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"
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 { .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 { .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, },
128 * Upon node removal below min_child, if child pool is filled
129 * beyond capacity, we roll back to pigeon.
131 { .type_class
= RCU_JA_PIGEON
, .min_child
= 89, .max_child
= ja_type_7_max_child
, .order
= 10, },
133 { .type_class
= RCU_JA_NULL
, .min_child
= 0, .max_child
= ja_type_8_max_child
, },
135 #else /* !(CAA_BITS_PER_LONG < 64) */
136 /* 64-bit pointers */
138 ja_type_0_max_child
= 1,
139 ja_type_1_max_child
= 3,
140 ja_type_2_max_child
= 7,
141 ja_type_3_max_child
= 14,
142 ja_type_4_max_child
= 28,
143 ja_type_5_max_child
= 54,
144 ja_type_6_max_child
= 104,
145 ja_type_7_max_child
= 256,
146 ja_type_8_max_child
= 256,
150 ja_type_0_max_linear_child
= 1,
151 ja_type_1_max_linear_child
= 3,
152 ja_type_2_max_linear_child
= 7,
153 ja_type_3_max_linear_child
= 14,
154 ja_type_4_max_linear_child
= 28,
155 ja_type_5_max_linear_child
= 27,
156 ja_type_6_max_linear_child
= 26,
160 ja_type_5_nr_pool_order
= 1,
161 ja_type_6_nr_pool_order
= 2,
164 const struct cds_ja_type ja_types
[] = {
165 { .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, },
166 { .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, },
167 { .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, },
168 { .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, },
169 { .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, },
171 /* Pools may fill sooner than max_child. */
172 { .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, },
173 { .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, },
176 * Upon node removal below min_child, if child pool is filled
177 * beyond capacity, we roll back to pigeon.
179 { .type_class
= RCU_JA_PIGEON
, .min_child
= 101, .max_child
= ja_type_7_max_child
, .order
= 11, },
181 { .type_class
= RCU_JA_NULL
, .min_child
= 0, .max_child
= ja_type_8_max_child
, },
183 #endif /* !(BITS_PER_LONG < 64) */
185 static inline __attribute__((unused
))
186 void static_array_size_check(void)
188 CAA_BUILD_BUG_ON(CAA_ARRAY_SIZE(ja_types
) < JA_TYPE_MAX_NR
);
192 * The cds_ja_node contains the compressed node data needed for
193 * read-side. For linear and pool node configurations, it starts with a
194 * byte counting the number of children in the node. Then, the
195 * node-specific data is placed.
196 * The node mutex, if any is needed, protecting concurrent updated of
197 * each node is placed in a separate hash table indexed by node address.
198 * For the pigeon configuration, the number of children is also kept in
199 * a separate hash table, indexed by node address, because it is only
200 * required for updates.
203 #define DECLARE_LINEAR_NODE(index) \
206 uint8_t child_value[ja_type_## index ##_max_linear_child]; \
207 struct cds_ja_inode_flag *child_ptr[ja_type_## index ##_max_linear_child]; \
210 #define DECLARE_POOL_NODE(index) \
214 uint8_t child_value[ja_type_## index ##_max_linear_child]; \
215 struct cds_ja_inode_flag *child_ptr[ja_type_## index ##_max_linear_child]; \
216 } linear[1U << ja_type_## index ##_nr_pool_order]; \
219 struct cds_ja_inode
{
221 /* Linear configuration */
222 DECLARE_LINEAR_NODE(0) conf_0
;
223 DECLARE_LINEAR_NODE(1) conf_1
;
224 DECLARE_LINEAR_NODE(2) conf_2
;
225 DECLARE_LINEAR_NODE(3) conf_3
;
226 DECLARE_LINEAR_NODE(4) conf_4
;
228 /* Pool configuration */
229 DECLARE_POOL_NODE(5) conf_5
;
230 DECLARE_POOL_NODE(6) conf_6
;
232 /* Pigeon configuration */
234 struct cds_ja_inode_flag
*child
[ja_type_7_max_child
];
236 /* data aliasing nodes for computed accesses */
237 uint8_t data
[sizeof(struct cds_ja_inode_flag
*) * ja_type_7_max_child
];
242 JA_RECOMPACT_ADD_SAME
,
243 JA_RECOMPACT_ADD_NEXT
,
248 unsigned long node_fallback_count_distribution
[JA_ENTRY_PER_NODE
];
250 unsigned long nr_nodes_allocated
, nr_nodes_freed
;
253 struct cds_ja_inode
*_ja_node_mask_ptr(struct cds_ja_inode_flag
*node
)
255 return (struct cds_ja_inode
*) (((unsigned long) node
) & JA_PTR_MASK
);
258 unsigned long ja_node_type(struct cds_ja_inode_flag
*node
)
262 if (_ja_node_mask_ptr(node
) == NULL
) {
263 return NODE_INDEX_NULL
;
265 type
= (unsigned int) ((unsigned long) node
& JA_TYPE_MASK
);
266 assert(type
< (1UL << JA_TYPE_BITS
));
270 struct cds_ja_inode
*ja_node_ptr(struct cds_ja_inode_flag
*node
)
272 unsigned long type_index
= ja_node_type(node
);
273 const struct cds_ja_type
*type
;
275 type
= &ja_types
[type_index
];
276 switch (type
->type_class
) {
278 case RCU_JA_PIGEON
: /* fall-through */
279 case RCU_JA_NULL
: /* fall-through */
280 default: /* fall-through */
281 return _ja_node_mask_ptr(node
);
283 switch (type
->nr_pool_order
) {
285 return (struct cds_ja_inode
*) (((unsigned long) node
) & ~(JA_POOL_1D_MASK
| JA_TYPE_MASK
));
287 return (struct cds_ja_inode
*) (((unsigned long) node
) & ~(JA_POOL_2D_MASK
| JA_POOL_1D_MASK
| JA_TYPE_MASK
));
294 struct cds_ja_inode
*alloc_cds_ja_node(const struct cds_ja_type
*ja_type
)
296 size_t len
= 1U << ja_type
->order
;
300 ret
= posix_memalign(&p
, len
, len
);
305 uatomic_inc(&nr_nodes_allocated
);
309 void free_cds_ja_node(struct cds_ja_inode
*node
)
313 uatomic_inc(&nr_nodes_freed
);
316 #define __JA_ALIGN_MASK(v, mask) (((v) + (mask)) & ~(mask))
317 #define JA_ALIGN(v, align) __JA_ALIGN_MASK(v, (typeof(v)) (align) - 1)
318 #define __JA_FLOOR_MASK(v, mask) ((v) & ~(mask))
319 #define JA_FLOOR(v, align) __JA_FLOOR_MASK(v, (typeof(v)) (align) - 1)
322 uint8_t *align_ptr_size(uint8_t *ptr
)
324 return (uint8_t *) JA_ALIGN((unsigned long) ptr
, sizeof(void *));
328 uint8_t ja_linear_node_get_nr_child(const struct cds_ja_type
*type
,
329 struct cds_ja_inode
*node
)
331 assert(type
->type_class
== RCU_JA_LINEAR
|| type
->type_class
== RCU_JA_POOL
);
332 return rcu_dereference(node
->u
.data
[0]);
336 * The order in which values and pointers are does does not matter: if
337 * a value is missing, we return NULL. If a value is there, but its
338 * associated pointers is still NULL, we return NULL too.
