptr
,
599 struct cds_ja_inode
*linear
;
601 assert(type
->type_class
== RCU_JA_POOL
);
602 linear
= (struct cds_ja_inode
*)
603 &node
->u
.data
[((unsigned long) n
>> (CHAR_BIT
- type
->nr_pool_order
)) << type
->pool_size_order
];
604 return ja_linear_node_clear_ptr(type
, linear
, shadow_node
, node_flag_ptr
);
608 int ja_pigeon_node_clear_ptr(const struct cds_ja_type
*type
,
609 struct cds_ja_inode
*node
,
610 struct cds_ja_shadow_node
*shadow_node
,
611 struct cds_ja_inode_flag
**node_flag_ptr
)
613 assert(type
->type_class
== RCU_JA_PIGEON
);
614 dbg_printf("ja_pigeon_node_clear_ptr: clearing ptr: %p\n", *node_flag_ptr
);
615 rcu_assign_pointer(*node_flag_ptr
, NULL
);
616 shadow_node
->nr_child
--;
621 * _ja_node_clear_ptr: clear ptr item within a node. Return an error
622 * (negative error value) if it is not found (-ENOENT).
625 int _ja_node_clear_ptr(const struct cds_ja_type
*type
,
626 struct cds_ja_inode
*node
,
627 struct cds_ja_shadow_node
*shadow_node
,
628 struct cds_ja_inode_flag
**node_flag_ptr
,
631 switch (type
->type_class
) {
633 return ja_linear_node_clear_ptr(type
, node
, shadow_node
, node_flag_ptr
);
635 return ja_pool_node_clear_ptr(type
, node
, shadow_node
, node_flag_ptr
, n
);
637 return ja_pigeon_node_clear_ptr(type
, node
, shadow_node
, node_flag_ptr
);
649 * ja_node_recompact_add: recompact a node, adding a new child.
650 * TODO: for pool type, take selection bit(s) into account.
651 * Return 0 on success, -EAGAIN if need to retry, or other negative
652 * error value otherwise.
655 int ja_node_recompact(enum ja_recompact mode
,
657 unsigned int old_type_index
,
658 const struct cds_ja_type
*old_type
,
659 struct cds_ja_inode
*old_node
,
660 struct cds_ja_shadow_node
*shadow_node
,
661 struct cds_ja_inode_flag
**old_node_flag_ptr
, uint8_t n
,
662 struct cds_ja_inode_flag
*child_node_flag
,
663 struct cds_ja_inode_flag
**nullify_node_flag_ptr
)
665 unsigned int new_type_index
;
666 struct cds_ja_inode
*new_node
;
667 struct cds_ja_shadow_node
*new_shadow_node
= NULL
;
668 const struct cds_ja_type
*new_type
;
669 struct cds_ja_inode_flag
*new_node_flag
, *old_node_flag
;
673 old_node_flag
= *old_node_flag_ptr
;
677 new_type_index
= old_type_index
;
679 case JA_RECOMPACT_ADD
:
680 if (!shadow_node
|| old_type_index
== NODE_INDEX_NULL
) {
683 new_type_index
= old_type_index
+ 1;
686 case JA_RECOMPACT_DEL
:
687 if (old_type_index
== 0) {
688 new_type_index
= NODE_INDEX_NULL
;
690 new_type_index
= old_type_index
- 1;
697 retry
: /* for fallback */
698 dbg_printf("Recompact from type %d to type %d\n",
699 old_type_index
, new_type_index
);
700 new_type
= &ja_types
[new_type_index
];
701 if (new_type_index
!= NODE_INDEX_NULL
) {
702 new_node
= alloc_cds_ja_node(new_type
);
705 new_node_flag
= ja_node_flag(new_node
, new_type_index
);
706 dbg_printf("Recompact inherit lock from %p\n", shadow_node
