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
28 #include <urcu/rcuja.h>
29 #include <urcu/compiler.h>
30 #include <urcu/arch.h>
32 #include <urcu-pointer.h>
33 #include <urcu/uatomic.h>
36 #include "rcuja-internal.h"
39 #define abs_int(a) ((int) (a) > 0 ? (int) (a) : -((int) (a)))
42 enum cds_ja_type_class
{
43 RCU_JA_LINEAR
= 0, /* Type A */
44 /* 32-bit: 1 to 25 children, 8 to 128 bytes */
45 /* 64-bit: 1 to 28 children, 16 to 256 bytes */
46 RCU_JA_POOL
= 1, /* Type B */
47 /* 32-bit: 26 to 100 children, 256 to 512 bytes */
48 /* 64-bit: 29 to 112 children, 512 to 1024 bytes */
49 RCU_JA_PIGEON
= 2, /* Type C */
50 /* 32-bit: 101 to 256 children, 1024 bytes */
51 /* 64-bit: 113 to 256 children, 2048 bytes */
52 /* Leaf nodes are implicit from their height in the tree */
55 RCU_JA_NULL
, /* not an encoded type, but keeps code regular */
59 enum cds_ja_type_class type_class
;
60 uint16_t min_child
; /* minimum number of children: 1 to 256 */
61 uint16_t max_child
; /* maximum number of children: 1 to 256 */
62 uint16_t max_linear_child
; /* per-pool max nr. children: 1 to 256 */
63 uint16_t order
; /* node size is (1 << order), in bytes */
64 uint16_t nr_pool_order
; /* number of pools */
65 uint16_t pool_size_order
; /* pool size */
69 * Iteration on the array to find the right node size for the number of
70 * children stops when it reaches .max_child == 256 (this is the largest
71 * possible node size, which contains 256 children).
72 * The min_child overlaps with the previous max_child to provide an
73 * hysteresis loop to reallocation for patterns of cyclic add/removal
74 * within the same node.
75 * The node the index within the following arrays is represented on 3
76 * bits. It identifies the node type, min/max number of children, and
78 * The max_child values for the RCU_JA_POOL below result from
79 * statistical approximation: over million populations, the max_child
80 * covers between 97% and 99% of the populations generated. Therefore, a
81 * fallback should exist to cover the rare extreme population unbalance
82 * cases, but it will not have a major impact on speed nor space
83 * consumption, since those are rare cases.
86 #if (CAA_BITS_PER_LONG < 64)
89 ja_type_0_max_child
= 1,
90 ja_type_1_max_child
= 3,
91 ja_type_2_max_child
= 6,
92 ja_type_3_max_child
= 12,
93 ja_type_4_max_child
= 25,
94 ja_type_5_max_child
= 48,
95 ja_type_6_max_child
= 92,
96 ja_type_7_max_child
= 256,
97 ja_type_8_max_child
= 0, /* NULL */
101 ja_type_0_max_linear_child
= 1,
102 ja_type_1_max_linear_child
= 3,
103 ja_type_2_max_linear_child
= 6,
104 ja_type_3_max_linear_child
= 12,
105 ja_type_4_max_linear_child
= 25,
106 ja_type_5_max_linear_child
= 24,
107 ja_type_6_max_linear_child
= 23,
111 ja_type_5_nr_pool_order
= 1,
112 ja_type_6_nr_pool_order
= 2,
115 const struct cds_ja_type ja_types
[] = {
116 { .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, },
117 { .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, },
118 { .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, },
119 { .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, },
120 { .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, },
122 /* Pools may fill sooner than max_child */
123 /* This pool is hardcoded at index 5. See ja_node_ptr(). */
124 { .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, },
125 /* This pool is hardcoded at index 6. See ja_node_ptr(). */
126 { .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, },
129 * Upon node removal below min_child, if child pool is filled
130 * beyond capacity, we roll back to pigeon.
132 { .type_class
= RCU_JA_PIGEON
, .min_child
= 83, .max_child
= ja_type_7_max_child
, .order
= 10, },
134 { .type_class
= RCU_JA_NULL
, .min_child
= 0, .max_child
= ja_type_8_max_child
, },
136 #else /* !(CAA_BITS_PER_LONG < 64) */
137 /* 64-bit pointers */
139 ja_type_0_max_child
= 1,
140 ja_type_1_max_child
= 3,
141 ja_type_2_max_child
= 7,
142 ja_type_3_max_child
= 14,
143 ja_type_4_max_child
= 28,
144 ja_type_5_max_child
= 54,
145 ja_type_6_max_child
= 104,
146 ja_type_7_max_child
= 256,
147 ja_type_8_max_child
= 256,
151 ja_type_0_max_linear_child
= 1,
152 ja_type_1_max_linear_child
= 3,
153 ja_type_2_max_linear_child
= 7,
154 ja_type_3_max_linear_child
= 14,
155 ja_type_4_max_linear_child
= 28,
156 ja_type_5_max_linear_child
= 27,
157 ja_type_6_max_linear_child
= 26,
161 ja_type_5_nr_pool_order
= 1,
162 ja_type_6_nr_pool_order
= 2,
165 const struct cds_ja_type ja_types
[] = {
166 { .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, },
167 { .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, },
168 { .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, },
169 { .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, },
170 { .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, },
172 /* Pools may fill sooner than max_child. */
173 /* This pool is hardcoded at index 5. See ja_node_ptr(). */
174 { .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, },
175 /* This pool is hardcoded at index 6. See ja_node_ptr(). */
176 { .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, },
179 * Upon node removal below min_child, if child pool is filled
180 * beyond capacity, we roll back to pigeon.
182 { .type_class
= RCU_JA_PIGEON
, .min_child
= 95, .max_child
= ja_type_7_max_child
, .order
= 11, },
184 { .type_class
= RCU_JA_NULL
, .min_child
= 0, .max_child
= ja_type_8_max_child
, },
186 #endif /* !(BITS_PER_LONG < 64) */
188 static inline __attribute__((unused
))
189 void static_array_size_check(void)
191 CAA_BUILD_BUG_ON(CAA_ARRAY_SIZE(ja_types
) < JA_TYPE_MAX_NR
);
195 * The cds_ja_node contains the compressed node data needed for
196 * read-side. For linear and pool node configurations, it starts with a
197 * byte counting the number of children in the node. Then, the
198 * node-specific data is placed.
199 * The node mutex, if any is needed, protecting concurrent updated of
200 * each node is placed in a separate hash table indexed by node address.
201 * For the pigeon configuration, the number of children is also kept in
202 * a separate hash table, indexed by node address, because it is only
203 * required for updates.
206 #define DECLARE_LINEAR_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]; \
213 #define DECLARE_POOL_NODE(index) \
217 uint8_t child_value[ja_type_## index ##_max_linear_child]; \
218 struct cds_ja_inode_flag *child_ptr[ja_type_## index ##_max_linear_child]; \
219 } linear[1U << ja_type_## index ##_nr_pool_order]; \
222 struct cds_ja_inode
{
224 /* Linear configuration */
225 DECLARE_LINEAR_NODE(0) conf_0
;
226 DECLARE_LINEAR_NODE(1) conf_1
;
227 DECLARE_LINEAR_NODE(2) conf_2
;
228 DECLARE_LINEAR_NODE(3) conf_3
;
229 DECLARE_LINEAR_NODE(4) conf_4
;
231 /* Pool configuration */
232 DECLARE_POOL_NODE(5) conf_5
;
233 DECLARE_POOL_NODE(6) conf_6
;
235 /* Pigeon configuration */
237 struct cds_ja_inode_flag
*child
[ja_type_7_max_child
];
239 /* data aliasing nodes for computed accesses */
240 uint8_t data
[sizeof(struct cds_ja_inode_flag
*) * ja_type_7_max_child
];
245 JA_RECOMPACT_ADD_SAME
,
246 JA_RECOMPACT_ADD_NEXT
,
251 struct cds_ja_inode
*_ja_node_mask_ptr(struct cds_ja_inode_flag
*node
)
253 return (struct cds_ja_inode
*) (((unsigned long) node
) & JA_PTR_MASK
);
256 unsigned long ja_node_type(struct cds_ja_inode_flag
*node
)
260 if (_ja_node_mask_ptr(node
) == NULL
) {
261 return NODE_INDEX_NULL
;
263 type
= (unsigned int) ((unsigned long) node
& JA_TYPE_MASK
);
264 assert(type
< (1UL << JA_TYPE_BITS
));
269 struct cds_ja_inode
*alloc_cds_ja_node(struct cds_ja
*ja
,
270 const struct cds_ja_type
*ja_type
)
272 size_t len
= 1U << ja_type
->order
;
276 ret
= posix_memalign(&p
, len
, len
);
281 uatomic_inc(&ja
->nr_nodes_allocated
);
285 void free_cds_ja_node(struct cds_ja
*ja
, struct cds_ja_inode
*node
)
289 uatomic_inc(&ja
->nr_nodes_freed
);
292 #define __JA_ALIGN_MASK(v, mask) (((v) + (mask)) & ~(mask))
293 #define JA_ALIGN(v, align) __JA_ALIGN_MASK(v, (typeof(v)) (align) - 1)
294 #define __JA_FLOOR_MASK(v, mask) ((v) & ~(mask))
295 #define JA_FLOOR(v, align) __JA_FLOOR_MASK(v, (typeof(v)) (align) - 1)
298 uint8_t *align_ptr_size(uint8_t *ptr
)
300 return (uint8_t *) JA_ALIGN((unsigned long) ptr
, sizeof(void *));
304 uint8_t ja_linear_node_get_nr_child(const struct cds_ja_type
*type
,
305 struct cds_ja_inode
*node
)
307 assert(type
->type_class
== RCU_JA_LINEAR
|| type
->type_class
== RCU_JA_POOL
);
308 return rcu_dereference(node
->u
.data
[0]);
312 * The order in which values and pointers are does does not matter: if
313 * a value is missing, we return NULL. If a value is there, but its
314 * associated pointers is still NULL, we return NULL too.
