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
**child_node_flag_v
,
281 struct cds_ja_inode_flag
***node_flag_ptr
,
286 struct cds_ja_inode_flag
**pointers
;
287 struct cds_ja_inode_flag
*ptr
;
290 assert(type
->type_class
== RCU_JA_LINEAR
|| type
->type_class
== RCU_JA_POOL
);
292 nr_child
= ja_linear_node_get_nr_child(type
, node
);
293 cmm_smp_rmb(); /* read nr_child before values and pointers */
294 assert(nr_child
<= type
->max_linear_child
);
295 assert(type
->type_class
!= RCU_JA_LINEAR
|| nr_child
>= type
->min_child
);
297 values
= &node
->u
.data
[1];
298 for (i
= 0; i
< nr_child
; i
++) {
299 if (CMM_LOAD_SHARED(values
[i
]) == n
)
303 if (caa_unlikely(node_flag_ptr
))
304 *node_flag_ptr
= NULL
;
307 pointers
= (struct cds_ja_inode_flag
**) align_ptr_size(&values
[type
->max_linear_child
]);
308 ptr
= rcu_dereference(pointers
[i
]);
309 if (caa_unlikely(child_node_flag_ptr
) && ptr
)
310 *child_node_flag_ptr
= &pointers
[i
];
311 if (caa_unlikely(child_node_flag_v
) && ptr
)
312 *child_node_flag_v
= ptr
;
313 if (caa_unlikely(node_flag_ptr
))
314 *node_flag_ptr
= &pointers
[i
];
319 void ja_linear_node_get_ith_pos(const struct cds_ja_type
*type
,
320 struct cds_ja_inode
*node
,
323 struct cds_ja_inode_flag
**iter
)
326 struct cds_ja_inode_flag
**pointers
;
328 assert(type
->type_class
== RCU_JA_LINEAR
|| type
->type_class
== RCU_JA_POOL
);
329 assert(i
< ja_linear_node_get_nr_child(type
, node
));
331 values
= &node
->u
.data
[1];
333 pointers
= (struct cds_ja_inode_flag
**) align_ptr_size(&values
[type
->max_linear_child
]);
338 struct cds_ja_inode_flag
*ja_pool_node_get_nth(const struct cds_ja_type
*type
,
339 struct cds_ja_inode
*node
,
340 struct cds_ja_inode_flag
***child_node_flag_ptr
,
341 struct cds_ja_inode_flag
**child_node_flag_v
,
342 struct cds_ja_inode_flag
***node_flag_ptr
,
345 struct cds_ja_inode
*linear
;
347 assert(type
->type_class
== RCU_JA_POOL
);
349 * TODO: currently, we select the pool by highest bits. We
350 * should support various encodings.
352 linear
= (struct cds_ja_inode
*)
353 &node
->u
.data
[((unsigned long) n
>> (CHAR_BIT
- type
->nr_pool_order
)) << type
->pool_size_order
];
354 return ja_linear_node_get_nth(type
, linear
, child_node_flag_ptr
,
355 child_node_flag_v
, node_flag_ptr
, n
);
359 struct cds_ja_inode
*ja_pool_node_get_ith_pool(const struct cds_ja_type
*type
,
360 struct cds_ja_inode
*node
,
363 assert(type
->type_class
== RCU_JA_POOL
);
364 return (struct cds_ja_inode
*)
365 &node
->u
.data
[(unsigned int) i
<< type
->pool_size_order
];
369 struct cds_ja_inode_flag
*ja_pigeon_node_get_nth(const struct cds_ja_type
*type
,
370 struct cds_ja_inode
*node
,
371 struct cds_ja_inode_flag
***child_node_flag_ptr
,
372 struct cds_ja_inode_flag
**child_node_flag_v
,
373 struct cds_ja_inode_flag
***node_flag_ptr
,
376 struct cds_ja_inode_flag
**child_node_flag
;
377 struct cds_ja_inode_flag
*child_node_flag_read
;
379 assert(type
->type_class
== RCU_JA_PIGEON
);
380 child_node_flag
= &((struct cds_ja_inode_flag
**) node
->u
.data
)[n
];
381 child_node_flag_read
= rcu_dereference(*child_node_flag
);
382 dbg_printf("ja_pigeon_node_get_nth child_node_flag_ptr %p\n",
384 if (caa_unlikely(child_node_flag_ptr
) && child_node_flag_read
)
385 *child_node_flag_ptr
= child_node_flag
;
386 if (caa_unlikely(child_node_flag_v
) && child_node_flag_read
)
387 *child_node_flag_v
= child_node_flag_read
;
388 if (caa_unlikely(node_flag_ptr
))
389 *node_flag_ptr
= child_node_flag
;
390 return child_node_flag_read
;
394 struct cds_ja_inode_flag
*ja_pigeon_node_get_ith_pos(const struct cds_ja_type
*type
,
395 struct cds_ja_inode
*node
,
398 return ja_pigeon_node_get_nth(type
, node
, NULL
, NULL
, NULL
, i
);
402 * ja_node_get_nth: get nth item from a node.
403 * node_flag is already rcu_dereference'd.