);
707 new_shadow_node
= rcuja_shadow_set(ja
->ht
, new_node_flag
, shadow_node
, ja
);
708 if (!new_shadow_node
) {
713 new_shadow_node
->fallback_removal_count
=
714 JA_FALLBACK_REMOVAL_COUNT
;
717 new_node_flag
= NULL
;
720 assert(mode
!= JA_RECOMPACT_ADD
|| old_type
->type_class
!= RCU_JA_PIGEON
);
722 if (new_type_index
== NODE_INDEX_NULL
)
725 switch (old_type
->type_class
) {
729 ja_linear_node_get_nr_child(old_type
, old_node
);
732 for (i
= 0; i
< nr_child
; i
++) {
733 struct cds_ja_inode_flag
*iter
;
736 ja_linear_node_get_ith_pos(old_type
, old_node
, i
, &v
, &iter
);
739 if (mode
== JA_RECOMPACT_DEL
&& *nullify_node_flag_ptr
== iter
)
741 ret
= _ja_node_set_nth(new_type
, new_node
,
744 if (new_type
->type_class
== RCU_JA_POOL
&& ret
) {
745 goto fallback_toosmall
;
753 unsigned int pool_nr
;
755 for (pool_nr
= 0; pool_nr
< (1U << old_type
->nr_pool_order
); pool_nr
++) {
756 struct cds_ja_inode
*pool
=
757 ja_pool_node_get_ith_pool(old_type
,
760 ja_linear_node_get_nr_child(old_type
, pool
);
763 for (j
= 0; j
< nr_child
; j
++) {
764 struct cds_ja_inode_flag
*iter
;
767 ja_linear_node_get_ith_pos(old_type
, pool
,
771 if (mode
== JA_RECOMPACT_DEL
&& *nullify_node_flag_ptr
== iter
)
773 ret
= _ja_node_set_nth(new_type
, new_node
,
776 if (new_type
->type_class
== RCU_JA_POOL
778 goto fallback_toosmall
;
786 assert(mode
== JA_RECOMPACT_ADD
);
793 assert(mode
== JA_RECOMPACT_DEL
);
794 nr_child
= shadow_node
->nr_child
;
795 for (i
= 0; i
< nr_child
; i
++) {
796 struct cds_ja_inode_flag
*iter
;
798 iter
= ja_pigeon_node_get_ith_pos(old_type
, old_node
, i
);
801 if (mode
== JA_RECOMPACT_DEL
&& *nullify_node_flag_ptr
== iter
)
803 ret
= _ja_node_set_nth(new_type
, new_node
,
806 if (new_type
->type_class
== RCU_JA_POOL
&& ret
) {
807 goto fallback_toosmall
;
820 if (mode
== JA_RECOMPACT_ADD
) {
822 ret
= _ja_node_set_nth(new_type
, new_node
,
827 /* Return pointer to new recompacted node through old_node_flag_ptr */
828 *old_node_flag_ptr
= new_node_flag
;
832 flags
= RCUJA_SHADOW_CLEAR_FREE_NODE
;
834 * It is OK to free the lock associated with a node
835 * going to NULL, since we are holding the parent lock.
836 * This synchronizes removal with re-add of that node.
838 if (new_type_index
== NODE_INDEX_NULL
)
839 flags
= RCUJA_SHADOW_CLEAR_FREE_LOCK
;
840 ret
= rcuja_shadow_clear(ja
->ht
, old_node_flag
, shadow_node
,
850 /* fallback if next pool is too small */
851 assert(new_shadow_node
);
852 ret
= rcuja_shadow_clear(ja
->ht
, new_node_flag
, new_shadow_node
,
853 RCUJA_SHADOW_CLEAR_FREE_NODE
);
856 /* Choose fallback type: pigeon */
857 new_type_index
= (1UL << JA_TYPE_BITS
) - 1;
858 dbg_printf("Fallback to type %d\n", new_type_index
);
859 uatomic_inc(&ja
->nr_fallback
);
865 * Return 0 on success, -EAGAIN if need to retry, or other negative
866 * error value otherwise.