317 struct cds_ja_inode_flag
*ja_linear_node_get_nth(const struct cds_ja_type
*type
,
318 struct cds_ja_inode
*node
,
319 struct cds_ja_inode_flag
***node_flag_ptr
,
324 struct cds_ja_inode_flag
**pointers
;
325 struct cds_ja_inode_flag
*ptr
;
328 assert(type
->type_class
== RCU_JA_LINEAR
|| type
->type_class
== RCU_JA_POOL
);
330 nr_child
= ja_linear_node_get_nr_child(type
, node
);
331 cmm_smp_rmb(); /* read nr_child before values and pointers */
332 assert(nr_child
<= type
->max_linear_child
);
333 assert(type
->type_class
!= RCU_JA_LINEAR
|| nr_child
>= type
->min_child
);
335 values
= &node
->u
.data
[1];
336 for (i
= 0; i
< nr_child
; i
++) {
337 if (CMM_LOAD_SHARED(values
[i
]) == n
)
341 if (caa_unlikely(node_flag_ptr
))
342 *node_flag_ptr
= NULL
;
345 pointers
= (struct cds_ja_inode_flag
**) align_ptr_size(&values
[type
->max_linear_child
]);
346 ptr
= rcu_dereference(pointers
[i
]);
347 if (caa_unlikely(node_flag_ptr
))
348 *node_flag_ptr
= &pointers
[i
];
353 struct cds_ja_inode_flag
*ja_linear_node_get_left(const struct cds_ja_type
*type
,
354 struct cds_ja_inode
*node
,
359 struct cds_ja_inode_flag
**pointers
;
360 struct cds_ja_inode_flag
*ptr
;
361 unsigned int i
, match_idx
;
364 assert(type
->type_class
== RCU_JA_LINEAR
|| type
->type_class
== RCU_JA_POOL
);
366 nr_child
= ja_linear_node_get_nr_child(type
, node
);
367 cmm_smp_rmb(); /* read nr_child before values and pointers */
368 assert(nr_child
<= type
->max_linear_child
);
369 assert(type
->type_class
!= RCU_JA_LINEAR
|| nr_child
>= type
->min_child
);
371 values
= &node
->u
.data
[1];
372 for (i
= 0; i
< nr_child
; i
++) {
375 v
= CMM_LOAD_SHARED(values
[i
]);
376 if (v
< n
&& (int) v
> match_v
) {
384 pointers
= (struct cds_ja_inode_flag
**) align_ptr_size(&values
[type
->max_linear_child
]);
385 ptr
= rcu_dereference(pointers
[match_idx
]);
390 void ja_linear_node_get_ith_pos(const struct cds_ja_type
*type
,
391 struct cds_ja_inode
*node
,
394 struct cds_ja_inode_flag
**iter
)
397 struct cds_ja_inode_flag
**pointers
;
399 assert(type
->type_class
== RCU_JA_LINEAR
|| type
->type_class
== RCU_JA_POOL
);
400 assert(i
< ja_linear_node_get_nr_child(type
, node
));
402 values
= &node
->u
.data
[1];
404 pointers
= (struct cds_ja_inode_flag
**) align_ptr_size(&values
[type
->max_linear_child
]);
409 struct cds_ja_inode_flag
*ja_pool_node_get_nth(const struct cds_ja_type
*type
,
410 struct cds_ja_inode
*node
,
411 struct cds_ja_inode_flag
*node_flag
,
412 struct cds_ja_inode_flag
***node_flag_ptr
,
415 struct cds_ja_inode
*linear
;
417 assert(type
->type_class
== RCU_JA_POOL
);
419 switch (type
->nr_pool_order
) {
422 unsigned long bitsel
, index
;
424 bitsel
= ja_node_pool_1d_bitsel(node_flag
);
425 assert(bitsel
< CHAR_BIT
);
426 index
= ((unsigned long) n
>> bitsel
) & 0x1;
427 linear
= (struct cds_ja_inode
*) &node
->u
.data
[index
<< type
->pool_size_order
];
432 unsigned long bitsel
[2], index
[2], rindex
;
434 ja_node_pool_2d_bitsel(node_flag
, bitsel
);
435 assert(bitsel
[0] < CHAR_BIT
);
436 assert(bitsel
[1] < CHAR_BIT
);
437 index
[0] = ((unsigned long) n
>> bitsel
[0]) & 0x1;
439 index
[1] = ((unsigned long) n
>> bitsel
[1]) & 0x1;
440 rindex
= index
[0] | index
[1];
441 linear
= (struct cds_ja_inode
*) &node
->u
.data
[rindex
<< type
->pool_size_order
];
448 return ja_linear_node_get_nth(type
, linear
, node_flag_ptr
, n
);
452 struct cds_ja_inode
*ja_pool_node_get_ith_pool(const struct cds_ja_type
*type
,
453 struct cds_ja_inode
*node
,
456 assert(type
->type_class
== RCU_JA_POOL
);
457 return (struct cds_ja_inode
*)
458 &node
->u
.data
[(unsigned int) i
<< type
->pool_size_order
];
462 struct cds_ja_inode_flag
*ja_pool_node_get_left(const struct cds_ja_type
*type
,
463 struct cds_ja_inode
*node
,
466 unsigned int pool_nr
;
468 struct cds_ja_inode_flag
*match_node_flag
= NULL
;
470 assert(type
->type_class
== RCU_JA_POOL
);
472 for (pool_nr
= 0; pool_nr
< (1U << type
->nr_pool_order
); pool_nr
++) {
473 struct cds_ja_inode
*pool
=
474 ja_pool_node_get_ith_pool(type
,
477 ja_linear_node_get_nr_child(type
, pool
);
480 for (j
= 0; j
< nr_child
; j
++) {
481 struct cds_ja_inode_flag
*iter
;
484 ja_linear_node_get_ith_pos(type
, pool
,
488 if (v
< n
&& (int) v
> match_v
) {
490 match_node_flag
= iter
;
494 return match_node_flag
;
498 struct cds_ja_inode_flag
*ja_pigeon_node_get_nth(const struct cds_ja_type
*type
,
499 struct cds_ja_inode
*node
,
500 struct cds_ja_inode_flag
***node_flag_ptr
,
503 struct cds_ja_inode_flag
**child_node_flag_ptr
;
504 struct cds_ja_inode_flag
*child_node_flag
;
506 assert(type
->type_class
== RCU_JA_PIGEON
);
507 child_node_flag_ptr
= &((struct cds_ja_inode_flag
**) node
->u
.data
)[n
];
508 child_node_flag
= rcu_dereference(*child_node_flag_ptr
);
509 dbg_printf("ja_pigeon_node_get_nth child_node_flag_ptr %p\n",
510 child_node_flag_ptr
);
511 if (caa_unlikely(node_flag_ptr
))
512 *node_flag_ptr
= child_node_flag_ptr
;
513 return child_node_flag
;
517 struct cds_ja_inode_flag
*ja_pigeon_node_get_left(const struct cds_ja_type
*type
,
518 struct cds_ja_inode
*node
,
521 struct cds_ja_inode_flag
**child_node_flag_ptr
;
522 struct cds_ja_inode_flag
*child_node_flag
;
525 assert(type
->type_class
== RCU_JA_PIGEON
);
527 /* n - 1 is first value left of n */
528 for (i
= n
- 1; i
>= 0; i
--) {
529 child_node_flag_ptr
= &((struct cds_ja_inode_flag
**) node
->u
.data
)[i
];
530 child_node_flag
= rcu_dereference(*child_node_flag_ptr
);
531 if (child_node_flag
) {
532 dbg_printf("ja_pigeon_node_get_left child_node_flag %p\n",
534 return child_node_flag
;
541 struct cds_ja_inode_flag
*ja_pigeon_node_get_ith_pos(const struct cds_ja_type
*type
,
542 struct cds_ja_inode
*node
,
545 return ja_pigeon_node_get_nth(type
, node
, NULL
, i
);
549 * ja_node_get_nth: get nth item from a node.
550 * node_flag is already rcu_dereference'd.
553 struct cds_ja_inode_flag
*ja_node_get_nth(struct cds_ja_inode_flag
*node_flag
,
554 struct cds_ja_inode_flag
***node_flag_ptr
,
557 unsigned int type_index
;
558 struct cds_ja_inode
*node
;
559 const struct cds_ja_type
*type
;
561 node
= ja_node_ptr(node_flag
);
562 assert(node
!= NULL
);
563 type_index
= ja_node_type(node_flag
);
564 type
= &ja_types
[type_index
];
566 switch (type
->type_class
) {
568 return ja_linear_node_get_nth(type
, node
,
571 return ja_pool_node_get_nth(type
, node
, node_flag
,
574 return ja_pigeon_node_get_nth(type
, node
,
578 return (void *) -1UL;
583 struct cds_ja_inode_flag
*ja_node_get_left(struct cds_ja_inode_flag
*node_flag
,
586 unsigned int type_index
;
587 struct cds_ja_inode
*node
;
588 const struct cds_ja_type
*type
;
590 node
= ja_node_ptr(node_flag
);
591 assert(node
!= NULL
);
592 type_index
= ja_node_type(node_flag
);
593 type
= &ja_types
[type_index
];
595 switch (type
->type_class
) {
597 return ja_linear_node_get_left(type
, node
, n
);
599 return ja_pool_node_get_left(type
, node
, n
);
601 return ja_pigeon_node_get_left(type
, node
, n
);
604 return (void *) -1UL;
609 struct cds_ja_inode_flag
*ja_node_get_rightmost(struct cds_ja_inode_flag
*node_flag
)
611 return ja_node_get_left(node_flag
, JA_ENTRY_PER_NODE
);
615 int ja_linear_node_set_nth(const struct cds_ja_type
*type
,
616 struct cds_ja_inode
*node
,
617 struct cds_ja_shadow_node
*shadow_node
,
619 struct cds_ja_inode_flag
*child_node_flag
)
622 uint8_t *values
, *nr_child_ptr
;
623 struct cds_ja_inode_flag
**pointers
;
624 unsigned int i
, unused
= 0;
626 assert(type
->type_class
== RCU_JA_LINEAR
|| type
->type_class
== RCU_JA_POOL
);
628 nr_child_ptr
= &node
->u
.data
[0];
629 dbg_printf("linear set nth: n %u, nr_child_ptr %p\n",
630 (unsigned int) n
, nr_child_ptr
);
631 nr_child
= *nr_child_ptr
;
632 assert(nr_child
<= type
->max_linear_child
);
634 values
= &node
->u
.data
[1];
635 pointers
= (struct cds_ja_inode_flag
**) align_ptr_size(&values
[type
->max_linear_child
]);
636 /* Check if node value is already populated */
637 for (i
= 0; i
< nr_child
; i
++) {
638 if (values
[i
] == n
) {
648 if (i
== nr_child
&& nr_child
>= type
->max_linear_child
) {
650 return -ERANGE
; /* recompact node */
652 return -ENOSPC
; /* No space left in this node type */
655 assert(pointers
[i
] == NULL
);
656 rcu_assign_pointer(pointers
[i
], child_node_flag
);
657 /* If we expanded the nr_child, increment it */
659 CMM_STORE_SHARED(values
[nr_child
], n
);
660 /* write pointer and value before nr_child */
662 CMM_STORE_SHARED(*nr_child_ptr
, nr_child
+ 1);
664 shadow_node
->nr_child
++;
665 dbg_printf("linear set nth: %u child, shadow: %u child, for node %p shadow %p\n",
666 (unsigned int) CMM_LOAD_SHARED(*nr_child_ptr
),
667 (unsigned int) shadow_node
->nr_child
,
674 int ja_pool_node_set_nth(const struct cds_ja_type
*type
,
675 struct cds_ja_inode
*node
,
676 struct cds_ja_inode_flag
*node_flag
,
677 struct cds_ja_shadow_node
*shadow_node
,
679 struct cds_ja_inode_flag
*child_node_flag
)
681 struct cds_ja_inode
*linear
;
683 assert(type
->type_class
== RCU_JA_POOL
);
685 switch (type
->nr_pool_order
) {
688 unsigned long bitsel
, index
;
690 bitsel
= ja_node_pool_1d_bitsel(node_flag
);
691 assert(bitsel
< CHAR_BIT
);
692 index
= ((unsigned long) n
>> bitsel
) & 0x1;
693 linear
= (struct cds_ja_inode
*) &node
->u
.data
[index
<< type
->pool_size_order
];
698 unsigned long bitsel
[2], index
[2], rindex
;
700 ja_node_pool_2d_bitsel(node_flag
, bitsel
);
701 assert(bitsel
[0] < CHAR_BIT
);
702 assert(bitsel
[1] < CHAR_BIT
);
703 index
[0] = ((unsigned long) n
>> bitsel
[0]) & 0x1;
705 index
[1] = ((unsigned long) n
>> bitsel
[1]) & 0x1;
706 rindex
= index
[0] | index
[1];
707 linear
= (struct cds_ja_inode
*) &node
->u
.data
[rindex
<< type
->pool_size_order
];
715 return ja_linear_node_set_nth(type
, linear
, shadow_node
,
720 int ja_pigeon_node_set_nth(const struct cds_ja_type
*type
,
721 struct cds_ja_inode
*node
,
722 struct cds_ja_shadow_node
*shadow_node
,
724 struct cds_ja_inode_flag
*child_node_flag
)
726 struct cds_ja_inode_flag
**ptr
;
728 assert(type
->type_class
== RCU_JA_PIGEON
);
729 ptr
= &((struct cds_ja_inode_flag
**) node
->u
.data
)[n
];
732 rcu_assign_pointer(*ptr
, child_node_flag
);
733 shadow_node
->nr_child
++;
738 * _ja_node_set_nth: set nth item within a node. Return an error
739 * (negative error value) if it is already there.