406 struct cds_ja_inode_flag
*ja_node_get_nth(struct cds_ja_inode_flag
*node_flag
,
407 struct cds_ja_inode_flag
***child_node_flag_ptr
,
408 struct cds_ja_inode_flag
**child_node_flag
,
409 struct cds_ja_inode_flag
***node_flag_ptr
,
412 unsigned int type_index
;
413 struct cds_ja_inode
*node
;
414 const struct cds_ja_type
*type
;
416 node
= ja_node_ptr(node_flag
);
417 assert(node
!= NULL
);
418 type_index
= ja_node_type(node_flag
);
419 type
= &ja_types
[type_index
];
421 switch (type
->type_class
) {
423 return ja_linear_node_get_nth(type
, node
,
424 child_node_flag_ptr
, child_node_flag
,
427 return ja_pool_node_get_nth(type
, node
,
428 child_node_flag_ptr
, child_node_flag
,
431 return ja_pigeon_node_get_nth(type
, node
,
432 child_node_flag_ptr
, child_node_flag
,
436 return (void *) -1UL;
441 int ja_linear_node_set_nth(const struct cds_ja_type
*type
,
442 struct cds_ja_inode
*node
,
443 struct cds_ja_shadow_node
*shadow_node
,
445 struct cds_ja_inode_flag
*child_node_flag
)
448 uint8_t *values
, *nr_child_ptr
;
449 struct cds_ja_inode_flag
**pointers
;
450 unsigned int i
, unused
= 0;
452 assert(type
->type_class
== RCU_JA_LINEAR
|| type
->type_class
== RCU_JA_POOL
);
454 nr_child_ptr
= &node
->u
.data
[0];
455 dbg_printf("linear set nth: nr_child_ptr %p\n", nr_child_ptr
);
456 nr_child
= *nr_child_ptr
;
457 assert(nr_child
<= type
->max_linear_child
);
459 values
= &node
->u
.data
[1];
460 pointers
= (struct cds_ja_inode_flag
**) align_ptr_size(&values
[type
->max_linear_child
]);
461 /* Check if node value is already populated */
462 for (i
= 0; i
< nr_child
; i
++) {
463 if (values
[i
] == n
) {
473 if (i
== nr_child
&& nr_child
>= type
->max_linear_child
) {
475 return -ERANGE
; /* recompact node */
477 return -ENOSPC
; /* No space left in this node type */
480 assert(pointers
[i
] == NULL
);
481 rcu_assign_pointer(pointers
[i
], child_node_flag
);
482 /* If we expanded the nr_child, increment it */
484 CMM_STORE_SHARED(values
[nr_child
], n
);
485 /* write pointer and value before nr_child */
487 CMM_STORE_SHARED(*nr_child_ptr
, nr_child
+ 1);
489 shadow_node
->nr_child
++;
490 dbg_printf("linear set nth: %u child, shadow: %u child, for node %p shadow %p\n",
491 (unsigned int) CMM_LOAD_SHARED(*nr_child_ptr
),
492 (unsigned int) shadow_node
->nr_child
,
499 int ja_pool_node_set_nth(const struct cds_ja_type
*type
,
500 struct cds_ja_inode
*node
,
501 struct cds_ja_shadow_node
*shadow_node
,
503 struct cds_ja_inode_flag
*child_node_flag
)
505 struct cds_ja_inode
*linear
;
507 assert(type
->type_class
== RCU_JA_POOL
);
508 linear
= (struct cds_ja_inode
*)
509 &node
->u
.data
[((unsigned long) n
>> (CHAR_BIT
- type
->nr_pool_order
)) << type
->pool_size_order
];
510 return ja_linear_node_set_nth(type
, linear
, shadow_node
,
515 int ja_pigeon_node_set_nth(const struct cds_ja_type
*type
,
516 struct cds_ja_inode
*node
,
517 struct cds_ja_shadow_node
*shadow_node
,
519 struct cds_ja_inode_flag
*child_node_flag
)
521 struct cds_ja_inode_flag
**ptr
;
523 assert(type
->type_class
== RCU_JA_PIGEON
);
524 ptr
= &((struct cds_ja_inode_flag
**) node
->u
.data
)[n
];
527 rcu_assign_pointer(*ptr
, child_node_flag
);
528 shadow_node
->nr_child
++;
533 * _ja_node_set_nth: set nth item within a node. Return an error
534 * (negative error value) if it is already there.
537 int _ja_node_set_nth(const struct cds_ja_type
*type
,
538 struct cds_ja_inode
*node
,
539 struct cds_ja_shadow_node
*shadow_node
,
541 struct cds_ja_inode_flag
*child_node_flag
)
543 switch (type
->type_class
) {
545 return ja_linear_node_set_nth(type
, node
, shadow_node
, n
,
548 return ja_pool_node_set_nth(type
, node
, shadow_node
, n
,
551 return ja_pigeon_node_set_nth(type
, node
, shadow_node
, n
,
564 int ja_linear_node_clear_ptr(const struct cds_ja_type
*type
,
565 struct cds_ja_inode
*node
,
566 struct cds_ja_shadow_node
*shadow_node
,
567 struct cds_ja_inode_flag
**node_flag_ptr
)
570 uint8_t *nr_child_ptr
;
572 assert(type
->type_class
== RCU_JA_LINEAR
|| type
->type_class
== RCU_JA_POOL
);
574 nr_child_ptr
= &node
->u
.data
[0];
575 dbg_printf("linear clear ptr: nr_child_ptr %p\n", nr_child_ptr
);
576 nr_child
= *nr_child_ptr
;
577 assert(nr_child
<= type
->max_linear_child
);
579 if (shadow_node
->fallback_removal_count
) {
580 shadow_node
->fallback_removal_count
--;
582 if (shadow_node
->nr_child
<= type
->min_child
) {
583 /* We need to try recompacting the node */
587 assert(*node_flag_ptr
!= NULL
);
588 rcu_assign_pointer(*node_flag_ptr
, NULL
);
590 * Value and nr_child are never changed (would cause ABA issue).
591 * Instead, we leave the pointer to NULL and recompact the node
592 * once in a while. It is allowed to set a NULL pointer to a new
593 * value without recompaction though.
594 * Only update the shadow node accounting.