869 int ja_node_set_nth(struct cds_ja
*ja
,
870 struct cds_ja_inode_flag
**node_flag
, uint8_t n
,
871 struct cds_ja_inode_flag
*child_node_flag
,
872 struct cds_ja_shadow_node
*shadow_node
)
875 unsigned int type_index
;
876 const struct cds_ja_type
*type
;
877 struct cds_ja_inode
*node
;
879 dbg_printf("ja_node_set_nth for n=%u, node %p, shadow %p\n",
880 (unsigned int) n
, ja_node_ptr(*node_flag
), shadow_node
);
882 node
= ja_node_ptr(*node_flag
);
883 type_index
= ja_node_type(*node_flag
);
884 type
= &ja_types
[type_index
];
885 ret
= _ja_node_set_nth(type
, node
, shadow_node
,
889 /* Not enough space in node, need to recompact. */
890 ret
= ja_node_recompact(JA_RECOMPACT_ADD
, ja
, type_index
, type
, node
,
891 shadow_node
, node_flag
, n
, child_node_flag
, NULL
);
894 /* Node needs to be recompacted. */
895 ret
= ja_node_recompact(JA_RECOMPACT
, ja
, type_index
, type
, node
,
896 shadow_node
, node_flag
, n
, child_node_flag
, NULL
);
903 * Return 0 on success, -EAGAIN if need to retry, or other negative
904 * error value otherwise.
907 int ja_node_clear_ptr(struct cds_ja
*ja
,
908 struct cds_ja_inode_flag
**node_flag_ptr
, /* Pointer to location to nullify */
909 struct cds_ja_inode_flag
**parent_node_flag_ptr
, /* Address of parent ptr in its parent */
910 struct cds_ja_shadow_node
*shadow_node
, /* of parent */
914 unsigned int type_index
;
915 const struct cds_ja_type
*type
;
916 struct cds_ja_inode
*node
;
918 dbg_printf("ja_node_clear_ptr for node %p, shadow %p, target ptr %p\n",
919 ja_node_ptr(*parent_node_flag_ptr
), shadow_node
, node_flag_ptr
);
921 node
= ja_node_ptr(*parent_node_flag_ptr
);
922 type_index
= ja_node_type(*parent_node_flag_ptr
);
923 type
= &ja_types
[type_index
];
924 ret
= _ja_node_clear_ptr(type
, node
, shadow_node
, node_flag_ptr
, n
);
926 /* Should to try recompaction. */
927 ret
= ja_node_recompact(JA_RECOMPACT_DEL
, ja
, type_index
, type
, node
,
928 shadow_node
, parent_node_flag_ptr
, n
, NULL
,
934 struct cds_hlist_head
cds_ja_lookup(struct cds_ja
*ja
, uint64_t key
)
936 unsigned int tree_depth
, i
;
937 struct cds_ja_inode_flag
*node_flag
;
938 struct cds_hlist_head head
= { NULL
};
940 if (caa_unlikely(key
> ja
->key_max
))
942 tree_depth
= ja
->tree_depth
;
943 node_flag
= rcu_dereference(ja
->root
);
945 /* level 0: root node */
946 if (!ja_node_ptr(node_flag
))
949 for (i
= 1; i
< tree_depth
; i
++) {
952 iter_key
= (uint8_t) (key
>> (JA_BITS_PER_BYTE
* (tree_depth
- i
- 1)));
953 node_flag
= ja_node_get_nth(node_flag
, NULL
, NULL
,
955 dbg_printf("cds_ja_lookup iter key lookup %u finds node_flag %p\n",
956 (unsigned int) iter_key
, node_flag
);
957 if (!ja_node_ptr(node_flag
))
961 /* Last level lookup succeded. We got an actual match. */
962 head
.next
= (struct cds_hlist_node
*) node_flag
;
967 * We reached an unpopulated node. Create it and the children we need,
968 * and then attach the entire branch to the current node. This may
969 * trigger recompaction of the current node. Locks needed: node lock
970 * (for add), and, possibly, parent node lock (to update pointer due to
971 * node recompaction).
973 * First take node lock, check if recompaction is needed, then take
974 * parent lock (if needed). Then we can proceed to create the new
975 * branch. Publish the new branch, and release locks.
976 * TODO: we currently always take the parent lock even when not needed.