742 int _ja_node_set_nth(const struct cds_ja_type
*type
,
743 struct cds_ja_inode
*node
,
744 struct cds_ja_inode_flag
*node_flag
,
745 struct cds_ja_shadow_node
*shadow_node
,
747 struct cds_ja_inode_flag
*child_node_flag
)
749 switch (type
->type_class
) {
751 return ja_linear_node_set_nth(type
, node
, shadow_node
, n
,
754 return ja_pool_node_set_nth(type
, node
, node_flag
, shadow_node
, n
,
757 return ja_pigeon_node_set_nth(type
, node
, shadow_node
, n
,
770 int ja_linear_node_clear_ptr(const struct cds_ja_type
*type
,
771 struct cds_ja_inode
*node
,
772 struct cds_ja_shadow_node
*shadow_node
,
773 struct cds_ja_inode_flag
**node_flag_ptr
)
776 uint8_t *nr_child_ptr
;
778 assert(type
->type_class
== RCU_JA_LINEAR
|| type
->type_class
== RCU_JA_POOL
);
780 nr_child_ptr
= &node
->u
.data
[0];
781 nr_child
= *nr_child_ptr
;
782 assert(nr_child
<= type
->max_linear_child
);
784 if (type
->type_class
== RCU_JA_LINEAR
) {
785 assert(!shadow_node
->fallback_removal_count
);
786 if (shadow_node
->nr_child
<= type
->min_child
) {
787 /* We need to try recompacting the node */
791 dbg_printf("linear clear ptr: nr_child_ptr %p\n", nr_child_ptr
);
792 assert(*node_flag_ptr
!= NULL
);
793 rcu_assign_pointer(*node_flag_ptr
, NULL
);
795 * Value and nr_child are never changed (would cause ABA issue).
796 * Instead, we leave the pointer to NULL and recompact the node
797 * once in a while. It is allowed to set a NULL pointer to a new
798 * value without recompaction though.
799 * Only update the shadow node accounting.
801 shadow_node
->nr_child
--;
802 dbg_printf("linear clear ptr: %u child, shadow: %u child, for node %p shadow %p\n",
803 (unsigned int) CMM_LOAD_SHARED(*nr_child_ptr
),
804 (unsigned int) shadow_node
->nr_child
,
810 int ja_pool_node_clear_ptr(const struct cds_ja_type
*type
,
811 struct cds_ja_inode
*node
,
812 struct cds_ja_inode_flag
*node_flag
,
813 struct cds_ja_shadow_node
*shadow_node
,
814 struct cds_ja_inode_flag
**node_flag_ptr
,
817 struct cds_ja_inode
*linear
;
819 assert(type
->type_class
== RCU_JA_POOL
);
821 if (shadow_node
->fallback_removal_count
) {
822 shadow_node
->fallback_removal_count
--;
824 /* We should try recompacting the node */
825 if (shadow_node
->nr_child
<= type
->min_child
)
829 switch (type
->nr_pool_order
) {
832 unsigned long bitsel
, index
;
834 bitsel
= ja_node_pool_1d_bitsel(node_flag
);
835 assert(bitsel
< CHAR_BIT
);
836 index
= ((unsigned long) n
>> bitsel
) & type
->nr_pool_order
;
837 linear
= (struct cds_ja_inode
*) &node
->u
.data
[index
<< type
->pool_size_order
];
842 unsigned long bitsel
[2], index
[2], rindex
;
844 ja_node_pool_2d_bitsel(node_flag
, bitsel
);
845 assert(bitsel
[0] < CHAR_BIT
);
846 assert(bitsel
[1] < CHAR_BIT
);
847 index
[0] = ((unsigned long) n
>> bitsel
[0]) & 0x1;
849 index
[1] = ((unsigned long) n
>> bitsel
[1]) & 0x1;
850 rindex
= index
[0] | index
[1];
851 linear
= (struct cds_ja_inode
*) &node
->u
.data
[rindex
<< type
->pool_size_order
];
859 return ja_linear_node_clear_ptr(type
, linear
, shadow_node
, node_flag_ptr
);
863 int ja_pigeon_node_clear_ptr(const struct cds_ja_type
*type
,
864 struct cds_ja_inode
*node
,
865 struct cds_ja_shadow_node
*shadow_node
,
866 struct cds_ja_inode_flag
**node_flag_ptr
)
868 assert(type
->type_class
== RCU_JA_PIGEON
);
870 if (shadow_node
->fallback_removal_count
) {
871 shadow_node
->fallback_removal_count
--;
873 /* We should try recompacting the node */
874 if (shadow_node
->nr_child
<= type
->min_child
)
877 dbg_printf("ja_pigeon_node_clear_ptr: clearing ptr: %p\n", *node_flag_ptr
);
878 rcu_assign_pointer(*node_flag_ptr
, NULL
);
879 shadow_node
->nr_child
--;
884 * _ja_node_clear_ptr: clear ptr item within a node. Return an error
885 * (negative error value) if it is not found (-ENOENT).
888 int _ja_node_clear_ptr(const struct cds_ja_type
*type
,
889 struct cds_ja_inode
*node
,
890 struct cds_ja_inode_flag
*node_flag
,
891 struct cds_ja_shadow_node
*shadow_node
,
892 struct cds_ja_inode_flag
**node_flag_ptr
,
895 switch (type
->type_class
) {
897 return ja_linear_node_clear_ptr(type
, node
, shadow_node
, node_flag_ptr
);
899 return ja_pool_node_clear_ptr(type
, node
, node_flag
, shadow_node
, node_flag_ptr
, n
);
901 return ja_pigeon_node_clear_ptr(type
, node
, shadow_node
, node_flag_ptr
);
913 * Calculate bit distribution. Returns the bit (0 to 7) that splits the
914 * distribution in two sub-distributions containing as much elements one
915 * compared to the other.
918 unsigned int ja_node_sum_distribution_1d(enum ja_recompact mode
,
920 unsigned int type_index
,
921 const struct cds_ja_type
*type
,
922 struct cds_ja_inode
*node
,
923 struct cds_ja_shadow_node
*shadow_node
,
925 struct cds_ja_inode_flag
*child_node_flag
,
926 struct cds_ja_inode_flag
**nullify_node_flag_ptr
)
928 uint8_t nr_one
[JA_BITS_PER_BYTE
];
929 unsigned int bitsel
= 0, bit_i
, overall_best_distance
= UINT_MAX
;
930 unsigned int distrib_nr_child
= 0;
932 memset(nr_one
, 0, sizeof(nr_one
));
934 switch (type
->type_class
) {
938 ja_linear_node_get_nr_child(type
, node
);
941 for (i
= 0; i
< nr_child
; i
++) {
942 struct cds_ja_inode_flag
*iter
;
945 ja_linear_node_get_ith_pos(type
, node
, i
, &v
, &iter
);
948 if (mode
== JA_RECOMPACT_DEL
&& *nullify_node_flag_ptr
== iter
)
950 for (bit_i
= 0; bit_i
< JA_BITS_PER_BYTE
; bit_i
++) {
951 if (v
& (1U << bit_i
))
960 unsigned int pool_nr
;
962 for (pool_nr
= 0; pool_nr
< (1U << type
->nr_pool_order
); pool_nr
++) {
963 struct cds_ja_inode
*pool
=
964 ja_pool_node_get_ith_pool(type
,
967 ja_linear_node_get_nr_child(type
, pool
);
970 for (j
= 0; j
< nr_child
; j
++) {
971 struct cds_ja_inode_flag
*iter
;
974 ja_linear_node_get_ith_pos(type
, pool
,
978 if (mode
== JA_RECOMPACT_DEL
&& *nullify_node_flag_ptr
== iter
)
980 for (bit_i
= 0; bit_i
< JA_BITS_PER_BYTE
; bit_i
++) {
981 if (v
& (1U << bit_i
))
993 assert(mode
== JA_RECOMPACT_DEL
);
994 for (i
= 0; i
< JA_ENTRY_PER_NODE
; i
++) {
995 struct cds_ja_inode_flag
*iter
;
997 iter
= ja_pigeon_node_get_ith_pos(type
, node
, i
);
1000 if (mode
== JA_RECOMPACT_DEL
&& *nullify_node_flag_ptr
== iter
)
1002 for (bit_i
= 0; bit_i
< JA_BITS_PER_BYTE
; bit_i
++) {
1003 if (i
& (1U << bit_i
))
1011 assert(mode
== JA_RECOMPACT_ADD_NEXT
);
1018 if (mode
== JA_RECOMPACT_ADD_NEXT
|| mode
== JA_RECOMPACT_ADD_SAME
) {
1019 for (bit_i
= 0; bit_i
< JA_BITS_PER_BYTE
; bit_i
++) {
1020 if (n
& (1U << bit_i
))
1027 * The best bit selector is that for which the number of ones is
1028 * closest to half of the number of children in the
1029 * distribution. We calculate the distance using the double of
1030 * the sub-distribution sizes to eliminate truncation error.