596 shadow_node
->nr_child
--;
597 dbg_printf("linear clear ptr: %u child, shadow: %u child, for node %p shadow %p\n",
598 (unsigned int) CMM_LOAD_SHARED(*nr_child_ptr
),
599 (unsigned int) shadow_node
->nr_child
,
606 int ja_pool_node_clear_ptr(const struct cds_ja_type
*type
,
607 struct cds_ja_inode
*node
,
608 struct cds_ja_shadow_node
*shadow_node
,
609 struct cds_ja_inode_flag
**node_flag_ptr
,
612 struct cds_ja_inode
*linear
;
614 assert(type
->type_class
== RCU_JA_POOL
);
615 linear
= (struct cds_ja_inode
*)
616 &node
->u
.data
[((unsigned long) n
>> (CHAR_BIT
- type
->nr_pool_order
)) << type
->pool_size_order
];
617 return ja_linear_node_clear_ptr(type
, linear
, shadow_node
, node_flag_ptr
);
621 int ja_pigeon_node_clear_ptr(const struct cds_ja_type
*type
,
622 struct cds_ja_inode
*node
,
623 struct cds_ja_shadow_node
*shadow_node
,
624 struct cds_ja_inode_flag
**node_flag_ptr
)
626 assert(type
->type_class
== RCU_JA_PIGEON
);
627 dbg_printf("ja_pigeon_node_clear_ptr: clearing ptr: %p\n", *node_flag_ptr
);
628 rcu_assign_pointer(*node_flag_ptr
, NULL
);
629 shadow_node
->nr_child
--;
634 * _ja_node_clear_ptr: clear ptr item within a node. Return an error
635 * (negative error value) if it is not found (-ENOENT).
638 int _ja_node_clear_ptr(const struct cds_ja_type
*type
,
639 struct cds_ja_inode
*node
,
640 struct cds_ja_shadow_node
*shadow_node
,
641 struct cds_ja_inode_flag
**node_flag_ptr
,
644 switch (type
->type_class
) {
646 return ja_linear_node_clear_ptr(type
, node
, shadow_node
, node_flag_ptr
);
648 return ja_pool_node_clear_ptr(type
, node
, shadow_node
, node_flag_ptr
, n
);
650 return ja_pigeon_node_clear_ptr(type
, node
, shadow_node
, node_flag_ptr
);
662 * ja_node_recompact_add: recompact a node, adding a new child.
663 * TODO: for pool type, take selection bit(s) into account.
664 * Return 0 on success, -EAGAIN if need to retry, or other negative
665 * error value otherwise.
668 int ja_node_recompact(enum ja_recompact mode
,
670 unsigned int old_type_index
,
671 const struct cds_ja_type
*old_type
,
672 struct cds_ja_inode
*old_node
,
673 struct cds_ja_shadow_node
*shadow_node
,
674 struct cds_ja_inode_flag
**old_node_flag_ptr
, uint8_t n
,
675 struct cds_ja_inode_flag
*child_node_flag
,
676 struct cds_ja_inode_flag
**nullify_node_flag_ptr
)
678 unsigned int new_type_index
;
679 struct cds_ja_inode
*new_node
;
680 struct cds_ja_shadow_node
*new_shadow_node
= NULL
;
681 const struct cds_ja_type
*new_type
;
682 struct cds_ja_inode_flag
*new_node_flag
, *old_node_flag
;
686 old_node_flag
= *old_node_flag_ptr
;
690 new_type_index
= old_type_index
;
692 case JA_RECOMPACT_ADD
:
693 if (!shadow_node
|| old_type_index
== NODE_INDEX_NULL
) {
696 new_type_index
= old_type_index
+ 1;
699 case JA_RECOMPACT_DEL
:
700 if (old_type_index
== 0) {
701 new_type_index
= NODE_INDEX_NULL
;
703 new_type_index
= old_type_index
- 1;
710 retry
: /* for fallback */
711 dbg_printf("Recompact from type %d to type %d\n",
712 old_type_index
, new_type_index
);
713 new_type
= &ja_types
[new_type_index
];
714 if (new_type_index
!= NODE_INDEX_NULL
) {
715 new_node
= alloc_cds_ja_node(new_type
);
718 new_node_flag
= ja_node_flag(new_node
, new_type_index
);
719 dbg_printf("Recompact inherit lock from %p\n", shadow_node
);
720 new_shadow_node
= rcuja_shadow_set(ja
->ht
, new_node_flag
, shadow_node
, ja
);
721 if (!new_shadow_node
) {
726 new_shadow_node
->fallback_removal_count
=
727 JA_FALLBACK_REMOVAL_COUNT
;
730 new_node_flag
= NULL
;
733 assert(mode
!= JA_RECOMPACT_ADD
|| old_type
->type_class
!= RCU_JA_PIGEON
);
735 if (new_type_index
== NODE_INDEX_NULL
)
738 switch (old_type
->type_class
) {
742 ja_linear_node_get_nr_child(old_type
, old_node
);
745 for (i
= 0; i
< nr_child
; i
++) {
746 struct cds_ja_inode_flag
*iter
;
749 ja_linear_node_get_ith_pos(old_type
, old_node
, i
, &v
, &iter
);
752 if (mode
== JA_RECOMPACT_DEL
&& *nullify_node_flag_ptr
== iter
)
754 ret
= _ja_node_set_nth(new_type
, new_node
,
757 if (new_type
->type_class
== RCU_JA_POOL
&& ret
) {
758 goto fallback_toosmall
;
766 unsigned int pool_nr
;
768 for (pool_nr
= 0; pool_nr
< (1U << old_type
->nr_pool_order
); pool_nr
++) {
769 struct cds_ja_inode
*pool
=
770 ja_pool_node_get_ith_pool(old_type
,
773 ja_linear_node_get_nr_child(old_type
, pool
);
776 for (j
= 0; j
< nr_child
; j
++) {
777 struct cds_ja_inode_flag
*iter
;
780 ja_linear_node_get_ith_pos(old_type
, pool
,
784 if (mode
== JA_RECOMPACT_DEL
&& *nullify_node_flag_ptr
== iter
)
786 ret
= _ja_node_set_nth(new_type
, new_node
,
789 if (new_type
->type_class
== RCU_JA_POOL
791 goto fallback_toosmall
;
799 assert(mode
== JA_RECOMPACT_ADD
);
806 assert(mode
== JA_RECOMPACT_DEL
);
807 nr_child
= shadow_node
->nr_child
;
808 for (i
= 0; i
< nr_child
; i
++) {
809 struct cds_ja_inode_flag
*iter
;
811 iter
= ja_pigeon_node_get_ith_pos(old_type
, old_node
, i
);
814 if (mode
== JA_RECOMPACT_DEL
&& *nullify_node_flag_ptr
== iter
)
816 ret
= _ja_node_set_nth(new_type
, new_node
,
819 if (new_type
->type_class
== RCU_JA_POOL
&& ret
) {
820 goto fallback_toosmall
;
833 if (mode
== JA_RECOMPACT_ADD
) {
835 ret
= _ja_node_set_nth(new_type
, new_node
,
838 if (new_type
->type_class
== RCU_JA_POOL
&& ret
) {
839 goto fallback_toosmall
;
843 /* Return pointer to new recompacted node through old_node_flag_ptr */
844 *old_node_flag_ptr
= new_node_flag
;
848 flags
= RCUJA_SHADOW_CLEAR_FREE_NODE
;
850 * It is OK to free the lock associated with a node
851 * going to NULL, since we are holding the parent lock.
