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
];
248 struct cds_ja_inode
*_ja_node_mask_ptr(struct cds_ja_inode_flag
*node
)
250 return (struct cds_ja_inode
*) (((unsigned long) node
) & JA_PTR_MASK
);
253 unsigned long ja_node_type(struct cds_ja_inode_flag
*node
)
257 if (_ja_node_mask_ptr(node
) == NULL
) {
258 return NODE_INDEX_NULL
;
260 type
= (unsigned int) ((unsigned long) node
& JA_TYPE_MASK
);
261 assert(type
< (1UL << JA_TYPE_BITS
));
265 struct cds_ja_inode
*ja_node_ptr(struct cds_ja_inode_flag
*node
)
267 unsigned long type_index
= ja_node_type(node
);
268 const struct cds_ja_type
*type
;
270 type
= &ja_types
[type_index
];
271 switch (type
->type_class
) {
273 case RCU_JA_PIGEON
: /* fall-through */
274 case RCU_JA_NULL
: /* fall-through */
275 default: /* fall-through */
276 return _ja_node_mask_ptr(node
);
278 switch (type
->nr_pool_order
) {
280 return (struct cds_ja_inode
*) (((unsigned long) node
) & ~(JA_POOL_1D_MASK
| JA_TYPE_MASK
));
282 return (struct cds_ja_inode
*) (((unsigned long) node
) & ~(JA_POOL_2D_MASK
| JA_POOL_1D_MASK
| JA_TYPE_MASK
));
289 struct cds_ja_inode
*alloc_cds_ja_node(const struct cds_ja_type
*ja_type
)
291 size_t len
= 1U << ja_type
->order
;
295 ret
= posix_memalign(&p
, len
, len
);
303 void free_cds_ja_node(struct cds_ja_inode
*node
)
308 #define __JA_ALIGN_MASK(v, mask) (((v) + (mask)) & ~(mask))
309 #define JA_ALIGN(v, align) __JA_ALIGN_MASK(v, (typeof(v)) (align) - 1)
310 #define __JA_FLOOR_MASK(v, mask) ((v) & ~(mask))
311 #define JA_FLOOR(v, align) __JA_FLOOR_MASK(v, (typeof(v)) (align) - 1)
314 uint8_t *align_ptr_size(uint8_t *ptr
)
316 return (uint8_t *) JA_ALIGN((unsigned long) ptr
, sizeof(void *));
320 uint8_t ja_linear_node_get_nr_child(const struct cds_ja_type
*type
,
321 struct cds_ja_inode
*node
)
323 assert(type
->type_class
== RCU_JA_LINEAR
|| type
->type_class
== RCU_JA_POOL
);
324 return rcu_dereference(node
->u
.data
[0]);
328 * The order in which values and pointers are does does not matter: if
329 * a value is missing, we return NULL. If a value is there, but its
330 * associated pointers is still NULL, we return NULL too.
333 struct cds_ja_inode_flag
*ja_linear_node_get_nth(const struct cds_ja_type
*type
,
334 struct cds_ja_inode
*node
,
335 struct cds_ja_inode_flag
***child_node_flag_ptr
,
336 struct cds_ja_inode_flag
**child_node_flag_v
,
337 struct cds_ja_inode_flag
***node_flag_ptr
,
342 struct cds_ja_inode_flag
**pointers
;
343 struct cds_ja_inode_flag
*ptr
;
346 assert(type
->type_class
== RCU_JA_LINEAR
|| type
->type_class
== RCU_JA_POOL
);
348 nr_child
= ja_linear_node_get_nr_child(type
, node
);
349 cmm_smp_rmb(); /* read nr_child before values and pointers */
350 assert(nr_child
<= type
->max_linear_child
);
351 assert(type
->type_class
!= RCU_JA_LINEAR
|| nr_child
>= type
->min_child
);
353 values
= &node
->u
.data
[1];
354 for (i
= 0; i
< nr_child
; i
++) {
355 if (CMM_LOAD_SHARED(values
[i
]) == n
)
359 if (caa_unlikely(node_flag_ptr
))
360 *node_flag_ptr
= NULL
;
363 pointers
= (struct cds_ja_inode_flag
**) align_ptr_size(&values
[type
->max_linear_child
]);
364 ptr
= rcu_dereference(pointers
[i
]);
365 if (caa_unlikely(child_node_flag_ptr
) && ptr
)
366 *child_node_flag_ptr
= &pointers
[i
];
367 if (caa_unlikely(child_node_flag_v
) && ptr
)
368 *child_node_flag_v
= ptr
;
369 if (caa_unlikely(node_flag_ptr
))
370 *node_flag_ptr
= &pointers
[i
];
375 void ja_linear_node_get_ith_pos(const struct cds_ja_type
*type
,
376 struct cds_ja_inode
*node
,
379 struct cds_ja_inode_flag
**iter
)
382 struct cds_ja_inode_flag
**pointers
;
384 assert(type
->type_class
== RCU_JA_LINEAR
|| type
->type_class
== RCU_JA_POOL
);
385 assert(i
< ja_linear_node_get_nr_child(type
, node
));
387 values
= &node
->u
.data
[1];
389 pointers
= (struct cds_ja_inode_flag
**) align_ptr_size(&values
[type
->max_linear_child
]);
394 struct cds_ja_inode_flag
*ja_pool_node_get_nth(const struct cds_ja_type
*type
,
395 struct cds_ja_inode
*node
,
396 struct cds_ja_inode_flag
*node_flag
,
397 struct cds_ja_inode_flag
***child_node_flag_ptr
,
398 struct cds_ja_inode_flag
**child_node_flag_v
,
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
) & type
->nr_pool_order
;
414 linear
= (struct cds_ja_inode
*) &node
->u
.data
[index
<< type
->pool_size_order
];
420 * TODO: currently, we select the pool by highest bits. We
421 * should support various encodings.
423 linear
= (struct cds_ja_inode
*)
424 &node
->u
.data
[((unsigned long) n
>> (CHAR_BIT
- type
->nr_pool_order
)) << type
->pool_size_order
];
431 return ja_linear_node_get_nth(type
, linear
, child_node_flag_ptr
,
432 child_node_flag_v
, node_flag_ptr
, n
);
436 struct cds_ja_inode
*ja_pool_node_get_ith_pool(const struct cds_ja_type
*type
,
437 struct cds_ja_inode
*node
,
440 assert(type
->type_class
== RCU_JA_POOL
);
441 return (struct cds_ja_inode
*)
442 &node
->u
.data
[(unsigned int) i
<< type
->pool_size_order
];
446 struct cds_ja_inode_flag
*ja_pigeon_node_get_nth(const struct cds_ja_type
*type
,
447 struct cds_ja_inode
*node
,
448 struct cds_ja_inode_flag
***child_node_flag_ptr
,
449 struct cds_ja_inode_flag
**child_node_flag_v
,
450 struct cds_ja_inode_flag
***node_flag_ptr
,
453 struct cds_ja_inode_flag
**child_node_flag
;
454 struct cds_ja_inode_flag
*child_node_flag_read
;
456 assert(type
->type_class
== RCU_JA_PIGEON
);
457 child_node_flag
= &((struct cds_ja_inode_flag
**) node
->u
.data
)[n
];
458 child_node_flag_read
= rcu_dereference(*child_node_flag
);
459 dbg_printf("ja_pigeon_node_get_nth child_node_flag_ptr %p\n",
461 if (caa_unlikely(child_node_flag_ptr
) && child_node_flag_read
)
462 *child_node_flag_ptr
= child_node_flag
;
463 if (caa_unlikely(child_node_flag_v
) && child_node_flag_read
)
464 *child_node_flag_v
= child_node_flag_read
;
465 if (caa_unlikely(node_flag_ptr
))
466 *node_flag_ptr
= child_node_flag
;
467 return child_node_flag_read
;
471 struct cds_ja_inode_flag
*ja_pigeon_node_get_ith_pos(const struct cds_ja_type
*type
,
472 struct cds_ja_inode
*node
,
475 return ja_pigeon_node_get_nth(type
, node
, NULL
, NULL
, NULL
, i
);
479 * ja_node_get_nth: get nth item from a node.