1032 for (bit_i
= 0; bit_i
< JA_BITS_PER_BYTE
; bit_i
++) {
1033 unsigned int distance_to_best
;
1035 distance_to_best
= abs_int(((unsigned int) nr_one
[bit_i
] << 1U) - distrib_nr_child
);
1036 if (distance_to_best
< overall_best_distance
) {
1037 overall_best_distance
= distance_to_best
;
1041 dbg_printf("1 dimension pool bit selection: (%u)\n", bitsel
);
1046 * Calculate bit distribution in two dimensions. Returns the two bits
1047 * (each 0 to 7) that splits the distribution in four sub-distributions
1048 * containing as much elements one compared to the other.
1051 void ja_node_sum_distribution_2d(enum ja_recompact mode
,
1053 unsigned int type_index
,
1054 const struct cds_ja_type
*type
,
1055 struct cds_ja_inode
*node
,
1056 struct cds_ja_shadow_node
*shadow_node
,
1058 struct cds_ja_inode_flag
*child_node_flag
,
1059 struct cds_ja_inode_flag
**nullify_node_flag_ptr
,
1060 unsigned int *_bitsel
)
1062 uint8_t nr_2d_11
[JA_BITS_PER_BYTE
][JA_BITS_PER_BYTE
],
1063 nr_2d_10
[JA_BITS_PER_BYTE
][JA_BITS_PER_BYTE
],
1064 nr_2d_01
[JA_BITS_PER_BYTE
][JA_BITS_PER_BYTE
],
1065 nr_2d_00
[JA_BITS_PER_BYTE
][JA_BITS_PER_BYTE
];
1066 unsigned int bitsel
[2] = { 0, 1 };
1067 unsigned int bit_i
, bit_j
;
1068 int overall_best_distance
= INT_MAX
;
1069 unsigned int distrib_nr_child
= 0;
1071 memset(nr_2d_11
, 0, sizeof(nr_2d_11
));
1072 memset(nr_2d_10
, 0, sizeof(nr_2d_10
));
1073 memset(nr_2d_01
, 0, sizeof(nr_2d_01
));
1074 memset(nr_2d_00
, 0, sizeof(nr_2d_00
));
1076 switch (type
->type_class
) {
1080 ja_linear_node_get_nr_child(type
, node
);
1083 for (i
= 0; i
< nr_child
; i
++) {
1084 struct cds_ja_inode_flag
*iter
;
1087 ja_linear_node_get_ith_pos(type
, node
, i
, &v
, &iter
);
1090 if (mode
== JA_RECOMPACT_DEL
&& *nullify_node_flag_ptr
== iter
)
1092 for (bit_i
= 0; bit_i
< JA_BITS_PER_BYTE
; bit_i
++) {
1093 for (bit_j
= 0; bit_j
< bit_i
; bit_j
++) {
1094 if (v
& (1U << bit_i
)) {
1095 if (v
& (1U << bit_j
)) {
1096 nr_2d_11
[bit_i
][bit_j
]++;
1098 nr_2d_10
[bit_i
][bit_j
]++;
1101 if (v
& (1U << bit_j
)) {
1102 nr_2d_01
[bit_i
][bit_j
]++;
1104 nr_2d_00
[bit_i
][bit_j
]++;
1115 unsigned int pool_nr
;
1117 for (pool_nr
= 0; pool_nr
< (1U << type
->nr_pool_order
); pool_nr
++) {
1118 struct cds_ja_inode
*pool
=
1119 ja_pool_node_get_ith_pool(type
,
1122 ja_linear_node_get_nr_child(type
, pool
);
1125 for (j
= 0; j
< nr_child
; j
++) {
1126 struct cds_ja_inode_flag
*iter
;
1129 ja_linear_node_get_ith_pos(type
, pool
,
1133 if (mode
== JA_RECOMPACT_DEL
&& *nullify_node_flag_ptr
== iter
)
1135 for (bit_i
= 0; bit_i
< JA_BITS_PER_BYTE
; bit_i
++) {
1136 for (bit_j
= 0; bit_j
< bit_i
; bit_j
++) {
1137 if (v
& (1U << bit_i
)) {
1138 if (v
& (1U << bit_j
)) {
1139 nr_2d_11
[bit_i
][bit_j
]++;
1141 nr_2d_10
[bit_i
][bit_j
]++;
1144 if (v
& (1U << bit_j
)) {
1145 nr_2d_01
[bit_i
][bit_j
]++;
1147 nr_2d_00
[bit_i
][bit_j
]++;
1161 assert(mode
== JA_RECOMPACT_DEL
);
1162 for (i
= 0; i
< JA_ENTRY_PER_NODE
; i
++) {
1163 struct cds_ja_inode_flag
*iter
;
1165 iter
= ja_pigeon_node_get_ith_pos(type
, node
, i
);
1168 if (mode
== JA_RECOMPACT_DEL
&& *nullify_node_flag_ptr
== iter
)
1170 for (bit_i
= 0; bit_i
< JA_BITS_PER_BYTE
; bit_i
++) {
1171 for (bit_j
= 0; bit_j
< bit_i
; bit_j
++) {
1172 if (i
& (1U << bit_i
)) {
1173 if (i
& (1U << bit_j
)) {
1174 nr_2d_11
[bit_i
][bit_j
]++;
1176 nr_2d_10
[bit_i
][bit_j
]++;
1179 if (i
& (1U << bit_j
)) {
1180 nr_2d_01
[bit_i
][bit_j
]++;
1182 nr_2d_00
[bit_i
][bit_j
]++;
1192 assert(mode
== JA_RECOMPACT_ADD_NEXT
);
1199 if (mode
== JA_RECOMPACT_ADD_NEXT
|| mode
== JA_RECOMPACT_ADD_SAME
) {
1200 for (bit_i
= 0; bit_i
< JA_BITS_PER_BYTE
; bit_i
++) {
1201 for (bit_j
= 0; bit_j
< bit_i
; bit_j
++) {
1202 if (n
& (1U << bit_i
)) {
1203 if (n
& (1U << bit_j
)) {
1204 nr_2d_11
[bit_i
][bit_j
]++;
1206 nr_2d_10
[bit_i
][bit_j
]++;
1209 if (n
& (1U << bit_j
)) {
1210 nr_2d_01
[bit_i
][bit_j
]++;
1212 nr_2d_00
[bit_i
][bit_j
]++;
1221 * The best bit selector is that for which the number of nodes
1222 * in each sub-class is closest to one-fourth of the number of
1223 * children in the distribution. We calculate the distance using
1224 * 4 times the size of the sub-distribution to eliminate
1227 for (bit_i
= 0; bit_i
< JA_BITS_PER_BYTE
; bit_i
++) {
1228 for (bit_j
= 0; bit_j
< bit_i
; bit_j
++) {
1229 int distance_to_best
[4];
1231 distance_to_best
[0] = ((unsigned int) nr_2d_11
[bit_i
][bit_j
] << 2U) - distrib_nr_child
;
1232 distance_to_best
[1] = ((unsigned int) nr_2d_10
[bit_i
][bit_j
] << 2U) - distrib_nr_child
;
1233 distance_to_best
[2] = ((unsigned int) nr_2d_01
[bit_i
][bit_j
] << 2U) - distrib_nr_child
;
1234 distance_to_best
[3] = ((unsigned int) nr_2d_00
[bit_i
][bit_j
] << 2U) - distrib_nr_child
;
1236 /* Consider worse distance above best */
1237 if (distance_to_best
[1] > 0 && distance_to_best
[1] > distance_to_best
[0])
1238 distance_to_best
[0] = distance_to_best
[1];
1239 if (distance_to_best
[2] > 0 && distance_to_best
[2] > distance_to_best
[0])
1240 distance_to_best
[0] = distance_to_best
[2];
1241 if (distance_to_best
[3] > 0 && distance_to_best
[3] > distance_to_best
[0])
1242 distance_to_best
[0] = distance_to_best
[3];
1245 * If our worse distance is better than overall,
1246 * we become new best candidate.
1248 if (distance_to_best
[0] < overall_best_distance
) {
1249 overall_best_distance
= distance_to_best
[0];
1256 dbg_printf("2 dimensions pool bit selection: (%u,%u)\n", bitsel
[0], bitsel
[1]);
1258 /* Return our bit selection */
1259 _bitsel
[0] = bitsel
[0];
1260 _bitsel
[1] = bitsel
[1];
1264 unsigned int find_nearest_type_index(unsigned int type_index
,
1265 unsigned int nr_nodes
)
1267 const struct cds_ja_type
*type
;
1269 assert(type_index
!= NODE_INDEX_NULL
);
1271 return NODE_INDEX_NULL
;
1273 type
= &ja_types
[type_index
];
1274 if (nr_nodes
< type
->min_child
)
1276 else if (nr_nodes
> type
->max_child
)
1285 * ja_node_recompact_add: recompact a node, adding a new child.
1286 * Return 0 on success, -EAGAIN if need to retry, or other negative
1287 * error value otherwise.
1290 int ja_node_recompact(enum ja_recompact mode
,
1292 unsigned int old_type_index
,
1293 const struct cds_ja_type
*old_type
,
1294 struct cds_ja_inode
*old_node
,
1295 struct cds_ja_shadow_node
*shadow_node
,
1296 struct cds_ja_inode_flag
**old_node_flag_ptr
, uint8_t n
,
1297 struct cds_ja_inode_flag
*child_node_flag
,
1298 struct cds_ja_inode_flag
**nullify_node_flag_ptr
,
1301 unsigned int new_type_index
;
1302 struct cds_ja_inode
*new_node
;
1303 struct cds_ja_shadow_node
*new_shadow_node
= NULL
;
1304 const struct cds_ja_type
*new_type
;
1305 struct cds_ja_inode_flag
*new_node_flag
, *old_node_flag
;
1309 old_node_flag
= *old_node_flag_ptr
;
1312 * Need to find nearest type index even for ADD_SAME, because
1313 * this recompaction, when applied to linear nodes, will garbage
1314 * collect dummy (NULL) entries, and can therefore cause a few
1315 * linear representations to be skipped.