852 * This synchronizes removal with re-add of that node.
854 if (new_type_index
== NODE_INDEX_NULL
)
855 flags
= RCUJA_SHADOW_CLEAR_FREE_LOCK
;
856 ret
= rcuja_shadow_clear(ja
->ht
, old_node_flag
, shadow_node
,
866 /* fallback if next pool is too small */
867 assert(new_shadow_node
);
868 ret
= rcuja_shadow_clear(ja
->ht
, new_node_flag
, new_shadow_node
,
869 RCUJA_SHADOW_CLEAR_FREE_NODE
);
872 /* Choose fallback type: pigeon */
873 new_type_index
= (1UL << JA_TYPE_BITS
) - 1;
874 dbg_printf("Fallback to type %d\n", new_type_index
);
875 uatomic_inc(&ja
->nr_fallback
);
881 * Return 0 on success, -EAGAIN if need to retry, or other negative
882 * error value otherwise.
885 int ja_node_set_nth(struct cds_ja
*ja
,
886 struct cds_ja_inode_flag
**node_flag
, uint8_t n
,
887 struct cds_ja_inode_flag
*child_node_flag
,
888 struct cds_ja_shadow_node
*shadow_node
)
891 unsigned int type_index
;
892 const struct cds_ja_type
*type
;
893 struct cds_ja_inode
*node
;
895 dbg_printf("ja_node_set_nth for n=%u, node %p, shadow %p\n",
896 (unsigned int) n
, ja_node_ptr(*node_flag
), shadow_node
);
898 node
= ja_node_ptr(*node_flag
);
899 type_index
= ja_node_type(*node_flag
);
900 type
= &ja_types
[type_index
];
901 ret
= _ja_node_set_nth(type
, node
, shadow_node
,
905 /* Not enough space in node, need to recompact. */
906 ret
= ja_node_recompact(JA_RECOMPACT_ADD
, ja
, type_index
, type
, node
,
907 shadow_node
, node_flag
, n
, child_node_flag
, NULL
);
910 /* Node needs to be recompacted. */
911 ret
= ja_node_recompact(JA_RECOMPACT
, ja
, type_index
, type
, node
,
912 shadow_node
, node_flag
, n
, child_node_flag
, NULL
);
919 * Return 0 on success, -EAGAIN if need to retry, or other negative
920 * error value otherwise.
923 int ja_node_clear_ptr(struct cds_ja
*ja
,
924 struct cds_ja_inode_flag
**node_flag_ptr
, /* Pointer to location to nullify */
925 struct cds_ja_inode_flag
**parent_node_flag_ptr
, /* Address of parent ptr in its parent */
926 struct cds_ja_shadow_node
*shadow_node
, /* of parent */
930 unsigned int type_index
;
931 const struct cds_ja_type
*type
;
932 struct cds_ja_inode
*node
;
934 dbg_printf("ja_node_clear_ptr for node %p, shadow %p, target ptr %p\n",
935 ja_node_ptr(*parent_node_flag_ptr
), shadow_node
, node_flag_ptr
);
937 node
= ja_node_ptr(*parent_node_flag_ptr
);
938 type_index
= ja_node_type(*parent_node_flag_ptr
);
939 type
= &ja_types
[type_index
];
940 ret
= _ja_node_clear_ptr(type
, node
, shadow_node
, node_flag_ptr
, n
);
942 /* Should to try recompaction. */
943 ret
= ja_node_recompact(JA_RECOMPACT_DEL
, ja
, type_index
, type
, node
,
944 shadow_node
, parent_node_flag_ptr
, n
, NULL
,
950 struct cds_hlist_head
cds_ja_lookup(struct cds_ja
*ja
, uint64_t key
)
952 unsigned int tree_depth
, i
;
953 struct cds_ja_inode_flag
*node_flag
;
954 struct cds_hlist_head head
= { NULL
};
956 if (caa_unlikely(key
> ja
->key_max
))
958 tree_depth
= ja
->tree_depth
;
959 node_flag
= rcu_dereference(ja
->root
);
961 /* level 0: root node */
962 if (!ja_node_ptr(node_flag
))
965 for (i
= 1; i
< tree_depth
; i
++) {
968 iter_key
= (uint8_t) (key
>> (JA_BITS_PER_BYTE
* (tree_depth
- i
- 1)));
969 node_flag
= ja_node_get_nth(node_flag
, NULL
, NULL
, NULL
,
971 dbg_printf("cds_ja_lookup iter key lookup %u finds node_flag %p\n",
972 (unsigned int) iter_key
, node_flag
);
973 if (!ja_node_ptr(node_flag
))
977 /* Last level lookup succeded. We got an actual match. */
978 head
.next
= (struct cds_hlist_node
*) node_flag
;
983 * We reached an unpopulated node. Create it and the children we need,
984 * and then attach the entire branch to the current node. This may
985 * trigger recompaction of the current node. Locks needed: node lock
986 * (for add), and, possibly, parent node lock (to update pointer due to
987 * node recompaction).