979 int ja_attach_node(struct cds_ja
*ja
,
980 struct cds_ja_inode_flag
**attach_node_flag_ptr
,
981 struct cds_ja_inode_flag
**node_flag_ptr
,
982 struct cds_ja_inode_flag
*node_flag
,
983 struct cds_ja_inode_flag
*parent_node_flag
,
986 struct cds_ja_node
*child_node
)
988 struct cds_ja_shadow_node
*shadow_node
= NULL
,
989 *parent_shadow_node
= NULL
;
990 struct cds_ja_inode
*node
= ja_node_ptr(node_flag
);
991 struct cds_ja_inode
*parent_node
= ja_node_ptr(parent_node_flag
);
992 struct cds_hlist_head head
;
993 struct cds_ja_inode_flag
*iter_node_flag
, *iter_dest_node_flag
;
995 struct cds_ja_inode_flag
*created_nodes
[JA_MAX_DEPTH
];
996 int nr_created_nodes
= 0;
998 dbg_printf("Attach node at level %u (node %p, node_flag %p)\n",
999 level
, node
, node_flag
);
1002 shadow_node
= rcuja_shadow_lookup_lock(ja
->ht
, node_flag
);
1008 parent_shadow_node
= rcuja_shadow_lookup_lock(ja
->ht
,
1010 if (!parent_shadow_node
) {
1016 if (node_flag_ptr
&& ja_node_ptr(*node_flag_ptr
)) {
1018 * Target node is non-NULL: it has been updated between
1019 * RCU lookup and lock acquisition. We need to re-try
1020 * lookup and attach.
1026 /* Create new branch, starting from bottom */
1027 CDS_INIT_HLIST_HEAD(&head
);
1028 cds_hlist_add_head_rcu(&child_node
->list
, &head
);
1029 iter_node_flag
= (struct cds_ja_inode_flag
*) head
.next
;
1031 for (i
= ja
->tree_depth
; i
> (int) level
; i
--) {
1034 iter_key
= (uint8_t) (key
>> (JA_BITS_PER_BYTE
* (ja
->tree_depth
- i
)));
1035 dbg_printf("branch creation level %d, key %u\n",
1036 i
- 1, (unsigned int) iter_key
);
1037 iter_dest_node_flag
= NULL
;
1038 ret
= ja_node_set_nth(ja
, &iter_dest_node_flag
,
1044 created_nodes
[nr_created_nodes
++] = iter_dest_node_flag
;
1045 iter_node_flag
= iter_dest_node_flag
;
1051 iter_key
= (uint8_t) (key
>> (JA_BITS_PER_BYTE
* (ja
->tree_depth
- level
)));
1052 /* We need to use set_nth on the previous level. */
1053 iter_dest_node_flag
= node_flag
;
1054 ret
= ja_node_set_nth(ja
, &iter_dest_node_flag
,
1060 created_nodes
[nr_created_nodes
++] = iter_dest_node_flag
;
1061 iter_node_flag
= iter_dest_node_flag
;
1064 /* Publish new branch */
1065 dbg_printf("Publish branch %p, replacing %p\n",
1066 iter_node_flag
, *attach_node_flag_ptr
);
1067 rcu_assign_pointer(*attach_node_flag_ptr
, iter_node_flag
);
1074 for (i
= 0; i
< nr_created_nodes
; i
++) {
1078 flags
= RCUJA_SHADOW_CLEAR_FREE_LOCK
;
1080 flags
|= RCUJA_SHADOW_CLEAR_FREE_NODE
;
1081 tmpret
= rcuja_shadow_clear(ja
->ht
,
1089 if (parent_shadow_node
)
1090 rcuja_shadow_unlock(parent_shadow_node
);
1093 rcuja_shadow_unlock(shadow_node
);
1099 * Lock the parent containing the hlist head pointer, and add node to list of
1100 * duplicates. Failure can happen if concurrent update changes the
1101 * parent before we get the lock. We return -EAGAIN in that case.
1102 * Return 0 on success, negative error value on failure.