1318 case JA_RECOMPACT_ADD_SAME
:
1319 new_type_index
= find_nearest_type_index(old_type_index
,
1320 shadow_node
->nr_child
+ 1);
1321 dbg_printf("Recompact for node with %u children\n",
1322 shadow_node
->nr_child
+ 1);
1324 case JA_RECOMPACT_ADD_NEXT
:
1325 if (!shadow_node
|| old_type_index
== NODE_INDEX_NULL
) {
1327 dbg_printf("Recompact for NULL\n");
1329 new_type_index
= find_nearest_type_index(old_type_index
,
1330 shadow_node
->nr_child
+ 1);
1331 dbg_printf("Recompact for node with %u children\n",
1332 shadow_node
->nr_child
+ 1);
1335 case JA_RECOMPACT_DEL
:
1336 new_type_index
= find_nearest_type_index(old_type_index
,
1337 shadow_node
->nr_child
- 1);
1338 dbg_printf("Recompact for node with %u children\n",
1339 shadow_node
->nr_child
- 1);
1345 retry
: /* for fallback */
1346 dbg_printf("Recompact from type %d to type %d\n",
1347 old_type_index
, new_type_index
);
1348 new_type
= &ja_types
[new_type_index
];
1349 if (new_type_index
!= NODE_INDEX_NULL
) {
1350 new_node
= alloc_cds_ja_node(ja
, new_type
);
1354 if (new_type
->type_class
== RCU_JA_POOL
) {
1355 switch (new_type
->nr_pool_order
) {
1358 unsigned int node_distrib_bitsel
;
1360 node_distrib_bitsel
=
1361 ja_node_sum_distribution_1d(mode
, ja
,
1362 old_type_index
, old_type
,
1363 old_node
, shadow_node
,
1365 nullify_node_flag_ptr
);
1366 assert(!((unsigned long) new_node
& JA_POOL_1D_MASK
));
1367 new_node_flag
= ja_node_flag_pool_1d(new_node
,
1368 new_type_index
, node_distrib_bitsel
);
1373 unsigned int node_distrib_bitsel
[2];
1375 ja_node_sum_distribution_2d(mode
, ja
,
1376 old_type_index
, old_type
,
1377 old_node
, shadow_node
,
1379 nullify_node_flag_ptr
,
1380 node_distrib_bitsel
);
1381 assert(!((unsigned long) new_node
& JA_POOL_1D_MASK
));
1382 assert(!((unsigned long) new_node
& JA_POOL_2D_MASK
));
1383 new_node_flag
= ja_node_flag_pool_2d(new_node
,
1384 new_type_index
, node_distrib_bitsel
);
1391 new_node_flag
= ja_node_flag(new_node
, new_type_index
);
1394 dbg_printf("Recompact inherit lock from %p\n", shadow_node
);
1395 new_shadow_node
= rcuja_shadow_set(ja
->ht
, new_node_flag
, shadow_node
, ja
, level
);
1396 if (!new_shadow_node
) {
1397 free_cds_ja_node(ja
, new_node
);
1401 new_shadow_node
->fallback_removal_count
=
1402 JA_FALLBACK_REMOVAL_COUNT
;
1405 new_node_flag
= NULL
;
1408 assert(mode
!= JA_RECOMPACT_ADD_NEXT
|| old_type
->type_class
!= RCU_JA_PIGEON
);
1410 if (new_type_index
== NODE_INDEX_NULL
)
1413 switch (old_type
->type_class
) {
1417 ja_linear_node_get_nr_child(old_type
, old_node
);
1420 for (i
= 0; i
< nr_child
; i
++) {
1421 struct cds_ja_inode_flag
*iter
;
1424 ja_linear_node_get_ith_pos(old_type
, old_node
, i
, &v
, &iter
);
1427 if (mode
== JA_RECOMPACT_DEL
&& *nullify_node_flag_ptr
== iter
)
1429 ret
= _ja_node_set_nth(new_type
, new_node
, new_node_flag
,
1432 if (new_type
->type_class
== RCU_JA_POOL
&& ret
) {
1433 goto fallback_toosmall
;
1441 unsigned int pool_nr
;
1443 for (pool_nr
= 0; pool_nr
< (1U << old_type
->nr_pool_order
); pool_nr
++) {
1444 struct cds_ja_inode
*pool
=
1445 ja_pool_node_get_ith_pool(old_type
,
1448 ja_linear_node_get_nr_child(old_type
, pool
);
1451 for (j
= 0; j
< nr_child
; j
++) {
1452 struct cds_ja_inode_flag
*iter
;
1455 ja_linear_node_get_ith_pos(old_type
, pool
,
1459 if (mode
== JA_RECOMPACT_DEL
&& *nullify_node_flag_ptr
== iter
)
1461 ret
= _ja_node_set_nth(new_type
, new_node
, new_node_flag
,
1464 if (new_type
->type_class
== RCU_JA_POOL
1466 goto fallback_toosmall
;
1474 assert(mode
== JA_RECOMPACT_ADD_NEXT
);
1480 assert(mode
== JA_RECOMPACT_DEL
);
1481 for (i
= 0; i
< JA_ENTRY_PER_NODE
; i
++) {
1482 struct cds_ja_inode_flag
*iter
;
1484 iter
= ja_pigeon_node_get_ith_pos(old_type
, old_node
, i
);
1487 if (mode
== JA_RECOMPACT_DEL
&& *nullify_node_flag_ptr
== iter
)
1489 ret
= _ja_node_set_nth(new_type
, new_node
, new_node_flag
,
1492 if (new_type
->type_class
== RCU_JA_POOL
&& ret
) {
1493 goto fallback_toosmall
;
1506 if (mode
== JA_RECOMPACT_ADD_NEXT
|| mode
== JA_RECOMPACT_ADD_SAME
) {
1508 ret
= _ja_node_set_nth(new_type
, new_node
, new_node_flag
,
1510 n
, child_node_flag
);
1511 if (new_type
->type_class
== RCU_JA_POOL
&& ret
) {
1512 goto fallback_toosmall
;
1518 dbg_printf("Using fallback for %u children, node type index: %u, mode %s\n",
1519 new_shadow_node
->nr_child
, old_type_index
, mode
== JA_RECOMPACT_ADD_NEXT
? "add_next" :
1520 (mode
== JA_RECOMPACT_DEL
? "del" : "add_same"));
1521 uatomic_inc(&ja
->node_fallback_count_distribution
[new_shadow_node
->nr_child
]);
1524 /* Return pointer to new recompacted node through old_node_flag_ptr */
1525 *old_node_flag_ptr
= new_node_flag
;
1529 flags
= RCUJA_SHADOW_CLEAR_FREE_NODE
;
1531 * It is OK to free the lock associated with a node
1532 * going to NULL, since we are holding the parent lock.
1533 * This synchronizes removal with re-add of that node.
1535 if (new_type_index
== NODE_INDEX_NULL
)
1536 flags
|= RCUJA_SHADOW_CLEAR_FREE_LOCK
;
1537 ret
= rcuja_shadow_clear(ja
->ht
, old_node_flag
, shadow_node
,
1547 /* fallback if next pool is too small */
1548 assert(new_shadow_node
);
1549 ret
= rcuja_shadow_clear(ja
->ht
, new_node_flag
, new_shadow_node
,
1550 RCUJA_SHADOW_CLEAR_FREE_NODE
);
1554 case JA_RECOMPACT_ADD_SAME
:
1556 * JA_RECOMPACT_ADD_SAME is only triggered if a linear
1557 * node within a pool has unused entries. It should
1558 * therefore _never_ be too small.
1563 case JA_RECOMPACT_ADD_NEXT
:
1565 const struct cds_ja_type
*next_type
;
1568 * Recompaction attempt on add failed. Should only
1569 * happen if target node type is pool. Caused by
1570 * hard-to-split distribution. Recompact using the next
1571 * distribution size.
1573 assert(new_type
->type_class
== RCU_JA_POOL
);
1574 next_type
= &ja_types
[new_type_index
+ 1];
1576 * Try going to the next pool size if our population
1577 * fits within its range. This is not flagged as a
1580 if (shadow_node
->nr_child
+ 1 >= next_type
->min_child
1581 && shadow_node
->nr_child
+ 1 <= next_type
->max_child
) {
1586 dbg_printf("Add fallback to type %d\n", new_type_index
);
1587 uatomic_inc(&ja
->nr_fallback
);
1593 case JA_RECOMPACT_DEL
:
1595 * Recompaction attempt on delete failed. Should only
1596 * happen if target node type is pool. This is caused by
1597 * a hard-to-split distribution. Recompact on same node
1598 * size, but flag current node as "fallback" to ensure
1599 * we don't attempt recompaction before some activity
1600 * has reshuffled our node.
1602 assert(new_type
->type_class
== RCU_JA_POOL
);
1603 new_type_index
= old_type_index
;
1604 dbg_printf("Delete fallback keeping type %d\n", new_type_index
);
1605 uatomic_inc(&ja
->nr_fallback
);
1614 * Last resort fallback: pigeon.
1616 new_type_index
= (1UL << JA_TYPE_BITS
) - 1;
1617 dbg_printf("Fallback to type %d\n", new_type_index
);
1618 uatomic_inc(&ja
->nr_fallback
);
1624 * Return 0 on success, -EAGAIN if need to retry, or other negative
1625 * error value otherwise.
1628 int ja_node_set_nth(struct cds_ja
*ja
,
1629 struct cds_ja_inode_flag
**node_flag
, uint8_t n
,
1630 struct cds_ja_inode_flag
*child_node_flag
,
1631 struct cds_ja_shadow_node
*shadow_node
,
1635 unsigned int type_index
;
1636 const struct cds_ja_type
*type
;
1637 struct cds_ja_inode
*node
;
1639 dbg_printf("ja_node_set_nth for n=%u, node %p, shadow %p\n",
1640 (unsigned int) n
, ja_node_ptr(*node_flag
), shadow_node
);
1642 node
= ja_node_ptr(*node_flag
);
1643 type_index
= ja_node_type(*node_flag
);
1644 type
= &ja_types
[type_index
];
1645 ret
= _ja_node_set_nth(type
, node
, *node_flag
, shadow_node
,
1646 n
, child_node_flag
);
1649 /* Not enough space in node, need to recompact to next type. */
1650 ret
= ja_node_recompact(JA_RECOMPACT_ADD_NEXT
, ja
, type_index
, type
, node
,
1651 shadow_node
, node_flag
, n
, child_node_flag
, NULL
, level
);
1654 /* Node needs to be recompacted. */
1655 ret
= ja_node_recompact(JA_RECOMPACT_ADD_SAME
, ja
, type_index
, type
, node
,
1656 shadow_node
, node_flag
, n
, child_node_flag
, NULL
, level
);
1663 * Return 0 on success, -EAGAIN if need to retry, or other negative
1664 * error value otherwise.