341 struct cds_ja_inode_flag
*ja_linear_node_get_nth(const struct cds_ja_type
*type
,
342 struct cds_ja_inode
*node
,
343 struct cds_ja_inode_flag
***node_flag_ptr
,
348 struct cds_ja_inode_flag
**pointers
;
349 struct cds_ja_inode_flag
*ptr
;
352 assert(type
->type_class
== RCU_JA_LINEAR
|| type
->type_class
== RCU_JA_POOL
);
354 nr_child
= ja_linear_node_get_nr_child(type
, node
);
355 cmm_smp_rmb(); /* read nr_child before values and pointers */
356 assert(nr_child
<= type
->max_linear_child
);
357 assert(type
->type_class
!= RCU_JA_LINEAR
|| nr_child
>= type
->min_child
);
359 values
= &node
->u
.data
[1];
360 for (i
= 0; i
< nr_child
; i
++) {
361 if (CMM_LOAD_SHARED(values
[i
]) == n
)
365 if (caa_unlikely(node_flag_ptr
))
366 *node_flag_ptr
= NULL
;
369 pointers
= (struct cds_ja_inode_flag
**) align_ptr_size(&values
[type
->max_linear_child
]);
370 ptr
= rcu_dereference(pointers
[i
]);
371 if (caa_unlikely(node_flag_ptr
))
372 *node_flag_ptr
= &pointers
[i
];
377 void ja_linear_node_get_ith_pos(const struct cds_ja_type
*type
,
378 struct cds_ja_inode
*node
,
381 struct cds_ja_inode_flag
**iter
)
384 struct cds_ja_inode_flag
**pointers
;
386 assert(type
->type_class
== RCU_JA_LINEAR
|| type
->type_class
== RCU_JA_POOL
);
387 assert(i
< ja_linear_node_get_nr_child(type
, node
));
389 values
= &node
->u
.data
[1];
391 pointers
= (struct cds_ja_inode_flag
**) align_ptr_size(&values
[type
->max_linear_child
]);
396 struct cds_ja_inode_flag
*ja_pool_node_get_nth(const struct cds_ja_type
*type
,
397 struct cds_ja_inode
*node
,
398 struct cds_ja_inode_flag
*node_flag
,
399 struct cds_ja_inode_flag
***node_flag_ptr
,
402 struct cds_ja_inode
*linear
;
404 assert(type
->type_class
== RCU_JA_POOL
);
406 switch (type
->nr_pool_order
) {
409 unsigned long bitsel
, index
;
411 bitsel
= ja_node_pool_1d_bitsel(node_flag
);
412 assert(bitsel
< CHAR_BIT
);
413 index
= ((unsigned long) n
>> bitsel
) & 0x1;
414 linear
= (struct cds_ja_inode
*) &node
->u
.data
[index
<< type
->pool_size_order
];
419 unsigned long bitsel
[2], index
[2], rindex
;
421 ja_node_pool_2d_bitsel(node_flag
, bitsel
);
422 assert(bitsel
[0] < CHAR_BIT
);
423 assert(bitsel
[1] < CHAR_BIT
);
424 index
[0] = ((unsigned long) n
>> bitsel
[0]) & 0x1;
426 index
[1] = ((unsigned long) n
>> bitsel
[1]) & 0x1;
427 rindex
= index
[0] | index
[1];
428 linear
= (struct cds_ja_inode
*) &node
->u
.data
[rindex
<< type
->pool_size_order
];
435 return ja_linear_node_get_nth(type
, linear
, node_flag_ptr
, n
);
439 struct cds_ja_inode
*ja_pool_node_get_ith_pool(const struct cds_ja_type
*type
,
440 struct cds_ja_inode
*node
,
443 assert(type
->type_class
== RCU_JA_POOL
);
444 return (struct cds_ja_inode
*)
445 &node
->u
.data
[(unsigned int) i
<< type
->pool_size_order
];
449 struct cds_ja_inode_flag
*ja_pigeon_node_get_nth(const struct cds_ja_type
*type
,
450 struct cds_ja_inode
*node
,
451 struct cds_ja_inode_flag
***node_flag_ptr
,
454 struct cds_ja_inode_flag
**child_node_flag_ptr
;
455 struct cds_ja_inode_flag
*child_node_flag
;
457 assert(type
->type_class
== RCU_JA_PIGEON
);
458 child_node_flag_ptr
= &((struct cds_ja_inode_flag
**) node
->u
.data
)[n
];
459 child_node_flag
= rcu_dereference(*child_node_flag_ptr
);
460 dbg_printf("ja_pigeon_node_get_nth child_node_flag_ptr %p\n",
461 child_node_flag_ptr
);
462 if (caa_unlikely(node_flag_ptr
))
463 *node_flag_ptr
= child_node_flag_ptr
;
464 return child_node_flag
;
468 struct cds_ja_inode_flag
*ja_pigeon_node_get_ith_pos(const struct cds_ja_type
*type
,
469 struct cds_ja_inode
*node
,
472 return ja_pigeon_node_get_nth(type
, node
, NULL
, i
);
476 * ja_node_get_nth: get nth item from a node.
477 * node_flag is already rcu_dereference'd.
480 struct cds_ja_inode_flag
*ja_node_get_nth(struct cds_ja_inode_flag
*node_flag
,
481 struct cds_ja_inode_flag
***node_flag_ptr
,
484 unsigned int type_index
;
485 struct cds_ja_inode
*node
;
486 const struct cds_ja_type
*type
;
488 node
= ja_node_ptr(node_flag
);
489 assert(node
!= NULL
);
490 type_index
= ja_node_type(node_flag
);
491 type
= &ja_types
[type_index
];
493 switch (type
->type_class
) {
495 return ja_linear_node_get_nth(type
, node
,
498 return ja_pool_node_get_nth(type
, node
, node_flag
,
501 return ja_pigeon_node_get_nth(type
, node
,
505 return (void *) -1UL;
510 int ja_linear_node_set_nth(const struct cds_ja_type
*type
,
511 struct cds_ja_inode
*node
,
512 struct cds_ja_shadow_node
*shadow_node
,
514 struct cds_ja_inode_flag
*child_node_flag
)
517 uint8_t *values
, *nr_child_ptr
;
518 struct cds_ja_inode_flag
**pointers
;
519 unsigned int i
, unused
= 0;
521 assert(type
->type_class
== RCU_JA_LINEAR
|| type
->type_class
== RCU_JA_POOL
);
523 nr_child_ptr
= &node
->u
.data
[0];
524 dbg_printf("linear set nth: n %u, nr_child_ptr %p\n",
525 (unsigned int) n
, nr_child_ptr
);
526 nr_child
= *nr_child_ptr
;
527 assert(nr_child
<= type
->max_linear_child
);
529 values
= &node
->u
.data
[1];
530 pointers
= (struct cds_ja_inode_flag
**) align_ptr_size(&values
[type
->max_linear_child
]);
531 /* Check if node value is already populated */
532 for (i
= 0; i
< nr_child
; i
++) {
533 if (values
[i
] == n
) {
543 if (i
== nr_child
&& nr_child
>= type
->max_linear_child
) {
545 return -ERANGE
; /* recompact node */
547 return -ENOSPC
; /* No space left in this node type */
550 assert(pointers
[i
] == NULL
);
551 rcu_assign_pointer(pointers
[i
], child_node_flag
);
552 /* If we expanded the nr_child, increment it */
554 CMM_STORE_SHARED(values
[nr_child
], n
);
555 /* write pointer and value before nr_child */
557 CMM_STORE_SHARED(*nr_child_ptr
, nr_child
+ 1);
559 shadow_node
->nr_child
++;
560 dbg_printf("linear set nth: %u child, shadow: %u child, for node %p shadow %p\n",
561 (unsigned int) CMM_LOAD_SHARED(*nr_child_ptr
),
562 (unsigned int) shadow_node
->nr_child
,
569 int ja_pool_node_set_nth(const struct cds_ja_type
*type
,
570 struct cds_ja_inode
*node
,
571 struct cds_ja_inode_flag
*node_flag
,
572 struct cds_ja_shadow_node
*shadow_node
,
574 struct cds_ja_inode_flag
*child_node_flag
)
576 struct cds_ja_inode
*linear
;
578 assert(type
->type_class
== RCU_JA_POOL
);
580 switch (type
->nr_pool_order
) {
583 unsigned long bitsel
, index
;
585 bitsel
= ja_node_pool_1d_bitsel(node_flag
);
586 assert(bitsel
< CHAR_BIT
);
587 index
= ((unsigned long) n
>> bitsel
) & 0x1;
588 linear
= (struct cds_ja_inode
*) &node
->u
.data
[index
<< type
->pool_size_order
];
593 unsigned long bitsel
[2], index
[2], rindex
;
595 ja_node_pool_2d_bitsel(node_flag
, bitsel
);
596 assert(bitsel
[0] < CHAR_BIT
);
597 assert(bitsel
[1] < CHAR_BIT
);
598 index
[0] = ((unsigned long) n
>> bitsel
[0]) & 0x1;
600 index
[1] = ((unsigned long) n
>> bitsel
[1]) & 0x1;
601 rindex
= index
[0] | index
[1];
602 linear
= (struct cds_ja_inode
*) &node
->u
.data
[rindex
<< type
->pool_size_order
];
610 return ja_linear_node_set_nth(type
, linear
, shadow_node
,
615 int ja_pigeon_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
)
621 struct cds_ja_inode_flag
**ptr
;
623 assert(type
->type_class
== RCU_JA_PIGEON
);
624 ptr
= &((struct cds_ja_inode_flag
**) node
->u
.data
)[n
];
627 rcu_assign_pointer(*ptr
, child_node_flag
);
628 shadow_node
->nr_child
++;
633 * _ja_node_set_nth: set nth item within a node. Return an error
634 * (negative error value) if it is already there.
637 int _ja_node_set_nth(const struct cds_ja_type
*type
,
638 struct cds_ja_inode
*node
,
639 struct cds_ja_inode_flag
*node_flag
,
640 struct cds_ja_shadow_node
*shadow_node
,
642 struct cds_ja_inode_flag
*child_node_flag
)
644 switch (type
->type_class
) {
646 return ja_linear_node_set_nth(type
, node
, shadow_node
, n
,
649 return ja_pool_node_set_nth(type
, node
, node_flag
, shadow_node
, n
,
652 return ja_pigeon_node_set_nth(type
, node
, shadow_node
, n
,
665 int ja_linear_node_clear_ptr(const struct cds_ja_type
*type
,
666 struct cds_ja_inode
*node
,
667 struct cds_ja_shadow_node
*shadow_node
,
668 struct cds_ja_inode_flag
**node_flag_ptr
)
671 uint8_t *nr_child_ptr
;
673 assert(type
->type_class
== RCU_JA_LINEAR
|| type
->type_class
== RCU_JA_POOL
);
675 nr_child_ptr
= &node
->u
.data
[0];
676 nr_child
= *nr_child_ptr
;
677 assert(nr_child
<= type
->max_linear_child
);
679 if (type
->type_class
== RCU_JA_LINEAR
) {
680 assert(!shadow_node
->fallback_removal_count
);
681 if (shadow_node
->nr_child
<= type
->min_child
) {
682 /* We need to try recompacting the node */
686 dbg_printf("linear clear ptr: nr_child_ptr %p\n", nr_child_ptr
);
687 assert(*node_flag_ptr
!= NULL
);
688 rcu_assign_pointer(*node_flag_ptr
, NULL
);
690 * Value and nr_child are never changed (would cause ABA issue).
691 * Instead, we leave the pointer to NULL and recompact the node
692 * once in a while. It is allowed to set a NULL pointer to a new
693 * value without recompaction though.
694 * Only update the shadow node accounting.