989 * First take node lock, check if recompaction is needed, then take
990 * parent lock (if needed). Then we can proceed to create the new
991 * branch. Publish the new branch, and release locks.
992 * TODO: we currently always take the parent lock even when not needed.
995 int ja_attach_node(struct cds_ja
*ja
,
996 struct cds_ja_inode_flag
**attach_node_flag_ptr
,
997 struct cds_ja_inode_flag
*attach_node_flag
,
998 struct cds_ja_inode_flag
**node_flag_ptr
,
999 struct cds_ja_inode_flag
*node_flag
,
1000 struct cds_ja_inode_flag
*parent_node_flag
,
1003 struct cds_ja_node
*child_node
)
1005 struct cds_ja_shadow_node
*shadow_node
= NULL
,
1006 *parent_shadow_node
= NULL
;
1007 struct cds_ja_inode
*node
= ja_node_ptr(node_flag
);
1008 struct cds_ja_inode
*parent_node
= ja_node_ptr(parent_node_flag
);
1009 struct cds_hlist_head head
;
1010 struct cds_ja_inode_flag
*iter_node_flag
, *iter_dest_node_flag
;
1012 struct cds_ja_inode_flag
*created_nodes
[JA_MAX_DEPTH
];
1013 int nr_created_nodes
= 0;
1015 dbg_printf("Attach node at level %u (node %p, node_flag %p)\n",
1016 level
, node
, node_flag
);
1019 shadow_node
= rcuja_shadow_lookup_lock(ja
->ht
, node_flag
);
1025 parent_shadow_node
= rcuja_shadow_lookup_lock(ja
->ht
,
1027 if (!parent_shadow_node
) {
1033 if (node_flag_ptr
&& ja_node_ptr(*node_flag_ptr
)) {
1035 * Target node has been updated between RCU lookup and
1036 * lock acquisition. We need to re-try lookup and
1043 if (attach_node_flag_ptr
&& ja_node_ptr(*attach_node_flag_ptr
) !=
1044 ja_node_ptr(attach_node_flag
)) {
1046 * Target node has been updated between RCU lookup and
1047 * lock acquisition. We need to re-try lookup and
1054 /* Create new branch, starting from bottom */
1055 CDS_INIT_HLIST_HEAD(&head
);
1056 cds_hlist_add_head_rcu(&child_node
->list
, &head
);
1057 iter_node_flag
= (struct cds_ja_inode_flag
*) head
.next
;
1059 for (i
= ja
->tree_depth
; i
> (int) level
; i
--) {
1062 iter_key
= (uint8_t) (key
>> (JA_BITS_PER_BYTE
* (ja
->tree_depth
- i
)));
1063 dbg_printf("branch creation level %d, key %u\n",
1064 i
- 1, (unsigned int) iter_key
);
1065 iter_dest_node_flag
= NULL
;
1066 ret
= ja_node_set_nth(ja
, &iter_dest_node_flag
,
1072 created_nodes
[nr_created_nodes
++] = iter_dest_node_flag
;
1073 iter_node_flag
= iter_dest_node_flag
;
1079 iter_key
= (uint8_t) (key
>> (JA_BITS_PER_BYTE
* (ja
->tree_depth
- level
)));
1080 /* We need to use set_nth on the previous level. */
1081 iter_dest_node_flag
= node_flag
;
1082 ret
= ja_node_set_nth(ja
, &iter_dest_node_flag
,
1088 created_nodes
[nr_created_nodes
++] = iter_dest_node_flag
;
1089 iter_node_flag
= iter_dest_node_flag
;
1092 /* Publish new branch */
1093 dbg_printf("Publish branch %p, replacing %p\n",
1094 iter_node_flag
, *attach_node_flag_ptr
);
1095 rcu_assign_pointer(*attach_node_flag_ptr
, iter_node_flag
);
1102 for (i
= 0; i
< nr_created_nodes
; i
++) {
1106 flags
= RCUJA_SHADOW_CLEAR_FREE_LOCK
;
1108 flags
|= RCUJA_SHADOW_CLEAR_FREE_NODE
;
1109 tmpret
= rcuja_shadow_clear(ja
->ht
,
1117 if (parent_shadow_node
)
1118 rcuja_shadow_unlock(parent_shadow_node
);
1121 rcuja_shadow_unlock(shadow_node
);
1127 * Lock the parent containing the hlist head pointer, and add node to list of
1128 * duplicates. Failure can happen if concurrent update changes the
1129 * parent before we get the lock. We return -EAGAIN in that case.
1130 * Return 0 on success, negative error value on failure.