480 * node_flag is already rcu_dereference'd.
483 struct cds_ja_inode_flag
*ja_node_get_nth(struct cds_ja_inode_flag
*node_flag
,
484 struct cds_ja_inode_flag
***child_node_flag_ptr
,
485 struct cds_ja_inode_flag
**child_node_flag
,
486 struct cds_ja_inode_flag
***node_flag_ptr
,
489 unsigned int type_index
;
490 struct cds_ja_inode
*node
;
491 const struct cds_ja_type
*type
;
493 node
= ja_node_ptr(node_flag
);
494 assert(node
!= NULL
);
495 type_index
= ja_node_type(node_flag
);
496 type
= &ja_types
[type_index
];
498 switch (type
->type_class
) {
500 return ja_linear_node_get_nth(type
, node
,
501 child_node_flag_ptr
, child_node_flag
,
504 return ja_pool_node_get_nth(type
, node
, node_flag
,
505 child_node_flag_ptr
, child_node_flag
,
508 return ja_pigeon_node_get_nth(type
, node
,
509 child_node_flag_ptr
, child_node_flag
,
513 return (void *) -1UL;
518 int ja_linear_node_set_nth(const struct cds_ja_type
*type
,
519 struct cds_ja_inode
*node
,
520 struct cds_ja_shadow_node
*shadow_node
,
522 struct cds_ja_inode_flag
*child_node_flag
)
525 uint8_t *values
, *nr_child_ptr
;
526 struct cds_ja_inode_flag
**pointers
;
527 unsigned int i
, unused
= 0;
529 assert(type
->type_class
== RCU_JA_LINEAR
|| type
->type_class
== RCU_JA_POOL
);
531 nr_child_ptr
= &node
->u
.data
[0];
532 dbg_printf("linear set nth: nr_child_ptr %p\n", nr_child_ptr
);
533 nr_child
= *nr_child_ptr
;
534 assert(nr_child
<= type
->max_linear_child
);
536 values
= &node
->u
.data
[1];
537 pointers
= (struct cds_ja_inode_flag
**) align_ptr_size(&values
[type
->max_linear_child
]);
538 /* Check if node value is already populated */
539 for (i
= 0; i
< nr_child
; i
++) {
540 if (values
[i
] == n
) {
550 if (i
== nr_child
&& nr_child
>= type
->max_linear_child
) {
552 return -ERANGE
; /* recompact node */
554 return -ENOSPC
; /* No space left in this node type */
557 assert(pointers
[i
] == NULL
);
558 rcu_assign_pointer(pointers
[i
], child_node_flag
);
559 /* If we expanded the nr_child, increment it */
561 CMM_STORE_SHARED(values
[nr_child
], n
);
562 /* write pointer and value before nr_child */
564 CMM_STORE_SHARED(*nr_child_ptr
, nr_child
+ 1);
566 shadow_node
->nr_child
++;
567 dbg_printf("linear set nth: %u child, shadow: %u child, for node %p shadow %p\n",
568 (unsigned int) CMM_LOAD_SHARED(*nr_child_ptr
),
569 (unsigned int) shadow_node
->nr_child
,
576 int ja_pool_node_set_nth(const struct cds_ja_type
*type
,
577 struct cds_ja_inode
*node
,
578 struct cds_ja_inode_flag
*node_flag
,
579 struct cds_ja_shadow_node
*shadow_node
,
581 struct cds_ja_inode_flag
*child_node_flag
)
583 struct cds_ja_inode
*linear
;
585 assert(type
->type_class
== RCU_JA_POOL
);
587 switch (type
->nr_pool_order
) {
590 unsigned long bitsel
, index
;
592 bitsel
= ja_node_pool_1d_bitsel(node_flag
);
593 assert(bitsel
< CHAR_BIT
);
594 index
= ((unsigned long) n
>> bitsel
) & type
->nr_pool_order
;
595 linear
= (struct cds_ja_inode
*) &node
->u
.data
[index
<< type
->pool_size_order
];
601 * TODO: currently, we select the pool by highest bits. We
602 * should support various encodings.
604 linear
= (struct cds_ja_inode
*)
605 &node
->u
.data
[((unsigned long) n
>> (CHAR_BIT
- type
->nr_pool_order
)) << type
->pool_size_order
];
613 return ja_linear_node_set_nth(type
, linear
, shadow_node
,
618 int ja_pigeon_node_set_nth(const struct cds_ja_type
*type
,
619 struct cds_ja_inode
*node
,
620 struct cds_ja_shadow_node
*shadow_node
,
622 struct cds_ja_inode_flag
*child_node_flag
)
624 struct cds_ja_inode_flag
**ptr
;
626 assert(type
->type_class
== RCU_JA_PIGEON
);
627 ptr
= &((struct cds_ja_inode_flag
**) node
->u
.data
)[n
];
630 rcu_assign_pointer(*ptr
, child_node_flag
);
631 shadow_node
->nr_child
++;
636 * _ja_node_set_nth: set nth item within a node. Return an error
637 * (negative error value) if it is already there.
640 int _ja_node_set_nth(const struct cds_ja_type
*type
,
641 struct cds_ja_inode
*node
,
642 struct cds_ja_inode_flag
*node_flag
,
643 struct cds_ja_shadow_node
*shadow_node
,
645 struct cds_ja_inode_flag
*child_node_flag
)
647 switch (type
->type_class
) {
649 return ja_linear_node_set_nth(type
, node
, shadow_node
, n
,
652 return ja_pool_node_set_nth(type
, node
, node_flag
, shadow_node
, n
,
655 return ja_pigeon_node_set_nth(type
, node
, shadow_node
, n
,
668 int ja_linear_node_clear_ptr(const struct cds_ja_type
*type
,
669 struct cds_ja_inode
*node
,
670 struct cds_ja_shadow_node
*shadow_node
,
671 struct cds_ja_inode_flag
**node_flag_ptr
)
674 uint8_t *nr_child_ptr
;
676 assert(type
->type_class
== RCU_JA_LINEAR
|| type
->type_class
== RCU_JA_POOL
);
678 nr_child_ptr
= &node
->u
.data
[0];
679 dbg_printf("linear clear ptr: nr_child_ptr %p\n", nr_child_ptr
);
680 nr_child
= *nr_child_ptr
;
681 assert(nr_child
<= type
->max_linear_child
);
683 if (shadow_node
->fallback_removal_count
) {
684 shadow_node
->fallback_removal_count
--;
686 if (shadow_node
->nr_child
<= type
->min_child
) {
687 /* We need to try recompacting the node */
691 assert(*node_flag_ptr
!= NULL
);
692 rcu_assign_pointer(*node_flag_ptr
, NULL
);
694 * Value and nr_child are never changed (would cause ABA issue).