1105 int ja_chain_node(struct cds_ja
*ja
,
1106 struct cds_ja_inode_flag
*parent_node_flag
,
1107 struct cds_ja_inode_flag
**node_flag_ptr
,
1108 struct cds_hlist_head
*head
,
1109 struct cds_ja_node
*node
)
1111 struct cds_ja_shadow_node
*shadow_node
;
1114 shadow_node
= rcuja_shadow_lookup_lock(ja
->ht
, parent_node_flag
);
1118 if (!ja_node_ptr(*node_flag_ptr
)) {
1122 cds_hlist_add_head_rcu(&node
->list
, head
);
1124 rcuja_shadow_unlock(shadow_node
);
1128 int cds_ja_add(struct cds_ja
*ja
, uint64_t key
,
1129 struct cds_ja_node
*new_node
)
1131 unsigned int tree_depth
, i
;
1132 struct cds_ja_inode_flag
**attach_node_flag_ptr
,
1134 struct cds_ja_inode_flag
*node_flag
,
1139 if (caa_unlikely(key
> ja
->key_max
)) {
1142 tree_depth
= ja
->tree_depth
;
1145 dbg_printf("cds_ja_add attempt: key %" PRIu64
", node %p\n",
1147 parent2_node_flag
= NULL
;
1149 (struct cds_ja_inode_flag
*) &ja
->root
; /* Use root ptr address as key for mutex */
1150 attach_node_flag_ptr
= &ja
->root
;
1151 node_flag_ptr
= &ja
->root
;
1152 node_flag
= rcu_dereference(ja
->root
);
1154 /* Iterate on all internal levels */
1155 for (i
= 1; i
< tree_depth
; i
++) {
1158 dbg_printf("cds_ja_add iter attach_node_flag_ptr %p node_flag_ptr %p node_flag %p\n",
1159 *attach_node_flag_ptr
, *node_flag_ptr
, node_flag
);
1160 if (!ja_node_ptr(node_flag
)) {
1161 ret
= ja_attach_node(ja
, attach_node_flag_ptr
,
1163 parent_node_flag
, parent2_node_flag
,
1165 if (ret
== -EAGAIN
|| ret
== -EEXIST
)
1170 iter_key
= (uint8_t) (key
>> (JA_BITS_PER_BYTE
* (tree_depth
- i
- 1)));
1171 parent2_node_flag
= parent_node_flag
;
1172 parent_node_flag
= node_flag
;
1173 node_flag
= ja_node_get_nth(node_flag
,
1174 &attach_node_flag_ptr
,
1177 dbg_printf("cds_ja_add iter key lookup %u finds node_flag %p attach_node_flag_ptr %p node_flag_ptr %p\n",
1178 (unsigned int) iter_key
, node_flag
,
1179 *attach_node_flag_ptr
,
1184 * We reached bottom of tree, simply add node to last internal
1185 * level, or chain it if key is already present.
1187 if (!ja_node_ptr(node_flag
)) {
1188 dbg_printf("cds_ja_add last attach_node_flag_ptr %p node_flag_ptr %p node_flag %p\n",
1189 *attach_node_flag_ptr
, *node_flag_ptr
, node_flag
);
1190 ret
= ja_attach_node(ja
, attach_node_flag_ptr
,
1191 node_flag_ptr
, parent_node_flag
,
1192 parent2_node_flag
, key
, i
, new_node
);
1194 ret
= ja_chain_node(ja
,
1197 (struct cds_hlist_head
*) attach_node_flag_ptr
,
1200 if (ret
== -EAGAIN
|| ret
== -EEXIST
)
1207 * Note: there is no need to lookup the pointer address associated with
1208 * each node's nth item after taking the lock: it's already been done by
1209 * cds_ja_del while holding the rcu read-side lock, and our node rules
1210 * ensure that when a match value -> pointer is found in a node, it is
1211 * _NEVER_ changed for that node without recompaction, and recompaction
1212 * reallocates the node.
1213 * However, when a child is removed from "linear" nodes, its pointer
1214 * is set to NULL. We therefore check, while holding the locks, if this
1215 * pointer is NULL, and return -ENOENT to the caller if it is the case.
1218 int ja_detach_node(struct cds_ja
*ja
,
1219 struct cds_ja_inode_flag
**snapshot
,
1220 struct cds_ja_inode_flag
***snapshot_ptr
,
1221 uint8_t *snapshot_n
,
1224 struct cds_ja_node
*node
)
1226 struct cds_ja_shadow_node
*shadow_nodes
[JA_MAX_DEPTH
];
1227 struct cds_ja_inode_flag
**node_flag_ptr
= NULL
,
1228 *parent_node_flag
= NULL
,
1229 **parent_node_flag_ptr
= NULL
;
1230 struct cds_ja_inode_flag
*iter_node_flag
;
1231 int ret
, i
, nr_shadow
= 0, nr_clear
= 0, nr_branch
= 0;
1234 assert(nr_snapshot
== ja
->tree_depth
+ 1);
1237 * From the last internal level node going up, get the node
1238 * lock, check if the node has only one child left. If it is the
1239 * case, we continue iterating upward. When we reach a node
1240 * which has more that one child left, we lock the parent, and
1241 * proceed to the node deletion (removing its children too).
