4 * Userspace RCU library - RCU Judy Array
6 * Copyright (C) 2000 - 2002 Hewlett-Packard Company
7 * Copyright 2012-2013 - Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
9 * This library is free software; you can redistribute it and/or
10 * modify it under the terms of the GNU Lesser General Public
11 * License as published by the Free Software Foundation; either
12 * version 2.1 of the License, or (at your option) any later version.
14 * This library is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
17 * Lesser General Public License for more details.
19 * You should have received a copy of the GNU Lesser General Public
20 * License along with this library; if not, write to the Free Software
21 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
29 #include <urcu/rcuja.h>
30 #include <urcu/compiler.h>
31 #include <urcu/arch.h>
33 #include <urcu-pointer.h>
34 #include <urcu/uatomic.h>
37 #include "rcuja-internal.h"
40 #define abs_int(a) ((int) (a) > 0 ? (int) (a) : -((int) (a)))
43 enum cds_ja_type_class
{
44 RCU_JA_LINEAR
= 0, /* Type A */
45 /* 32-bit: 1 to 25 children, 8 to 128 bytes */
46 /* 64-bit: 1 to 28 children, 16 to 256 bytes */
47 RCU_JA_POOL
= 1, /* Type B */
48 /* 32-bit: 26 to 100 children, 256 to 512 bytes */
49 /* 64-bit: 29 to 112 children, 512 to 1024 bytes */
50 RCU_JA_PIGEON
= 2, /* Type C */
51 /* 32-bit: 101 to 256 children, 1024 bytes */
52 /* 64-bit: 113 to 256 children, 2048 bytes */
53 /* Leaf nodes are implicit from their height in the tree */
56 RCU_JA_NULL
, /* not an encoded type, but keeps code regular */
60 enum cds_ja_type_class type_class
;
61 uint16_t min_child
; /* minimum number of children: 1 to 256 */
62 uint16_t max_child
; /* maximum number of children: 1 to 256 */
63 uint16_t max_linear_child
; /* per-pool max nr. children: 1 to 256 */
64 uint16_t order
; /* node size is (1 << order), in bytes */
65 uint16_t nr_pool_order
; /* number of pools */
66 uint16_t pool_size_order
; /* pool size */
70 * Iteration on the array to find the right node size for the number of
71 * children stops when it reaches .max_child == 256 (this is the largest
72 * possible node size, which contains 256 children).
73 * The min_child overlaps with the previous max_child to provide an
74 * hysteresis loop to reallocation for patterns of cyclic add/removal
75 * within the same node.
76 * The node the index within the following arrays is represented on 3
77 * bits. It identifies the node type, min/max number of children, and
79 * The max_child values for the RCU_JA_POOL below result from
80 * statistical approximation: over million populations, the max_child
81 * covers between 97% and 99% of the populations generated. Therefore, a
82 * fallback should exist to cover the rare extreme population unbalance
83 * cases, but it will not have a major impact on speed nor space
84 * consumption, since those are rare cases.
87 #if (CAA_BITS_PER_LONG < 64)
90 ja_type_0_max_child
= 1,
91 ja_type_1_max_child
= 3,
92 ja_type_2_max_child
= 6,
93 ja_type_3_max_child
= 12,
94 ja_type_4_max_child
= 25,
95 ja_type_5_max_child
= 48,
96 ja_type_6_max_child
= 92,
97 ja_type_7_max_child
= 256,
98 ja_type_8_max_child
= 0, /* NULL */
102 ja_type_0_max_linear_child
= 1,
103 ja_type_1_max_linear_child
= 3,
104 ja_type_2_max_linear_child
= 6,
105 ja_type_3_max_linear_child
= 12,
106 ja_type_4_max_linear_child
= 25,
107 ja_type_5_max_linear_child
= 24,
108 ja_type_6_max_linear_child
= 23,
112 ja_type_5_nr_pool_order
= 1,
113 ja_type_6_nr_pool_order
= 2,
116 const struct cds_ja_type ja_types
[] = {
117 { .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, },
118 { .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, },
119 { .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, },
120 { .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, },
121 { .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, },
123 /* Pools may fill sooner than max_child */
124 /* This pool is hardcoded at index 5. See ja_node_ptr(). */
125 { .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, },
126 /* This pool is hardcoded at index 6. See ja_node_ptr(). */
127 { .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, },
130 * Upon node removal below min_child, if child pool is filled
131 * beyond capacity, we roll back to pigeon.
133 { .type_class
= RCU_JA_PIGEON
, .min_child
= 83, .max_child
= ja_type_7_max_child
, .order
= 10, },
135 { .type_class
= RCU_JA_NULL
, .min_child
= 0, .max_child
= ja_type_8_max_child
, },
137 #else /* !(CAA_BITS_PER_LONG < 64) */
138 /* 64-bit pointers */
140 ja_type_0_max_child
= 1,
141 ja_type_1_max_child
= 3,
142 ja_type_2_max_child
= 7,
143 ja_type_3_max_child
= 14,
144 ja_type_4_max_child
= 28,
145 ja_type_5_max_child
= 54,
146 ja_type_6_max_child
= 104,
147 ja_type_7_max_child
= 256,
148 ja_type_8_max_child
= 256,
152 ja_type_0_max_linear_child
= 1,
153 ja_type_1_max_linear_child
= 3,
154 ja_type_2_max_linear_child
= 7,
155 ja_type_3_max_linear_child
= 14,
156 ja_type_4_max_linear_child
= 28,
157 ja_type_5_max_linear_child
= 27,
158 ja_type_6_max_linear_child
= 26,
162 ja_type_5_nr_pool_order
= 1,
163 ja_type_6_nr_pool_order
= 2,
166 const struct cds_ja_type ja_types
[] = {
167 { .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, },
168 { .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, },
169 { .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, },
170 { .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, },
171 { .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, },
173 /* Pools may fill sooner than max_child. */
174 /* This pool is hardcoded at index 5. See ja_node_ptr(). */
175 { .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, },
176 /* This pool is hardcoded at index 6. See ja_node_ptr(). */
177 { .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, },
180 * Upon node removal below min_child, if child pool is filled
181 * beyond capacity, we roll back to pigeon.
183 { .type_class
= RCU_JA_PIGEON
, .min_child
= 95, .max_child
= ja_type_7_max_child
, .order
= 11, },
185 { .type_class
= RCU_JA_NULL
, .min_child
= 0, .max_child
= ja_type_8_max_child
, },
187 #endif /* !(BITS_PER_LONG < 64) */
189 static inline __attribute__((unused
))
190 void static_array_size_check(void)
192 CAA_BUILD_BUG_ON(CAA_ARRAY_SIZE(ja_types
) < JA_TYPE_MAX_NR
);
196 * The cds_ja_node contains the compressed node data needed for
197 * read-side. For linear and pool node configurations, it starts with a
198 * byte counting the number of children in the node. Then, the
199 * node-specific data is placed.
200 * The node mutex, if any is needed, protecting concurrent updated of
201 * each node is placed in a separate hash table indexed by node address.
202 * For the pigeon configuration, the number of children is also kept in
203 * a separate hash table, indexed by node address, because it is only
204 * required for updates.
207 #define DECLARE_LINEAR_NODE(index) \
210 uint8_t child_value[ja_type_## index ##_max_linear_child]; \
211 struct cds_ja_inode_flag *child_ptr[ja_type_## index ##_max_linear_child]; \
214 #define DECLARE_POOL_NODE(index) \
218 uint8_t child_value[ja_type_## index ##_max_linear_child]; \
219 struct cds_ja_inode_flag *child_ptr[ja_type_## index ##_max_linear_child]; \
220 } linear[1U << ja_type_## index ##_nr_pool_order]; \
223 struct cds_ja_inode
{
225 /* Linear configuration */
226 DECLARE_LINEAR_NODE(0) conf_0
;
227 DECLARE_LINEAR_NODE(1) conf_1
;
228 DECLARE_LINEAR_NODE(2) conf_2
;
229 DECLARE_LINEAR_NODE(3) conf_3
;
230 DECLARE_LINEAR_NODE(4) conf_4
;
232 /* Pool configuration */
233 DECLARE_POOL_NODE(5) conf_5
;
234 DECLARE_POOL_NODE(6) conf_6
;
236 /* Pigeon configuration */
238 struct cds_ja_inode_flag
*child
[ja_type_7_max_child
];
240 /* data aliasing nodes for computed accesses */
241 uint8_t data
[sizeof(struct cds_ja_inode_flag
*) * ja_type_7_max_child
];
246 JA_RECOMPACT_ADD_SAME
,
247 JA_RECOMPACT_ADD_NEXT
,
251 enum ja_lookup_inequality
{
264 struct cds_ja_inode
*_ja_node_mask_ptr(struct cds_ja_inode_flag
*node
)
266 return (struct cds_ja_inode
*) (((unsigned long) node
) & JA_PTR_MASK
);
269 unsigned long ja_node_type(struct cds_ja_inode_flag
*node
)
273 if (_ja_node_mask_ptr(node
) == NULL
) {
274 return NODE_INDEX_NULL
;
276 type
= (unsigned int) ((unsigned long) node
& JA_TYPE_MASK
);
277 assert(type
< (1UL << JA_TYPE_BITS
));
282 struct cds_ja_inode
*alloc_cds_ja_node(struct cds_ja
*ja
,
283 const struct cds_ja_type
*ja_type
)
285 size_t len
= 1U << ja_type
->order
;
289 ret
= posix_memalign(&p
, len
, len
);
294 uatomic_inc(&ja
->nr_nodes_allocated
);
298 void free_cds_ja_node(struct cds_ja
*ja
, struct cds_ja_inode
*node
)
302 uatomic_inc(&ja
->nr_nodes_freed
);
305 #define __JA_ALIGN_MASK(v, mask) (((v) + (mask)) & ~(mask))
306 #define JA_ALIGN(v, align) __JA_ALIGN_MASK(v, (typeof(v)) (align) - 1)
307 #define __JA_FLOOR_MASK(v, mask) ((v) & ~(mask))
308 #define JA_FLOOR(v, align) __JA_FLOOR_MASK(v, (typeof(v)) (align) - 1)
311 uint8_t *align_ptr_size(uint8_t *ptr
)
313 return (uint8_t *) JA_ALIGN((unsigned long) ptr
, sizeof(void *));
317 uint8_t ja_linear_node_get_nr_child(const struct cds_ja_type
*type
,
318 struct cds_ja_inode
*node
)
320 assert(type
->type_class
== RCU_JA_LINEAR
|| type
->type_class
== RCU_JA_POOL
);
321 return rcu_dereference(node
->u
.data
[0]);
325 * The order in which values and pointers are does does not matter: if
326 * a value is missing, we return NULL. If a value is there, but its
327 * associated pointers is still NULL, we return NULL too.