1667 int ja_node_clear_ptr(struct cds_ja
*ja
,
1668 struct cds_ja_inode_flag
**node_flag_ptr
, /* Pointer to location to nullify */
1669 struct cds_ja_inode_flag
**parent_node_flag_ptr
, /* Address of parent ptr in its parent */
1670 struct cds_ja_shadow_node
*shadow_node
, /* of parent */
1671 uint8_t n
, int level
)
1674 unsigned int type_index
;
1675 const struct cds_ja_type
*type
;
1676 struct cds_ja_inode
*node
;
1678 dbg_printf("ja_node_clear_ptr for node %p, shadow %p, target ptr %p\n",
1679 ja_node_ptr(*parent_node_flag_ptr
), shadow_node
, node_flag_ptr
);
1681 node
= ja_node_ptr(*parent_node_flag_ptr
);
1682 type_index
= ja_node_type(*parent_node_flag_ptr
);
1683 type
= &ja_types
[type_index
];
1684 ret
= _ja_node_clear_ptr(type
, node
, *parent_node_flag_ptr
, shadow_node
, node_flag_ptr
, n
);
1685 if (ret
== -EFBIG
) {
1686 /* Should try recompaction. */
1687 ret
= ja_node_recompact(JA_RECOMPACT_DEL
, ja
, type_index
, type
, node
,
1688 shadow_node
, parent_node_flag_ptr
, n
, NULL
,
1689 node_flag_ptr
, level
);
1694 struct cds_ja_node
*cds_ja_lookup(struct cds_ja
*ja
, uint64_t key
)
1696 unsigned int tree_depth
, i
;
1697 struct cds_ja_inode_flag
*node_flag
;
1699 if (caa_unlikely(key
> ja
->key_max
))
1701 tree_depth
= ja
->tree_depth
;
1702 node_flag
= rcu_dereference(ja
->root
);
1704 /* level 0: root node */
1705 if (!ja_node_ptr(node_flag
))
1708 for (i
= 1; i
< tree_depth
; i
++) {
1711 iter_key
= (uint8_t) (key
>> (JA_BITS_PER_BYTE
* (tree_depth
- i
- 1)));
1712 node_flag
= ja_node_get_nth(node_flag
, NULL
, iter_key
);
1713 dbg_printf("cds_ja_lookup iter key lookup %u finds node_flag %p\n",
1714 (unsigned int) iter_key
, node_flag
);
1715 if (!ja_node_ptr(node_flag
))
1719 /* Last level lookup succeded. We got an actual match. */
1720 return (struct cds_ja_node
*) node_flag
;
1723 struct cds_ja_node
*cds_ja_lookup_lower_equal(struct cds_ja
*ja
, uint64_t key
)
1725 int tree_depth
, level
;
1726 struct cds_ja_inode_flag
*node_flag
, *cur_node_depth
[JA_MAX_DEPTH
];
1728 if (caa_unlikely(key
> ja
->key_max
|| !key
))
1731 memset(cur_node_depth
, 0, sizeof(cur_node_depth
));
1732 tree_depth
= ja
->tree_depth
;
1733 node_flag
= rcu_dereference(ja
->root
);
1734 cur_node_depth
[0] = node_flag
;
1736 /* level 0: root node */
1737 if (!ja_node_ptr(node_flag
))
1740 for (level
= 1; level
< tree_depth
; level
++) {
1743 iter_key
= (uint8_t) (key
>> (JA_BITS_PER_BYTE
* (tree_depth
- level
- 1)));
1744 node_flag
= ja_node_get_nth(node_flag
, NULL
, iter_key
);
1745 if (!ja_node_ptr(node_flag
))
1747 cur_node_depth
[level
] = node_flag
;
1748 dbg_printf("cds_ja_lookup iter key lookup %u finds node_flag %p\n",
1749 (unsigned int) iter_key
, node_flag
);
1752 if (level
== tree_depth
) {
1753 /* Last level lookup succeded. We got an equal match. */
1754 return (struct cds_ja_node
*) node_flag
;
1758 * Find highest value left of current node.
1759 * Current node is cur_node_depth[level].
1760 * Start at current level. If we cannot find any key left of
1761 * ours, go one level up, seek highest value left of current
1762 * (recursively), and when we find one, get the rightmost child
1763 * of its rightmost child (recursively).
1765 for (; level
> 0; level
--) {
1768 iter_key
= (uint8_t) (key
>> (JA_BITS_PER_BYTE
* (tree_depth
- level
- 1)));
1769 node_flag
= ja_node_get_left(cur_node_depth
[level
- 1],
1771 /* If found left sibling, find rightmost child. */
1772 if (ja_node_ptr(node_flag
))
1777 /* Reached the root and could not find a left sibling. */
1784 * From this point, we are guaranteed to be able to find a
1785 * "lower than" match. ja_attach_node() and ja_detach_node()
1786 * both guarantee that it is not possible for a lookup to reach
1790 /* Find rightmost child of rightmost child (recursively). */
1791 for (; level
< tree_depth
; level
++) {
1792 node_flag
= ja_node_get_rightmost(node_flag
);
1793 /* If found left sibling, find rightmost child. */
1794 if (!ja_node_ptr(node_flag
))
1798 assert(level
== tree_depth
);
1800 return (struct cds_ja_node
*) node_flag
;
1804 * We reached an unpopulated node. Create it and the children we need,
1805 * and then attach the entire branch to the current node. This may
1806 * trigger recompaction of the current node. Locks needed: node lock
1807 * (for add), and, possibly, parent node lock (to update pointer due to
1808 * node recompaction).
1810 * First take node lock, check if recompaction is needed, then take
1811 * parent lock (if needed). Then we can proceed to create the new
1812 * branch. Publish the new branch, and release locks.
1813 * TODO: we currently always take the parent lock even when not needed.
1815 * ja_attach_node() ensures that a lookup will _never_ see a branch that
1816 * leads to a dead-end: before attaching a branch, the entire content of
1817 * the new branch is populated, thus creating a cluster, before
1818 * attaching the cluster to the rest of the tree, thus making it visible
1822 int ja_attach_node(struct cds_ja
*ja
,
1823 struct cds_ja_inode_flag
**attach_node_flag_ptr
,
1824 struct cds_ja_inode_flag
*attach_node_flag
,
1825 struct cds_ja_inode_flag
*parent_attach_node_flag
,
1826 struct cds_ja_inode_flag
**old_node_flag_ptr
,
1827 struct cds_ja_inode_flag
*old_node_flag
,
1830 struct cds_ja_node
*child_node
)
1832 struct cds_ja_shadow_node
*shadow_node
= NULL
,
1833 *parent_shadow_node
= NULL
;
1834 struct cds_ja_inode_flag
*iter_node_flag
, *iter_dest_node_flag
;
1836 struct cds_ja_inode_flag
*created_nodes
[JA_MAX_DEPTH
];
1837 int nr_created_nodes
= 0;
1839 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",
1840 level
, old_node_flag
, attach_node_flag_ptr
, attach_node_flag
, parent_attach_node_flag
);
1842 assert(!old_node_flag
);
1843 if (attach_node_flag
) {
1844 shadow_node
= rcuja_shadow_lookup_lock(ja
->ht
, attach_node_flag
);
1850 if (parent_attach_node_flag
) {
1851 parent_shadow_node
= rcuja_shadow_lookup_lock(ja
->ht
,
1852 parent_attach_node_flag
);
1853 if (!parent_shadow_node
) {
1859 if (old_node_flag_ptr
&& ja_node_ptr(*old_node_flag_ptr
)) {
1861 * Target node has been updated between RCU lookup and
1862 * lock acquisition. We need to re-try lookup and
1870 * Perform a lookup query to handle the case where
1871 * old_node_flag_ptr is NULL. We cannot use it to check if the
1872 * node has been populated between RCU lookup and mutex
1875 if (!old_node_flag_ptr
) {
1877 struct cds_ja_inode_flag
*lookup_node_flag
;
1878 struct cds_ja_inode_flag
**lookup_node_flag_ptr
;
1880 iter_key
= (uint8_t) (key
>> (JA_BITS_PER_BYTE
* (ja
->tree_depth
- level
)));
1881 lookup_node_flag
= ja_node_get_nth(attach_node_flag
,
1882 &lookup_node_flag_ptr
,
1884 if (lookup_node_flag
) {
1890 if (attach_node_flag_ptr
&& ja_node_ptr(*attach_node_flag_ptr
) !=
1891 ja_node_ptr(attach_node_flag
)) {
1893 * Target node has been updated between RCU lookup and
1894 * lock acquisition. We need to re-try lookup and
1901 /* Create new branch, starting from bottom */
1902 iter_node_flag
= (struct cds_ja_inode_flag
*) child_node
;
1904 for (i
= ja
->tree_depth
- 1; i
>= (int) level
; i
--) {
1907 iter_key
= (uint8_t) (key
>> (JA_BITS_PER_BYTE
* (ja
->tree_depth
- i
- 1)));
1908 dbg_printf("branch creation level %d, key %u\n",
1909 i
, (unsigned int) iter_key
);
1910 iter_dest_node_flag
= NULL
;
1911 ret
= ja_node_set_nth(ja
, &iter_dest_node_flag
,
1916 dbg_printf("branch creation error %d\n", ret
);
1919 created_nodes
[nr_created_nodes
++] = iter_dest_node_flag
;
1920 iter_node_flag
= iter_dest_node_flag
;
1924 /* Publish branch */
1927 * Attaching to root node.
1929 rcu_assign_pointer(ja
->root
, iter_node_flag
);
1933 iter_key
= (uint8_t) (key
>> (JA_BITS_PER_BYTE
* (ja
->tree_depth
- level
)));
1934 dbg_printf("publish branch at level %d, key %u\n",
1935 level
- 1, (unsigned int) iter_key
);
1936 /* We need to use set_nth on the previous level. */
1937 iter_dest_node_flag
= attach_node_flag
;
1938 ret
= ja_node_set_nth(ja
, &iter_dest_node_flag
,
1941 shadow_node
, level
- 1);
1943 dbg_printf("branch publish error %d\n", ret
);
1949 rcu_assign_pointer(*attach_node_flag_ptr
, iter_dest_node_flag
);
1957 for (i
= 0; i
< nr_created_nodes
; i
++) {
1961 flags
= RCUJA_SHADOW_CLEAR_FREE_LOCK
;
1963 flags
|= RCUJA_SHADOW_CLEAR_FREE_NODE
;
1964 tmpret
= rcuja_shadow_clear(ja
->ht
,
1972 if (parent_shadow_node
)
1973 rcuja_shadow_unlock(parent_shadow_node
);
1976 rcuja_shadow_unlock(shadow_node
);
1982 * Lock the parent containing the pointer to list of duplicates, and add
1983 * node to this list. Failure can happen if concurrent update changes
1984 * the parent before we get the lock. We return -EAGAIN in that case.
1985 * Return 0 on success, negative error value on failure.