695 * Instead, we leave the pointer to NULL and recompact the node
696 * once in a while. It is allowed to set a NULL pointer to a new
697 * value without recompaction though.
698 * Only update the shadow node accounting.
700 shadow_node
->nr_child
--;
701 dbg_printf("linear clear ptr: %u child, shadow: %u child, for node %p shadow %p\n",
702 (unsigned int) CMM_LOAD_SHARED(*nr_child_ptr
),
703 (unsigned int) shadow_node
->nr_child
,
710 int ja_pool_node_clear_ptr(const struct cds_ja_type
*type
,
711 struct cds_ja_inode
*node
,
712 struct cds_ja_shadow_node
*shadow_node
,
713 struct cds_ja_inode_flag
**node_flag_ptr
,
716 struct cds_ja_inode
*linear
;
718 assert(type
->type_class
== RCU_JA_POOL
);
719 linear
= (struct cds_ja_inode
*)
720 &node
->u
.data
[((unsigned long) n
>> (CHAR_BIT
- type
->nr_pool_order
)) << type
->pool_size_order
];
721 return ja_linear_node_clear_ptr(type
, linear
, shadow_node
, node_flag_ptr
);
725 int ja_pigeon_node_clear_ptr(const struct cds_ja_type
*type
,
726 struct cds_ja_inode
*node
,
727 struct cds_ja_shadow_node
*shadow_node
,
728 struct cds_ja_inode_flag
**node_flag_ptr
)
730 assert(type
->type_class
== RCU_JA_PIGEON
);
731 dbg_printf("ja_pigeon_node_clear_ptr: clearing ptr: %p\n", *node_flag_ptr
);
732 rcu_assign_pointer(*node_flag_ptr
, NULL
);
733 shadow_node
->nr_child
--;
738 * _ja_node_clear_ptr: clear ptr item within a node. Return an error
739 * (negative error value) if it is not found (-ENOENT).
742 int _ja_node_clear_ptr(const struct cds_ja_type
*type
,
743 struct cds_ja_inode
*node
,
744 struct cds_ja_shadow_node
*shadow_node
,
745 struct cds_ja_inode_flag
**node_flag_ptr
,
748 switch (type
->type_class
) {
750 return ja_linear_node_clear_ptr(type
, node
, shadow_node
, node_flag_ptr
);
752 return ja_pool_node_clear_ptr(type
, node
, shadow_node
, node_flag_ptr
, n
);
754 return ja_pigeon_node_clear_ptr(type
, node
, shadow_node
, node_flag_ptr
);
766 * Calculate bit distribution. Returns the bit (0 to 7) that splits the
767 * distribution in two sub-distributions containing as much elements one
768 * compared to the other.
771 unsigned int ja_node_sum_distribution_1d(enum ja_recompact mode
,
773 unsigned int type_index
,
774 const struct cds_ja_type
*type
,
775 struct cds_ja_inode
*node
,
776 struct cds_ja_shadow_node
*shadow_node
,
778 struct cds_ja_inode_flag
*child_node_flag
,
779 struct cds_ja_inode_flag
**nullify_node_flag_ptr
)
781 uint8_t nr_one
[JA_BITS_PER_BYTE
];
782 unsigned int bitsel
= 0, bit_i
, overall_best_distance
= UINT_MAX
;
783 unsigned int distrib_nr_child
= 0;
785 memset(nr_one
, 0, sizeof(nr_one
));
787 switch (type
->type_class
) {
791 ja_linear_node_get_nr_child(type
, node
);
794 for (i
= 0; i
< nr_child
; i
++) {
795 struct cds_ja_inode_flag
*iter
;
798 ja_linear_node_get_ith_pos(type
, node
, i
, &v
, &iter
);
801 if (mode
== JA_RECOMPACT_DEL
&& *nullify_node_flag_ptr
== iter
)
803 for (bit_i
= 0; bit_i
< JA_BITS_PER_BYTE
; bit_i
++) {
804 if (v
& (1U << bit_i
))
813 unsigned int pool_nr
;
815 for (pool_nr
= 0; pool_nr
< (1U << type
->nr_pool_order
); pool_nr
++) {
816 struct cds_ja_inode
*pool
=
817 ja_pool_node_get_ith_pool(type
,
820 ja_linear_node_get_nr_child(type
, pool
);
823 for (j
= 0; j
< nr_child
; j
++) {
824 struct cds_ja_inode_flag
*iter
;
827 ja_linear_node_get_ith_pos(type
, pool
,
831 if (mode
== JA_RECOMPACT_DEL
&& *nullify_node_flag_ptr
== iter
)
833 for (bit_i
= 0; bit_i
< JA_BITS_PER_BYTE
; bit_i
++) {
834 if (v
& (1U << bit_i
))
847 assert(mode
== JA_RECOMPACT_DEL
);
848 nr_child
= shadow_node
->nr_child
;
849 for (i
= 0; i
< nr_child
; i
++) {
850 struct cds_ja_inode_flag
*iter
;
852 iter
= ja_pigeon_node_get_ith_pos(type
, node
, i
);
855 if (mode
== JA_RECOMPACT_DEL
&& *nullify_node_flag_ptr
== iter
)
857 for (bit_i
= 0; bit_i
< JA_BITS_PER_BYTE
; bit_i
++) {
858 if (i
& (1U << bit_i
))
866 assert(mode
== JA_RECOMPACT_ADD
);
873 if (mode
== JA_RECOMPACT_ADD
) {
874 for (bit_i
= 0; bit_i
< JA_BITS_PER_BYTE
; bit_i
++) {
875 if (n
& (1U << bit_i
))
882 * The best bit selector is that for which the number of ones is
883 * closest to half of the number of children in the
884 * distribution. We calculate the distance using the double of
885 * the sub-distribution sizes to eliminate truncation error.
887 for (bit_i
= 0; bit_i
< JA_BITS_PER_BYTE
; bit_i
++) {
888 unsigned int distance_to_best
;
890 distance_to_best
= abs_int((nr_one
[bit_i
] << 1U) - distrib_nr_child
);
891 if (distance_to_best
< overall_best_distance
) {
892 overall_best_distance
= distance_to_best
;
896 dbg_printf("1 dimension pool bit selection: (%u)\n", bitsel
);
901 * ja_node_recompact_add: recompact a node, adding a new child.