330 struct cds_ja_inode_flag
*ja_linear_node_get_nth(const struct cds_ja_type
*type
,
331 struct cds_ja_inode
*node
,
332 struct cds_ja_inode_flag
***node_flag_ptr
,
337 struct cds_ja_inode_flag
**pointers
;
338 struct cds_ja_inode_flag
*ptr
;
341 assert(type
->type_class
== RCU_JA_LINEAR
|| type
->type_class
== RCU_JA_POOL
);
343 nr_child
= ja_linear_node_get_nr_child(type
, node
);
344 cmm_smp_rmb(); /* read nr_child before values and pointers */
345 assert(nr_child
<= type
->max_linear_child
);
346 assert(type
->type_class
!= RCU_JA_LINEAR
|| nr_child
>= type
->min_child
);
348 values
= &node
->u
.data
[1];
349 for (i
= 0; i
< nr_child
; i
++) {
350 if (CMM_LOAD_SHARED(values
[i
]) == n
)
354 if (caa_unlikely(node_flag_ptr
))
355 *node_flag_ptr
= NULL
;
358 pointers
= (struct cds_ja_inode_flag
**) align_ptr_size(&values
[type
->max_linear_child
]);
359 ptr
= rcu_dereference(pointers
[i
]);
360 if (caa_unlikely(node_flag_ptr
))
361 *node_flag_ptr
= &pointers
[i
];
366 struct cds_ja_inode_flag
*ja_linear_node_get_direction(const struct cds_ja_type
*type
,
367 struct cds_ja_inode
*node
,
368 int n
, uint8_t *result_key
,
369 enum ja_direction dir
)
373 struct cds_ja_inode_flag
**pointers
;
374 struct cds_ja_inode_flag
*ptr
= NULL
;
376 int match_idx
= -1, match_v
;
378 assert(type
->type_class
== RCU_JA_LINEAR
|| type
->type_class
== RCU_JA_POOL
);
379 assert(dir
== JA_LEFT
|| dir
== JA_RIGHT
);
381 if (dir
== JA_LEFT
) {
384 match_v
= JA_ENTRY_PER_NODE
;
387 nr_child
= ja_linear_node_get_nr_child(type
, node
);
388 cmm_smp_rmb(); /* read nr_child before values and pointers */
389 assert(nr_child
<= type
->max_linear_child
);
390 assert(type
->type_class
!= RCU_JA_LINEAR
|| nr_child
>= type
->min_child
);
392 values
= &node
->u
.data
[1];
393 pointers
= (struct cds_ja_inode_flag
**) align_ptr_size(&values
[type
->max_linear_child
]);
394 for (i
= 0; i
< nr_child
; i
++) {
397 v
= CMM_LOAD_SHARED(values
[i
]);
398 ptr
= CMM_LOAD_SHARED(pointers
[i
]);
401 if (dir
== JA_LEFT
) {
402 if ((int) v
< n
&& (int) v
> match_v
) {
407 if ((int) v
> n
&& (int) v
< match_v
) {
417 assert(match_v
>= 0 && match_v
< JA_ENTRY_PER_NODE
);
419 *result_key
= (uint8_t) match_v
;
420 ptr
= rcu_dereference(pointers
[match_idx
]);
425 void ja_linear_node_get_ith_pos(const struct cds_ja_type
*type
,
426 struct cds_ja_inode
*node
,
429 struct cds_ja_inode_flag
**iter
)
432 struct cds_ja_inode_flag
**pointers
;
434 assert(type
->type_class
== RCU_JA_LINEAR
|| type
->type_class
== RCU_JA_POOL
);
435 assert(i
< ja_linear_node_get_nr_child(type
, node
));
437 values
= &node
->u
.data
[1];
439 pointers
= (struct cds_ja_inode_flag
**) align_ptr_size(&values
[type
->max_linear_child
]);
444 struct cds_ja_inode_flag
*ja_pool_node_get_nth(const struct cds_ja_type
*type
,
445 struct cds_ja_inode
*node
,
446 struct cds_ja_inode_flag
*node_flag
,
447 struct cds_ja_inode_flag
***node_flag_ptr
,
450 struct cds_ja_inode
*linear
;
452 assert(type
->type_class
== RCU_JA_POOL
);
454 switch (type
->nr_pool_order
) {
457 unsigned long bitsel
, index
;
459 bitsel
= ja_node_pool_1d_bitsel(node_flag
);
460 assert(bitsel
< CHAR_BIT
);
461 index
= ((unsigned long) n
>> bitsel
) & 0x1;
462 linear
= (struct cds_ja_inode
*) &node
->u
.data
[index
<< type
->pool_size_order
];
467 unsigned long bitsel
[2], index
[2], rindex
;
469 ja_node_pool_2d_bitsel(node_flag
, bitsel
);
470 assert(bitsel
[0] < CHAR_BIT
);
471 assert(bitsel
[1] < CHAR_BIT
);
472 index
[0] = ((unsigned long) n
>> bitsel
[0]) & 0x1;
474 index
[1] = ((unsigned long) n
>> bitsel
[1]) & 0x1;
475 rindex
= index
[0] | index
[1];
476 linear
= (struct cds_ja_inode
*) &node
->u
.data
[rindex
<< type
->pool_size_order
];
483 return ja_linear_node_get_nth(type
, linear
, node_flag_ptr
, n
);
487 struct cds_ja_inode
*ja_pool_node_get_ith_pool(const struct cds_ja_type
*type
,
488 struct cds_ja_inode
*node
,
491 assert(type
->type_class
== RCU_JA_POOL
);
492 return (struct cds_ja_inode
*)
493 &node
->u
.data
[(unsigned int) i
<< type
->pool_size_order
];
497 struct cds_ja_inode_flag
*ja_pool_node_get_direction(const struct cds_ja_type
*type
,
498 struct cds_ja_inode
*node
,
499 int n
, uint8_t *result_key
,
500 enum ja_direction dir
)
502 unsigned int pool_nr
;
504 struct cds_ja_inode_flag
*match_node_flag
= NULL
;
506 assert(type
->type_class
== RCU_JA_POOL
);
507 assert(dir
== JA_LEFT
|| dir
== JA_RIGHT
);
509 if (dir
== JA_LEFT
) {
512 match_v
= JA_ENTRY_PER_NODE
;
515 for (pool_nr
= 0; pool_nr
< (1U << type
->nr_pool_order
); pool_nr
++) {
516 struct cds_ja_inode
*pool
=
517 ja_pool_node_get_ith_pool(type
,
520 ja_linear_node_get_nr_child(type
, pool
);
523 for (j
= 0; j
< nr_child
; j
++) {
524 struct cds_ja_inode_flag
*iter
;
527 ja_linear_node_get_ith_pos(type
, pool
,
531 if (dir
== JA_LEFT
) {
532 if ((int) v
< n
&& (int) v
> match_v
) {
534 match_node_flag
= iter
;
537 if ((int) v
> n
&& (int) v
< match_v
) {
539 match_node_flag
= iter
;
545 *result_key
= (uint8_t) match_v
;
546 return match_node_flag
;
550 struct cds_ja_inode_flag
*ja_pigeon_node_get_nth(const struct cds_ja_type
*type
,
551 struct cds_ja_inode
*node
,
552 struct cds_ja_inode_flag
***node_flag_ptr
,
555 struct cds_ja_inode_flag
**child_node_flag_ptr
;
556 struct cds_ja_inode_flag
*child_node_flag
;
558 assert(type
->type_class
== RCU_JA_PIGEON
);
559 child_node_flag_ptr
= &((struct cds_ja_inode_flag
**) node
->u
.data
)[n
];
560 child_node_flag
= rcu_dereference(*child_node_flag_ptr
);
561 dbg_printf("ja_pigeon_node_get_nth child_node_flag_ptr %p\n",
562 child_node_flag_ptr
);
563 if (caa_unlikely(node_flag_ptr
))
564 *node_flag_ptr
= child_node_flag_ptr
;
565 return child_node_flag
;
569 struct cds_ja_inode_flag
*ja_pigeon_node_get_direction(const struct cds_ja_type
*type
,
570 struct cds_ja_inode
*node
,
571 int n
, uint8_t *result_key
,
572 enum ja_direction dir
)
574 struct cds_ja_inode_flag
**child_node_flag_ptr
;
575 struct cds_ja_inode_flag
*child_node_flag
;
578 assert(type
->type_class
== RCU_JA_PIGEON
);
579 assert(dir
== JA_LEFT
|| dir
== JA_RIGHT
);
581 if (dir
== JA_LEFT
) {
582 /* n - 1 is first value left of n */
583 for (i
= n
- 1; i
>= 0; i
--) {
584 child_node_flag_ptr
= &((struct cds_ja_inode_flag
**) node
->u
.data
)[i
];
585 child_node_flag
= rcu_dereference(*child_node_flag_ptr
);
586 if (child_node_flag
) {
587 dbg_printf("ja_pigeon_node_get_left child_node_flag %p\n",
589 *result_key
= (uint8_t) i
;
590 return child_node_flag
;
594 /* n + 1 is first value right of n */
595 for (i
= n
+ 1; i
< JA_ENTRY_PER_NODE
; i
++) {
596 child_node_flag_ptr
= &((struct cds_ja_inode_flag
**) node
->u
.data
)[i
];
597 child_node_flag
= rcu_dereference(*child_node_flag_ptr
);
598 if (child_node_flag
) {
599 dbg_printf("ja_pigeon_node_get_right child_node_flag %p\n",
601 *result_key
= (uint8_t) i
;
602 return child_node_flag
;
610 struct cds_ja_inode_flag
*ja_pigeon_node_get_ith_pos(const struct cds_ja_type
*type
,
611 struct cds_ja_inode
*node
,
614 return ja_pigeon_node_get_nth(type
, node
, NULL
, i
);
618 * ja_node_get_nth: get nth item from a node.
619 * node_flag is already rcu_dereference'd.