1243 for (i
= nr_snapshot
- 2; i
>= 1; i
--) {
1244 struct cds_ja_shadow_node
*shadow_node
;
1246 shadow_node
= rcuja_shadow_lookup_lock(ja
->ht
,
1252 assert(shadow_node
->nr_child
> 0);
1253 shadow_nodes
[nr_shadow
++] = shadow_node
;
1254 if (shadow_node
->nr_child
== 1 && i
> 1)
1257 if (shadow_node
->nr_child
> 1 || i
== 1) {
1258 /* Lock parent and break */
1259 shadow_node
= rcuja_shadow_lookup_lock(ja
->ht
,
1265 shadow_nodes
[nr_shadow
++] = shadow_node
;
1266 node_flag_ptr
= snapshot_ptr
[i
+ 1];
1267 n
= snapshot_n
[i
+ 1];
1268 parent_node_flag_ptr
= snapshot_ptr
[i
];
1269 parent_node_flag
= snapshot
[i
];
1272 * Lock parent's parent, in case we need
1273 * to recompact parent.
1275 shadow_node
= rcuja_shadow_lookup_lock(ja
->ht
,
1281 shadow_nodes
[nr_shadow
++] = shadow_node
;
1288 * Check if node has been removed between RCU lookup and lock
1291 assert(node_flag_ptr
);
1292 if (!ja_node_ptr(*node_flag_ptr
)) {
1298 * At this point, we want to delete all nodes that are about to
1299 * be removed from shadow_nodes (except the last one, which is
1300 * either the root or the parent of the upmost node with 1
1301 * child). OK to as to free lock here, because RCU read lock is
1302 * held, and free only performed in call_rcu.
1305 for (i
= 0; i
< nr_clear
; i
++) {
1306 ret
= rcuja_shadow_clear(ja
->ht
,
1307 shadow_nodes
[i
]->node_flag
,
1309 RCUJA_SHADOW_CLEAR_FREE_NODE
1310 | RCUJA_SHADOW_CLEAR_FREE_LOCK
);
1314 iter_node_flag
= parent_node_flag
;
1315 /* Remove from parent */
1316 ret
= ja_node_clear_ptr(ja
,
1317 node_flag_ptr
, /* Pointer to location to nullify */
1318 &iter_node_flag
, /* Old new parent ptr in its parent */
1319 shadow_nodes
[nr_branch
- 1], /* of parent */
1324 dbg_printf("ja_detach_node: publish %p instead of %p\n",
1325 iter_node_flag
, *parent_node_flag_ptr
);
1326 /* Update address of parent ptr in its parent */
1327 rcu_assign_pointer(*parent_node_flag_ptr
, iter_node_flag
);
1330 for (i
= 0; i
< nr_shadow
; i
++)
1331 rcuja_shadow_unlock(shadow_nodes
[i
]);
1336 int ja_unchain_node(struct cds_ja
*ja
,
1337 struct cds_ja_inode_flag
*parent_node_flag
,
1338 struct cds_ja_inode_flag
**node_flag_ptr
,
1339 struct cds_hlist_head
*head
,
1340 struct cds_ja_node
*node
)
1342 struct cds_ja_shadow_node
*shadow_node
;
1343 struct cds_hlist_node
*hlist_node
;
1344 int ret
= 0, count
= 0;
1346 shadow_node
= rcuja_shadow_lookup_lock(ja
->ht
, parent_node_flag
);
1349 if (!ja_node_ptr(*node_flag_ptr
)) {
1354 * Retry if another thread removed all but one of duplicates
1355 * since check (this check was performed without lock).
1357 cds_hlist_for_each_rcu(hlist_node
, head
, list
) {
1365 cds_hlist_del_rcu(&node
->list
);
1367 rcuja_shadow_unlock(shadow_node
);
1372 * Called with RCU read lock held.