902 * Return 0 on success, -EAGAIN if need to retry, or other negative
903 * error value otherwise.
906 int ja_node_recompact(enum ja_recompact mode
,
908 unsigned int old_type_index
,
909 const struct cds_ja_type
*old_type
,
910 struct cds_ja_inode
*old_node
,
911 struct cds_ja_shadow_node
*shadow_node
,
912 struct cds_ja_inode_flag
**old_node_flag_ptr
, uint8_t n
,
913 struct cds_ja_inode_flag
*child_node_flag
,
914 struct cds_ja_inode_flag
**nullify_node_flag_ptr
)
916 unsigned int new_type_index
;
917 struct cds_ja_inode
*new_node
;
918 struct cds_ja_shadow_node
*new_shadow_node
= NULL
;
919 const struct cds_ja_type
*new_type
;
920 struct cds_ja_inode_flag
*new_node_flag
, *old_node_flag
;
924 old_node_flag
= *old_node_flag_ptr
;
928 new_type_index
= old_type_index
;
930 case JA_RECOMPACT_ADD
:
931 if (!shadow_node
|| old_type_index
== NODE_INDEX_NULL
) {
934 new_type_index
= old_type_index
+ 1;
937 case JA_RECOMPACT_DEL
:
938 if (old_type_index
== 0) {
939 new_type_index
= NODE_INDEX_NULL
;
941 new_type_index
= old_type_index
- 1;
948 retry
: /* for fallback */
949 dbg_printf("Recompact from type %d to type %d\n",
950 old_type_index
, new_type_index
);
951 new_type
= &ja_types
[new_type_index
];
952 if (new_type_index
!= NODE_INDEX_NULL
) {
953 new_node
= alloc_cds_ja_node(new_type
);
957 if (new_type
->type_class
== RCU_JA_POOL
) {
958 switch (new_type
->nr_pool_order
) {
961 unsigned int node_distrib_bitsel
= 0;
962 node_distrib_bitsel
=
963 ja_node_sum_distribution_1d(mode
, ja
,
964 old_type_index
, old_type
,
965 old_node
, shadow_node
,
967 nullify_node_flag_ptr
);
968 assert(!((unsigned long) new_node
& JA_POOL_1D_MASK
));
969 new_node_flag
= ja_node_flag_pool_1d(new_node
,
970 new_type_index
, node_distrib_bitsel
);
975 /* TODO: pool order 2 in 2d */
976 assert(!((unsigned long) new_node
& JA_POOL_1D_MASK
));
977 assert(!((unsigned long) new_node
& JA_POOL_2D_MASK
));
978 new_node_flag
= ja_node_flag(new_node
, new_type_index
);
985 new_node_flag
= ja_node_flag(new_node
, new_type_index
);
988 dbg_printf("Recompact inherit lock from %p\n", shadow_node
);
989 new_shadow_node
= rcuja_shadow_set(ja
->ht
, new_node_flag
, shadow_node
, ja
);
990 if (!new_shadow_node
) {
995 new_shadow_node
->fallback_removal_count
=
996 JA_FALLBACK_REMOVAL_COUNT
;
999 new_node_flag
= NULL
;
1002 assert(mode
!= JA_RECOMPACT_ADD
|| old_type
->type_class
!= RCU_JA_PIGEON
);
1004 if (new_type_index
== NODE_INDEX_NULL
)
1007 switch (old_type
->type_class
) {
1011 ja_linear_node_get_nr_child(old_type
, old_node
);
1014 for (i
= 0; i
< nr_child
; i
++) {
1015 struct cds_ja_inode_flag
*iter
;
1018 ja_linear_node_get_ith_pos(old_type
, old_node
, i
, &v
, &iter
);
1021 if (mode
== JA_RECOMPACT_DEL
&& *nullify_node_flag_ptr
== iter
)
1023 ret
= _ja_node_set_nth(new_type
, new_node
, new_node_flag
,
1026 if (new_type
->type_class
== RCU_JA_POOL
&& ret
) {
1027 goto fallback_toosmall
;
1035 unsigned int pool_nr
;
1037 for (pool_nr
= 0; pool_nr
< (1U << old_type
->nr_pool_order
); pool_nr
++) {
1038 struct cds_ja_inode
*pool
=
1039 ja_pool_node_get_ith_pool(old_type
,
1042 ja_linear_node_get_nr_child(old_type
, pool
);
1045 for (j
= 0; j
< nr_child
; j
++) {
1046 struct cds_ja_inode_flag
*iter
;
1049 ja_linear_node_get_ith_pos(old_type
, pool
,
1053 if (mode
== JA_RECOMPACT_DEL
&& *nullify_node_flag_ptr
== iter
)
1055 ret
= _ja_node_set_nth(new_type
, new_node
, new_node_flag
,
1058 if (new_type
->type_class
== RCU_JA_POOL
1060 goto fallback_toosmall
;
1068 assert(mode
== JA_RECOMPACT_ADD
);
1075 assert(mode
== JA_RECOMPACT_DEL
);
1076 nr_child
= shadow_node
->nr_child
;
1077 for (i
= 0; i
< nr_child
; i
++) {
1078 struct cds_ja_inode_flag
*iter
;
1080 iter
= ja_pigeon_node_get_ith_pos(old_type
, old_node
, i
);
1083 if (mode
== JA_RECOMPACT_DEL
&& *nullify_node_flag_ptr
== iter
)
1085 ret
= _ja_node_set_nth(new_type
, new_node
, new_node_flag
,
1088 if (new_type
->type_class
== RCU_JA_POOL
&& ret
) {
1089 goto fallback_toosmall
;
1102 if (mode
== JA_RECOMPACT_ADD
) {
1104 ret
= _ja_node_set_nth(new_type
, new_node
, new_node_flag
,
1106 n
, child_node_flag
);
1107 if (new_type
->type_class
== RCU_JA_POOL
&& ret
) {
1108 goto fallback_toosmall
;
1112 /* Return pointer to new recompacted node through old_node_flag_ptr */
1113 *old_node_flag_ptr
= new_node_flag
;
1117 flags
= RCUJA_SHADOW_CLEAR_FREE_NODE
;
1119 * It is OK to free the lock associated with a node
1120 * going to NULL, since we are holding the parent lock.
1121 * This synchronizes removal with re-add of that node.
1123 if (new_type_index
== NODE_INDEX_NULL
)
1124 flags
= RCUJA_SHADOW_CLEAR_FREE_LOCK
;
1125 ret
= rcuja_shadow_clear(ja
->ht
, old_node_flag
, shadow_node
,
1135 /* fallback if next pool is too small */
1136 assert(new_shadow_node
);
1137 ret
= rcuja_shadow_clear(ja
->ht
, new_node_flag
, new_shadow_node
,
1138 RCUJA_SHADOW_CLEAR_FREE_NODE
);
1141 /* Choose fallback type: pigeon */
1142 new_type_index
= (1UL << JA_TYPE_BITS
) - 1;
1143 dbg_printf("Fallback to type %d\n", new_type_index
);
1144 uatomic_inc(&ja
->nr_fallback
);
1150 * Return 0 on success, -EAGAIN if need to retry, or other negative
1151 * error value otherwise.