622 struct cds_ja_inode_flag
*ja_node_get_nth(struct cds_ja_inode_flag
*node_flag
,
623 struct cds_ja_inode_flag
***node_flag_ptr
,
626 unsigned int type_index
;
627 struct cds_ja_inode
*node
;
628 const struct cds_ja_type
*type
;
630 node
= ja_node_ptr(node_flag
);
631 assert(node
!= NULL
);
632 type_index
= ja_node_type(node_flag
);
633 type
= &ja_types
[type_index
];
635 switch (type
->type_class
) {
637 return ja_linear_node_get_nth(type
, node
,
640 return ja_pool_node_get_nth(type
, node
, node_flag
,
643 return ja_pigeon_node_get_nth(type
, node
,
647 return (void *) -1UL;
652 struct cds_ja_inode_flag
*ja_node_get_direction(struct cds_ja_inode_flag
*node_flag
,
653 int n
, uint8_t *result_key
,
654 enum ja_direction dir
)
656 unsigned int type_index
;
657 struct cds_ja_inode
*node
;
658 const struct cds_ja_type
*type
;
660 node
= ja_node_ptr(node_flag
);
661 assert(node
!= NULL
);
662 type_index
= ja_node_type(node_flag
);
663 type
= &ja_types
[type_index
];
665 switch (type
->type_class
) {
667 return ja_linear_node_get_direction(type
, node
, n
, result_key
, dir
);
669 return ja_pool_node_get_direction(type
, node
, n
, result_key
, dir
);
671 return ja_pigeon_node_get_direction(type
, node
, n
, result_key
, dir
);
674 return (void *) -1UL;
679 struct cds_ja_inode_flag
*ja_node_get_leftright(struct cds_ja_inode_flag
*node_flag
,
680 unsigned int n
, uint8_t *result_key
,
681 enum ja_direction dir
)
683 return ja_node_get_direction(node_flag
, n
, result_key
, dir
);
687 struct cds_ja_inode_flag
*ja_node_get_minmax(struct cds_ja_inode_flag
*node_flag
,
689 enum ja_direction dir
)
693 return ja_node_get_direction(node_flag
,
694 -1, result_key
, JA_RIGHT
);
696 return ja_node_get_direction(node_flag
,
697 JA_ENTRY_PER_NODE
, result_key
, JA_LEFT
);
704 int ja_linear_node_set_nth(const struct cds_ja_type
*type
,
705 struct cds_ja_inode
*node
,
706 struct cds_ja_shadow_node
*shadow_node
,
708 struct cds_ja_inode_flag
*child_node_flag
)
711 uint8_t *values
, *nr_child_ptr
;
712 struct cds_ja_inode_flag
**pointers
;
713 unsigned int i
, unused
= 0;
715 assert(type
->type_class
== RCU_JA_LINEAR
|| type
->type_class
== RCU_JA_POOL
);
717 nr_child_ptr
= &node
->u
.data
[0];
718 dbg_printf("linear set nth: n %u, nr_child_ptr %p\n",
719 (unsigned int) n
, nr_child_ptr
);
720 nr_child
= *nr_child_ptr
;
721 assert(nr_child
<= type
->max_linear_child
);
723 values
= &node
->u
.data
[1];
724 pointers
= (struct cds_ja_inode_flag
**) align_ptr_size(&values
[type
->max_linear_child
]);
725 /* Check if node value is already populated */
726 for (i
= 0; i
< nr_child
; i
++) {
727 if (values
[i
] == n
) {
737 if (i
== nr_child
&& nr_child
>= type
->max_linear_child
) {
739 return -ERANGE
; /* recompact node */
741 return -ENOSPC
; /* No space left in this node type */
744 assert(pointers
[i
] == NULL
);
745 rcu_assign_pointer(pointers
[i
], child_node_flag
);
746 /* If we expanded the nr_child, increment it */
748 CMM_STORE_SHARED(values
[nr_child
], n
);
749 /* write pointer and value before nr_child */
751 CMM_STORE_SHARED(*nr_child_ptr
, nr_child
+ 1);
753 shadow_node
->nr_child
++;
754 dbg_printf("linear set nth: %u child, shadow: %u child, for node %p shadow %p\n",
755 (unsigned int) CMM_LOAD_SHARED(*nr_child_ptr
),
756 (unsigned int) shadow_node
->nr_child
,
763 int ja_pool_node_set_nth(const struct cds_ja_type
*type
,
764 struct cds_ja_inode
*node
,
765 struct cds_ja_inode_flag
*node_flag
,
766 struct cds_ja_shadow_node
*shadow_node
,
768 struct cds_ja_inode_flag
*child_node_flag
)
770 struct cds_ja_inode
*linear
;
772 assert(type
->type_class
== RCU_JA_POOL
);
774 switch (type
->nr_pool_order
) {
777 unsigned long bitsel
, index
;
779 bitsel
= ja_node_pool_1d_bitsel(node_flag
);
780 assert(bitsel
< CHAR_BIT
);
781 index
= ((unsigned long) n
>> bitsel
) & 0x1;
782 linear
= (struct cds_ja_inode
*) &node
->u
.data
[index
<< type
->pool_size_order
];
787 unsigned long bitsel
[2], index
[2], rindex
;
789 ja_node_pool_2d_bitsel(node_flag
, bitsel
);
790 assert(bitsel
[0] < CHAR_BIT
);
791 assert(bitsel
[1] < CHAR_BIT
);
792 index
[0] = ((unsigned long) n
>> bitsel
[0]) & 0x1;
794 index
[1] = ((unsigned long) n
>> bitsel
[1]) & 0x1;
795 rindex
= index
[0] | index
[1];
796 linear
= (struct cds_ja_inode
*) &node
->u
.data
[rindex
<< type
->pool_size_order
];
804 return ja_linear_node_set_nth(type
, linear
, shadow_node
,
809 int ja_pigeon_node_set_nth(const struct cds_ja_type
*type
,
810 struct cds_ja_inode
*node
,
811 struct cds_ja_shadow_node
*shadow_node
,
813 struct cds_ja_inode_flag
*child_node_flag
)
815 struct cds_ja_inode_flag
**ptr
;
817 assert(type
->type_class
== RCU_JA_PIGEON
);
818 ptr
= &((struct cds_ja_inode_flag
**) node
->u
.data
)[n
];
821 rcu_assign_pointer(*ptr
, child_node_flag
);
822 shadow_node
->nr_child
++;
827 * _ja_node_set_nth: set nth item within a node. Return an error
828 * (negative error value) if it is already there.
831 int _ja_node_set_nth(const struct cds_ja_type
*type
,
832 struct cds_ja_inode
*node
,
833 struct cds_ja_inode_flag
*node_flag
,
834 struct cds_ja_shadow_node
*shadow_node
,
836 struct cds_ja_inode_flag
*child_node_flag
)
838 switch (type
->type_class
) {
840 return ja_linear_node_set_nth(type
, node
, shadow_node
, n
,
843 return ja_pool_node_set_nth(type
, node
, node_flag
, shadow_node
, n
,
846 return ja_pigeon_node_set_nth(type
, node
, shadow_node
, n
,
859 int ja_linear_node_clear_ptr(const struct cds_ja_type
*type
,
860 struct cds_ja_inode
*node
,
861 struct cds_ja_shadow_node
*shadow_node
,
862 struct cds_ja_inode_flag
**node_flag_ptr
)
865 uint8_t *nr_child_ptr
;
867 assert(type
->type_class
== RCU_JA_LINEAR
|| type
->type_class
== RCU_JA_POOL
);
869 nr_child_ptr
= &node
->u
.data
[0];
870 nr_child
= *nr_child_ptr
;
871 assert(nr_child
<= type
->max_linear_child
);
873 if (type
->type_class
== RCU_JA_LINEAR
) {
874 assert(!shadow_node
->fallback_removal_count
);
875 if (shadow_node
->nr_child
<= type
->min_child
) {
876 /* We need to try recompacting the node */
880 dbg_printf("linear clear ptr: nr_child_ptr %p\n", nr_child_ptr
);
881 assert(*node_flag_ptr
!= NULL
);
882 rcu_assign_pointer(*node_flag_ptr
, NULL
);
884 * Value and nr_child are never changed (would cause ABA issue).
885 * Instead, we leave the pointer to NULL and recompact the node
886 * once in a while. It is allowed to set a NULL pointer to a new
887 * value without recompaction though.
888 * Only update the shadow node accounting.
890 shadow_node
->nr_child
--;
891 dbg_printf("linear clear ptr: %u child, shadow: %u child, for node %p shadow %p\n",
892 (unsigned int) CMM_LOAD_SHARED(*nr_child_ptr
),
893 (unsigned int) shadow_node
->nr_child
,
899 int ja_pool_node_clear_ptr(const struct cds_ja_type
*type
,
900 struct cds_ja_inode
*node
,
901 struct cds_ja_inode_flag
*node_flag
,
902 struct cds_ja_shadow_node
*shadow_node
,
903 struct cds_ja_inode_flag
**node_flag_ptr
,
906 struct cds_ja_inode
*linear
;
908 assert(type
->type_class
== RCU_JA_POOL
);
910 if (shadow_node
->fallback_removal_count
) {
911 shadow_node
->fallback_removal_count
--;
913 /* We should try recompacting the node */
914 if (shadow_node
->nr_child
<= type
->min_child
)
918 switch (type
->nr_pool_order
) {
921 unsigned long bitsel
, index
;
923 bitsel
= ja_node_pool_1d_bitsel(node_flag
);
924 assert(bitsel
< CHAR_BIT
);
925 index
= ((unsigned long) n
>> bitsel
) & type
->nr_pool_order
;
926 linear
= (struct cds_ja_inode
*) &node
->u
.data
[index
<< type
->pool_size_order
];
931 unsigned long bitsel
[2], index
[2], rindex
;
933 ja_node_pool_2d_bitsel(node_flag
, bitsel
);
934 assert(bitsel
[0] < CHAR_BIT
);
935 assert(bitsel
[1] < CHAR_BIT
);
936 index
[0] = ((unsigned long) n
>> bitsel
[0]) & 0x1;
938 index
[1] = ((unsigned long) n
>> bitsel
[1]) & 0x1;
939 rindex
= index
[0] | index
[1];
940 linear
= (struct cds_ja_inode
*) &node
->u
.data
[rindex
<< type
->pool_size_order
];
948 return ja_linear_node_clear_ptr(type
, linear
, shadow_node
, node_flag_ptr
);
952 int ja_pigeon_node_clear_ptr(const struct cds_ja_type
*type
,
953 struct cds_ja_inode
*node
,
954 struct cds_ja_shadow_node
*shadow_node
,
955 struct cds_ja_inode_flag
**node_flag_ptr
)
957 assert(type
->type_class
== RCU_JA_PIGEON
);
959 if (shadow_node
->fallback_removal_count
) {
960 shadow_node
->fallback_removal_count
--;
962 /* We should try recompacting the node */
963 if (shadow_node
->nr_child
<= type
->min_child
)
966 dbg_printf("ja_pigeon_node_clear_ptr: clearing ptr: %p\n", *node_flag_ptr
);
967 rcu_assign_pointer(*node_flag_ptr
, NULL
);
968 shadow_node
->nr_child
--;
973 * _ja_node_clear_ptr: clear ptr item within a node. Return an error
974 * (negative error value) if it is not found (-ENOENT).
977 int _ja_node_clear_ptr(const struct cds_ja_type
*type
,
978 struct cds_ja_inode
*node
,
979 struct cds_ja_inode_flag
*node_flag
,
980 struct cds_ja_shadow_node
*shadow_node
,
981 struct cds_ja_inode_flag
**node_flag_ptr
,
984 switch (type
->type_class
) {
986 return ja_linear_node_clear_ptr(type
, node
, shadow_node
, node_flag_ptr
);
988 return ja_pool_node_clear_ptr(type
, node
, node_flag
, shadow_node
, node_flag_ptr
, n
);
990 return ja_pigeon_node_clear_ptr(type
, node
, shadow_node
, node_flag_ptr
);
1002 * Calculate bit distribution. Returns the bit (0 to 7) that splits the
1003 * distribution in two sub-distributions containing as much elements one
1004 * compared to the other.