1374 int cds_ja_del(struct cds_ja
*ja
, uint64_t key
,
1375 struct cds_ja_node
*node
)
1377 unsigned int tree_depth
, i
;
1378 struct cds_ja_inode_flag
*snapshot
[JA_MAX_DEPTH
];
1379 struct cds_ja_inode_flag
**snapshot_ptr
[JA_MAX_DEPTH
];
1380 uint8_t snapshot_n
[JA_MAX_DEPTH
];
1381 struct cds_ja_inode_flag
*node_flag
;
1382 struct cds_ja_inode_flag
**prev_node_flag_ptr
,
1387 if (caa_unlikely(key
> ja
->key_max
))
1389 tree_depth
= ja
->tree_depth
;
1393 dbg_printf("cds_ja_del attempt: key %" PRIu64
", node %p\n",
1396 /* snapshot for level 0 is only for shadow node lookup */
1399 snapshot_ptr
[nr_snapshot
] = NULL
;
1400 snapshot
[nr_snapshot
++] = (struct cds_ja_inode_flag
*) &ja
->root
;
1401 node_flag
= rcu_dereference(ja
->root
);
1402 prev_node_flag_ptr
= &ja
->root
;
1403 node_flag_ptr
= &ja
->root
;
1405 /* Iterate on all internal levels */
1406 for (i
= 1; i
< tree_depth
; i
++) {
1409 dbg_printf("cds_ja_del iter node_flag %p\n",
1411 if (!ja_node_ptr(node_flag
)) {
1414 iter_key
= (uint8_t) (key
>> (JA_BITS_PER_BYTE
* (tree_depth
- i
- 1)));
1415 snapshot_n
[nr_snapshot
+ 1] = iter_key
;
1416 snapshot_ptr
[nr_snapshot
] = prev_node_flag_ptr
;
1417 snapshot
[nr_snapshot
++] = node_flag
;
1418 node_flag
= ja_node_get_nth(node_flag
,
1419 &prev_node_flag_ptr
,
1422 dbg_printf("cds_ja_del iter key lookup %u finds node_flag %p, prev_node_flag_ptr %p\n",
1423 (unsigned int) iter_key
, node_flag
,
1424 prev_node_flag_ptr
);
1428 * We reached bottom of tree, try to find the node we are trying
1429 * to remove. Fail if we cannot find it.
1431 if (!ja_node_ptr(node_flag
)) {
1432 dbg_printf("cds_ja_del: no node found for key %" PRIu64
"\n",
1436 struct cds_hlist_head hlist_head
;
1437 struct cds_hlist_node
*hlist_node
;
1438 struct cds_ja_node
*entry
, *match
= NULL
;
1442 (struct cds_hlist_node
*) ja_node_ptr(node_flag
);
1443 cds_hlist_for_each_entry_rcu(entry
,
1447 dbg_printf("cds_ja_del: compare %p with entry %p\n", node
, entry
);
1453 dbg_printf("cds_ja_del: no node match for node %p key %" PRIu64
"\n", node
, key
);
1459 * Removing last of duplicates. Last snapshot
1460 * does not have a shadow node (external leafs).
1462 snapshot_ptr
[nr_snapshot
] = prev_node_flag_ptr
;
1463 snapshot
[nr_snapshot
++] = node_flag
;
1464 ret
= ja_detach_node(ja
, snapshot
, snapshot_ptr
,
1465 snapshot_n
, nr_snapshot
, key
, node
);
1467 ret
= ja_unchain_node(ja
, snapshot
[nr_snapshot
- 1],
1468 node_flag_ptr
, &hlist_head
, match
);
1472 * Explanation of -ENOENT handling: caused by concurrent delete
1473 * between RCU lookup and actual removal. Need to re-do the
1474 * lookup and removal attempt.
1476 if (ret
== -EAGAIN
|| ret
== -ENOENT
)
1481 struct cds_ja
*_cds_ja_new(unsigned int key_bits
,
1482 const struct rcu_flavor_struct
*flavor
)
1486 struct cds_ja_shadow_node
*root_shadow_node
;
1488 ja
= calloc(sizeof(*ja
), 1);
1500 ja
->key_max
= (1ULL << key_bits
) - 1;
1503 ja
->key_max
= UINT64_MAX
;
1509 /* ja->root is NULL */
1510 /* tree_depth 0 is for pointer to root node */
1511 ja
->tree_depth
= (key_bits
>> JA_LOG2_BITS_PER_BYTE
) + 1;
1512 assert(ja
->tree_depth
<= JA_MAX_DEPTH
);
1513 ja
->ht
= rcuja_create_ht(flavor
);
1518 * Note: we should not free this node until judy array destroy.