696 shadow_node
->nr_child
--;
697 dbg_printf("linear clear ptr: %u child, shadow: %u child, for node %p shadow %p\n",
698 (unsigned int) CMM_LOAD_SHARED(*nr_child_ptr
),
699 (unsigned int) shadow_node
->nr_child
,
705 int ja_pool_node_clear_ptr(const struct cds_ja_type
*type
,
706 struct cds_ja_inode
*node
,
707 struct cds_ja_inode_flag
*node_flag
,
708 struct cds_ja_shadow_node
*shadow_node
,
709 struct cds_ja_inode_flag
**node_flag_ptr
,
712 struct cds_ja_inode
*linear
;
714 assert(type
->type_class
== RCU_JA_POOL
);
716 if (shadow_node
->fallback_removal_count
) {
717 shadow_node
->fallback_removal_count
--;
719 /* We should try recompacting the node */
720 if (shadow_node
->nr_child
<= type
->min_child
)
724 switch (type
->nr_pool_order
) {
727 unsigned long bitsel
, index
;
729 bitsel
= ja_node_pool_1d_bitsel(node_flag
);
730 assert(bitsel
< CHAR_BIT
);
731 index
= ((unsigned long) n
>> bitsel
) & type
->nr_pool_order
;
732 linear
= (struct cds_ja_inode
*) &node
->u
.data
[index
<< type
->pool_size_order
];
737 unsigned long bitsel
[2], index
[2], rindex
;
739 ja_node_pool_2d_bitsel(node_flag
, bitsel
);
740 assert(bitsel
[0] < CHAR_BIT
);
741 assert(bitsel
[1] < CHAR_BIT
);
742 index
[0] = ((unsigned long) n
>> bitsel
[0]) & 0x1;
744 index
[1] = ((unsigned long) n
>> bitsel
[1]) & 0x1;
745 rindex
= index
[0] | index
[1];
746 linear
= (struct cds_ja_inode
*) &node
->u
.data
[rindex
<< type
->pool_size_order
];
754 return ja_linear_node_clear_ptr(type
, linear
, shadow_node
, node_flag_ptr
);
758 int ja_pigeon_node_clear_ptr(const struct cds_ja_type
*type
,
759 struct cds_ja_inode
*node
,
760 struct cds_ja_shadow_node
*shadow_node
,
761 struct cds_ja_inode_flag
**node_flag_ptr
)
763 assert(type
->type_class
== RCU_JA_PIGEON
);
765 if (shadow_node
->fallback_removal_count
) {
766 shadow_node
->fallback_removal_count
--;
768 /* We should try recompacting the node */
769 if (shadow_node
->nr_child
<= type
->min_child
)
772 dbg_printf("ja_pigeon_node_clear_ptr: clearing ptr: %p\n", *node_flag_ptr
);
773 rcu_assign_pointer(*node_flag_ptr
, NULL
);
774 shadow_node
->nr_child
--;
779 * _ja_node_clear_ptr: clear ptr item within a node. Return an error
780 * (negative error value) if it is not found (-ENOENT).
783 int _ja_node_clear_ptr(const struct cds_ja_type
*type
,
784 struct cds_ja_inode
*node
,
785 struct cds_ja_inode_flag
*node_flag
,
786 struct cds_ja_shadow_node
*shadow_node
,
787 struct cds_ja_inode_flag
**node_flag_ptr
,
790 switch (type
->type_class
) {
792 return ja_linear_node_clear_ptr(type
, node
, shadow_node
, node_flag_ptr
);
794 return ja_pool_node_clear_ptr(type
, node
, node_flag
, shadow_node
, node_flag_ptr
, n
);
796 return ja_pigeon_node_clear_ptr(type
, node
, shadow_node
, node_flag_ptr
);
808 * Calculate bit distribution. Returns the bit (0 to 7) that splits the
809 * distribution in two sub-distributions containing as much elements one
810 * compared to the other.
813 unsigned int ja_node_sum_distribution_1d(enum ja_recompact mode
,
815 unsigned int type_index
,
816 const struct cds_ja_type
*type
,
817 struct cds_ja_inode
*node
,
818 struct cds_ja_shadow_node
*shadow_node
,
820 struct cds_ja_inode_flag
*child_node_flag
,
821 struct cds_ja_inode_flag
**nullify_node_flag_ptr
)
823 uint8_t nr_one
[JA_BITS_PER_BYTE
];
824 unsigned int bitsel
= 0, bit_i
, overall_best_distance
= UINT_MAX
;
825 unsigned int distrib_nr_child
= 0;
827 memset(nr_one
, 0, sizeof(nr_one
));
829 switch (type
->type_class
) {
833 ja_linear_node_get_nr_child(type
, node
);
836 for (i
= 0; i
< nr_child
; i
++) {
837 struct cds_ja_inode_flag
*iter
;
840 ja_linear_node_get_ith_pos(type
, node
, i
, &v
, &iter
);
843 if (mode
== JA_RECOMPACT_DEL
&& *nullify_node_flag_ptr
== iter
)
845 for (bit_i
= 0; bit_i
< JA_BITS_PER_BYTE
; bit_i
++) {
846 if (v
& (1U << bit_i
))
855 unsigned int pool_nr
;
857 for (pool_nr
= 0; pool_nr
< (1U << type
->nr_pool_order
); pool_nr
++) {
858 struct cds_ja_inode
*pool
=
859 ja_pool_node_get_ith_pool(type
,
862 ja_linear_node_get_nr_child(type
, pool
);
865 for (j
= 0; j
< nr_child
; j
++) {
866 struct cds_ja_inode_flag
*iter
;
869 ja_linear_node_get_ith_pos(type
, pool
,
873 if (mode
== JA_RECOMPACT_DEL
&& *nullify_node_flag_ptr
== iter
)
875 for (bit_i
= 0; bit_i
< JA_BITS_PER_BYTE
; bit_i
++) {
876 if (v
& (1U << bit_i
))
888 assert(mode
== JA_RECOMPACT_DEL
);
889 for (i
= 0; i
< JA_ENTRY_PER_NODE
; i
++) {
890 struct cds_ja_inode_flag
*iter
;
892 iter
= ja_pigeon_node_get_ith_pos(type
, node
, i
);
895 if (mode
== JA_RECOMPACT_DEL
&& *nullify_node_flag_ptr
== iter
)
897 for (bit_i
= 0; bit_i
< JA_BITS_PER_BYTE
; bit_i
++) {
898 if (i
& (1U << bit_i
))
906 assert(mode
== JA_RECOMPACT_ADD_NEXT
);
913 if (mode
== JA_RECOMPACT_ADD_NEXT
|| mode
== JA_RECOMPACT_ADD_SAME
) {
914 for (bit_i
= 0; bit_i
< JA_BITS_PER_BYTE
; bit_i
++) {
915 if (n
& (1U << bit_i
))
922 * The best bit selector is that for which the number of ones is
923 * closest to half of the number of children in the
924 * distribution. We calculate the distance using the double of
925 * the sub-distribution sizes to eliminate truncation error.
927 for (bit_i
= 0; bit_i
< JA_BITS_PER_BYTE
; bit_i
++) {
928 unsigned int distance_to_best
;
930 distance_to_best
= abs_int((nr_one
[bit_i
] << 1U) - distrib_nr_child
);
931 if (distance_to_best
< overall_best_distance
) {
932 overall_best_distance
= distance_to_best
;
936 dbg_printf("1 dimension pool bit selection: (%u)\n", bitsel
);
941 * Calculate bit distribution in two dimensions. Returns the two bits
942 * (each 0 to 7) that splits the distribution in four sub-distributions
943 * containing as much elements one compared to the other.