1007 unsigned int ja_node_sum_distribution_1d(enum ja_recompact mode
,
1009 unsigned int type_index
,
1010 const struct cds_ja_type
*type
,
1011 struct cds_ja_inode
*node
,
1012 struct cds_ja_shadow_node
*shadow_node
,
1014 struct cds_ja_inode_flag
*child_node_flag
,
1015 struct cds_ja_inode_flag
**nullify_node_flag_ptr
)
1017 uint8_t nr_one
[JA_BITS_PER_BYTE
];
1018 unsigned int bitsel
= 0, bit_i
, overall_best_distance
= UINT_MAX
;
1019 unsigned int distrib_nr_child
= 0;
1021 memset(nr_one
, 0, sizeof(nr_one
));
1023 switch (type
->type_class
) {
1027 ja_linear_node_get_nr_child(type
, node
);
1030 for (i
= 0; i
< nr_child
; i
++) {
1031 struct cds_ja_inode_flag
*iter
;
1034 ja_linear_node_get_ith_pos(type
, node
, i
, &v
, &iter
);
1037 if (mode
== JA_RECOMPACT_DEL
&& *nullify_node_flag_ptr
== iter
)
1039 for (bit_i
= 0; bit_i
< JA_BITS_PER_BYTE
; bit_i
++) {
1040 if (v
& (1U << bit_i
))
1049 unsigned int pool_nr
;
1051 for (pool_nr
= 0; pool_nr
< (1U << type
->nr_pool_order
); pool_nr
++) {
1052 struct cds_ja_inode
*pool
=
1053 ja_pool_node_get_ith_pool(type
,
1056 ja_linear_node_get_nr_child(type
, pool
);
1059 for (j
= 0; j
< nr_child
; j
++) {
1060 struct cds_ja_inode_flag
*iter
;
1063 ja_linear_node_get_ith_pos(type
, pool
,
1067 if (mode
== JA_RECOMPACT_DEL
&& *nullify_node_flag_ptr
== iter
)
1069 for (bit_i
= 0; bit_i
< JA_BITS_PER_BYTE
; bit_i
++) {
1070 if (v
& (1U << bit_i
))
1082 assert(mode
== JA_RECOMPACT_DEL
);
1083 for (i
= 0; i
< JA_ENTRY_PER_NODE
; i
++) {
1084 struct cds_ja_inode_flag
*iter
;
1086 iter
= ja_pigeon_node_get_ith_pos(type
, node
, i
);
1089 if (mode
== JA_RECOMPACT_DEL
&& *nullify_node_flag_ptr
== iter
)
1091 for (bit_i
= 0; bit_i
< JA_BITS_PER_BYTE
; bit_i
++) {
1092 if (i
& (1U << bit_i
))
1100 assert(mode
== JA_RECOMPACT_ADD_NEXT
);
1107 if (mode
== JA_RECOMPACT_ADD_NEXT
|| mode
== JA_RECOMPACT_ADD_SAME
) {
1108 for (bit_i
= 0; bit_i
< JA_BITS_PER_BYTE
; bit_i
++) {
1109 if (n
& (1U << bit_i
))
1116 * The best bit selector is that for which the number of ones is
1117 * closest to half of the number of children in the
1118 * distribution. We calculate the distance using the double of
1119 * the sub-distribution sizes to eliminate truncation error.
1121 for (bit_i
= 0; bit_i
< JA_BITS_PER_BYTE
; bit_i
++) {
1122 unsigned int distance_to_best
;
1124 distance_to_best
= abs_int(((unsigned int) nr_one
[bit_i
] << 1U) - distrib_nr_child
);
1125 if (distance_to_best
< overall_best_distance
) {
1126 overall_best_distance
= distance_to_best
;
1130 dbg_printf("1 dimension pool bit selection: (%u)\n", bitsel
);
1135 * Calculate bit distribution in two dimensions. Returns the two bits
1136 * (each 0 to 7) that splits the distribution in four sub-distributions
1137 * containing as much elements one compared to the other.
1140 void ja_node_sum_distribution_2d(enum ja_recompact mode
,
1142 unsigned int type_index
,
1143 const struct cds_ja_type
*type
,
1144 struct cds_ja_inode
*node
,
1145 struct cds_ja_shadow_node
*shadow_node
,
1147 struct cds_ja_inode_flag
*child_node_flag
,
1148 struct cds_ja_inode_flag
**nullify_node_flag_ptr
,
1149 unsigned int *_bitsel
)
1151 uint8_t nr_2d_11
[JA_BITS_PER_BYTE
][JA_BITS_PER_BYTE
],
1152 nr_2d_10
[JA_BITS_PER_BYTE
][JA_BITS_PER_BYTE
],
1153 nr_2d_01
[JA_BITS_PER_BYTE
][JA_BITS_PER_BYTE
],
1154 nr_2d_00
[JA_BITS_PER_BYTE
][JA_BITS_PER_BYTE
];
1155 unsigned int bitsel
[2] = { 0, 1 };
1156 unsigned int bit_i
, bit_j
;
1157 int overall_best_distance
= INT_MAX
;
1158 unsigned int distrib_nr_child
= 0;
1160 memset(nr_2d_11
, 0, sizeof(nr_2d_11
));
1161 memset(nr_2d_10
, 0, sizeof(nr_2d_10
));
1162 memset(nr_2d_01
, 0, sizeof(nr_2d_01
));
1163 memset(nr_2d_00
, 0, sizeof(nr_2d_00
));
1165 switch (type
->type_class
) {
1169 ja_linear_node_get_nr_child(type
, node
);
1172 for (i
= 0; i
< nr_child
; i
++) {
1173 struct cds_ja_inode_flag
*iter
;
1176 ja_linear_node_get_ith_pos(type
, node
, i
, &v
, &iter
);
1179 if (mode
== JA_RECOMPACT_DEL
&& *nullify_node_flag_ptr
== iter
)
1181 for (bit_i
= 0; bit_i
< JA_BITS_PER_BYTE
; bit_i
++) {
1182 for (bit_j
= 0; bit_j
< bit_i
; bit_j
++) {
1183 if (v
& (1U << bit_i
)) {
1184 if (v
& (1U << bit_j
)) {
1185 nr_2d_11
[bit_i
][bit_j
]++;
1187 nr_2d_10
[bit_i
][bit_j
]++;
1190 if (v
& (1U << bit_j
)) {
1191 nr_2d_01
[bit_i
][bit_j
]++;
1193 nr_2d_00
[bit_i
][bit_j
]++;
1204 unsigned int pool_nr
;
1206 for (pool_nr
= 0; pool_nr
< (1U << type
->nr_pool_order
); pool_nr
++) {
1207 struct cds_ja_inode
*pool
=
1208 ja_pool_node_get_ith_pool(type
,
1211 ja_linear_node_get_nr_child(type
, pool
);
1214 for (j
= 0; j
< nr_child
; j
++) {
1215 struct cds_ja_inode_flag
*iter
;
1218 ja_linear_node_get_ith_pos(type
, pool
,
1222 if (mode
== JA_RECOMPACT_DEL
&& *nullify_node_flag_ptr
== iter
)
1224 for (bit_i
= 0; bit_i
< JA_BITS_PER_BYTE
; bit_i
++) {
1225 for (bit_j
= 0; bit_j
< bit_i
; bit_j
++) {
1226 if (v
& (1U << bit_i
)) {
1227 if (v
& (1U << bit_j
)) {
1228 nr_2d_11
[bit_i
][bit_j
]++;
1230 nr_2d_10
[bit_i
][bit_j
]++;
1233 if (v
& (1U << bit_j
)) {
1234 nr_2d_01
[bit_i
][bit_j
]++;
1236 nr_2d_00
[bit_i
][bit_j
]++;
1250 assert(mode
== JA_RECOMPACT_DEL
);
1251 for (i
= 0; i
< JA_ENTRY_PER_NODE
; i
++) {
1252 struct cds_ja_inode_flag
*iter
;
1254 iter
= ja_pigeon_node_get_ith_pos(type
, node
, i
);
1257 if (mode
== JA_RECOMPACT_DEL
&& *nullify_node_flag_ptr
== iter
)
1259 for (bit_i
= 0; bit_i
< JA_BITS_PER_BYTE
; bit_i
++) {
1260 for (bit_j
= 0; bit_j
< bit_i
; bit_j
++) {
1261 if (i
& (1U << bit_i
)) {
1262 if (i
& (1U << bit_j
)) {
1263 nr_2d_11
[bit_i
][bit_j
]++;
1265 nr_2d_10
[bit_i
][bit_j
]++;
1268 if (i
& (1U << bit_j
)) {
1269 nr_2d_01
[bit_i
][bit_j
]++;
1271 nr_2d_00
[bit_i
][bit_j
]++;
1281 assert(mode
== JA_RECOMPACT_ADD_NEXT
);
1288 if (mode
== JA_RECOMPACT_ADD_NEXT
|| mode
== JA_RECOMPACT_ADD_SAME
) {
1289 for (bit_i
= 0; bit_i
< JA_BITS_PER_BYTE
; bit_i
++) {
1290 for (bit_j
= 0; bit_j
< bit_i
; bit_j
++) {
1291 if (n
& (1U << bit_i
)) {
1292 if (n
& (1U << bit_j
)) {
1293 nr_2d_11
[bit_i
][bit_j
]++;
1295 nr_2d_10
[bit_i
][bit_j
]++;
1298 if (n
& (1U << bit_j
)) {
1299 nr_2d_01
[bit_i
][bit_j
]++;
1301 nr_2d_00
[bit_i
][bit_j
]++;
1310 * The best bit selector is that for which the number of nodes
1311 * in each sub-class is closest to one-fourth of the number of
1312 * children in the distribution. We calculate the distance using
1313 * 4 times the size of the sub-distribution to eliminate
1316 for (bit_i
= 0; bit_i
< JA_BITS_PER_BYTE
; bit_i
++) {
1317 for (bit_j
= 0; bit_j
< bit_i
; bit_j
++) {
1318 int distance_to_best
[4];
1320 distance_to_best
[0] = ((unsigned int) nr_2d_11
[bit_i
][bit_j
] << 2U) - distrib_nr_child
;
1321 distance_to_best
[1] = ((unsigned int) nr_2d_10
[bit_i
][bit_j
] << 2U) - distrib_nr_child
;
1322 distance_to_best
[2] = ((unsigned int) nr_2d_01
[bit_i
][bit_j
] << 2U) - distrib_nr_child
;
1323 distance_to_best
[3] = ((unsigned int) nr_2d_00
[bit_i
][bit_j
] << 2U) - distrib_nr_child
;
1325 /* Consider worse distance above best */
1326 if (distance_to_best
[1] > 0 && distance_to_best
[1] > distance_to_best
[0])
1327 distance_to_best
[0] = distance_to_best
[1];
1328 if (distance_to_best
[2] > 0 && distance_to_best
[2] > distance_to_best
[0])
1329 distance_to_best
[0] = distance_to_best
[2];
1330 if (distance_to_best
[3] > 0 && distance_to_best
[3] > distance_to_best
[0])
1331 distance_to_best
[0] = distance_to_best
[3];
1334 * If our worse distance is better than overall,
1335 * we become new best candidate.
1337 if (distance_to_best
[0] < overall_best_distance
) {
1338 overall_best_distance
= distance_to_best
[0];
1345 dbg_printf("2 dimensions pool bit selection: (%u,%u)\n", bitsel
[0], bitsel
[1]);
1347 /* Return our bit selection */
1348 _bitsel
[0] = bitsel
[0];
1349 _bitsel
[1] = bitsel
[1];
1353 unsigned int find_nearest_type_index(unsigned int type_index
,
1354 unsigned int nr_nodes
)
1356 const struct cds_ja_type
*type
;
1358 assert(type_index
!= NODE_INDEX_NULL
);
1360 return NODE_INDEX_NULL
;
1362 type
= &ja_types
[type_index
];
1363 if (nr_nodes
< type
->min_child
)
1365 else if (nr_nodes
> type
->max_child
)
1374 * ja_node_recompact_add: recompact a node, adding a new child.