1154 int ja_node_set_nth(struct cds_ja
*ja
,
1155 struct cds_ja_inode_flag
**node_flag
, uint8_t n
,
1156 struct cds_ja_inode_flag
*child_node_flag
,
1157 struct cds_ja_shadow_node
*shadow_node
)
1160 unsigned int type_index
;
1161 const struct cds_ja_type
*type
;
1162 struct cds_ja_inode
*node
;
1164 dbg_printf("ja_node_set_nth for n=%u, node %p, shadow %p\n",
1165 (unsigned int) n
, ja_node_ptr(*node_flag
), shadow_node
);
1167 node
= ja_node_ptr(*node_flag
);
1168 type_index
= ja_node_type(*node_flag
);
1169 type
= &ja_types
[type_index
];
1170 ret
= _ja_node_set_nth(type
, node
, *node_flag
, shadow_node
,
1171 n
, child_node_flag
);
1174 /* Not enough space in node, need to recompact. */
1175 ret
= ja_node_recompact(JA_RECOMPACT_ADD
, ja
, type_index
, type
, node
,
1176 shadow_node
, node_flag
, n
, child_node_flag
, NULL
);
1179 /* Node needs to be recompacted. */
1180 ret
= ja_node_recompact(JA_RECOMPACT
, ja
, type_index
, type
, node
,
1181 shadow_node
, node_flag
, n
, child_node_flag
, NULL
);
1188 * Return 0 on success, -EAGAIN if need to retry, or other negative
1189 * error value otherwise.
1192 int ja_node_clear_ptr(struct cds_ja
*ja
,
1193 struct cds_ja_inode_flag
**node_flag_ptr
, /* Pointer to location to nullify */
1194 struct cds_ja_inode_flag
**parent_node_flag_ptr
, /* Address of parent ptr in its parent */
1195 struct cds_ja_shadow_node
*shadow_node
, /* of parent */
1199 unsigned int type_index
;
1200 const struct cds_ja_type
*type
;
1201 struct cds_ja_inode
*node
;
1203 dbg_printf("ja_node_clear_ptr for node %p, shadow %p, target ptr %p\n",
1204 ja_node_ptr(*parent_node_flag_ptr
), shadow_node
, node_flag_ptr
);
1206 node
= ja_node_ptr(*parent_node_flag_ptr
);
1207 type_index
= ja_node_type(*parent_node_flag_ptr
);
1208 type
= &ja_types
[type_index
];
1209 ret
= _ja_node_clear_ptr(type
, node
, shadow_node
, node_flag_ptr
, n
);
1210 if (ret
== -EFBIG
) {
1211 /* Should to try recompaction. */
1212 ret
= ja_node_recompact(JA_RECOMPACT_DEL
, ja
, type_index
, type
, node
,
1213 shadow_node
, parent_node_flag_ptr
, n
, NULL
,
1219 struct cds_hlist_head
cds_ja_lookup(struct cds_ja
*ja
, uint64_t key
)
1221 unsigned int tree_depth
, i
;
1222 struct cds_ja_inode_flag
*node_flag
;
1223 struct cds_hlist_head head
= { NULL
};
1225 if (caa_unlikely(key
> ja
->key_max
))
1227 tree_depth
= ja
->tree_depth
;
1228 node_flag
= rcu_dereference(ja
->root
);
1230 /* level 0: root node */
1231 if (!ja_node_ptr(node_flag
))
1234 for (i
= 1; i
< tree_depth
; i
++) {
1237 iter_key
= (uint8_t) (key
>> (JA_BITS_PER_BYTE
* (tree_depth
- i
- 1)));
1238 node_flag
= ja_node_get_nth(node_flag
, NULL
, NULL
, NULL
,
1240 dbg_printf("cds_ja_lookup iter key lookup %u finds node_flag %p\n",
1241 (unsigned int) iter_key
, node_flag
);
1242 if (!ja_node_ptr(node_flag
))
1246 /* Last level lookup succeded. We got an actual match. */
1247 head
.next
= (struct cds_hlist_node
*) node_flag
;
1252 * We reached an unpopulated node. Create it and the children we need,
1253 * and then attach the entire branch to the current node. This may
1254 * trigger recompaction of the current node. Locks needed: node lock
1255 * (for add), and, possibly, parent node lock (to update pointer due to
1256 * node recompaction).
1258 * First take node lock, check if recompaction is needed, then take
1259 * parent lock (if needed). Then we can proceed to create the new
1260 * branch. Publish the new branch, and release locks.
1261 * TODO: we currently always take the parent lock even when not needed.
1264 int ja_attach_node(struct cds_ja
*ja
,
1265 struct cds_ja_inode_flag
**attach_node_flag_ptr
,
1266 struct cds_ja_inode_flag
*attach_node_flag
,
1267 struct cds_ja_inode_flag
**node_flag_ptr
,
1268 struct cds_ja_inode_flag
*node_flag
,
1269 struct cds_ja_inode_flag
*parent_node_flag
,
1272 struct cds_ja_node
*child_node
)
1274 struct cds_ja_shadow_node
*shadow_node
= NULL
,
1275 *parent_shadow_node
= NULL
;
1276 struct cds_ja_inode
*node
= ja_node_ptr(node_flag
);
1277 struct cds_ja_inode
*parent_node
= ja_node_ptr(parent_node_flag
);
1278 struct cds_hlist_head head
;
1279 struct cds_ja_inode_flag
*iter_node_flag
, *iter_dest_node_flag
;
1281 struct cds_ja_inode_flag
*created_nodes
[JA_MAX_DEPTH
];
1282 int nr_created_nodes
= 0;
1284 dbg_printf("Attach node at level %u (node %p, node_flag %p)\n",
1285 level
, node
, node_flag
);
1288 shadow_node
= rcuja_shadow_lookup_lock(ja
->ht
, node_flag
);
1294 parent_shadow_node
= rcuja_shadow_lookup_lock(ja
->ht
,
1296 if (!parent_shadow_node
) {
1302 if (node_flag_ptr
&& ja_node_ptr(*node_flag_ptr
)) {
1304 * Target node has been updated between RCU lookup and
1305 * lock acquisition. We need to re-try lookup and
1312 if (attach_node_flag_ptr
&& ja_node_ptr(*attach_node_flag_ptr
) !=
1313 ja_node_ptr(attach_node_flag
)) {
1315 * Target node has been updated between RCU lookup and
1316 * lock acquisition. We need to re-try lookup and
1323 /* Create new branch, starting from bottom */
1324 CDS_INIT_HLIST_HEAD(&head
);
1325 cds_hlist_add_head_rcu(&child_node
->list
, &head
);
1326 iter_node_flag
= (struct cds_ja_inode_flag
*) head
.next
;
1328 for (i
= ja
->tree_depth
; i
> (int) level
; i
--) {
1331 iter_key
= (uint8_t) (key
>> (JA_BITS_PER_BYTE
* (ja
->tree_depth
- i
)));
1332 dbg_printf("branch creation level %d, key %u\n",
1333 i
- 1, (unsigned int) iter_key
);
1334 iter_dest_node_flag
= NULL
;
1335 ret
= ja_node_set_nth(ja
, &iter_dest_node_flag
,
1341 created_nodes
[nr_created_nodes
++] = iter_dest_node_flag
;
1342 iter_node_flag
= iter_dest_node_flag
;
1348 iter_key
= (uint8_t) (key
>> (JA_BITS_PER_BYTE
* (ja
->tree_depth
- level
)));
1349 /* We need to use set_nth on the previous level. */
1350 iter_dest_node_flag
= node_flag
;
1351 ret
= ja_node_set_nth(ja
, &iter_dest_node_flag
,
1357 created_nodes
[nr_created_nodes
++] = iter_dest_node_flag
;
1358 iter_node_flag
= iter_dest_node_flag
;
1361 /* Publish new branch */
1362 dbg_printf("Publish branch %p, replacing %p\n",
1363 iter_node_flag
, *attach_node_flag_ptr
);
1364 rcu_assign_pointer(*attach_node_flag_ptr
, iter_node_flag
);
1371 for (i
= 0; i
< nr_created_nodes
; i
++) {
1375 flags
= RCUJA_SHADOW_CLEAR_FREE_LOCK
;
1377 flags
|= RCUJA_SHADOW_CLEAR_FREE_NODE
;
1378 tmpret
= rcuja_shadow_clear(ja
->ht
,
1386 if (parent_shadow_node
)
1387 rcuja_shadow_unlock(parent_shadow_node
);
1390 rcuja_shadow_unlock(shadow_node
);
1396 * Lock the parent containing the hlist head pointer, and add node to list of
1397 * duplicates. Failure can happen if concurrent update changes the
1398 * parent before we get the lock. We return -EAGAIN in that case.