1133 int ja_chain_node(struct cds_ja
*ja
,
1134 struct cds_ja_inode_flag
*parent_node_flag
,
1135 struct cds_ja_inode_flag
**node_flag_ptr
,
1136 struct cds_ja_inode_flag
*node_flag
,
1137 struct cds_hlist_head
*head
,
1138 struct cds_ja_node
*node
)
1140 struct cds_ja_shadow_node
*shadow_node
;
1143 shadow_node
= rcuja_shadow_lookup_lock(ja
->ht
, parent_node_flag
);
1147 if (ja_node_ptr(*node_flag_ptr
) != ja_node_ptr(node_flag
)) {
1151 cds_hlist_add_head_rcu(&node
->list
, head
);
1153 rcuja_shadow_unlock(shadow_node
);
1157 int cds_ja_add(struct cds_ja
*ja
, uint64_t key
,
1158 struct cds_ja_node
*new_node
)
1160 unsigned int tree_depth
, i
;
1161 struct cds_ja_inode_flag
**attach_node_flag_ptr
,
1163 struct cds_ja_inode_flag
*node_flag
,
1169 if (caa_unlikely(key
> ja
->key_max
)) {
1172 tree_depth
= ja
->tree_depth
;
1175 dbg_printf("cds_ja_add attempt: key %" PRIu64
", node %p\n",
1177 parent2_node_flag
= NULL
;
1179 (struct cds_ja_inode_flag
*) &ja
->root
; /* Use root ptr address as key for mutex */
1180 attach_node_flag_ptr
= &ja
->root
;
1181 attach_node_flag
= rcu_dereference(ja
->root
);
1182 node_flag_ptr
= &ja
->root
;
1183 node_flag
= rcu_dereference(ja
->root
);
1185 /* Iterate on all internal levels */
1186 for (i
= 1; i
< tree_depth
; i
++) {
1189 dbg_printf("cds_ja_add iter attach_node_flag_ptr %p node_flag_ptr %p node_flag %p\n",
1190 attach_node_flag_ptr
, node_flag_ptr
, node_flag
);
1191 if (!ja_node_ptr(node_flag
)) {
1192 ret
= ja_attach_node(ja
, attach_node_flag_ptr
,
1198 if (ret
== -EAGAIN
|| ret
== -EEXIST
)
1203 iter_key
= (uint8_t) (key
>> (JA_BITS_PER_BYTE
* (tree_depth
- i
- 1)));
1204 parent2_node_flag
= parent_node_flag
;
1205 parent_node_flag
= node_flag
;
1206 node_flag
= ja_node_get_nth(node_flag
,
1207 &attach_node_flag_ptr
,
1211 dbg_printf("cds_ja_add iter key lookup %u finds node_flag %p attach_node_flag_ptr %p node_flag_ptr %p\n",
1212 (unsigned int) iter_key
, node_flag
,
1213 attach_node_flag_ptr
,
1218 * We reached bottom of tree, simply add node to last internal
1219 * level, or chain it if key is already present.
1221 if (!ja_node_ptr(node_flag
)) {
1222 dbg_printf("cds_ja_add attach_node_flag_ptr %p node_flag_ptr %p node_flag %p\n",
1223 attach_node_flag_ptr
, node_flag_ptr
, node_flag
);
1224 ret
= ja_attach_node(ja
, attach_node_flag_ptr
,
1226 node_flag_ptr
, parent_node_flag
,
1227 parent2_node_flag
, key
, i
, new_node
);
1229 ret
= ja_chain_node(ja
,
1233 (struct cds_hlist_head
*) attach_node_flag_ptr
,
1236 if (ret
== -EAGAIN
|| ret
== -EEXIST
)
1243 * Note: there is no need to lookup the pointer address associated with
1244 * each node's nth item after taking the lock: it's already been done by
1245 * cds_ja_del while holding the rcu read-side lock, and our node rules
1246 * ensure that when a match value -> pointer is found in a node, it is
1247 * _NEVER_ changed for that node without recompaction, and recompaction
1248 * reallocates the node.
1249 * However, when a child is removed from "linear" nodes, its pointer
1250 * is set to NULL. We therefore check, while holding the locks, if this
1251 * pointer is NULL, and return -ENOENT to the caller if it is the case.
1254 int ja_detach_node(struct cds_ja
*ja
,
1255 struct cds_ja_inode_flag
**snapshot
,
1256 struct cds_ja_inode_flag
***snapshot_ptr
,
1257 uint8_t *snapshot_n
,
1260 struct cds_ja_node
*node
)
1262 struct cds_ja_shadow_node
*shadow_nodes
[JA_MAX_DEPTH
];
1263 struct cds_ja_inode_flag
**node_flag_ptr
= NULL
,
1264 *parent_node_flag
= NULL
,
1265 **parent_node_flag_ptr
= NULL
;
1266 struct cds_ja_inode_flag
*iter_node_flag
;
1267 int ret
, i
, nr_shadow
= 0, nr_clear
= 0, nr_branch
= 0;
1270 assert(nr_snapshot
== ja
->tree_depth
+ 1);
1273 * From the last internal level node going up, get the node
1274 * lock, check if the node has only one child left. If it is the
1275 * case, we continue iterating upward. When we reach a node
1276 * which has more that one child left, we lock the parent, and
1277 * proceed to the node deletion (removing its children too).
1279 for (i
= nr_snapshot
- 2; i
>= 1; i
--) {
1280 struct cds_ja_shadow_node
*shadow_node
;
1282 shadow_node
= rcuja_shadow_lookup_lock(ja
->ht
,
1288 shadow_nodes
[nr_shadow
++] = shadow_node
;
1291 * Check if node has been removed between RCU
1292 * lookup and lock acquisition.
1294 assert(snapshot_ptr
[i
+ 1]);
1295 if (ja_node_ptr(*snapshot_ptr
[i
+ 1])
1296 != ja_node_ptr(snapshot
[i
+ 1])) {
1301 assert(shadow_node
->nr_child
> 0);
1302 if (shadow_node
->nr_child
== 1 && i
> 1)
1305 if (shadow_node
->nr_child
> 1 || i
== 1) {
1306 /* Lock parent and break */
1307 shadow_node
= rcuja_shadow_lookup_lock(ja
->ht
,
1313 shadow_nodes
[nr_shadow
++] = shadow_node
;
1316 * Check if node has been removed between RCU
1317 * lookup and lock acquisition.
1319 assert(snapshot_ptr
[i
]);
1320 if (ja_node_ptr(*snapshot_ptr
[i
])
1321 != ja_node_ptr(snapshot
[i
])) {
1326 node_flag_ptr
= snapshot_ptr
[i
+ 1];
1327 n
= snapshot_n
[i
+ 1];
1328 parent_node_flag_ptr
= snapshot_ptr
[i
];
1329 parent_node_flag
= snapshot
[i
];
1333 * Lock parent's parent, in case we need
1334 * to recompact parent.
1336 shadow_node
= rcuja_shadow_lookup_lock(ja
->ht
,
1342 shadow_nodes
[nr_shadow
++] = shadow_node
;
1345 * Check if node has been removed between RCU
1346 * lookup and lock acquisition.