946 void ja_node_sum_distribution_2d(enum ja_recompact mode
,
948 unsigned int type_index
,
949 const struct cds_ja_type
*type
,
950 struct cds_ja_inode
*node
,
951 struct cds_ja_shadow_node
*shadow_node
,
953 struct cds_ja_inode_flag
*child_node_flag
,
954 struct cds_ja_inode_flag
**nullify_node_flag_ptr
,
955 unsigned int *_bitsel
)
957 uint8_t nr_2d_11
[JA_BITS_PER_BYTE
][JA_BITS_PER_BYTE
],
958 nr_2d_10
[JA_BITS_PER_BYTE
][JA_BITS_PER_BYTE
],
959 nr_2d_01
[JA_BITS_PER_BYTE
][JA_BITS_PER_BYTE
],
960 nr_2d_00
[JA_BITS_PER_BYTE
][JA_BITS_PER_BYTE
];
961 unsigned int bitsel
[2] = { 0, 1 };
962 unsigned int bit_i
, bit_j
;
963 int overall_best_distance
= INT_MAX
;
964 unsigned int distrib_nr_child
= 0;
966 memset(nr_2d_11
, 0, sizeof(nr_2d_11
));
967 memset(nr_2d_10
, 0, sizeof(nr_2d_10
));
968 memset(nr_2d_01
, 0, sizeof(nr_2d_01
));
969 memset(nr_2d_00
, 0, sizeof(nr_2d_00
));
971 switch (type
->type_class
) {
975 ja_linear_node_get_nr_child(type
, node
);
978 for (i
= 0; i
< nr_child
; i
++) {
979 struct cds_ja_inode_flag
*iter
;
982 ja_linear_node_get_ith_pos(type
, node
, i
, &v
, &iter
);
985 if (mode
== JA_RECOMPACT_DEL
&& *nullify_node_flag_ptr
== iter
)
987 for (bit_i
= 0; bit_i
< JA_BITS_PER_BYTE
; bit_i
++) {
988 for (bit_j
= 0; bit_j
< bit_i
; bit_j
++) {
989 if ((v
& (1U << bit_i
)) && (v
& (1U << bit_j
))) {
990 nr_2d_11
[bit_i
][bit_j
]++;
992 if ((v
& (1U << bit_i
)) && !(v
& (1U << bit_j
))) {
993 nr_2d_10
[bit_i
][bit_j
]++;
995 if (!(v
& (1U << bit_i
)) && (v
& (1U << bit_j
))) {
996 nr_2d_01
[bit_i
][bit_j
]++;
998 if (!(v
& (1U << bit_i
)) && !(v
& (1U << bit_j
))) {
999 nr_2d_00
[bit_i
][bit_j
]++;
1009 unsigned int pool_nr
;
1011 for (pool_nr
= 0; pool_nr
< (1U << type
->nr_pool_order
); pool_nr
++) {
1012 struct cds_ja_inode
*pool
=
1013 ja_pool_node_get_ith_pool(type
,
1016 ja_linear_node_get_nr_child(type
, pool
);
1019 for (j
= 0; j
< nr_child
; j
++) {
1020 struct cds_ja_inode_flag
*iter
;
1023 ja_linear_node_get_ith_pos(type
, pool
,
1027 if (mode
== JA_RECOMPACT_DEL
&& *nullify_node_flag_ptr
== iter
)
1029 for (bit_i
= 0; bit_i
< JA_BITS_PER_BYTE
; bit_i
++) {
1030 for (bit_j
= 0; bit_j
< bit_i
; bit_j
++) {
1031 if ((v
& (1U << bit_i
)) && (v
& (1U << bit_j
))) {
1032 nr_2d_11
[bit_i
][bit_j
]++;
1034 if ((v
& (1U << bit_i
)) && !(v
& (1U << bit_j
))) {
1035 nr_2d_10
[bit_i
][bit_j
]++;
1037 if (!(v
& (1U << bit_i
)) && (v
& (1U << bit_j
))) {
1038 nr_2d_01
[bit_i
][bit_j
]++;
1040 if (!(v
& (1U << bit_i
)) && !(v
& (1U << bit_j
))) {
1041 nr_2d_00
[bit_i
][bit_j
]++;
1054 assert(mode
== JA_RECOMPACT_DEL
);
1055 for (i
= 0; i
< JA_ENTRY_PER_NODE
; i
++) {
1056 struct cds_ja_inode_flag
*iter
;
1058 iter
= ja_pigeon_node_get_ith_pos(type
, node
, i
);
1061 if (mode
== JA_RECOMPACT_DEL
&& *nullify_node_flag_ptr
== iter
)
1063 for (bit_i
= 0; bit_i
< JA_BITS_PER_BYTE
; bit_i
++) {
1064 for (bit_j
= 0; bit_j
< bit_i
; bit_j
++) {
1065 if ((i
& (1U << bit_i
)) && (i
& (1U << bit_j
))) {
1066 nr_2d_11
[bit_i
][bit_j
]++;
1068 if ((i
& (1U << bit_i
)) && !(i
& (1U << bit_j
))) {
1069 nr_2d_10
[bit_i
][bit_j
]++;
1071 if (!(i
& (1U << bit_i
)) && (i
& (1U << bit_j
))) {
1072 nr_2d_01
[bit_i
][bit_j
]++;
1074 if (!(i
& (1U << bit_i
)) && !(i
& (1U << bit_j
))) {
1075 nr_2d_00
[bit_i
][bit_j
]++;
1084 assert(mode
== JA_RECOMPACT_ADD_NEXT
);
1091 if (mode
== JA_RECOMPACT_ADD_NEXT
|| mode
== JA_RECOMPACT_ADD_SAME
) {
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 ((n
& (1U << bit_i
)) && (n
& (1U << bit_j
))) {
1095 nr_2d_11
[bit_i
][bit_j
]++;
1097 if ((n
& (1U << bit_i
)) && !(n
& (1U << bit_j
))) {
1098 nr_2d_10
[bit_i
][bit_j
]++;
1100 if (!(n
& (1U << bit_i
)) && (n
& (1U << bit_j
))) {
1101 nr_2d_01
[bit_i
][bit_j
]++;
1103 if (!(n
& (1U << bit_i
)) && !(n
& (1U << bit_j
))) {
1104 nr_2d_00
[bit_i
][bit_j
]++;
1112 * The best bit selector is that for which the number of nodes
1113 * in each sub-class is closest to one-fourth of the number of
1114 * children in the distribution. We calculate the distance using
1115 * 4 times the size of the sub-distribution to eliminate
1118 for (bit_i
= 0; bit_i
< JA_BITS_PER_BYTE
; bit_i
++) {
1119 for (bit_j
= 0; bit_j
< bit_i
; bit_j
++) {
1120 int distance_to_best
[4];
1122 distance_to_best
[0] = (nr_2d_11
[bit_i
][bit_j
] << 2U) - distrib_nr_child
;
1123 distance_to_best
[1] = (nr_2d_10
[bit_i
][bit_j
] << 2U) - distrib_nr_child
;
1124 distance_to_best
[2] = (nr_2d_01
[bit_i
][bit_j
] << 2U) - distrib_nr_child
;
1125 distance_to_best
[3] = (nr_2d_00
[bit_i
][bit_j
] << 2U) - distrib_nr_child
;
1127 /* Consider worse distance above best */
1128 if (distance_to_best
[1] > 0 && distance_to_best
[1] > distance_to_best
[0])
1129 distance_to_best
[0] = distance_to_best
[1];
1130 if (distance_to_best
[2] > 0 && distance_to_best
[2] > distance_to_best
[0])
1131 distance_to_best
[0] = distance_to_best
[2];
1132 if (distance_to_best
[3] > 0 && distance_to_best
[3] > distance_to_best
[0])
1133 distance_to_best
[0] = distance_to_best
[3];
1136 * If our worse distance is better than overall,
1137 * we become new best candidate.
1139 if (distance_to_best
[0] < overall_best_distance
) {
1140 overall_best_distance
= distance_to_best
[0];
1147 dbg_printf("2 dimensions pool bit selection: (%u,%u)\n", bitsel
[0], bitsel
[1]);
1149 /* Return our bit selection */
1150 _bitsel
[0] = bitsel
[0];
1151 _bitsel
[1] = bitsel
[1];
1155 unsigned int find_nearest_type_index(unsigned int type_index
,
1156 unsigned int nr_nodes
)
1158 const struct cds_ja_type
*type
;
1160 assert(type_index
!= NODE_INDEX_NULL
);
1162 return NODE_INDEX_NULL
;
1164 type
= &ja_types
[type_index
];
1165 if (nr_nodes
< type
->min_child
)
1167 else if (nr_nodes
> type
->max_child
)
1176 * ja_node_recompact_add: recompact a node, adding a new child.
1177 * Return 0 on success, -EAGAIN if need to retry, or other negative
1178 * error value otherwise.
1181 int ja_node_recompact(enum ja_recompact mode
,
1183 unsigned int old_type_index
,
1184 const struct cds_ja_type
*old_type
,
1185 struct cds_ja_inode
*old_node
,
1186 struct cds_ja_shadow_node
*shadow_node
,
1187 struct cds_ja_inode_flag
**old_node_flag_ptr
, uint8_t n
,
1188 struct cds_ja_inode_flag
*child_node_flag
,
1189 struct cds_ja_inode_flag
**nullify_node_flag_ptr
,
1192 unsigned int new_type_index
;
1193 struct cds_ja_inode
*new_node
;
1194 struct cds_ja_shadow_node
*new_shadow_node
= NULL
;
1195 const struct cds_ja_type
*new_type
;
1196 struct cds_ja_inode_flag
*new_node_flag
, *old_node_flag
;
1200 old_node_flag
= *old_node_flag_ptr
;
1203 * Need to find nearest type index even for ADD_SAME, because
1204 * this recompaction, when applied to linear nodes, will garbage
1205 * collect dummy (NULL) entries, and can therefore cause a few
1206 * linear representations to be skipped.