1399 * Return 0 on success, negative error value on failure.
1402 int ja_chain_node(struct cds_ja
*ja
,
1403 struct cds_ja_inode_flag
*parent_node_flag
,
1404 struct cds_ja_inode_flag
**node_flag_ptr
,
1405 struct cds_ja_inode_flag
*node_flag
,
1406 struct cds_hlist_head
*head
,
1407 struct cds_ja_node
*node
)
1409 struct cds_ja_shadow_node
*shadow_node
;
1412 shadow_node
= rcuja_shadow_lookup_lock(ja
->ht
, parent_node_flag
);
1416 if (ja_node_ptr(*node_flag_ptr
) != ja_node_ptr(node_flag
)) {
1420 cds_hlist_add_head_rcu(&node
->list
, head
);
1422 rcuja_shadow_unlock(shadow_node
);
1426 int cds_ja_add(struct cds_ja
*ja
, uint64_t key
,
1427 struct cds_ja_node
*new_node
)
1429 unsigned int tree_depth
, i
;
1430 struct cds_ja_inode_flag
**attach_node_flag_ptr
,
1432 struct cds_ja_inode_flag
*node_flag
,
1438 if (caa_unlikely(key
> ja
->key_max
)) {
1441 tree_depth
= ja
->tree_depth
;
1444 dbg_printf("cds_ja_add attempt: key %" PRIu64
", node %p\n",
1446 parent2_node_flag
= NULL
;
1448 (struct cds_ja_inode_flag
*) &ja
->root
; /* Use root ptr address as key for mutex */
1449 attach_node_flag_ptr
= &ja
->root
;
1450 attach_node_flag
= rcu_dereference(ja
->root
);
1451 node_flag_ptr
= &ja
->root
;
1452 node_flag
= rcu_dereference(ja
->root
);
1454 /* Iterate on all internal levels */
1455 for (i
= 1; i
< tree_depth
; i
++) {
1458 dbg_printf("cds_ja_add iter attach_node_flag_ptr %p node_flag_ptr %p node_flag %p\n",
1459 attach_node_flag_ptr
, node_flag_ptr
, node_flag
);
1460 if (!ja_node_ptr(node_flag
)) {
1461 ret
= ja_attach_node(ja
, attach_node_flag_ptr
,
1467 if (ret
== -EAGAIN
|| ret
== -EEXIST
)
1472 iter_key
= (uint8_t) (key
>> (JA_BITS_PER_BYTE
* (tree_depth
- i
- 1)));
1473 parent2_node_flag
= parent_node_flag
;
1474 parent_node_flag
= node_flag
;
1475 node_flag
= ja_node_get_nth(node_flag
,
1476 &attach_node_flag_ptr
,
1480 dbg_printf("cds_ja_add iter key lookup %u finds node_flag %p attach_node_flag_ptr %p node_flag_ptr %p\n",
1481 (unsigned int) iter_key
, node_flag
,
1482 attach_node_flag_ptr
,
1487 * We reached bottom of tree, simply add node to last internal
1488 * level, or chain it if key is already present.
1490 if (!ja_node_ptr(node_flag
)) {
1491 dbg_printf("cds_ja_add attach_node_flag_ptr %p node_flag_ptr %p node_flag %p\n",
1492 attach_node_flag_ptr
, node_flag_ptr
, node_flag
);
1493 ret
= ja_attach_node(ja
, attach_node_flag_ptr
,
1495 node_flag_ptr
, parent_node_flag
,
1496 parent2_node_flag
, key
, i
, new_node
);
1498 ret
= ja_chain_node(ja
,
1502 (struct cds_hlist_head
*) attach_node_flag_ptr
,
1505 if (ret
== -EAGAIN
|| ret
== -EEXIST
)
1512 * Note: there is no need to lookup the pointer address associated with
1513 * each node's nth item after taking the lock: it's already been done by
1514 * cds_ja_del while holding the rcu read-side lock, and our node rules
1515 * ensure that when a match value -> pointer is found in a node, it is
1516 * _NEVER_ changed for that node without recompaction, and recompaction
1517 * reallocates the node.
1518 * However, when a child is removed from "linear" nodes, its pointer
1519 * is set to NULL. We therefore check, while holding the locks, if this
1520 * pointer is NULL, and return -ENOENT to the caller if it is the case.
1523 int ja_detach_node(struct cds_ja
*ja
,
1524 struct cds_ja_inode_flag
**snapshot
,
1525 struct cds_ja_inode_flag
***snapshot_ptr
,
1526 uint8_t *snapshot_n
,
1529 struct cds_ja_node
*node
)
1531 struct cds_ja_shadow_node
*shadow_nodes
[JA_MAX_DEPTH
];
1532 struct cds_ja_inode_flag
**node_flag_ptr
= NULL
,
1533 *parent_node_flag
= NULL
,
1534 **parent_node_flag_ptr
= NULL
;
1535 struct cds_ja_inode_flag
*iter_node_flag
;
1536 int ret
, i
, nr_shadow
= 0, nr_clear
= 0, nr_branch
= 0;
1539 assert(nr_snapshot
== ja
->tree_depth
+ 1);
1542 * From the last internal level node going up, get the node
1543 * lock, check if the node has only one child left. If it is the
1544 * case, we continue iterating upward. When we reach a node
1545 * which has more that one child left, we lock the parent, and
1546 * proceed to the node deletion (removing its children too).