1348 assert(snapshot_ptr
[i
- 1]);
1349 if (ja_node_ptr(*snapshot_ptr
[i
- 1])
1350 != ja_node_ptr(snapshot
[i
- 1])) {
1361 * At this point, we want to delete all nodes that are about to
1362 * be removed from shadow_nodes (except the last one, which is
1363 * either the root or the parent of the upmost node with 1
1364 * child). OK to free lock here, because RCU read lock is held,
1365 * and free only performed in call_rcu.
1368 for (i
= 0; i
< nr_clear
; i
++) {
1369 ret
= rcuja_shadow_clear(ja
->ht
,
1370 shadow_nodes
[i
]->node_flag
,
1372 RCUJA_SHADOW_CLEAR_FREE_NODE
1373 | RCUJA_SHADOW_CLEAR_FREE_LOCK
);
1377 iter_node_flag
= parent_node_flag
;
1378 /* Remove from parent */
1379 ret
= ja_node_clear_ptr(ja
,
1380 node_flag_ptr
, /* Pointer to location to nullify */
1381 &iter_node_flag
, /* Old new parent ptr in its parent */
1382 shadow_nodes
[nr_branch
- 1], /* of parent */
1387 dbg_printf("ja_detach_node: publish %p instead of %p\n",
1388 iter_node_flag
, *parent_node_flag_ptr
);
1389 /* Update address of parent ptr in its parent */
1390 rcu_assign_pointer(*parent_node_flag_ptr
, iter_node_flag
);
1393 for (i
= 0; i
< nr_shadow
; i
++)
1394 rcuja_shadow_unlock(shadow_nodes
[i
]);
1399 int ja_unchain_node(struct cds_ja
*ja
,
1400 struct cds_ja_inode_flag
*parent_node_flag
,
1401 struct cds_ja_inode_flag
**node_flag_ptr
,
1402 struct cds_ja_inode_flag
*node_flag
,
1403 struct cds_ja_node
*node
)
1405 struct cds_ja_shadow_node
*shadow_node
;
1406 struct cds_hlist_node
*hlist_node
;
1407 struct cds_hlist_head hlist_head
;
1408 int ret
= 0, count
= 0, found
= 0;
1410 shadow_node
= rcuja_shadow_lookup_lock(ja
->ht
, parent_node_flag
);
1413 if (ja_node_ptr(*node_flag_ptr
) != ja_node_ptr(node_flag
)) {
1417 hlist_head
.next
= (struct cds_hlist_node
*) ja_node_ptr(node_flag
);
1419 * Retry if another thread removed all but one of duplicates
1420 * since check (this check was performed without lock).
1421 * Ensure that the node we are about to remove is still in the
1422 * list (while holding lock).
1424 cds_hlist_for_each_rcu(hlist_node
, &hlist_head
) {
1426 /* FIXME: currently a work-around */
1427 hlist_node
->prev
= (struct cds_hlist_node
*) node_flag_ptr
;
1430 if (hlist_node
== &node
->list
)
1434 if (!found
|| count
== 1) {
1438 cds_hlist_del_rcu(&node
->list
);
1440 * Validate that we indeed removed the node from linked list.
1442 assert(ja_node_ptr(*node_flag_ptr
) != (struct cds_ja_inode
*) node
);
1444 rcuja_shadow_unlock(shadow_node
);
1449 * Called with RCU read lock held.
1451 int cds_ja_del(struct cds_ja
*ja
, uint64_t key
,
1452 struct cds_ja_node
*node
)
1454 unsigned int tree_depth
, i
;
1455 struct cds_ja_inode_flag
*snapshot
[JA_MAX_DEPTH
];
1456 struct cds_ja_inode_flag
**snapshot_ptr
[JA_MAX_DEPTH
];
1457 uint8_t snapshot_n
[JA_MAX_DEPTH
];
1458 struct cds_ja_inode_flag
*node_flag
;
1459 struct cds_ja_inode_flag
**prev_node_flag_ptr
,
1464 if (caa_unlikely(key
> ja
->key_max
))
1466 tree_depth
= ja
->tree_depth
;
1470 dbg_printf("cds_ja_del attempt: key %" PRIu64
", node %p\n",
1473 /* snapshot for level 0 is only for shadow node lookup */
1476 snapshot_ptr
[nr_snapshot
] = NULL
;
1477 snapshot
[nr_snapshot
++] = (struct cds_ja_inode_flag
*) &ja
->root
;
1478 node_flag
= rcu_dereference(ja
->root
);
1479 prev_node_flag_ptr
= &ja
->root
;
1480 node_flag_ptr
= &ja
->root
;
1482 /* Iterate on all internal levels */
1483 for (i
= 1; i
< tree_depth
; i
++) {
1486 dbg_printf("cds_ja_del iter node_flag %p\n",
1488 if (!ja_node_ptr(node_flag
)) {
1491 iter_key
= (uint8_t) (key
>> (JA_BITS_PER_BYTE
* (tree_depth
- i
- 1)));
1492 snapshot_n
[nr_snapshot
+ 1] = iter_key
;
1493 snapshot_ptr
[nr_snapshot
] = prev_node_flag_ptr
;
1494 snapshot
[nr_snapshot
++] = node_flag
;
1495 node_flag
= ja_node_get_nth(node_flag
,
1496 &prev_node_flag_ptr
,
1500 dbg_printf("cds_ja_del iter key lookup %u finds node_flag %p, prev_node_flag_ptr %p\n",
1501 (unsigned int) iter_key
, node_flag
,
1502 prev_node_flag_ptr
);
1505 * We reached bottom of tree, try to find the node we are trying
1506 * to remove. Fail if we cannot find it.
1508 if (!ja_node_ptr(node_flag
)) {
1509 dbg_printf("cds_ja_del: no node found for key %" PRIu64
"\n",
1513 struct cds_hlist_head hlist_head
;
1514 struct cds_hlist_node
*hlist_node
;
1515 struct cds_ja_node
*entry
, *match
= NULL
;
1519 (struct cds_hlist_node
*) ja_node_ptr(node_flag
);
1520 cds_hlist_for_each_entry_rcu(entry
,
1524 dbg_printf("cds_ja_del: compare %p with entry %p\n", node
, entry
);
1530 dbg_printf("cds_ja_del: no node match for node %p key %" PRIu64
"\n", node
, key
);
1536 * Removing last of duplicates. Last snapshot
1537 * does not have a shadow node (external leafs).