1209 case JA_RECOMPACT_ADD_SAME
:
1210 new_type_index
= find_nearest_type_index(old_type_index
,
1211 shadow_node
->nr_child
+ 1);
1212 dbg_printf("Recompact for node with %u children\n",
1213 shadow_node
->nr_child
+ 1);
1215 case JA_RECOMPACT_ADD_NEXT
:
1216 if (!shadow_node
|| old_type_index
== NODE_INDEX_NULL
) {
1218 dbg_printf("Recompact for NULL\n");
1220 new_type_index
= find_nearest_type_index(old_type_index
,
1221 shadow_node
->nr_child
+ 1);
1222 dbg_printf("Recompact for node with %u children\n",
1223 shadow_node
->nr_child
+ 1);
1226 case JA_RECOMPACT_DEL
:
1227 new_type_index
= find_nearest_type_index(old_type_index
,
1228 shadow_node
->nr_child
- 1);
1229 dbg_printf("Recompact for node with %u children\n",
1230 shadow_node
->nr_child
- 1);
1236 retry
: /* for fallback */
1237 dbg_printf("Recompact from type %d to type %d\n",
1238 old_type_index
, new_type_index
);
1239 new_type
= &ja_types
[new_type_index
];
1240 if (new_type_index
!= NODE_INDEX_NULL
) {
1241 new_node
= alloc_cds_ja_node(new_type
);
1245 if (new_type
->type_class
== RCU_JA_POOL
) {
1246 switch (new_type
->nr_pool_order
) {
1249 unsigned int node_distrib_bitsel
;
1251 node_distrib_bitsel
=
1252 ja_node_sum_distribution_1d(mode
, ja
,
1253 old_type_index
, old_type
,
1254 old_node
, shadow_node
,
1256 nullify_node_flag_ptr
);
1257 assert(!((unsigned long) new_node
& JA_POOL_1D_MASK
));
1258 new_node_flag
= ja_node_flag_pool_1d(new_node
,
1259 new_type_index
, node_distrib_bitsel
);
1264 unsigned int node_distrib_bitsel
[2];
1266 ja_node_sum_distribution_2d(mode
, ja
,
1267 old_type_index
, old_type
,
1268 old_node
, shadow_node
,
1270 nullify_node_flag_ptr
,
1271 node_distrib_bitsel
);
1272 assert(!((unsigned long) new_node
& JA_POOL_1D_MASK
));
1273 assert(!((unsigned long) new_node
& JA_POOL_2D_MASK
));
1274 new_node_flag
= ja_node_flag_pool_2d(new_node
,
1275 new_type_index
, node_distrib_bitsel
);
1282 new_node_flag
= ja_node_flag(new_node
, new_type_index
);
1285 dbg_printf("Recompact inherit lock from %p\n", shadow_node
);
1286 new_shadow_node
= rcuja_shadow_set(ja
->ht
, new_node_flag
, shadow_node
, ja
, level
);
1287 if (!new_shadow_node
) {
1288 free_cds_ja_node(new_node
);
1292 new_shadow_node
->fallback_removal_count
=
1293 JA_FALLBACK_REMOVAL_COUNT
;
1296 new_node_flag
= NULL
;
1299 assert(mode
!= JA_RECOMPACT_ADD_NEXT
|| old_type
->type_class
!= RCU_JA_PIGEON
);
1301 if (new_type_index
== NODE_INDEX_NULL
)
1304 switch (old_type
->type_class
) {
1308 ja_linear_node_get_nr_child(old_type
, old_node
);
1311 for (i
= 0; i
< nr_child
; i
++) {
1312 struct cds_ja_inode_flag
*iter
;
1315 ja_linear_node_get_ith_pos(old_type
, old_node
, i
, &v
, &iter
);
1318 if (mode
== JA_RECOMPACT_DEL
&& *nullify_node_flag_ptr
== iter
)
1320 ret
= _ja_node_set_nth(new_type
, new_node
, new_node_flag
,
1323 if (new_type
->type_class
== RCU_JA_POOL
&& ret
) {
1324 goto fallback_toosmall
;
1332 unsigned int pool_nr
;
1334 for (pool_nr
= 0; pool_nr
< (1U << old_type
->nr_pool_order
); pool_nr
++) {
1335 struct cds_ja_inode
*pool
=
1336 ja_pool_node_get_ith_pool(old_type
,
1339 ja_linear_node_get_nr_child(old_type
, pool
);
1342 for (j
= 0; j
< nr_child
; j
++) {
1343 struct cds_ja_inode_flag
*iter
;
1346 ja_linear_node_get_ith_pos(old_type
, pool
,
1350 if (mode
== JA_RECOMPACT_DEL
&& *nullify_node_flag_ptr
== iter
)
1352 ret
= _ja_node_set_nth(new_type
, new_node
, new_node_flag
,
1355 if (new_type
->type_class
== RCU_JA_POOL
1357 goto fallback_toosmall
;
1365 assert(mode
== JA_RECOMPACT_ADD_NEXT
);
1371 assert(mode
== JA_RECOMPACT_DEL
);
1372 for (i
= 0; i
< JA_ENTRY_PER_NODE
; i
++) {
1373 struct cds_ja_inode_flag
*iter
;
1375 iter
= ja_pigeon_node_get_ith_pos(old_type
, old_node
, i
);
1378 if (mode
== JA_RECOMPACT_DEL
&& *nullify_node_flag_ptr
== iter
)
1380 ret
= _ja_node_set_nth(new_type
, new_node
, new_node_flag
,
1383 if (new_type
->type_class
== RCU_JA_POOL
&& ret
) {
1384 goto fallback_toosmall
;
1397 if (mode
== JA_RECOMPACT_ADD_NEXT
|| mode
== JA_RECOMPACT_ADD_SAME
) {
1399 ret
= _ja_node_set_nth(new_type
, new_node
, new_node_flag
,
1401 n
, child_node_flag
);
1402 if (new_type
->type_class
== RCU_JA_POOL
&& ret
) {
1403 goto fallback_toosmall
;
1409 dbg_printf("Using fallback for %u children, node type index: %u, mode %s\n",
1410 new_shadow_node
->nr_child
, old_type_index
, mode
== JA_RECOMPACT_ADD_NEXT
? "add_next" :
1411 (mode
== JA_RECOMPACT_DEL
? "del" : "add_same"));
1412 uatomic_inc(&node_fallback_count_distribution
[new_shadow_node
->nr_child
]);
1415 /* Return pointer to new recompacted node through old_node_flag_ptr */
1416 *old_node_flag_ptr
= new_node_flag
;
1420 flags
= RCUJA_SHADOW_CLEAR_FREE_NODE
;
1422 * It is OK to free the lock associated with a node
1423 * going to NULL, since we are holding the parent lock.
1424 * This synchronizes removal with re-add of that node.
1426 if (new_type_index
== NODE_INDEX_NULL
)
1427 flags
|= RCUJA_SHADOW_CLEAR_FREE_LOCK
;
1428 ret
= rcuja_shadow_clear(ja
->ht
, old_node_flag
, shadow_node
,
1438 /* fallback if next pool is too small */
1439 assert(new_shadow_node
);
1440 ret
= rcuja_shadow_clear(ja
->ht
, new_node_flag
, new_shadow_node
,
1441 RCUJA_SHADOW_CLEAR_FREE_NODE
);
1445 case JA_RECOMPACT_ADD_SAME
:
1447 * JA_RECOMPACT_ADD_SAME is only triggered if a linear
1448 * node within a pool has unused entries. It should
1449 * therefore _never_ be too small.
1454 case JA_RECOMPACT_ADD_NEXT
:
1456 const struct cds_ja_type
*next_type
;
1459 * Recompaction attempt on add failed. Should only
1460 * happen if target node type is pool. Caused by
1461 * hard-to-split distribution. Recompact using the next
1462 * distribution size.
1464 assert(new_type
->type_class
== RCU_JA_POOL
);
1465 next_type
= &ja_types
[new_type_index
+ 1];
1467 * Try going to the next pool size if our population
1468 * fits within its range. This is not flagged as a
1471 if (shadow_node
->nr_child
+ 1 >= next_type
->min_child
1472 && shadow_node
->nr_child
+ 1 <= next_type
->max_child
) {
1477 dbg_printf("Add fallback to type %d\n", new_type_index
);
1478 uatomic_inc(&ja
->nr_fallback
);
1484 case JA_RECOMPACT_DEL
:
1486 * Recompaction attempt on delete failed. Should only
1487 * happen if target node type is pool. This is caused by
1488 * a hard-to-split distribution. Recompact on same node
1489 * size, but flag current node as "fallback" to ensure
1490 * we don't attempt recompaction before some activity
1491 * has reshuffled our node.
1493 assert(new_type
->type_class
== RCU_JA_POOL
);
1494 new_type_index
= old_type_index
;
1495 dbg_printf("Delete fallback keeping type %d\n", new_type_index
);
1496 uatomic_inc(&ja
->nr_fallback
);
1505 * Last resort fallback: pigeon.
1507 new_type_index
= (1UL << JA_TYPE_BITS
) - 1;
1508 dbg_printf("Fallback to type %d\n", new_type_index
);
1509 uatomic_inc(&ja
->nr_fallback
);
1515 * Return 0 on success, -EAGAIN if need to retry, or other negative
1516 * error value otherwise.
1519 int ja_node_set_nth(struct cds_ja
*ja
,
1520 struct cds_ja_inode_flag
**node_flag
, uint8_t n
,
1521 struct cds_ja_inode_flag
*child_node_flag
,
1522 struct cds_ja_shadow_node
*shadow_node
,
1526 unsigned int type_index
;
1527 const struct cds_ja_type
*type
;
1528 struct cds_ja_inode
*node
;
1530 dbg_printf("ja_node_set_nth for n=%u, node %p, shadow %p\n",
1531 (unsigned int) n
, ja_node_ptr(*node_flag
), shadow_node
);
1533 node
= ja_node_ptr(*node_flag
);
1534 type_index
= ja_node_type(*node_flag
);
1535 type
= &ja_types
[type_index
];
1536 ret
= _ja_node_set_nth(type
, node
, *node_flag
, shadow_node
,
1537 n
, child_node_flag
);
1540 /* Not enough space in node, need to recompact to next type. */
1541 ret
= ja_node_recompact(JA_RECOMPACT_ADD_NEXT
, ja
, type_index
, type
, node
,
1542 shadow_node
, node_flag
, n
, child_node_flag
, NULL
, level
);
1545 /* Node needs to be recompacted. */
1546 ret
= ja_node_recompact(JA_RECOMPACT_ADD_SAME
, ja
, type_index
, type
, node
,
1547 shadow_node
, node_flag
, n
, child_node_flag
, NULL
, level
);
1554 * Return 0 on success, -EAGAIN if need to retry, or other negative
1555 * error value otherwise.