1988 int ja_chain_node(struct cds_ja
*ja
,
1989 struct cds_ja_inode_flag
*parent_node_flag
,
1990 struct cds_ja_inode_flag
**node_flag_ptr
,
1991 struct cds_ja_inode_flag
*node_flag
,
1992 struct cds_ja_node
*node
)
1994 struct cds_ja_shadow_node
*shadow_node
;
1997 shadow_node
= rcuja_shadow_lookup_lock(ja
->ht
, parent_node_flag
);
2001 if (ja_node_ptr(*node_flag_ptr
) != ja_node_ptr(node_flag
)) {
2006 * Add node to head of list. Safe against concurrent RCU read
2009 node
->next
= (struct cds_ja_node
*) node_flag
;
2010 rcu_assign_pointer(*node_flag_ptr
, (struct cds_ja_inode_flag
*) node
);
2012 rcuja_shadow_unlock(shadow_node
);
2017 int _cds_ja_add(struct cds_ja
*ja
, uint64_t key
,
2018 struct cds_ja_node
*node
,
2019 struct cds_ja_node
**unique_node_ret
)
2021 unsigned int tree_depth
, i
;
2022 struct cds_ja_inode_flag
*attach_node_flag
,
2026 *parent_attach_node_flag
;
2027 struct cds_ja_inode_flag
**attach_node_flag_ptr
,
2028 **parent_node_flag_ptr
,
2032 if (caa_unlikely(key
> ja
->key_max
)) {
2035 tree_depth
= ja
->tree_depth
;
2038 dbg_printf("cds_ja_add attempt: key %" PRIu64
", node %p\n",
2040 parent2_node_flag
= NULL
;
2042 (struct cds_ja_inode_flag
*) &ja
->root
; /* Use root ptr address as key for mutex */
2043 parent_node_flag_ptr
= NULL
;
2044 node_flag
= rcu_dereference(ja
->root
);
2045 node_flag_ptr
= &ja
->root
;
2047 /* Iterate on all internal levels */
2048 for (i
= 1; i
< tree_depth
; i
++) {
2051 if (!ja_node_ptr(node_flag
))
2053 dbg_printf("cds_ja_add iter parent2_node_flag %p parent_node_flag %p node_flag_ptr %p node_flag %p\n",
2054 parent2_node_flag
, parent_node_flag
, node_flag_ptr
, node_flag
);
2055 iter_key
= (uint8_t) (key
>> (JA_BITS_PER_BYTE
* (tree_depth
- i
- 1)));
2056 parent2_node_flag
= parent_node_flag
;
2057 parent_node_flag
= node_flag
;
2058 parent_node_flag_ptr
= node_flag_ptr
;
2059 node_flag
= ja_node_get_nth(node_flag
,
2065 * We reached either bottom of tree or internal NULL node,
2066 * simply add node to last internal level, or chain it if key is
2069 if (!ja_node_ptr(node_flag
)) {
2070 dbg_printf("cds_ja_add NULL parent2_node_flag %p parent_node_flag %p node_flag_ptr %p node_flag %p\n",
2071 parent2_node_flag
, parent_node_flag
, node_flag_ptr
, node_flag
);
2073 attach_node_flag
= parent_node_flag
;
2074 attach_node_flag_ptr
= parent_node_flag_ptr
;
2075 parent_attach_node_flag
= parent2_node_flag
;
2077 ret
= ja_attach_node(ja
, attach_node_flag_ptr
,
2079 parent_attach_node_flag
,
2084 if (unique_node_ret
) {
2085 *unique_node_ret
= (struct cds_ja_node
*) ja_node_ptr(node_flag
);
2089 dbg_printf("cds_ja_add duplicate parent2_node_flag %p parent_node_flag %p node_flag_ptr %p node_flag %p\n",
2090 parent2_node_flag
, parent_node_flag
, node_flag_ptr
, node_flag
);
2092 attach_node_flag
= node_flag
;
2093 attach_node_flag_ptr
= node_flag_ptr
;
2094 parent_attach_node_flag
= parent_node_flag
;
2096 ret
= ja_chain_node(ja
,
2097 parent_attach_node_flag
,
2098 attach_node_flag_ptr
,
2102 if (ret
== -EAGAIN
|| ret
== -EEXIST
)
2108 int cds_ja_add(struct cds_ja
*ja
, uint64_t key
,
2109 struct cds_ja_node
*node
)
2111 return _cds_ja_add(ja
, key
, node
, NULL
);
2114 struct cds_ja_node
*cds_ja_add_unique(struct cds_ja
*ja
, uint64_t key
,
2115 struct cds_ja_node
*node
)
2118 struct cds_ja_node
*ret_node
;
2120 ret
= _cds_ja_add(ja
, key
, node
, &ret_node
);
2128 * Note: there is no need to lookup the pointer address associated with
2129 * each node's nth item after taking the lock: it's already been done by
2130 * cds_ja_del while holding the rcu read-side lock, and our node rules
2131 * ensure that when a match value -> pointer is found in a node, it is
2132 * _NEVER_ changed for that node without recompaction, and recompaction
2133 * reallocates the node.
2134 * However, when a child is removed from "linear" nodes, its pointer
2135 * is set to NULL. We therefore check, while holding the locks, if this
2136 * pointer is NULL, and return -ENOENT to the caller if it is the case.
2138 * ja_detach_node() ensures that a lookup will _never_ see a branch that
2139 * leads to a dead-end: when removing branch, it makes sure to perform
2140 * the "cut" at the highest node that has only one child, effectively
2141 * replacing it with a NULL pointer.
2144 int ja_detach_node(struct cds_ja
*ja
,
2145 struct cds_ja_inode_flag
**snapshot
,
2146 struct cds_ja_inode_flag
***snapshot_ptr
,
2147 uint8_t *snapshot_n
,
2150 struct cds_ja_node
*node
)
2152 struct cds_ja_shadow_node
*shadow_nodes
[JA_MAX_DEPTH
];
2153 struct cds_ja_inode_flag
**node_flag_ptr
= NULL
,
2154 *parent_node_flag
= NULL
,
2155 **parent_node_flag_ptr
= NULL
;
2156 struct cds_ja_inode_flag
*iter_node_flag
;
2157 int ret
, i
, nr_shadow
= 0, nr_clear
= 0, nr_branch
= 0;
2160 assert(nr_snapshot
== ja
->tree_depth
+ 1);
2163 * From the last internal level node going up, get the node
2164 * lock, check if the node has only one child left. If it is the
2165 * case, we continue iterating upward. When we reach a node
2166 * which has more that one child left, we lock the parent, and
2167 * proceed to the node deletion (removing its children too).
2169 for (i
= nr_snapshot
- 2; i
>= 1; i
--) {
2170 struct cds_ja_shadow_node
*shadow_node
;
2172 shadow_node
= rcuja_shadow_lookup_lock(ja
->ht
,
2178 shadow_nodes
[nr_shadow
++] = shadow_node
;
2181 * Check if node has been removed between RCU
2182 * lookup and lock acquisition.
2184 assert(snapshot_ptr
[i
+ 1]);
2185 if (ja_node_ptr(*snapshot_ptr
[i
+ 1])
2186 != ja_node_ptr(snapshot
[i
+ 1])) {
2191 assert(shadow_node
->nr_child
> 0);
2192 if (shadow_node
->nr_child
== 1 && i
> 1)
2195 if (shadow_node
->nr_child
> 1 || i
== 1) {
2196 /* Lock parent and break */
2197 shadow_node
= rcuja_shadow_lookup_lock(ja
->ht
,
2203 shadow_nodes
[nr_shadow
++] = shadow_node
;
2206 * Check if node has been removed between RCU
2207 * lookup and lock acquisition.
2209 assert(snapshot_ptr
[i
]);
2210 if (ja_node_ptr(*snapshot_ptr
[i
])
2211 != ja_node_ptr(snapshot
[i
])) {
2216 node_flag_ptr
= snapshot_ptr
[i
+ 1];
2217 n
= snapshot_n
[i
+ 1];
2218 parent_node_flag_ptr
= snapshot_ptr
[i
];
2219 parent_node_flag
= snapshot
[i
];
2223 * Lock parent's parent, in case we need
2224 * to recompact parent.
2226 shadow_node
= rcuja_shadow_lookup_lock(ja
->ht
,
2232 shadow_nodes
[nr_shadow
++] = shadow_node
;
2235 * Check if node has been removed between RCU
2236 * lookup and lock acquisition.
2238 assert(snapshot_ptr
[i
- 1]);
2239 if (ja_node_ptr(*snapshot_ptr
[i
- 1])
2240 != ja_node_ptr(snapshot
[i
- 1])) {
2251 * At this point, we want to delete all nodes that are about to
2252 * be removed from shadow_nodes (except the last one, which is
2253 * either the root or the parent of the upmost node with 1
2254 * child). OK to free lock here, because RCU read lock is held,
2255 * and free only performed in call_rcu.
2258 for (i
= 0; i
< nr_clear
; i
++) {
2259 ret
= rcuja_shadow_clear(ja
->ht
,
2260 shadow_nodes
[i
]->node_flag
,
2262 RCUJA_SHADOW_CLEAR_FREE_NODE
2263 | RCUJA_SHADOW_CLEAR_FREE_LOCK
);
2267 iter_node_flag
= parent_node_flag
;
2268 /* Remove from parent */
2269 ret
= ja_node_clear_ptr(ja
,
2270 node_flag_ptr
, /* Pointer to location to nullify */
2271 &iter_node_flag
, /* Old new parent ptr in its parent */
2272 shadow_nodes
[nr_branch
- 1], /* of parent */
2277 dbg_printf("ja_detach_node: publish %p instead of %p\n",
2278 iter_node_flag
, *parent_node_flag_ptr
);
2279 /* Update address of parent ptr in its parent */
2280 rcu_assign_pointer(*parent_node_flag_ptr
, iter_node_flag
);
2283 for (i
= 0; i
< nr_shadow
; i
++)
2284 rcuja_shadow_unlock(shadow_nodes
[i
]);
2289 int ja_unchain_node(struct cds_ja
*ja
,
2290 struct cds_ja_inode_flag
*parent_node_flag
,
2291 struct cds_ja_inode_flag
**node_flag_ptr
,
2292 struct cds_ja_inode_flag
*node_flag
,
2293 struct cds_ja_node
*node
)
2295 struct cds_ja_shadow_node
*shadow_node
;
2296 struct cds_ja_node
*iter_node
, **iter_node_ptr
, **prev_node_ptr
= NULL
;
2297 int ret
= 0, count
= 0, found
= 0;
2299 shadow_node
= rcuja_shadow_lookup_lock(ja
->ht
, parent_node_flag
);
2302 if (ja_node_ptr(*node_flag_ptr
) != ja_node_ptr(node_flag
)) {
2307 * Find the previous node's next pointer pointing to our node,
2308 * so we can update it. Retry if another thread removed all but
2309 * one of duplicates since check (this check was performed
2310 * without lock). Ensure that the node we are about to remove is
2311 * still in the list (while holding lock). No need for RCU
2312 * traversal here since we hold the lock on the parent.