1539 snapshot_ptr
[nr_snapshot
] = prev_node_flag_ptr
;
1540 snapshot
[nr_snapshot
++] = node_flag
;
1541 ret
= ja_detach_node(ja
, snapshot
, snapshot_ptr
,
1542 snapshot_n
, nr_snapshot
, key
, node
);
1544 ret
= ja_unchain_node(ja
, snapshot
[nr_snapshot
- 1],
1545 node_flag_ptr
, node_flag
, match
);
1549 * Explanation of -ENOENT handling: caused by concurrent delete
1550 * between RCU lookup and actual removal. Need to re-do the
1551 * lookup and removal attempt.
1553 if (ret
== -EAGAIN
|| ret
== -ENOENT
)
1558 struct cds_ja
*_cds_ja_new(unsigned int key_bits
,
1559 const struct rcu_flavor_struct
*flavor
)
1563 struct cds_ja_shadow_node
*root_shadow_node
;
1565 ja
= calloc(sizeof(*ja
), 1);
1577 ja
->key_max
= (1ULL << key_bits
) - 1;
1580 ja
->key_max
= UINT64_MAX
;
1586 /* ja->root is NULL */
1587 /* tree_depth 0 is for pointer to root node */
1588 ja
->tree_depth
= (key_bits
>> JA_LOG2_BITS_PER_BYTE
) + 1;
1589 assert(ja
->tree_depth
<= JA_MAX_DEPTH
);
1590 ja
->ht
= rcuja_create_ht(flavor
);
1595 * Note: we should not free this node until judy array destroy.
1597 root_shadow_node
= rcuja_shadow_set(ja
->ht
,
1598 (struct cds_ja_inode_flag
*) &ja
->root
,
1600 if (!root_shadow_node
) {
1604 root_shadow_node
->level
= 0;
1609 ret
= rcuja_delete_ht(ja
->ht
);
1619 * Called from RCU read-side CS.
1621 __attribute__((visibility("protected")))
1622 void rcuja_free_all_children(struct cds_ja_shadow_node
*shadow_node
,
1623 struct cds_ja_inode_flag
*node_flag
,
1624 void (*free_node_cb
)(struct rcu_head
*head
))
1626 const struct rcu_flavor_struct
*flavor
;
1627 unsigned int type_index
;
1628 struct cds_ja_inode
*node
;
1629 const struct cds_ja_type
*type
;
1631 flavor
= cds_lfht_rcu_flavor(shadow_node
->ja
->ht
);
1632 node
= ja_node_ptr(node_flag
);
1633 assert(node
!= NULL
);
1634 type_index
= ja_node_type(node_flag
);
1635 type
= &ja_types
[type_index
];
1637 switch (type
->type_class
) {
1641 ja_linear_node_get_nr_child(type
, node
);
1644 for (i
= 0; i
< nr_child
; i
++) {
1645 struct cds_ja_inode_flag
*iter
;
1646 struct cds_hlist_head head
;
1647 struct cds_ja_node
*entry
;
1648 struct cds_hlist_node
*pos
;
1651 ja_linear_node_get_ith_pos(type
, node
, i
, &v
, &iter
);
1654 head
.next
= (struct cds_hlist_node
*) iter
;
1655 cds_hlist_for_each_entry_rcu(entry
, pos
, &head
, list
) {
1656 flavor
->update_call_rcu(&entry
->head
, free_node_cb
);
1663 unsigned int pool_nr
;
1665 for (pool_nr
= 0; pool_nr
< (1U << type
->nr_pool_order
); pool_nr
++) {
1666 struct cds_ja_inode
*pool
=
1667 ja_pool_node_get_ith_pool(type
, node
, pool_nr
);
1669 ja_linear_node_get_nr_child(type
, pool
);
1672 for (j
= 0; j
< nr_child
; j
++) {
1673 struct cds_ja_inode_flag
*iter
;
1674 struct cds_hlist_head head
;
1675 struct cds_ja_node
*entry
;
1676 struct cds_hlist_node
*pos
;
1679 ja_linear_node_get_ith_pos(type
, node
, j
, &v
, &iter
);
1682 head
.next
= (struct cds_hlist_node
*) iter
;
1683 cds_hlist_for_each_entry_rcu(entry
, pos
, &head
, list
) {
1684 flavor
->update_call_rcu(&entry
->head
, free_node_cb
);
1697 nr_child
= shadow_node
->nr_child
;
1698 for (i
= 0; i
< nr_child
; i
++) {
1699 struct cds_ja_inode_flag
*iter
;
1700 struct cds_hlist_head head
;
1701 struct cds_ja_node
*entry
;
1702 struct cds_hlist_node
*pos
;
1704 iter
= ja_pigeon_node_get_ith_pos(type
, node
, i
);
1707 head
.next
= (struct cds_hlist_node
*) iter
;
1708 cds_hlist_for_each_entry_rcu(entry
, pos
, &head
, list
) {
1709 flavor
->update_call_rcu(&entry
->head
, free_node_cb
);
1720 * There should be no more concurrent add to the judy array while it is
1721 * being destroyed (ensured by the caller).
1723 int cds_ja_destroy(struct cds_ja
*ja
,
1724 void (*free_node_cb
)(struct rcu_head
*head
))
1728 rcuja_shadow_prune(ja
->ht
,
1729 RCUJA_SHADOW_CLEAR_FREE_NODE
| RCUJA_SHADOW_CLEAR_FREE_LOCK
,
1731 ret
= rcuja_delete_ht(ja
->ht
);
1734 if (uatomic_read(&ja
->nr_fallback
))
1736 "[warning] RCU Judy Array used %lu fallback node(s)\n",
1737 uatomic_read(&ja
->nr_fallback
));