1375 * Return 0 on success, -EAGAIN if need to retry, or other negative
1376 * error value otherwise.
1379 int ja_node_recompact(enum ja_recompact mode
,
1381 unsigned int old_type_index
,
1382 const struct cds_ja_type
*old_type
,
1383 struct cds_ja_inode
*old_node
,
1384 struct cds_ja_shadow_node
*shadow_node
,
1385 struct cds_ja_inode_flag
**old_node_flag_ptr
, uint8_t n
,
1386 struct cds_ja_inode_flag
*child_node_flag
,
1387 struct cds_ja_inode_flag
**nullify_node_flag_ptr
,
1390 unsigned int new_type_index
;
1391 struct cds_ja_inode
*new_node
;
1392 struct cds_ja_shadow_node
*new_shadow_node
= NULL
;
1393 const struct cds_ja_type
*new_type
;
1394 struct cds_ja_inode_flag
*new_node_flag
, *old_node_flag
;
1398 old_node_flag
= *old_node_flag_ptr
;
1401 * Need to find nearest type index even for ADD_SAME, because
1402 * this recompaction, when applied to linear nodes, will garbage
1403 * collect dummy (NULL) entries, and can therefore cause a few
1404 * linear representations to be skipped.
1407 case JA_RECOMPACT_ADD_SAME
:
1408 new_type_index
= find_nearest_type_index(old_type_index
,
1409 shadow_node
->nr_child
+ 1);
1410 dbg_printf("Recompact for node with %u children\n",
1411 shadow_node
->nr_child
+ 1);
1413 case JA_RECOMPACT_ADD_NEXT
:
1414 if (!shadow_node
|| old_type_index
== NODE_INDEX_NULL
) {
1416 dbg_printf("Recompact for NULL\n");
1418 new_type_index
= find_nearest_type_index(old_type_index
,
1419 shadow_node
->nr_child
+ 1);
1420 dbg_printf("Recompact for node with %u children\n",
1421 shadow_node
->nr_child
+ 1);
1424 case JA_RECOMPACT_DEL
:
1425 new_type_index
= find_nearest_type_index(old_type_index
,
1426 shadow_node
->nr_child
- 1);
1427 dbg_printf("Recompact for node with %u children\n",
1428 shadow_node
->nr_child
- 1);
1434 retry
: /* for fallback */
1435 dbg_printf("Recompact from type %d to type %d\n",
1436 old_type_index
, new_type_index
);
1437 new_type
= &ja_types
[new_type_index
];
1438 if (new_type_index
!= NODE_INDEX_NULL
) {
1439 new_node
= alloc_cds_ja_node(ja
, new_type
);
1443 if (new_type
->type_class
== RCU_JA_POOL
) {
1444 switch (new_type
->nr_pool_order
) {
1447 unsigned int node_distrib_bitsel
;
1449 node_distrib_bitsel
=
1450 ja_node_sum_distribution_1d(mode
, ja
,
1451 old_type_index
, old_type
,
1452 old_node
, shadow_node
,
1454 nullify_node_flag_ptr
);
1455 assert(!((unsigned long) new_node
& JA_POOL_1D_MASK
));
1456 new_node_flag
= ja_node_flag_pool_1d(new_node
,
1457 new_type_index
, node_distrib_bitsel
);
1462 unsigned int node_distrib_bitsel
[2];
1464 ja_node_sum_distribution_2d(mode
, ja
,
1465 old_type_index
, old_type
,
1466 old_node
, shadow_node
,
1468 nullify_node_flag_ptr
,
1469 node_distrib_bitsel
);
1470 assert(!((unsigned long) new_node
& JA_POOL_1D_MASK
));
1471 assert(!((unsigned long) new_node
& JA_POOL_2D_MASK
));
1472 new_node_flag
= ja_node_flag_pool_2d(new_node
,
1473 new_type_index
, node_distrib_bitsel
);
1480 new_node_flag
= ja_node_flag(new_node
, new_type_index
);
1483 dbg_printf("Recompact inherit lock from %p\n", shadow_node
);
1484 new_shadow_node
= rcuja_shadow_set(ja
->ht
, new_node_flag
, shadow_node
, ja
, level
);
1485 if (!new_shadow_node
) {
1486 free_cds_ja_node(ja
, new_node
);
1490 new_shadow_node
->fallback_removal_count
=
1491 JA_FALLBACK_REMOVAL_COUNT
;
1494 new_node_flag
= NULL
;
1497 assert(mode
!= JA_RECOMPACT_ADD_NEXT
|| old_type
->type_class
!= RCU_JA_PIGEON
);
1499 if (new_type_index
== NODE_INDEX_NULL
)
1502 switch (old_type
->type_class
) {
1506 ja_linear_node_get_nr_child(old_type
, old_node
);
1509 for (i
= 0; i
< nr_child
; i
++) {
1510 struct cds_ja_inode_flag
*iter
;
1513 ja_linear_node_get_ith_pos(old_type
, old_node
, i
, &v
, &iter
);
1516 if (mode
== JA_RECOMPACT_DEL
&& *nullify_node_flag_ptr
== iter
)
1518 ret
= _ja_node_set_nth(new_type
, new_node
, new_node_flag
,
1521 if (new_type
->type_class
== RCU_JA_POOL
&& ret
) {
1522 goto fallback_toosmall
;
1530 unsigned int pool_nr
;
1532 for (pool_nr
= 0; pool_nr
< (1U << old_type
->nr_pool_order
); pool_nr
++) {
1533 struct cds_ja_inode
*pool
=
1534 ja_pool_node_get_ith_pool(old_type
,
1537 ja_linear_node_get_nr_child(old_type
, pool
);
1540 for (j
= 0; j
< nr_child
; j
++) {
1541 struct cds_ja_inode_flag
*iter
;
1544 ja_linear_node_get_ith_pos(old_type
, pool
,
1548 if (mode
== JA_RECOMPACT_DEL
&& *nullify_node_flag_ptr
== iter
)
1550 ret
= _ja_node_set_nth(new_type
, new_node
, new_node_flag
,
1553 if (new_type
->type_class
== RCU_JA_POOL
1555 goto fallback_toosmall
;
1563 assert(mode
== JA_RECOMPACT_ADD_NEXT
);
1569 assert(mode
== JA_RECOMPACT_DEL
);
1570 for (i
= 0; i
< JA_ENTRY_PER_NODE
; i
++) {
1571 struct cds_ja_inode_flag
*iter
;
1573 iter
= ja_pigeon_node_get_ith_pos(old_type
, old_node
, i
);
1576 if (mode
== JA_RECOMPACT_DEL
&& *nullify_node_flag_ptr
== iter
)
1578 ret
= _ja_node_set_nth(new_type
, new_node
, new_node_flag
,
1581 if (new_type
->type_class
== RCU_JA_POOL
&& ret
) {
1582 goto fallback_toosmall
;
1595 if (mode
== JA_RECOMPACT_ADD_NEXT
|| mode
== JA_RECOMPACT_ADD_SAME
) {
1597 ret
= _ja_node_set_nth(new_type
, new_node
, new_node_flag
,
1599 n
, child_node_flag
);
1600 if (new_type
->type_class
== RCU_JA_POOL
&& ret
) {
1601 goto fallback_toosmall
;
1607 dbg_printf("Using fallback for %u children, node type index: %u, mode %s\n",
1608 new_shadow_node
->nr_child
, old_type_index
, mode
== JA_RECOMPACT_ADD_NEXT
? "add_next" :
1609 (mode
== JA_RECOMPACT_DEL
? "del" : "add_same"));
1610 uatomic_inc(&ja
->node_fallback_count_distribution
[new_shadow_node
->nr_child
]);
1613 /* Return pointer to new recompacted node through old_node_flag_ptr */
1614 *old_node_flag_ptr
= new_node_flag
;
1618 flags
= RCUJA_SHADOW_CLEAR_FREE_NODE
;
1620 * It is OK to free the lock associated with a node
1621 * going to NULL, since we are holding the parent lock.
1622 * This synchronizes removal with re-add of that node.
1624 if (new_type_index
== NODE_INDEX_NULL
)
1625 flags
|= RCUJA_SHADOW_CLEAR_FREE_LOCK
;
1626 ret
= rcuja_shadow_clear(ja
->ht
, old_node_flag
, shadow_node
,
1636 /* fallback if next pool is too small */
1637 assert(new_shadow_node
);
1638 ret
= rcuja_shadow_clear(ja
->ht
, new_node_flag
, new_shadow_node
,
1639 RCUJA_SHADOW_CLEAR_FREE_NODE
);
1643 case JA_RECOMPACT_ADD_SAME
:
1645 * JA_RECOMPACT_ADD_SAME is only triggered if a linear
1646 * node within a pool has unused entries. It should
1647 * therefore _never_ be too small.
1652 case JA_RECOMPACT_ADD_NEXT
:
1654 const struct cds_ja_type
*next_type
;
1657 * Recompaction attempt on add failed. Should only
1658 * happen if target node type is pool. Caused by
1659 * hard-to-split distribution. Recompact using the next
1660 * distribution size.
1662 assert(new_type
->type_class
== RCU_JA_POOL
);
1663 next_type
= &ja_types
[new_type_index
+ 1];
1665 * Try going to the next pool size if our population
1666 * fits within its range. This is not flagged as a
1669 if (shadow_node
->nr_child
+ 1 >= next_type
->min_child
1670 && shadow_node
->nr_child
+ 1 <= next_type
->max_child
) {
1675 dbg_printf("Add fallback to type %d\n", new_type_index
);
1676 uatomic_inc(&ja
->nr_fallback
);
1682 case JA_RECOMPACT_DEL
:
1684 * Recompaction attempt on delete failed. Should only
1685 * happen if target node type is pool. This is caused by
1686 * a hard-to-split distribution. Recompact on same node
1687 * size, but flag current node as "fallback" to ensure
1688 * we don't attempt recompaction before some activity
1689 * has reshuffled our node.
1691 assert(new_type
->type_class
== RCU_JA_POOL
);
1692 new_type_index
= old_type_index
;
1693 dbg_printf("Delete fallback keeping type %d\n", new_type_index
);
1694 uatomic_inc(&ja
->nr_fallback
);
1703 * Last resort fallback: pigeon.
1705 new_type_index
= (1UL << JA_TYPE_BITS
) - 1;
1706 dbg_printf("Fallback to type %d\n", new_type_index
);
1707 uatomic_inc(&ja
->nr_fallback
);
1713 * Return 0 on success, -EAGAIN if need to retry, or other negative
1714 * error value otherwise.