2314 iter_node_ptr
= (struct cds_ja_node
**) node_flag_ptr
;
2315 iter_node
= (struct cds_ja_node
*) ja_node_ptr(node_flag
);
2316 cds_ja_for_each_duplicate(iter_node
) {
2318 if (iter_node
== node
) {
2319 prev_node_ptr
= iter_node_ptr
;
2322 iter_node_ptr
= &iter_node
->next
;
2325 if (!found
|| count
== 1) {
2329 CMM_STORE_SHARED(*prev_node_ptr
, node
->next
);
2331 * Validate that we indeed removed the node from linked list.
2333 assert(ja_node_ptr(*node_flag_ptr
) != (struct cds_ja_inode
*) node
);
2335 rcuja_shadow_unlock(shadow_node
);
2340 * Called with RCU read lock held.
2342 int cds_ja_del(struct cds_ja
*ja
, uint64_t key
,
2343 struct cds_ja_node
*node
)
2345 unsigned int tree_depth
, i
;
2346 struct cds_ja_inode_flag
*snapshot
[JA_MAX_DEPTH
];
2347 struct cds_ja_inode_flag
**snapshot_ptr
[JA_MAX_DEPTH
];
2348 uint8_t snapshot_n
[JA_MAX_DEPTH
];
2349 struct cds_ja_inode_flag
*node_flag
;
2350 struct cds_ja_inode_flag
**prev_node_flag_ptr
,
2355 if (caa_unlikely(key
> ja
->key_max
))
2357 tree_depth
= ja
->tree_depth
;
2361 dbg_printf("cds_ja_del attempt: key %" PRIu64
", node %p\n",
2364 /* snapshot for level 0 is only for shadow node lookup */
2367 snapshot_ptr
[nr_snapshot
] = NULL
;
2368 snapshot
[nr_snapshot
++] = (struct cds_ja_inode_flag
*) &ja
->root
;
2369 node_flag
= rcu_dereference(ja
->root
);
2370 prev_node_flag_ptr
= &ja
->root
;
2371 node_flag_ptr
= &ja
->root
;
2373 /* Iterate on all internal levels */
2374 for (i
= 1; i
< tree_depth
; i
++) {
2377 dbg_printf("cds_ja_del iter node_flag %p\n",
2379 if (!ja_node_ptr(node_flag
)) {
2382 iter_key
= (uint8_t) (key
>> (JA_BITS_PER_BYTE
* (tree_depth
- i
- 1)));
2383 snapshot_n
[nr_snapshot
+ 1] = iter_key
;
2384 snapshot_ptr
[nr_snapshot
] = prev_node_flag_ptr
;
2385 snapshot
[nr_snapshot
++] = node_flag
;
2386 node_flag
= ja_node_get_nth(node_flag
,
2390 prev_node_flag_ptr
= node_flag_ptr
;
2391 dbg_printf("cds_ja_del iter key lookup %u finds node_flag %p, prev_node_flag_ptr %p\n",
2392 (unsigned int) iter_key
, node_flag
,
2393 prev_node_flag_ptr
);
2396 * We reached bottom of tree, try to find the node we are trying
2397 * to remove. Fail if we cannot find it.
2399 if (!ja_node_ptr(node_flag
)) {
2400 dbg_printf("cds_ja_del: no node found for key %" PRIu64
"\n",
2404 struct cds_ja_node
*iter_node
, *match
= NULL
;
2407 iter_node
= (struct cds_ja_node
*) ja_node_ptr(node_flag
);
2408 cds_ja_for_each_duplicate_rcu(iter_node
) {
2409 dbg_printf("cds_ja_del: compare %p with iter_node %p\n", node
, iter_node
);
2410 if (iter_node
== node
)
2416 dbg_printf("cds_ja_del: no node match for node %p key %" PRIu64
"\n", node
, key
);
2422 * Removing last of duplicates. Last snapshot
2423 * does not have a shadow node (external leafs).
2425 snapshot_ptr
[nr_snapshot
] = prev_node_flag_ptr
;
2426 snapshot
[nr_snapshot
++] = node_flag
;
2427 ret
= ja_detach_node(ja
, snapshot
, snapshot_ptr
,
2428 snapshot_n
, nr_snapshot
, key
, node
);
2430 ret
= ja_unchain_node(ja
, snapshot
[nr_snapshot
- 1],
2431 node_flag_ptr
, node_flag
, match
);
2435 * Explanation of -ENOENT handling: caused by concurrent delete
2436 * between RCU lookup and actual removal. Need to re-do the
2437 * lookup and removal attempt.
2439 if (ret
== -EAGAIN
|| ret
== -ENOENT
)
2444 struct cds_ja
*_cds_ja_new(unsigned int key_bits
,
2445 const struct rcu_flavor_struct
*flavor
)
2449 struct cds_ja_shadow_node
*root_shadow_node
;
2451 ja
= calloc(sizeof(*ja
), 1);
2463 ja
->key_max
= (1ULL << key_bits
) - 1;
2466 ja
->key_max
= UINT64_MAX
;
2472 /* ja->root is NULL */
2473 /* tree_depth 0 is for pointer to root node */
2474 ja
->tree_depth
= (key_bits
>> JA_LOG2_BITS_PER_BYTE
) + 1;
2475 assert(ja
->tree_depth
<= JA_MAX_DEPTH
);
2476 ja
->ht
= rcuja_create_ht(flavor
);
2481 * Note: we should not free this node until judy array destroy.
2483 root_shadow_node
= rcuja_shadow_set(ja
->ht
,
2484 (struct cds_ja_inode_flag
*) &ja
->root
,
2486 if (!root_shadow_node
) {
2494 ret
= rcuja_delete_ht(ja
->ht
);
2504 * Called from RCU read-side CS.
2506 __attribute__((visibility("protected")))
2507 void rcuja_free_all_children(struct cds_ja_shadow_node
*shadow_node
,
2508 struct cds_ja_inode_flag
*node_flag
,
2509 void (*rcu_free_node
)(struct cds_ja_node
*node
))
2511 unsigned int type_index
;
2512 struct cds_ja_inode
*node
;
2513 const struct cds_ja_type
*type
;
2515 node
= ja_node_ptr(node_flag
);
2516 assert(node
!= NULL
);
2517 type_index
= ja_node_type(node_flag
);
2518 type
= &ja_types
[type_index
];
2520 switch (type
->type_class
) {
2524 ja_linear_node_get_nr_child(type
, node
);
2527 for (i
= 0; i
< nr_child
; i
++) {
2528 struct cds_ja_inode_flag
*iter
;
2529 struct cds_ja_node
*node_iter
, *n
;
2532 ja_linear_node_get_ith_pos(type
, node
, i
, &v
, &iter
);
2533 node_iter
= (struct cds_ja_node
*) iter
;
2534 cds_ja_for_each_duplicate_safe(node_iter
, n
) {
2535 rcu_free_node(node_iter
);
2542 unsigned int pool_nr
;
2544 for (pool_nr
= 0; pool_nr
< (1U << type
->nr_pool_order
); pool_nr
++) {
2545 struct cds_ja_inode
*pool
=
2546 ja_pool_node_get_ith_pool(type
, node
, pool_nr
);
2548 ja_linear_node_get_nr_child(type
, pool
);
2551 for (j
= 0; j
< nr_child
; j
++) {
2552 struct cds_ja_inode_flag
*iter
;
2553 struct cds_ja_node
*node_iter
, *n
;
2556 ja_linear_node_get_ith_pos(type
, pool
, j
, &v
, &iter
);
2557 node_iter
= (struct cds_ja_node
*) iter
;
2558 cds_ja_for_each_duplicate_safe(node_iter
, n
) {
2559 rcu_free_node(node_iter
);
2571 for (i
= 0; i
< JA_ENTRY_PER_NODE
; i
++) {
2572 struct cds_ja_inode_flag
*iter
;
2573 struct cds_ja_node
*node_iter
, *n
;
2575 iter
= ja_pigeon_node_get_ith_pos(type
, node
, i
);
2576 node_iter
= (struct cds_ja_node
*) iter
;
2577 cds_ja_for_each_duplicate_safe(node_iter
, n
) {
2578 rcu_free_node(node_iter
);
2589 void print_debug_fallback_distribution(struct cds_ja
*ja
)
2593 fprintf(stderr
, "Fallback node distribution:\n");
2594 for (i
= 0; i
< JA_ENTRY_PER_NODE
; i
++) {
2595 if (!ja
->node_fallback_count_distribution
[i
])
2597 fprintf(stderr
, " %3u: %4lu\n",
2598 i
, ja
->node_fallback_count_distribution
[i
]);
2603 int ja_final_checks(struct cds_ja
*ja
)
2605 double fallback_ratio
;
2606 unsigned long na
, nf
, nr_fallback
;
2609 fallback_ratio
= (double) uatomic_read(&ja
->nr_fallback
);
2610 fallback_ratio
/= (double) uatomic_read(&ja
->nr_nodes_allocated
);
2611 nr_fallback
= uatomic_read(&ja
->nr_fallback
);
2614 "[warning] RCU Judy Array used %lu fallback node(s) (ratio: %g)\n",
2615 uatomic_read(&ja
->nr_fallback
),
2618 na
= uatomic_read(&ja
->nr_nodes_allocated
);
2619 nf
= uatomic_read(&ja
->nr_nodes_freed
);
2620 dbg_printf("Nodes allocated: %lu, Nodes freed: %lu.\n", na
, nf
);
2622 print_debug_fallback_distribution(ja
);
2625 fprintf(stderr
, "[error] Judy array leaked %ld nodes. Allocated: %lu, freed: %lu.\n",
2626 (long) na
- nf
, na
, nf
);
2633 * There should be no more concurrent add to the judy array while it is
2634 * being destroyed (ensured by the caller).
2636 int cds_ja_destroy(struct cds_ja
*ja
,
2637 void (*rcu_free_node
)(struct cds_ja_node
*node
))
2639 const struct rcu_flavor_struct
*flavor
;
2642 flavor
= cds_lfht_rcu_flavor(ja
->ht
);
2643 rcuja_shadow_prune(ja
->ht
,
2644 RCUJA_SHADOW_CLEAR_FREE_NODE
| RCUJA_SHADOW_CLEAR_FREE_LOCK
,
2646 flavor
->thread_offline();
2647 ret
= rcuja_delete_ht(ja
->ht
);
2651 /* Wait for in-flight call_rcu free to complete. */
2654 flavor
->thread_online();
2655 ret
= ja_final_checks(ja
);