1548 for (i
= nr_snapshot
- 2; i
>= 1; i
--) {
1549 struct cds_ja_shadow_node
*shadow_node
;
1551 shadow_node
= rcuja_shadow_lookup_lock(ja
->ht
,
1557 shadow_nodes
[nr_shadow
++] = shadow_node
;
1560 * Check if node has been removed between RCU
1561 * lookup and lock acquisition.
1563 assert(snapshot_ptr
[i
+ 1]);
1564 if (ja_node_ptr(*snapshot_ptr
[i
+ 1])
1565 != ja_node_ptr(snapshot
[i
+ 1])) {
1570 assert(shadow_node
->nr_child
> 0);
1571 if (shadow_node
->nr_child
== 1 && i
> 1)
1574 if (shadow_node
->nr_child
> 1 || i
== 1) {
1575 /* Lock parent and break */
1576 shadow_node
= rcuja_shadow_lookup_lock(ja
->ht
,
1582 shadow_nodes
[nr_shadow
++] = shadow_node
;
1585 * Check if node has been removed between RCU
1586 * lookup and lock acquisition.
1588 assert(snapshot_ptr
[i
]);
1589 if (ja_node_ptr(*snapshot_ptr
[i
])
1590 != ja_node_ptr(snapshot
[i
])) {
1595 node_flag_ptr
= snapshot_ptr
[i
+ 1];
1596 n
= snapshot_n
[i
+ 1];
1597 parent_node_flag_ptr
= snapshot_ptr
[i
];
1598 parent_node_flag
= snapshot
[i
];
1602 * Lock parent's parent, in case we need
1603 * to recompact parent.
1605 shadow_node
= rcuja_shadow_lookup_lock(ja
->ht
,
1611 shadow_nodes
[nr_shadow
++] = shadow_node
;
1614 * Check if node has been removed between RCU
1615 * lookup and lock acquisition.
1617 assert(snapshot_ptr
[i
- 1]);
1618 if (ja_node_ptr(*snapshot_ptr
[i
- 1])
1619 != ja_node_ptr(snapshot
[i
- 1])) {
1630 * At this point, we want to delete all nodes that are about to
1631 * be removed from shadow_nodes (except the last one, which is
1632 * either the root or the parent of the upmost node with 1
1633 * child). OK to free lock here, because RCU read lock is held,
1634 * and free only performed in call_rcu.
1637 for (i
= 0; i
< nr_clear
; i
++) {
1638 ret
= rcuja_shadow_clear(ja
->ht
,
1639 shadow_nodes
[i
]->node_flag
,
1641 RCUJA_SHADOW_CLEAR_FREE_NODE
1642 | RCUJA_SHADOW_CLEAR_FREE_LOCK
);
1646 iter_node_flag
= parent_node_flag
;
1647 /* Remove from parent */
1648 ret
= ja_node_clear_ptr(ja
,
1649 node_flag_ptr
, /* Pointer to location to nullify */
1650 &iter_node_flag
, /* Old new parent ptr in its parent */
1651 shadow_nodes
[nr_branch
- 1], /* of parent */
1656 dbg_printf("ja_detach_node: publish %p instead of %p\n",
1657 iter_node_flag
, *parent_node_flag_ptr
);
1658 /* Update address of parent ptr in its parent */
1659 rcu_assign_pointer(*parent_node_flag_ptr
, iter_node_flag
);
1662 for (i
= 0; i
< nr_shadow
; i
++)
1663 rcuja_shadow_unlock(shadow_nodes
[i
]);
1668 int ja_unchain_node(struct cds_ja
*ja
,
1669 struct cds_ja_inode_flag
*parent_node_flag
,
1670 struct cds_ja_inode_flag
**node_flag_ptr
,
1671 struct cds_ja_inode_flag
*node_flag
,
1672 struct cds_ja_node
*node
)
1674 struct cds_ja_shadow_node
*shadow_node
;
1675 struct cds_hlist_node
*hlist_node
;
1676 struct cds_hlist_head hlist_head
;
1677 int ret
= 0, count
= 0, found
= 0;
1679 shadow_node
= rcuja_shadow_lookup_lock(ja
->ht
, parent_node_flag
);
1682 if (ja_node_ptr(*node_flag_ptr
) != ja_node_ptr(node_flag
)) {
1686 hlist_head
.next
= (struct cds_hlist_node
*) ja_node_ptr(node_flag
);
1688 * Retry if another thread removed all but one of duplicates
1689 * since check (this check was performed without lock).
1690 * Ensure that the node we are about to remove is still in the
1691 * list (while holding lock).
1693 cds_hlist_for_each_rcu(hlist_node
, &hlist_head
) {
1695 /* FIXME: currently a work-around */
1696 hlist_node
->prev
= (struct cds_hlist_node
*) node_flag_ptr
;
1699 if (hlist_node
== &node
->list
)
1703 if (!found
|| count
== 1) {
1707 cds_hlist_del_rcu(&node
->list
);
1709 * Validate that we indeed removed the node from linked list.
1711 assert(ja_node_ptr(*node_flag_ptr
) != (struct cds_ja_inode
*) node
);
1713 rcuja_shadow_unlock(shadow_node
);
1718 * Called with RCU read lock held.
1720 int cds_ja_del(struct cds_ja
*ja
, uint64_t key
,
1721 struct cds_ja_node
*node
)
1723 unsigned int tree_depth
, i
;
1724 struct cds_ja_inode_flag
*snapshot
[JA_MAX_DEPTH
];
1725 struct cds_ja_inode_flag
**snapshot_ptr
[JA_MAX_DEPTH
];
1726 uint8_t snapshot_n
[JA_MAX_DEPTH
];
1727 struct cds_ja_inode_flag
*node_flag
;
1728 struct cds_ja_inode_flag
**prev_node_flag_ptr
,
1733 if (caa_unlikely(key
> ja
->key_max
))
1735 tree_depth
= ja
->tree_depth
;
1739 dbg_printf("cds_ja_del attempt: key %" PRIu64
", node %p\n",
1742 /* snapshot for level 0 is only for shadow node lookup */
1745 snapshot_ptr
[nr_snapshot
] = NULL
;
1746 snapshot
[nr_snapshot
++] = (struct cds_ja_inode_flag
*) &ja
->root
;
1747 node_flag
= rcu_dereference(ja
->root
);
1748 prev_node_flag_ptr
= &ja
->root
;
1749 node_flag_ptr
= &ja
->root
;
1751 /* Iterate on all internal levels */
1752 for (i
= 1; i
< tree_depth
; i
++) {
1755 dbg_printf("cds_ja_del iter node_flag %p\n",
1757 if (!ja_node_ptr(node_flag
)) {
1760 iter_key
= (uint8_t) (key
>> (JA_BITS_PER_BYTE
* (tree_depth
- i
- 1)));
1761 snapshot_n
[nr_snapshot
+ 1] = iter_key
;
1762 snapshot_ptr
[nr_snapshot
] = prev_node_flag_ptr
;
1763 snapshot
[nr_snapshot
++] = node_flag
;
1764 node_flag
= ja_node_get_nth(node_flag
,
1765 &prev_node_flag_ptr
,
1769 dbg_printf("cds_ja_del iter key lookup %u finds node_flag %p, prev_node_flag_ptr %p\n",
1770 (unsigned int) iter_key
, node_flag
,
1771 prev_node_flag_ptr
);
1774 * We reached bottom of tree, try to find the node we are trying
1775 * to remove. Fail if we cannot find it.