1717 int ja_node_set_nth(struct cds_ja
*ja
,
1718 struct cds_ja_inode_flag
**node_flag
, uint8_t n
,
1719 struct cds_ja_inode_flag
*child_node_flag
,
1720 struct cds_ja_shadow_node
*shadow_node
,
1724 unsigned int type_index
;
1725 const struct cds_ja_type
*type
;
1726 struct cds_ja_inode
*node
;
1728 dbg_printf("ja_node_set_nth for n=%u, node %p, shadow %p\n",
1729 (unsigned int) n
, ja_node_ptr(*node_flag
), shadow_node
);
1731 node
= ja_node_ptr(*node_flag
);
1732 type_index
= ja_node_type(*node_flag
);
1733 type
= &ja_types
[type_index
];
1734 ret
= _ja_node_set_nth(type
, node
, *node_flag
, shadow_node
,
1735 n
, child_node_flag
);
1738 /* Not enough space in node, need to recompact to next type. */
1739 ret
= ja_node_recompact(JA_RECOMPACT_ADD_NEXT
, ja
, type_index
, type
, node
,
1740 shadow_node
, node_flag
, n
, child_node_flag
, NULL
, level
);
1743 /* Node needs to be recompacted. */
1744 ret
= ja_node_recompact(JA_RECOMPACT_ADD_SAME
, ja
, type_index
, type
, node
,
1745 shadow_node
, node_flag
, n
, child_node_flag
, NULL
, level
);
1752 * Return 0 on success, -EAGAIN if need to retry, or other negative
1753 * error value otherwise.
1756 int ja_node_clear_ptr(struct cds_ja
*ja
,
1757 struct cds_ja_inode_flag
**node_flag_ptr
, /* Pointer to location to nullify */
1758 struct cds_ja_inode_flag
**parent_node_flag_ptr
, /* Address of parent ptr in its parent */
1759 struct cds_ja_shadow_node
*shadow_node
, /* of parent */
1760 uint8_t n
, int level
)
1763 unsigned int type_index
;
1764 const struct cds_ja_type
*type
;
1765 struct cds_ja_inode
*node
;
1767 dbg_printf("ja_node_clear_ptr for node %p, shadow %p, target ptr %p\n",
1768 ja_node_ptr(*parent_node_flag_ptr
), shadow_node
, node_flag_ptr
);
1770 node
= ja_node_ptr(*parent_node_flag_ptr
);
1771 type_index
= ja_node_type(*parent_node_flag_ptr
);
1772 type
= &ja_types
[type_index
];
1773 ret
= _ja_node_clear_ptr(type
, node
, *parent_node_flag_ptr
, shadow_node
, node_flag_ptr
, n
);
1774 if (ret
== -EFBIG
) {
1775 /* Should try recompaction. */
1776 ret
= ja_node_recompact(JA_RECOMPACT_DEL
, ja
, type_index
, type
, node
,
1777 shadow_node
, parent_node_flag_ptr
, n
, NULL
,
1778 node_flag_ptr
, level
);
1783 struct cds_ja_node
*cds_ja_lookup(struct cds_ja
*ja
, uint64_t key
)
1785 unsigned int tree_depth
, i
;
1786 struct cds_ja_inode_flag
*node_flag
;
1788 if (caa_unlikely(key
> ja
->key_max
))
1790 tree_depth
= ja
->tree_depth
;
1791 node_flag
= rcu_dereference(ja
->root
);
1793 /* level 0: root node */
1794 if (!ja_node_ptr(node_flag
))
1797 for (i
= 1; i
< tree_depth
; i
++) {
1800 iter_key
= (uint8_t) (key
>> (JA_BITS_PER_BYTE
* (tree_depth
- i
- 1)));
1801 node_flag
= ja_node_get_nth(node_flag
, NULL
, iter_key
);
1802 dbg_printf("cds_ja_lookup iter key lookup %u finds node_flag %p\n",
1803 (unsigned int) iter_key
, node_flag
);
1804 if (!ja_node_ptr(node_flag
))
1808 /* Last level lookup succeded. We got an actual match. */
1809 return (struct cds_ja_node
*) node_flag
;
1813 struct cds_ja_node
*cds_ja_lookup_inequality(struct cds_ja
*ja
, uint64_t key
,
1814 uint64_t *result_key
, enum ja_lookup_inequality mode
)
1816 int tree_depth
, level
;
1817 struct cds_ja_inode_flag
*node_flag
, *cur_node_depth
[JA_MAX_DEPTH
];
1818 uint8_t cur_key
[JA_MAX_DEPTH
];
1819 uint64_t _result_key
= 0;
1820 enum ja_direction dir
;
1824 if (caa_unlikely(key
> ja
->key_max
|| key
== 0))
1828 if (caa_unlikely(key
>= ja
->key_max
))
1835 memset(cur_node_depth
, 0, sizeof(cur_node_depth
));
1836 memset(cur_key
, 0, sizeof(cur_key
));
1837 tree_depth
= ja
->tree_depth
;
1838 node_flag
= rcu_dereference(ja
->root
);
1839 cur_node_depth
[0] = node_flag
;
1841 /* level 0: root node */
1842 if (!ja_node_ptr(node_flag
))
1845 for (level
= 1; level
< tree_depth
; level
++) {
1848 iter_key
= (uint8_t) (key
>> (JA_BITS_PER_BYTE
* (tree_depth
- level
- 1)));
1849 node_flag
= ja_node_get_nth(node_flag
, NULL
, iter_key
);
1850 if (!ja_node_ptr(node_flag
))
1852 cur_key
[level
- 1] = iter_key
;
1853 cur_node_depth
[level
] = node_flag
;
1854 dbg_printf("cds_ja_lookup_inequality iter key lookup %u finds node_flag %p\n",
1855 (unsigned int) iter_key
, node_flag
);
1858 if (level
== tree_depth
) {
1859 /* Last level lookup succeded. We got an equal match. */
1862 return (struct cds_ja_node
*) node_flag
;
1866 * Find highest value left/right of current node.
1867 * Current node is cur_node_depth[level].
1868 * Start at current level. If we cannot find any key left/right
1869 * of ours, go one level up, seek highest value left/right of
1870 * current (recursively), and when we find one, get the
1871 * rightmost/leftmost child of its rightmost/leftmost child
1884 for (; level
> 0; level
--) {
1887 iter_key
= (uint8_t) (key
>> (JA_BITS_PER_BYTE
* (tree_depth
- level
- 1)));
1888 node_flag
= ja_node_get_leftright(cur_node_depth
[level
- 1],
1889 iter_key
, &cur_key
[level
- 1], dir
);
1890 dbg_printf("cds_ja_lookup_inequality find sibling from %u at %u finds node_flag %p\n",
1891 (unsigned int) iter_key
, (unsigned int) cur_key
[level
- 1],
1893 /* If found left/right sibling, find rightmost/leftmost child. */
1894 if (ja_node_ptr(node_flag
))
1899 /* Reached the root and could not find a left/right sibling. */
1906 * From this point, we are guaranteed to be able to find a
1907 * "below than"/"above than" match. ja_attach_node() and
1908 * ja_detach_node() both guarantee that it is not possible for a
1909 * lookup to reach a dead-end.
1913 * Find rightmost/leftmost child of rightmost/leftmost child
1926 for (; level
< tree_depth
; level
++) {
1927 node_flag
= ja_node_get_minmax(node_flag
, &cur_key
[level
- 1], dir
);
1928 dbg_printf("cds_ja_lookup_inequality find minmax at %u finds node_flag %p\n",
1929 (unsigned int) cur_key
[level
- 1],
1931 if (!ja_node_ptr(node_flag
))
1935 assert(level
== tree_depth
);
1938 for (level
= 1; level
< tree_depth
; level
++) {
1939 _result_key
|= ((uint64_t) cur_key
[level
- 1])
1940 << (JA_BITS_PER_BYTE
* (tree_depth
- level
- 1));
1942 *result_key
= _result_key
;
1944 return (struct cds_ja_node
*) node_flag
;
1947 struct cds_ja_node
*cds_ja_lookup_below_equal(struct cds_ja
*ja
,
1948 uint64_t key
, uint64_t *result_key
)
1950 dbg_printf("cds_ja_lookup_below_equal key %" PRIu64
"\n", key
);
1951 return cds_ja_lookup_inequality(ja
, key
, result_key
, JA_LOOKUP_BE
);
1954 struct cds_ja_node
*cds_ja_lookup_above_equal(struct cds_ja
*ja
,
1955 uint64_t key
, uint64_t *result_key
)
1957 dbg_printf("cds_ja_lookup_above_equal key %" PRIu64
"\n", key
);
1958 return cds_ja_lookup_inequality(ja
, key
, result_key
, JA_LOOKUP_AE
);
1962 * We reached an unpopulated node. Create it and the children we need,
1963 * and then attach the entire branch to the current node. This may
1964 * trigger recompaction of the current node. Locks needed: node lock
1965 * (for add), and, possibly, parent node lock (to update pointer due to
1966 * node recompaction).
1968 * First take node lock, check if recompaction is needed, then take
1969 * parent lock (if needed). Then we can proceed to create the new
1970 * branch. Publish the new branch, and release locks.
1971 * TODO: we currently always take the parent lock even when not needed.
1973 * ja_attach_node() ensures that a lookup will _never_ see a branch that
1974 * leads to a dead-end: before attaching a branch, the entire content of
1975 * the new branch is populated, thus creating a cluster, before
1976 * attaching the cluster to the rest of the tree, thus making it visible
1980 int ja_attach_node(struct cds_ja
*ja
,
1981 struct cds_ja_inode_flag
**attach_node_flag_ptr
,
1982 struct cds_ja_inode_flag
*attach_node_flag
,
1983 struct cds_ja_inode_flag
*parent_attach_node_flag
,
1984 struct cds_ja_inode_flag
**old_node_flag_ptr
,
1985 struct cds_ja_inode_flag
*old_node_flag
,
1988 struct cds_ja_node
*child_node
)
1990 struct cds_ja_shadow_node
*shadow_node
= NULL
,
1991 *parent_shadow_node
= NULL
;
1992 struct cds_ja_inode_flag
*iter_node_flag
, *iter_dest_node_flag
;
1994 struct cds_ja_inode_flag
*created_nodes
[JA_MAX_DEPTH
];
1995 int nr_created_nodes
= 0;
1997 dbg_printf("Attach node at level %u (old_node_flag %p, attach_node_flag_ptr %p attach_node_flag %p, parent_attach_node_flag %p)\n",
1998 level
, old_node_flag
, attach_node_flag_ptr
, attach_node_flag
, parent_attach_node_flag
);
2000 assert(!old_node_flag
);
2001 if (attach_node_flag
) {
2002 shadow_node
= rcuja_shadow_lookup_lock(ja
->ht
, attach_node_flag
);
2008 if (parent_attach_node_flag
) {
2009 parent_shadow_node
= rcuja_shadow_lookup_lock(ja
->ht
,
2010 parent_attach_node_flag
);
2011 if (!parent_shadow_node
) {
2017 if (old_node_flag_ptr
&& ja_node_ptr(*old_node_flag_ptr
)) {
2019 * Target node has been updated between RCU lookup and
2020 * lock acquisition. We need to re-try lookup and
2028 * Perform a lookup query to handle the case where
2029 * old_node_flag_ptr is NULL. We cannot use it to check if the
2030 * node has been populated between RCU lookup and mutex
2033 if (!old_node_flag_ptr
) {
2035 struct cds_ja_inode_flag
*lookup_node_flag
;
2036 struct cds_ja_inode_flag
**lookup_node_flag_ptr
;
2038 iter_key
= (uint8_t) (key
>> (JA_BITS_PER_BYTE
* (ja
->tree_depth
- level
)));
2039 lookup_node_flag
= ja_node_get_nth(attach_node_flag
,
2040 &lookup_node_flag_ptr
,
2042 if (lookup_node_flag
) {
2048 if (attach_node_flag_ptr
&& ja_node_ptr(*attach_node_flag_ptr
) !=
2049 ja_node_ptr(attach_node_flag
)) {
2051 * Target node has been updated between RCU lookup and
2052 * lock acquisition. We need to re-try lookup and
2059 /* Create new branch, starting from bottom */
2060 iter_node_flag
= (struct cds_ja_inode_flag
*) child_node
;
2062 for (i
= ja
->tree_depth
- 1; i
>= (int) level
; i
--) {
2065 iter_key
= (uint8_t) (key
>> (JA_BITS_PER_BYTE
* (ja
->tree_depth
- i
- 1)));
2066 dbg_printf("branch creation level %d, key %u\n",
2067 i
, (unsigned int) iter_key
);
2068 iter_dest_node_flag
= NULL
;
2069 ret
= ja_node_set_nth(ja
, &iter_dest_node_flag
,
2074 dbg_printf("branch creation error %d\n", ret
);
2077 created_nodes
[nr_created_nodes
++] = iter_dest_node_flag
;
2078 iter_node_flag
= iter_dest_node_flag
;
2082 /* Publish branch */
2085 * Attaching to root node.