1520 root_shadow_node
= rcuja_shadow_set(ja
->ht
,
1521 (struct cds_ja_inode_flag
*) &ja
->root
,
1523 if (!root_shadow_node
) {
1527 root_shadow_node
->level
= 0;
1532 ret
= rcuja_delete_ht(ja
->ht
);
1542 * Called from RCU read-side CS.
1544 __attribute__((visibility("protected")))
1545 void rcuja_free_all_children(struct cds_ja_shadow_node
*shadow_node
,
1546 struct cds_ja_inode_flag
*node_flag
,
1547 void (*free_node_cb
)(struct rcu_head
*head
))
1549 const struct rcu_flavor_struct
*flavor
;
1550 unsigned int type_index
;
1551 struct cds_ja_inode
*node
;
1552 const struct cds_ja_type
*type
;
1554 flavor
= cds_lfht_rcu_flavor(shadow_node
->ja
->ht
);
1555 node
= ja_node_ptr(node_flag
);
1556 assert(node
!= NULL
);
1557 type_index
= ja_node_type(node_flag
);
1558 type
= &ja_types
[type_index
];
1560 switch (type
->type_class
) {
1564 ja_linear_node_get_nr_child(type
, node
);
1567 for (i
= 0; i
< nr_child
; i
++) {
1568 struct cds_ja_inode_flag
*iter
;
1569 struct cds_hlist_head head
;
1570 struct cds_ja_node
*entry
;
1571 struct cds_hlist_node
*pos
;
1574 ja_linear_node_get_ith_pos(type
, node
, i
, &v
, &iter
);
1577 head
.next
= (struct cds_hlist_node
*) iter
;
1578 cds_hlist_for_each_entry_rcu(entry
, pos
, &head
, list
) {
1579 flavor
->update_call_rcu(&entry
->head
, free_node_cb
);
1586 unsigned int pool_nr
;
1588 for (pool_nr
= 0; pool_nr
< (1U << type
->nr_pool_order
); pool_nr
++) {
1589 struct cds_ja_inode
*pool
=
1590 ja_pool_node_get_ith_pool(type
, node
, pool_nr
);
1592 ja_linear_node_get_nr_child(type
, pool
);
1595 for (j
= 0; j
< nr_child
; j
++) {
1596 struct cds_ja_inode_flag
*iter
;
1597 struct cds_hlist_head head
;
1598 struct cds_ja_node
*entry
;
1599 struct cds_hlist_node
*pos
;
1602 ja_linear_node_get_ith_pos(type
, node
, j
, &v
, &iter
);
1605 head
.next
= (struct cds_hlist_node
*) iter
;
1606 cds_hlist_for_each_entry_rcu(entry
, pos
, &head
, list
) {
1607 flavor
->update_call_rcu(&entry
->head
, free_node_cb
);
1620 nr_child
= shadow_node
->nr_child
;
1621 for (i
= 0; i
< nr_child
; i
++) {
1622 struct cds_ja_inode_flag
*iter
;
1623 struct cds_hlist_head head
;
1624 struct cds_ja_node
*entry
;
1625 struct cds_hlist_node
*pos
;
1627 iter
= ja_pigeon_node_get_ith_pos(type
, node
, i
);
1630 head
.next
= (struct cds_hlist_node
*) iter
;
1631 cds_hlist_for_each_entry_rcu(entry
, pos
, &head
, list
) {
1632 flavor
->update_call_rcu(&entry
->head
, free_node_cb
);
1643 * There should be no more concurrent add to the judy array while it is
1644 * being destroyed (ensured by the caller).
1646 int cds_ja_destroy(struct cds_ja
*ja
,
1647 void (*free_node_cb
)(struct rcu_head
*head
))
1651 rcuja_shadow_prune(ja
->ht
,
1652 RCUJA_SHADOW_CLEAR_FREE_NODE
| RCUJA_SHADOW_CLEAR_FREE_LOCK
,
1654 ret
= rcuja_delete_ht(ja
->ht
);
1657 if (uatomic_read(&ja
->nr_fallback
))
1659 "[warning] RCU Judy Array used %lu fallback node(s)\n",
1660 uatomic_read(&ja
->nr_fallback
));