1777 if (!ja_node_ptr(node_flag
)) {
1778 dbg_printf("cds_ja_del: no node found for key %" PRIu64
"\n",
1782 struct cds_hlist_head hlist_head
;
1783 struct cds_hlist_node
*hlist_node
;
1784 struct cds_ja_node
*entry
, *match
= NULL
;
1788 (struct cds_hlist_node
*) ja_node_ptr(node_flag
);
1789 cds_hlist_for_each_entry_rcu(entry
,
1793 dbg_printf("cds_ja_del: compare %p with entry %p\n", node
, entry
);
1799 dbg_printf("cds_ja_del: no node match for node %p key %" PRIu64
"\n", node
, key
);
1805 * Removing last of duplicates. Last snapshot
1806 * does not have a shadow node (external leafs).
1808 snapshot_ptr
[nr_snapshot
] = prev_node_flag_ptr
;
1809 snapshot
[nr_snapshot
++] = node_flag
;
1810 ret
= ja_detach_node(ja
, snapshot
, snapshot_ptr
,
1811 snapshot_n
, nr_snapshot
, key
, node
);
1813 ret
= ja_unchain_node(ja
, snapshot
[nr_snapshot
- 1],
1814 node_flag_ptr
, node_flag
, match
);
1818 * Explanation of -ENOENT handling: caused by concurrent delete
1819 * between RCU lookup and actual removal. Need to re-do the
1820 * lookup and removal attempt.
1822 if (ret
== -EAGAIN
|| ret
== -ENOENT
)
1827 struct cds_ja
*_cds_ja_new(unsigned int key_bits
,
1828 const struct rcu_flavor_struct
*flavor
)
1832 struct cds_ja_shadow_node
*root_shadow_node
;
1834 ja
= calloc(sizeof(*ja
), 1);
1846 ja
->key_max
= (1ULL << key_bits
) - 1;
1849 ja
->key_max
= UINT64_MAX
;
1855 /* ja->root is NULL */
1856 /* tree_depth 0 is for pointer to root node */
1857 ja
->tree_depth
= (key_bits
>> JA_LOG2_BITS_PER_BYTE
) + 1;
1858 assert(ja
->tree_depth
<= JA_MAX_DEPTH
);
1859 ja
->ht
= rcuja_create_ht(flavor
);
1864 * Note: we should not free this node until judy array destroy.
1866 root_shadow_node
= rcuja_shadow_set(ja
->ht
,
1867 (struct cds_ja_inode_flag
*) &ja
->root
,
1869 if (!root_shadow_node
) {
1873 root_shadow_node
->level
= 0;
1878 ret
= rcuja_delete_ht(ja
->ht
);
1888 * Called from RCU read-side CS.
1890 __attribute__((visibility("protected")))
1891 void rcuja_free_all_children(struct cds_ja_shadow_node
*shadow_node
,
1892 struct cds_ja_inode_flag
*node_flag
,
1893 void (*free_node_cb
)(struct rcu_head
*head
))
1895 const struct rcu_flavor_struct
*flavor
;
1896 unsigned int type_index
;
1897 struct cds_ja_inode
*node
;
1898 const struct cds_ja_type
*type
;
1900 flavor
= cds_lfht_rcu_flavor(shadow_node
->ja
->ht
);
1901 node
= ja_node_ptr(node_flag
);
1902 assert(node
!= NULL
);
1903 type_index
= ja_node_type(node_flag
);
1904 type
= &ja_types
[type_index
];
1906 switch (type
->type_class
) {
1910 ja_linear_node_get_nr_child(type
, node
);
1913 for (i
= 0; i
< nr_child
; i
++) {
1914 struct cds_ja_inode_flag
*iter
;
1915 struct cds_hlist_head head
;
1916 struct cds_ja_node
*entry
;
1917 struct cds_hlist_node
*pos
;
1920 ja_linear_node_get_ith_pos(type
, node
, i
, &v
, &iter
);
1923 head
.next
= (struct cds_hlist_node
*) iter
;
1924 cds_hlist_for_each_entry_rcu(entry
, pos
, &head
, list
) {
1925 flavor
->update_call_rcu(&entry
->head
, free_node_cb
);
1932 unsigned int pool_nr
;
1934 for (pool_nr
= 0; pool_nr
< (1U << type
->nr_pool_order
); pool_nr
++) {
1935 struct cds_ja_inode
*pool
=
1936 ja_pool_node_get_ith_pool(type
, node
, pool_nr
);
1938 ja_linear_node_get_nr_child(type
, pool
);
1941 for (j
= 0; j
< nr_child
; j
++) {
1942 struct cds_ja_inode_flag
*iter
;
1943 struct cds_hlist_head head
;
1944 struct cds_ja_node
*entry
;
1945 struct cds_hlist_node
*pos
;
1948 ja_linear_node_get_ith_pos(type
, node
, j
, &v
, &iter
);
1951 head
.next
= (struct cds_hlist_node
*) iter
;
1952 cds_hlist_for_each_entry_rcu(entry
, pos
, &head
, list
) {
1953 flavor
->update_call_rcu(&entry
->head
, free_node_cb
);
1966 nr_child
= shadow_node
->nr_child
;
1967 for (i
= 0; i
< nr_child
; i
++) {
1968 struct cds_ja_inode_flag
*iter
;
1969 struct cds_hlist_head head
;
1970 struct cds_ja_node
*entry
;
1971 struct cds_hlist_node
*pos
;
1973 iter
= ja_pigeon_node_get_ith_pos(type
, node
, i
);
1976 head
.next
= (struct cds_hlist_node
*) iter
;
1977 cds_hlist_for_each_entry_rcu(entry
, pos
, &head
, list
) {
1978 flavor
->update_call_rcu(&entry
->head
, free_node_cb
);
1989 * There should be no more concurrent add to the judy array while it is
1990 * being destroyed (ensured by the caller).
1992 int cds_ja_destroy(struct cds_ja
*ja
,
1993 void (*free_node_cb
)(struct rcu_head
*head
))
1997 rcuja_shadow_prune(ja
->ht
,
1998 RCUJA_SHADOW_CLEAR_FREE_NODE
| RCUJA_SHADOW_CLEAR_FREE_LOCK
,
2000 ret
= rcuja_delete_ht(ja
->ht
);
2003 if (uatomic_read(&ja
->nr_fallback
))
2005 "[warning] RCU Judy Array used %lu fallback node(s)\n",
2006 uatomic_read(&ja
->nr_fallback
));