2087 rcu_assign_pointer(ja
->root
, iter_node_flag
);
2091 iter_key
= (uint8_t) (key
>> (JA_BITS_PER_BYTE
* (ja
->tree_depth
- level
)));
2092 dbg_printf("publish branch at level %d, key %u\n",
2093 level
- 1, (unsigned int) iter_key
);
2094 /* We need to use set_nth on the previous level. */
2095 iter_dest_node_flag
= attach_node_flag
;
2096 ret
= ja_node_set_nth(ja
, &iter_dest_node_flag
,
2099 shadow_node
, level
- 1);
2101 dbg_printf("branch publish error %d\n", ret
);
2107 rcu_assign_pointer(*attach_node_flag_ptr
, iter_dest_node_flag
);
2115 for (i
= 0; i
< nr_created_nodes
; i
++) {
2119 flags
= RCUJA_SHADOW_CLEAR_FREE_LOCK
;
2121 flags
|= RCUJA_SHADOW_CLEAR_FREE_NODE
;
2122 tmpret
= rcuja_shadow_clear(ja
->ht
,
2130 if (parent_shadow_node
)
2131 rcuja_shadow_unlock(parent_shadow_node
);
2134 rcuja_shadow_unlock(shadow_node
);
2140 * Lock the parent containing the pointer to list of duplicates, and add
2141 * node to this list. Failure can happen if concurrent update changes
2142 * the parent before we get the lock. We return -EAGAIN in that case.
2143 * Return 0 on success, negative error value on failure.
2146 int ja_chain_node(struct cds_ja
*ja
,
2147 struct cds_ja_inode_flag
*parent_node_flag
,
2148 struct cds_ja_inode_flag
**node_flag_ptr
,
2149 struct cds_ja_inode_flag
*node_flag
,
2150 struct cds_ja_node
*node
)
2152 struct cds_ja_shadow_node
*shadow_node
;
2155 shadow_node
= rcuja_shadow_lookup_lock(ja
->ht
, parent_node_flag
);
2159 if (ja_node_ptr(*node_flag_ptr
) != ja_node_ptr(node_flag
)) {
2164 * Add node to head of list. Safe against concurrent RCU read
2167 node
->next
= (struct cds_ja_node
*) node_flag
;
2168 rcu_assign_pointer(*node_flag_ptr
, (struct cds_ja_inode_flag
*) node
);
2170 rcuja_shadow_unlock(shadow_node
);
2175 int _cds_ja_add(struct cds_ja
*ja
, uint64_t key
,
2176 struct cds_ja_node
*node
,
2177 struct cds_ja_node
**unique_node_ret
)
2179 unsigned int tree_depth
, i
;
2180 struct cds_ja_inode_flag
*attach_node_flag
,
2184 *parent_attach_node_flag
;
2185 struct cds_ja_inode_flag
**attach_node_flag_ptr
,
2186 **parent_node_flag_ptr
,
2190 if (caa_unlikely(key
> ja
->key_max
)) {
2193 tree_depth
= ja
->tree_depth
;
2196 dbg_printf("cds_ja_add attempt: key %" PRIu64
", node %p\n",
2198 parent2_node_flag
= NULL
;
2200 (struct cds_ja_inode_flag
*) &ja
->root
; /* Use root ptr address as key for mutex */
2201 parent_node_flag_ptr
= NULL
;
2202 node_flag
= rcu_dereference(ja
->root
);
2203 node_flag_ptr
= &ja
->root
;
2205 /* Iterate on all internal levels */
2206 for (i
= 1; i
< tree_depth
; i
++) {
2209 if (!ja_node_ptr(node_flag
))
2211 dbg_printf("cds_ja_add iter parent2_node_flag %p parent_node_flag %p node_flag_ptr %p node_flag %p\n",
2212 parent2_node_flag
, parent_node_flag
, node_flag_ptr
, node_flag
);
2213 iter_key
= (uint8_t) (key
>> (JA_BITS_PER_BYTE
* (tree_depth
- i
- 1)));
2214 parent2_node_flag
= parent_node_flag
;
2215 parent_node_flag
= node_flag
;
2216 parent_node_flag_ptr
= node_flag_ptr
;
2217 node_flag
= ja_node_get_nth(node_flag
,
2223 * We reached either bottom of tree or internal NULL node,
2224 * simply add node to last internal level, or chain it if key is
2227 if (!ja_node_ptr(node_flag
)) {
2228 dbg_printf("cds_ja_add NULL parent2_node_flag %p parent_node_flag %p node_flag_ptr %p node_flag %p\n",
2229 parent2_node_flag
, parent_node_flag
, node_flag_ptr
, node_flag
);
2231 attach_node_flag
= parent_node_flag
;
2232 attach_node_flag_ptr
= parent_node_flag_ptr
;
2233 parent_attach_node_flag
= parent2_node_flag
;
2235 ret
= ja_attach_node(ja
, attach_node_flag_ptr
,
2237 parent_attach_node_flag
,
2242 if (unique_node_ret
) {
2243 *unique_node_ret
= (struct cds_ja_node
*) ja_node_ptr(node_flag
);
2247 dbg_printf("cds_ja_add duplicate parent2_node_flag %p parent_node_flag %p node_flag_ptr %p node_flag %p\n",
2248 parent2_node_flag
, parent_node_flag
, node_flag_ptr
, node_flag
);
2250 attach_node_flag
= node_flag
;
2251 attach_node_flag_ptr
= node_flag_ptr
;
2252 parent_attach_node_flag
= parent_node_flag
;
2254 ret
= ja_chain_node(ja
,
2255 parent_attach_node_flag
,
2256 attach_node_flag_ptr
,
2260 if (ret
== -EAGAIN
|| ret
== -EEXIST
)
2266 int cds_ja_add(struct cds_ja
*ja
, uint64_t key
,
2267 struct cds_ja_node
*node
)
2269 return _cds_ja_add(ja
, key
, node
, NULL
);
2272 struct cds_ja_node
*cds_ja_add_unique(struct cds_ja
*ja
, uint64_t key
,
2273 struct cds_ja_node
*node
)
2276 struct cds_ja_node
*ret_node
;
2278 ret
= _cds_ja_add(ja
, key
, node
, &ret_node
);
2286 * Note: there is no need to lookup the pointer address associated with
2287 * each node's nth item after taking the lock: it's already been done by
2288 * cds_ja_del while holding the rcu read-side lock, and our node rules
2289 * ensure that when a match value -> pointer is found in a node, it is
2290 * _NEVER_ changed for that node without recompaction, and recompaction
2291 * reallocates the node.
2292 * However, when a child is removed from "linear" nodes, its pointer
2293 * is set to NULL. We therefore check, while holding the locks, if this
2294 * pointer is NULL, and return -ENOENT to the caller if it is the case.
2296 * ja_detach_node() ensures that a lookup will _never_ see a branch that
2297 * leads to a dead-end: when removing branch, it makes sure to perform
2298 * the "cut" at the highest node that has only one child, effectively
2299 * replacing it with a NULL pointer.
2302 int ja_detach_node(struct cds_ja
*ja
,
2303 struct cds_ja_inode_flag
**snapshot
,
2304 struct cds_ja_inode_flag
***snapshot_ptr
,
2305 uint8_t *snapshot_n
,
2308 struct cds_ja_node
*node
)
2310 struct cds_ja_shadow_node
*shadow_nodes
[JA_MAX_DEPTH
];
2311 struct cds_ja_inode_flag
**node_flag_ptr
= NULL
,
2312 *parent_node_flag
= NULL
,
2313 **parent_node_flag_ptr
= NULL
;
2314 struct cds_ja_inode_flag
*iter_node_flag
;
2315 int ret
, i
, nr_shadow
= 0, nr_clear
= 0, nr_branch
= 0;
2318 assert(nr_snapshot
== ja
->tree_depth
+ 1);
2321 * From the last internal level node going up, get the node
2322 * lock, check if the node has only one child left. If it is the
2323 * case, we continue iterating upward. When we reach a node
2324 * which has more that one child left, we lock the parent, and
2325 * proceed to the node deletion (removing its children too).
2327 for (i
= nr_snapshot
- 2; i
>= 1; i
--) {
2328 struct cds_ja_shadow_node
*shadow_node
;
2330 shadow_node
= rcuja_shadow_lookup_lock(ja
->ht
,
2336 shadow_nodes
[nr_shadow
++] = shadow_node
;
2339 * Check if node has been removed between RCU
2340 * lookup and lock acquisition.
2342 assert(snapshot_ptr
[i
+ 1]);
2343 if (ja_node_ptr(*snapshot_ptr
[i
+ 1])
2344 != ja_node_ptr(snapshot
[i
+ 1])) {
2349 assert(shadow_node
->nr_child
> 0);
2350 if (shadow_node
->nr_child
== 1 && i
> 1)
2353 if (shadow_node
->nr_child
> 1 || i
== 1) {
2354 /* Lock parent and break */
2355 shadow_node
= rcuja_shadow_lookup_lock(ja
->ht
,
2361 shadow_nodes
[nr_shadow
++] = shadow_node
;
2364 * Check if node has been removed between RCU
2365 * lookup and lock acquisition.