1558 int ja_node_clear_ptr(struct cds_ja
*ja
,
1559 struct cds_ja_inode_flag
**node_flag_ptr
, /* Pointer to location to nullify */
1560 struct cds_ja_inode_flag
**parent_node_flag_ptr
, /* Address of parent ptr in its parent */
1561 struct cds_ja_shadow_node
*shadow_node
, /* of parent */
1562 uint8_t n
, int level
)
1565 unsigned int type_index
;
1566 const struct cds_ja_type
*type
;
1567 struct cds_ja_inode
*node
;
1569 dbg_printf("ja_node_clear_ptr for node %p, shadow %p, target ptr %p\n",
1570 ja_node_ptr(*parent_node_flag_ptr
), shadow_node
, node_flag_ptr
);
1572 node
= ja_node_ptr(*parent_node_flag_ptr
);
1573 type_index
= ja_node_type(*parent_node_flag_ptr
);
1574 type
= &ja_types
[type_index
];
1575 ret
= _ja_node_clear_ptr(type
, node
, *parent_node_flag_ptr
, shadow_node
, node_flag_ptr
, n
);
1576 if (ret
== -EFBIG
) {
1577 /* Should try recompaction. */
1578 ret
= ja_node_recompact(JA_RECOMPACT_DEL
, ja
, type_index
, type
, node
,
1579 shadow_node
, parent_node_flag_ptr
, n
, NULL
,
1580 node_flag_ptr
, level
);
1585 struct cds_hlist_head
cds_ja_lookup(struct cds_ja
*ja
, uint64_t key
)
1587 unsigned int tree_depth
, i
;
1588 struct cds_ja_inode_flag
*node_flag
;
1589 struct cds_hlist_head head
= { NULL
};
1591 if (caa_unlikely(key
> ja
->key_max
))
1593 tree_depth
= ja
->tree_depth
;
1594 node_flag
= rcu_dereference(ja
->root
);
1596 /* level 0: root node */
1597 if (!ja_node_ptr(node_flag
))
1600 for (i
= 1; i
< tree_depth
; i
++) {
1603 iter_key
= (uint8_t) (key
>> (JA_BITS_PER_BYTE
* (tree_depth
- i
- 1)));
1604 node_flag
= ja_node_get_nth(node_flag
, NULL
, iter_key
);
1605 dbg_printf("cds_ja_lookup iter key lookup %u finds node_flag %p\n",
1606 (unsigned int) iter_key
, node_flag
);
1607 if (!ja_node_ptr(node_flag
))
1611 /* Last level lookup succeded. We got an actual match. */
1612 head
.next
= (struct cds_hlist_node
*) node_flag
;
1617 * We reached an unpopulated node. Create it and the children we need,
1618 * and then attach the entire branch to the current node. This may
1619 * trigger recompaction of the current node. Locks needed: node lock
1620 * (for add), and, possibly, parent node lock (to update pointer due to
1621 * node recompaction).
1623 * First take node lock, check if recompaction is needed, then take
1624 * parent lock (if needed). Then we can proceed to create the new
1625 * branch. Publish the new branch, and release locks.
1626 * TODO: we currently always take the parent lock even when not needed.
1629 int ja_attach_node(struct cds_ja
*ja
,
1630 struct cds_ja_inode_flag
**attach_node_flag_ptr
,
1631 struct cds_ja_inode_flag
*attach_node_flag
,
1632 struct cds_ja_inode_flag
*parent_attach_node_flag
,
1633 struct cds_ja_inode_flag
**old_node_flag_ptr
,
1634 struct cds_ja_inode_flag
*old_node_flag
,
1637 struct cds_ja_node
*child_node
)
1639 struct cds_ja_shadow_node
*shadow_node
= NULL
,
1640 *parent_shadow_node
= NULL
;
1641 struct cds_hlist_head head
;
1642 struct cds_ja_inode_flag
*iter_node_flag
, *iter_dest_node_flag
;
1644 struct cds_ja_inode_flag
*created_nodes
[JA_MAX_DEPTH
];
1645 int nr_created_nodes
= 0;
1647 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",
1648 level
, old_node_flag
, attach_node_flag_ptr
, attach_node_flag
, parent_attach_node_flag
);
1650 assert(!old_node_flag
);
1651 if (attach_node_flag
) {
1652 shadow_node
= rcuja_shadow_lookup_lock(ja
->ht
, attach_node_flag
);
1658 if (parent_attach_node_flag
) {
1659 parent_shadow_node
= rcuja_shadow_lookup_lock(ja
->ht
,
1660 parent_attach_node_flag
);
1661 if (!parent_shadow_node
) {
1667 if (old_node_flag_ptr
&& ja_node_ptr(*old_node_flag_ptr
)) {
1669 * Target node has been updated between RCU lookup and
1670 * lock acquisition. We need to re-try lookup and
1678 * Perform a lookup query to handle the case where
1679 * old_node_flag_ptr is NULL. We cannot use it to check if the
1680 * node has been populated between RCU lookup and mutex
1683 if (!old_node_flag_ptr
) {
1685 struct cds_ja_inode_flag
*lookup_node_flag
;
1686 struct cds_ja_inode_flag
**lookup_node_flag_ptr
;
1688 iter_key
= (uint8_t) (key
>> (JA_BITS_PER_BYTE
* (ja
->tree_depth
- level
)));
1689 lookup_node_flag
= ja_node_get_nth(attach_node_flag
,
1690 &lookup_node_flag_ptr
,
1692 if (lookup_node_flag
) {
1698 if (attach_node_flag_ptr
&& ja_node_ptr(*attach_node_flag_ptr
) !=
1699 ja_node_ptr(attach_node_flag
)) {
1701 * Target node has been updated between RCU lookup and
1702 * lock acquisition. We need to re-try lookup and
1709 /* Create new branch, starting from bottom */
1710 CDS_INIT_HLIST_HEAD(&head
);
1711 cds_hlist_add_head_rcu(&child_node
->list
, &head
);
1712 iter_node_flag
= (struct cds_ja_inode_flag
*) head
.next
;
1714 for (i
= ja
->tree_depth
- 1; i
>= (int) level
; i
--) {
1717 iter_key
= (uint8_t) (key
>> (JA_BITS_PER_BYTE
* (ja
->tree_depth
- i
- 1)));
1718 dbg_printf("branch creation level %d, key %u\n",
1719 i
, (unsigned int) iter_key
);
1720 iter_dest_node_flag
= NULL
;
1721 ret
= ja_node_set_nth(ja
, &iter_dest_node_flag
,
1726 dbg_printf("branch creation error %d\n", ret
);
1729 created_nodes
[nr_created_nodes
++] = iter_dest_node_flag
;
1730 iter_node_flag
= iter_dest_node_flag
;
1734 /* Publish branch */
1737 * Attaching to root node.
1739 rcu_assign_pointer(ja
->root
, iter_node_flag
);
1743 iter_key
= (uint8_t) (key
>> (JA_BITS_PER_BYTE
* (ja
->tree_depth
- level
)));
1744 dbg_printf("publish branch at level %d, key %u\n",
1745 level
- 1, (unsigned int) iter_key
);
1746 /* We need to use set_nth on the previous level. */
1747 iter_dest_node_flag
= attach_node_flag
;
1748 ret
= ja_node_set_nth(ja
, &iter_dest_node_flag
,
1751 shadow_node
, level
- 1);
1753 dbg_printf("branch publish error %d\n", ret
);
1759 rcu_assign_pointer(*attach_node_flag_ptr
, iter_dest_node_flag
);
1767 for (i
= 0; i
< nr_created_nodes
; i
++) {
1771 flags
= RCUJA_SHADOW_CLEAR_FREE_LOCK
;
1773 flags
|= RCUJA_SHADOW_CLEAR_FREE_NODE
;
1774 tmpret
= rcuja_shadow_clear(ja
->ht
,
1782 if (parent_shadow_node
)
1783 rcuja_shadow_unlock(parent_shadow_node
);
1786 rcuja_shadow_unlock(shadow_node
);
1792 * Lock the parent containing the hlist head pointer, and add node to list of
1793 * duplicates. Failure can happen if concurrent update changes the
1794 * parent before we get the lock. We return -EAGAIN in that case.
1795 * Return 0 on success, negative error value on failure.
1798 int ja_chain_node(struct cds_ja
*ja
,
1799 struct cds_ja_inode_flag
*parent_node_flag
,
1800 struct cds_ja_inode_flag
**node_flag_ptr
,
1801 struct cds_ja_inode_flag
*node_flag
,
1802 struct cds_ja_node
*node
)
1804 struct cds_ja_shadow_node
*shadow_node
;
1807 shadow_node
= rcuja_shadow_lookup_lock(ja
->ht
, parent_node_flag
);
1811 if (ja_node_ptr(*node_flag_ptr
) != ja_node_ptr(node_flag
)) {
1815 cds_hlist_add_head_rcu(&node
->list
, (struct cds_hlist_head
*) node_flag_ptr
);
1817 rcuja_shadow_unlock(shadow_node
);
1822 int _cds_ja_add(struct cds_ja
*ja
, uint64_t key
,
1823 struct cds_ja_node
*new_node
,
1824 struct cds_ja_node
**unique_node_ret
)
1826 unsigned int tree_depth
, i
;
1827 struct cds_ja_inode_flag
*attach_node_flag
,
1831 *parent_attach_node_flag
;
1832 struct cds_ja_inode_flag
**attach_node_flag_ptr
,
1833 **parent_node_flag_ptr
,
1837 if (caa_unlikely(key
> ja
->key_max
)) {
1840 tree_depth
= ja
->tree_depth
;
1843 dbg_printf("cds_ja_add attempt: key %" PRIu64
", node %p\n",
1845 parent2_node_flag
= NULL
;
1847 (struct cds_ja_inode_flag
*) &ja
->root
; /* Use root ptr address as key for mutex */
1848 parent_node_flag_ptr
= NULL
;
1849 node_flag
= rcu_dereference(ja
->root
);
1850 node_flag_ptr
= &ja
->root
;
1852 /* Iterate on all internal levels */
1853 for (i
= 1; i
< tree_depth
; i
++) {
1856 if (!ja_node_ptr(node_flag
))
1858 dbg_printf("cds_ja_add iter parent2_node_flag %p parent_node_flag %p node_flag_ptr %p node_flag %p\n",
1859 parent2_node_flag
, parent_node_flag
, node_flag_ptr
, node_flag
);
1860 iter_key
= (uint8_t) (key
>> (JA_BITS_PER_BYTE
* (tree_depth
- i
- 1)));
1861 parent2_node_flag
= parent_node_flag
;
1862 parent_node_flag
= node_flag
;
1863 parent_node_flag_ptr
= node_flag_ptr
;
1864 node_flag
= ja_node_get_nth(node_flag
,
1870 * We reached either bottom of tree or internal NULL node,
1871 * simply add node to last internal level, or chain it if key is
1874 if (!ja_node_ptr(node_flag
)) {
1875 dbg_printf("cds_ja_add NULL parent2_node_flag %p parent_node_flag %p node_flag_ptr %p node_flag %p\n",
1876 parent2_node_flag
, parent_node_flag
, node_flag_ptr
, node_flag
);
1878 attach_node_flag
= parent_node_flag
;
1879 attach_node_flag_ptr
= parent_node_flag_ptr
;
1880 parent_attach_node_flag
= parent2_node_flag
;
1882 ret
= ja_attach_node(ja
, attach_node_flag_ptr
,
1884 parent_attach_node_flag
,
1889 if (unique_node_ret
) {
1890 *unique_node_ret
= (struct cds_ja_node
*) ja_node_ptr(node_flag
);
1894 dbg_printf("cds_ja_add duplicate parent2_node_flag %p parent_node_flag %p node_flag_ptr %p node_flag %p\n",
1895 parent2_node_flag
, parent_node_flag
, node_flag_ptr
, node_flag
);
1897 attach_node_flag
= node_flag
;
1898 attach_node_flag_ptr
= node_flag_ptr
;
1899 parent_attach_node_flag
= parent_node_flag
;
1901 ret
= ja_chain_node(ja
,
1902 parent_attach_node_flag
,
1903 attach_node_flag_ptr
,
1907 if (ret
== -EAGAIN
|| ret
== -EEXIST
)
1913 int cds_ja_add(struct cds_ja
*ja
, uint64_t key
,
1914 struct cds_ja_node
*new_node
)
1916 return _cds_ja_add(ja
, key
, new_node
, NULL
);
1919 struct cds_ja_node
*cds_ja_add_unique(struct cds_ja
*ja
, uint64_t key
,
1920 struct cds_ja_node
*new_node
)
1923 struct cds_ja_node
*ret_node
;
1925 ret
= _cds_ja_add(ja
, key
, new_node
, &ret_node
);
1933 * Note: there is no need to lookup the pointer address associated with
1934 * each node's nth item after taking the lock: it's already been done by
1935 * cds_ja_del while holding the rcu read-side lock, and our node rules
1936 * ensure that when a match value -> pointer is found in a node, it is
1937 * _NEVER_ changed for that node without recompaction, and recompaction
1938 * reallocates the node.
1939 * However, when a child is removed from "linear" nodes, its pointer
1940 * is set to NULL. We therefore check, while holding the locks, if this
1941 * pointer is NULL, and return -ENOENT to the caller if it is the case.
1944 int ja_detach_node(struct cds_ja
*ja
,
1945 struct cds_ja_inode_flag
**snapshot
,
1946 struct cds_ja_inode_flag
***snapshot_ptr
,
1947 uint8_t *snapshot_n
,
1950 struct cds_ja_node
*node
)
1952 struct cds_ja_shadow_node
*shadow_nodes
[JA_MAX_DEPTH
];
1953 struct cds_ja_inode_flag
**node_flag_ptr
= NULL
,
1954 *parent_node_flag
= NULL
,
1955 **parent_node_flag_ptr
= NULL
;
1956 struct cds_ja_inode_flag
*iter_node_flag
;
1957 int ret
, i
, nr_shadow
= 0, nr_clear
= 0, nr_branch
= 0;
1960 assert(nr_snapshot
== ja
->tree_depth
+ 1);
1963 * From the last internal level node going up, get the node
1964 * lock, check if the node has only one child left. If it is the
1965 * case, we continue iterating upward. When we reach a node
1966 * which has more that one child left, we lock the parent, and
1967 * proceed to the node deletion (removing its children too).
1969 for (i
= nr_snapshot
- 2; i
>= 1; i
--) {
1970 struct cds_ja_shadow_node
*shadow_node
;
1972 shadow_node
= rcuja_shadow_lookup_lock(ja
->ht
,
1978 shadow_nodes
[nr_shadow
++] = shadow_node
;
1981 * Check if node has been removed between RCU
1982 * lookup and lock acquisition.
1984 assert(snapshot_ptr
[i
+ 1]);
1985 if (ja_node_ptr(*snapshot_ptr
[i
+ 1])
1986 != ja_node_ptr(snapshot
[i
+ 1])) {
1991 assert(shadow_node
->nr_child
> 0);
1992 if (shadow_node
->nr_child
== 1 && i
> 1)
1995 if (shadow_node
->nr_child
> 1 || i
== 1) {
1996 /* Lock parent and break */
1997 shadow_node
= rcuja_shadow_lookup_lock(ja
->ht
,
2003 shadow_nodes
[nr_shadow
++] = shadow_node
;
2006 * Check if node has been removed between RCU
2007 * lookup and lock acquisition.
2009 assert(snapshot_ptr
[i
]);
2010 if (ja_node_ptr(*snapshot_ptr
[i
])
2011 != ja_node_ptr(snapshot
[i
])) {
2016 node_flag_ptr
= snapshot_ptr
[i
+ 1];
2017 n
= snapshot_n
[i
+ 1];
2018 parent_node_flag_ptr
= snapshot_ptr
[i
];
2019 parent_node_flag
= snapshot
[i
];
2023 * Lock parent's parent, in case we need
2024 * to recompact parent.
2026 shadow_node
= rcuja_shadow_lookup_lock(ja
->ht
,
2032 shadow_nodes
[nr_shadow
++] = shadow_node
;
2035 * Check if node has been removed between RCU
2036 * lookup and lock acquisition.
2038 assert(snapshot_ptr
[i
- 1]);
2039 if (ja_node_ptr(*snapshot_ptr
[i
- 1])
2040 != ja_node_ptr(snapshot
[i
- 1])) {
2051 * At this point, we want to delete all nodes that are about to
2052 * be removed from shadow_nodes (except the last one, which is
2053 * either the root or the parent of the upmost node with 1
2054 * child). OK to free lock here, because RCU read lock is held,
2055 * and free only performed in call_rcu.
2058 for (i
= 0; i
< nr_clear
; i
++) {
2059 ret
= rcuja_shadow_clear(ja
->ht
,
2060 shadow_nodes
[i
]->node_flag
,
2062 RCUJA_SHADOW_CLEAR_FREE_NODE
2063 | RCUJA_SHADOW_CLEAR_FREE_LOCK
);
2067 iter_node_flag
= parent_node_flag
;
2068 /* Remove from parent */
2069 ret
= ja_node_clear_ptr(ja
,
2070 node_flag_ptr
, /* Pointer to location to nullify */
2071 &iter_node_flag
, /* Old new parent ptr in its parent */
2072 shadow_nodes
[nr_branch
- 1], /* of parent */
2077 dbg_printf("ja_detach_node: publish %p instead of %p\n",
2078 iter_node_flag
, *parent_node_flag_ptr
);
2079 /* Update address of parent ptr in its parent */
2080 rcu_assign_pointer(*parent_node_flag_ptr
, iter_node_flag
);
2083 for (i
= 0; i
< nr_shadow
; i
++)
2084 rcuja_shadow_unlock(shadow_nodes
[i
]);
2089 int ja_unchain_node(struct cds_ja
*ja
,
2090 struct cds_ja_inode_flag
*parent_node_flag
,
2091 struct cds_ja_inode_flag
**node_flag_ptr
,
2092 struct cds_ja_inode_flag
*node_flag
,
2093 struct cds_ja_node
*node
)
2095 struct cds_ja_shadow_node
*shadow_node
;
2096 struct cds_hlist_node
*hlist_node
;
2097 struct cds_hlist_head hlist_head
;
2098 int ret
= 0, count
= 0, found
= 0;
2100 shadow_node
= rcuja_shadow_lookup_lock(ja
->ht
, parent_node_flag
);
2103 if (ja_node_ptr(*node_flag_ptr
) != ja_node_ptr(node_flag
)) {
2107 hlist_head
.next
= (struct cds_hlist_node
*) ja_node_ptr(node_flag
);
2109 * Retry if another thread removed all but one of duplicates
2110 * since check (this check was performed without lock).
2111 * Ensure that the node we are about to remove is still in the
2112 * list (while holding lock).
2114 cds_hlist_for_each_rcu(hlist_node
, &hlist_head
) {
2116 /* FIXME: currently a work-around */
2117 hlist_node
->prev
= (struct cds_hlist_node
*) node_flag_ptr
;
2120 if (hlist_node
== &node
->list
)
2124 if (!found
|| count
== 1) {
2128 cds_hlist_del_rcu(&node
->list
);
2130 * Validate that we indeed removed the node from linked list.
2132 assert(ja_node_ptr(*node_flag_ptr
) != (struct cds_ja_inode
*) node
);
2134 rcuja_shadow_unlock(shadow_node
);
2139 * Called with RCU read lock held.