2367 assert(snapshot_ptr
[i
]);
2368 if (ja_node_ptr(*snapshot_ptr
[i
])
2369 != ja_node_ptr(snapshot
[i
])) {
2374 node_flag_ptr
= snapshot_ptr
[i
+ 1];
2375 n
= snapshot_n
[i
+ 1];
2376 parent_node_flag_ptr
= snapshot_ptr
[i
];
2377 parent_node_flag
= snapshot
[i
];
2381 * Lock parent's parent, in case we need
2382 * to recompact parent.
2384 shadow_node
= rcuja_shadow_lookup_lock(ja
->ht
,
2390 shadow_nodes
[nr_shadow
++] = shadow_node
;
2393 * Check if node has been removed between RCU
2394 * lookup and lock acquisition.
2396 assert(snapshot_ptr
[i
- 1]);
2397 if (ja_node_ptr(*snapshot_ptr
[i
- 1])
2398 != ja_node_ptr(snapshot
[i
- 1])) {
2409 * At this point, we want to delete all nodes that are about to
2410 * be removed from shadow_nodes (except the last one, which is
2411 * either the root or the parent of the upmost node with 1
2412 * child). OK to free lock here, because RCU read lock is held,
2413 * and free only performed in call_rcu.
2416 for (i
= 0; i
< nr_clear
; i
++) {
2417 ret
= rcuja_shadow_clear(ja
->ht
,
2418 shadow_nodes
[i
]->node_flag
,
2420 RCUJA_SHADOW_CLEAR_FREE_NODE
2421 | RCUJA_SHADOW_CLEAR_FREE_LOCK
);
2425 iter_node_flag
= parent_node_flag
;
2426 /* Remove from parent */
2427 ret
= ja_node_clear_ptr(ja
,
2428 node_flag_ptr
, /* Pointer to location to nullify */
2429 &iter_node_flag
, /* Old new parent ptr in its parent */
2430 shadow_nodes
[nr_branch
- 1], /* of parent */
2435 dbg_printf("ja_detach_node: publish %p instead of %p\n",
2436 iter_node_flag
, *parent_node_flag_ptr
);
2437 /* Update address of parent ptr in its parent */
2438 rcu_assign_pointer(*parent_node_flag_ptr
, iter_node_flag
);
2441 for (i
= 0; i
< nr_shadow
; i
++)
2442 rcuja_shadow_unlock(shadow_nodes
[i
]);
2447 int ja_unchain_node(struct cds_ja
*ja
,
2448 struct cds_ja_inode_flag
*parent_node_flag
,
2449 struct cds_ja_inode_flag
**node_flag_ptr
,
2450 struct cds_ja_inode_flag
*node_flag
,
2451 struct cds_ja_node
*node
)
2453 struct cds_ja_shadow_node
*shadow_node
;
2454 struct cds_ja_node
*iter_node
, **iter_node_ptr
, **prev_node_ptr
= NULL
;
2455 int ret
= 0, count
= 0, found
= 0;
2457 shadow_node
= rcuja_shadow_lookup_lock(ja
->ht
, parent_node_flag
);
2460 if (ja_node_ptr(*node_flag_ptr
) != ja_node_ptr(node_flag
)) {
2465 * Find the previous node's next pointer pointing to our node,
2466 * so we can update it. Retry if another thread removed all but
2467 * one of duplicates since check (this check was performed
2468 * without lock). Ensure that the node we are about to remove is
2469 * still in the list (while holding lock). No need for RCU
2470 * traversal here since we hold the lock on the parent.
2472 iter_node_ptr
= (struct cds_ja_node
**) node_flag_ptr
;
2473 iter_node
= (struct cds_ja_node
*) ja_node_ptr(node_flag
);
2474 cds_ja_for_each_duplicate(iter_node
) {
2476 if (iter_node
== node
) {
2477 prev_node_ptr
= iter_node_ptr
;
2480 iter_node_ptr
= &iter_node
->next
;
2483 if (!found
|| count
== 1) {
2487 CMM_STORE_SHARED(*prev_node_ptr
, node
->next
);
2489 * Validate that we indeed removed the node from linked list.
2491 assert(ja_node_ptr(*node_flag_ptr
) != (struct cds_ja_inode
*) node
);
2493 rcuja_shadow_unlock(shadow_node
);
2498 * Called with RCU read lock held.
2500 int cds_ja_del(struct cds_ja
*ja
, uint64_t key
,
2501 struct cds_ja_node
*node
)
2503 unsigned int tree_depth
, i
;
2504 struct cds_ja_inode_flag
*snapshot
[JA_MAX_DEPTH
];
2505 struct cds_ja_inode_flag
**snapshot_ptr
[JA_MAX_DEPTH
];
2506 uint8_t snapshot_n
[JA_MAX_DEPTH
];
2507 struct cds_ja_inode_flag
*node_flag
;
2508 struct cds_ja_inode_flag
**prev_node_flag_ptr
,
2513 if (caa_unlikely(key
> ja
->key_max
))
2515 tree_depth
= ja
->tree_depth
;
2519 dbg_printf("cds_ja_del attempt: key %" PRIu64
", node %p\n",
2522 /* snapshot for level 0 is only for shadow node lookup */
2525 snapshot_ptr
[nr_snapshot
] = NULL
;
2526 snapshot
[nr_snapshot
++] = (struct cds_ja_inode_flag
*) &ja
->root
;
2527 node_flag
= rcu_dereference(ja
->root
);
2528 prev_node_flag_ptr
= &ja
->root
;
2529 node_flag_ptr
= &ja
->root
;
2531 /* Iterate on all internal levels */
2532 for (i
= 1; i
< tree_depth
; i
++) {
2535 dbg_printf("cds_ja_del iter node_flag %p\n",
2537 if (!ja_node_ptr(node_flag
)) {
2540 iter_key
= (uint8_t) (key
>> (JA_BITS_PER_BYTE
* (tree_depth
- i
- 1)));
2541 snapshot_n
[nr_snapshot
+ 1] = iter_key
;
2542 snapshot_ptr
[nr_snapshot
] = prev_node_flag_ptr
;
2543 snapshot
[nr_snapshot
++] = node_flag
;
2544 node_flag
= ja_node_get_nth(node_flag
,
2548 prev_node_flag_ptr
= node_flag_ptr
;
2549 dbg_printf("cds_ja_del iter key lookup %u finds node_flag %p, prev_node_flag_ptr %p\n",
2550 (unsigned int) iter_key
, node_flag
,
2551 prev_node_flag_ptr
);
2554 * We reached bottom of tree, try to find the node we are trying
2555 * to remove. Fail if we cannot find it.
2557 if (!ja_node_ptr(node_flag
)) {
2558 dbg_printf("cds_ja_del: no node found for key %" PRIu64
"\n",
2562 struct cds_ja_node
*iter_node
, *match
= NULL
;
2565 iter_node
= (struct cds_ja_node
*) ja_node_ptr(node_flag
);
2566 cds_ja_for_each_duplicate_rcu(iter_node
) {
2567 dbg_printf("cds_ja_del: compare %p with iter_node %p\n", node
, iter_node
);
2568 if (iter_node
== node
)
2574 dbg_printf("cds_ja_del: no node match for node %p key %" PRIu64
"\n", node
, key
);
2580 * Removing last of duplicates. Last snapshot
2581 * does not have a shadow node (external leafs).
2583 snapshot_ptr
[nr_snapshot
] = prev_node_flag_ptr
;
2584 snapshot
[nr_snapshot
++] = node_flag
;
2585 ret
= ja_detach_node(ja
, snapshot
, snapshot_ptr
,
2586 snapshot_n
, nr_snapshot
, key
, node
);
2588 ret
= ja_unchain_node(ja
, snapshot
[nr_snapshot
- 1],
2589 node_flag_ptr
, node_flag
, match
);
2593 * Explanation of -ENOENT handling: caused by concurrent delete
2594 * between RCU lookup and actual removal. Need to re-do the
2595 * lookup and removal attempt.
2597 if (ret
== -EAGAIN
|| ret
== -ENOENT
)
2602 struct cds_ja
*_cds_ja_new(unsigned int key_bits
,
2603 const struct rcu_flavor_struct
*flavor
)
2607 struct cds_ja_shadow_node
*root_shadow_node
;
2609 ja
= calloc(sizeof(*ja
), 1);
2621 ja
->key_max
= (1ULL << key_bits
) - 1;
2624 ja
->key_max
= UINT64_MAX
;
2630 /* ja->root is NULL */
2631 /* tree_depth 0 is for pointer to root node */
2632 ja
->tree_depth
= (key_bits
>> JA_LOG2_BITS_PER_BYTE
) + 1;
2633 assert(ja
->tree_depth
<= JA_MAX_DEPTH
);
2634 ja
->ht
= rcuja_create_ht(flavor
);
2639 * Note: we should not free this node until judy array destroy.
2641 root_shadow_node
= rcuja_shadow_set(ja
->ht
,
2642 (struct cds_ja_inode_flag
*) &ja
->root
,
2644 if (!root_shadow_node
) {
2652 ret
= rcuja_delete_ht(ja
->ht
);
2662 void print_debug_fallback_distribution(struct cds_ja
*ja
)
2666 fprintf(stderr
, "Fallback node distribution:\n");
2667 for (i
= 0; i
< JA_ENTRY_PER_NODE
; i
++) {
2668 if (!ja
->node_fallback_count_distribution
[i
])
2670 fprintf(stderr
, " %3u: %4lu\n",
2671 i
, ja
->node_fallback_count_distribution
[i
]);
2676 int ja_final_checks(struct cds_ja
*ja
)
2678 double fallback_ratio
;
2679 unsigned long na
, nf
, nr_fallback
;
2682 fallback_ratio
= (double) uatomic_read(&ja
->nr_fallback
);
2683 fallback_ratio
/= (double) uatomic_read(&ja
->nr_nodes_allocated
);
2684 nr_fallback
= uatomic_read(&ja
->nr_fallback
);
2687 "[warning] RCU Judy Array used %lu fallback node(s) (ratio: %g)\n",
2688 uatomic_read(&ja
->nr_fallback
),
2691 na
= uatomic_read(&ja
->nr_nodes_allocated
);
2692 nf
= uatomic_read(&ja
->nr_nodes_freed
);
2693 dbg_printf("Nodes allocated: %lu, Nodes freed: %lu.\n", na
, nf
);
2695 print_debug_fallback_distribution(ja
);
2698 fprintf(stderr
, "[error] Judy array leaked %ld nodes. Allocated: %lu, freed: %lu.\n",
2699 (long) na
- nf
, na
, nf
);
2706 * There should be no more concurrent add, delete, nor look-up performed
2707 * on the Judy array while it is being destroyed (ensured by the
2710 int cds_ja_destroy(struct cds_ja
*ja
)
2712 const struct rcu_flavor_struct
*flavor
;
2715 flavor
= cds_lfht_rcu_flavor(ja
->ht
);
2716 rcuja_shadow_prune(ja
->ht
,
2717 RCUJA_SHADOW_CLEAR_FREE_NODE
| RCUJA_SHADOW_CLEAR_FREE_LOCK
);
2718 flavor
->thread_offline();
2719 ret
= rcuja_delete_ht(ja
->ht
);
2723 /* Wait for in-flight call_rcu free to complete. */
2726 flavor
->thread_online();
2727 ret
= ja_final_checks(ja
);