2141 int cds_ja_del(struct cds_ja
*ja
, uint64_t key
,
2142 struct cds_ja_node
*node
)
2144 unsigned int tree_depth
, i
;
2145 struct cds_ja_inode_flag
*snapshot
[JA_MAX_DEPTH
];
2146 struct cds_ja_inode_flag
**snapshot_ptr
[JA_MAX_DEPTH
];
2147 uint8_t snapshot_n
[JA_MAX_DEPTH
];
2148 struct cds_ja_inode_flag
*node_flag
;
2149 struct cds_ja_inode_flag
**prev_node_flag_ptr
,
2154 if (caa_unlikely(key
> ja
->key_max
))
2156 tree_depth
= ja
->tree_depth
;
2160 dbg_printf("cds_ja_del attempt: key %" PRIu64
", node %p\n",
2163 /* snapshot for level 0 is only for shadow node lookup */
2166 snapshot_ptr
[nr_snapshot
] = NULL
;
2167 snapshot
[nr_snapshot
++] = (struct cds_ja_inode_flag
*) &ja
->root
;
2168 node_flag
= rcu_dereference(ja
->root
);
2169 prev_node_flag_ptr
= &ja
->root
;
2170 node_flag_ptr
= &ja
->root
;
2172 /* Iterate on all internal levels */
2173 for (i
= 1; i
< tree_depth
; i
++) {
2176 dbg_printf("cds_ja_del iter node_flag %p\n",
2178 if (!ja_node_ptr(node_flag
)) {
2181 iter_key
= (uint8_t) (key
>> (JA_BITS_PER_BYTE
* (tree_depth
- i
- 1)));
2182 snapshot_n
[nr_snapshot
+ 1] = iter_key
;
2183 snapshot_ptr
[nr_snapshot
] = prev_node_flag_ptr
;
2184 snapshot
[nr_snapshot
++] = node_flag
;
2185 node_flag
= ja_node_get_nth(node_flag
,
2189 prev_node_flag_ptr
= node_flag_ptr
;
2190 dbg_printf("cds_ja_del iter key lookup %u finds node_flag %p, prev_node_flag_ptr %p\n",
2191 (unsigned int) iter_key
, node_flag
,
2192 prev_node_flag_ptr
);
2195 * We reached bottom of tree, try to find the node we are trying
2196 * to remove. Fail if we cannot find it.
2198 if (!ja_node_ptr(node_flag
)) {
2199 dbg_printf("cds_ja_del: no node found for key %" PRIu64
"\n",
2203 struct cds_hlist_head hlist_head
;
2204 struct cds_hlist_node
*hlist_node
;
2205 struct cds_ja_node
*entry
, *match
= NULL
;
2209 (struct cds_hlist_node
*) ja_node_ptr(node_flag
);
2210 cds_hlist_for_each_entry_rcu(entry
,
2214 dbg_printf("cds_ja_del: compare %p with entry %p\n", node
, entry
);
2220 dbg_printf("cds_ja_del: no node match for node %p key %" PRIu64
"\n", node
, key
);
2226 * Removing last of duplicates. Last snapshot
2227 * does not have a shadow node (external leafs).
2229 snapshot_ptr
[nr_snapshot
] = prev_node_flag_ptr
;
2230 snapshot
[nr_snapshot
++] = node_flag
;
2231 ret
= ja_detach_node(ja
, snapshot
, snapshot_ptr
,
2232 snapshot_n
, nr_snapshot
, key
, node
);
2234 ret
= ja_unchain_node(ja
, snapshot
[nr_snapshot
- 1],
2235 node_flag_ptr
, node_flag
, match
);
2239 * Explanation of -ENOENT handling: caused by concurrent delete
2240 * between RCU lookup and actual removal. Need to re-do the
2241 * lookup and removal attempt.
2243 if (ret
== -EAGAIN
|| ret
== -ENOENT
)
2248 struct cds_ja
*_cds_ja_new(unsigned int key_bits
,
2249 const struct rcu_flavor_struct
*flavor
)
2253 struct cds_ja_shadow_node
*root_shadow_node
;
2255 ja
= calloc(sizeof(*ja
), 1);
2267 ja
->key_max
= (1ULL << key_bits
) - 1;
2270 ja
->key_max
= UINT64_MAX
;
2276 /* ja->root is NULL */
2277 /* tree_depth 0 is for pointer to root node */
2278 ja
->tree_depth
= (key_bits
>> JA_LOG2_BITS_PER_BYTE
) + 1;
2279 assert(ja
->tree_depth
<= JA_MAX_DEPTH
);
2280 ja
->ht
= rcuja_create_ht(flavor
);
2285 * Note: we should not free this node until judy array destroy.
2287 root_shadow_node
= rcuja_shadow_set(ja
->ht
,
2288 (struct cds_ja_inode_flag
*) &ja
->root
,
2290 if (!root_shadow_node
) {
2298 ret
= rcuja_delete_ht(ja
->ht
);
2308 * Called from RCU read-side CS.
2310 __attribute__((visibility("protected")))
2311 void rcuja_free_all_children(struct cds_ja_shadow_node
*shadow_node
,
2312 struct cds_ja_inode_flag
*node_flag
,
2313 void (*free_node_cb
)(struct rcu_head
*head
))
2315 const struct rcu_flavor_struct
*flavor
;
2316 unsigned int type_index
;
2317 struct cds_ja_inode
*node
;
2318 const struct cds_ja_type
*type
;
2320 flavor
= cds_lfht_rcu_flavor(shadow_node
->ja
->ht
);
2321 node
= ja_node_ptr(node_flag
);
2322 assert(node
!= NULL
);
2323 type_index
= ja_node_type(node_flag
);
2324 type
= &ja_types
[type_index
];
2326 switch (type
->type_class
) {
2330 ja_linear_node_get_nr_child(type
, node
);
2333 for (i
= 0; i
< nr_child
; i
++) {
2334 struct cds_ja_inode_flag
*iter
;
2335 struct cds_hlist_head head
;
2336 struct cds_ja_node
*entry
;
2337 struct cds_hlist_node
*pos
, *tmp
;
2340 ja_linear_node_get_ith_pos(type
, node
, i
, &v
, &iter
);
2343 head
.next
= (struct cds_hlist_node
*) iter
;
2344 cds_hlist_for_each_entry_safe(entry
, pos
, tmp
, &head
, list
) {
2345 flavor
->update_call_rcu(&entry
->head
, free_node_cb
);
2352 unsigned int pool_nr
;
2354 for (pool_nr
= 0; pool_nr
< (1U << type
->nr_pool_order
); pool_nr
++) {
2355 struct cds_ja_inode
*pool
=
2356 ja_pool_node_get_ith_pool(type
, node
, pool_nr
);
2358 ja_linear_node_get_nr_child(type
, pool
);
2361 for (j
= 0; j
< nr_child
; j
++) {
2362 struct cds_ja_inode_flag
*iter
;
2363 struct cds_hlist_head head
;
2364 struct cds_ja_node
*entry
;
2365 struct cds_hlist_node
*pos
, *tmp
;
2368 ja_linear_node_get_ith_pos(type
, pool
, j
, &v
, &iter
);
2371 head
.next
= (struct cds_hlist_node
*) iter
;
2372 cds_hlist_for_each_entry_safe(entry
, pos
, tmp
, &head
, list
) {
2373 flavor
->update_call_rcu(&entry
->head
, free_node_cb
);
2385 for (i
= 0; i
< JA_ENTRY_PER_NODE
; i
++) {
2386 struct cds_ja_inode_flag
*iter
;
2387 struct cds_hlist_head head
;
2388 struct cds_ja_node
*entry
;
2389 struct cds_hlist_node
*pos
, *tmp
;
2391 iter
= ja_pigeon_node_get_ith_pos(type
, node
, i
);
2394 head
.next
= (struct cds_hlist_node
*) iter
;
2395 cds_hlist_for_each_entry_safe(entry
, pos
, tmp
, &head
, list
) {
2396 flavor
->update_call_rcu(&entry
->head
, free_node_cb
);
2407 void print_debug_fallback_distribution(void)
2411 fprintf(stderr
, "Fallback node distribution:\n");
2412 for (i
= 0; i
< JA_ENTRY_PER_NODE
; i
++) {
2413 if (!node_fallback_count_distribution
[i
])
2415 fprintf(stderr
, " %3u: %4lu\n",
2416 i
, node_fallback_count_distribution
[i
]);
2421 * There should be no more concurrent add to the judy array while it is
2422 * being destroyed (ensured by the caller).
2424 int cds_ja_destroy(struct cds_ja
*ja
,
2425 void (*free_node_cb
)(struct rcu_head
*head
))
2427 const struct rcu_flavor_struct
*flavor
;
2430 flavor
= cds_lfht_rcu_flavor(ja
->ht
);
2431 rcuja_shadow_prune(ja
->ht
,
2432 RCUJA_SHADOW_CLEAR_FREE_NODE
| RCUJA_SHADOW_CLEAR_FREE_LOCK
,
2434 flavor
->thread_offline();
2435 ret
= rcuja_delete_ht(ja
->ht
);
2438 flavor
->thread_online();
2439 if (uatomic_read(&ja
->nr_fallback
))
2441 "[warning] RCU Judy Array used %lu fallback node(s)\n",
2442 uatomic_read(&ja
->nr_fallback
));
2443 fprintf(stderr
, "Nodes allocated: %lu, Nodes freed: %lu. Fallback ratio: %g\n",
2444 uatomic_read(&nr_nodes_allocated
),
2445 uatomic_read(&nr_nodes_freed
),
2446 (double) uatomic_read(&ja
->nr_fallback
) / (double) uatomic_read(&nr_nodes_allocated
));
2447 print_debug_fallback_distribution();