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
;
799 ja_linear_node_get_ith_pos(type
, node
, i
, &v
, &iter
);
802 if (mode
== JA_RECOMPACT_DEL
&& *nullify_node_flag_ptr
== iter
)
804 for (bit
= 0; bit
< JA_BITS_PER_BYTE
; bit
++) {
814 unsigned int pool_nr
;
816 for (pool_nr
= 0; pool_nr
< (1U << type
->nr_pool_order
); pool_nr
++) {
817 struct cds_ja_inode
*pool
=
818 ja_pool_node_get_ith_pool(type
,
821 ja_linear_node_get_nr_child(type
, pool
);
824 for (j
= 0; j
< nr_child
; j
++) {
825 struct cds_ja_inode_flag
*iter
;
829 ja_linear_node_get_ith_pos(type
, pool
,
833 if (mode
== JA_RECOMPACT_DEL
&& *nullify_node_flag_ptr
== iter
)
835 for (bit
= 0; bit
< JA_BITS_PER_BYTE
; bit
++) {
849 assert(mode
== JA_RECOMPACT_DEL
);
850 nr_child
= shadow_node
->nr_child
;
851 for (i
= 0; i
< nr_child
; i
++) {
852 struct cds_ja_inode_flag
*iter
;
855 iter
= ja_pigeon_node_get_ith_pos(type
, node
, i
);
858 if (mode
== JA_RECOMPACT_DEL
&& *nullify_node_flag_ptr
== iter
)
860 for (bit
= 0; bit
< JA_BITS_PER_BYTE
; bit
++) {
869 assert(mode
== JA_RECOMPACT_ADD
);
876 if (mode
== JA_RECOMPACT_ADD
) {
879 for (bit
= 0; bit
< JA_BITS_PER_BYTE
; bit
++) {
887 * The best bit selector is that for which the number of ones is
888 * closest to half of the number of children in the
891 for (bit_i
= 0; bit_i
< JA_BITS_PER_BYTE
; bit_i
++) {
892 unsigned int distance_to_best
;
894 distance_to_best
= abs_int(nr_one
[bit_i
] - (distrib_nr_child
>> 1U));
895 if (distance_to_best
< overall_best_distance
) {
896 overall_best_distance
= distance_to_best
;
900 dbg_printf("1 dimension pool bit selection: (%u)\n", bitsel
);
905 * ja_node_recompact_add: recompact a node, adding a new child.
906 * Return 0 on success, -EAGAIN if need to retry, or other negative
907 * error value otherwise.
910 int ja_node_recompact(enum ja_recompact mode
,
912 unsigned int old_type_index
,
913 const struct cds_ja_type
*old_type
,
914 struct cds_ja_inode
*old_node
,
915 struct cds_ja_shadow_node
*shadow_node
,
916 struct cds_ja_inode_flag
**old_node_flag_ptr
, uint8_t n
,
917 struct cds_ja_inode_flag
*child_node_flag
,
918 struct cds_ja_inode_flag
**nullify_node_flag_ptr
)
920 unsigned int new_type_index
;
921 struct cds_ja_inode
*new_node
;
922 struct cds_ja_shadow_node
*new_shadow_node
= NULL
;
923 const struct cds_ja_type
*new_type
;
924 struct cds_ja_inode_flag
*new_node_flag
, *old_node_flag
;
928 old_node_flag
= *old_node_flag_ptr
;
932 new_type_index
= old_type_index
;
934 case JA_RECOMPACT_ADD
:
935 if (!shadow_node
|| old_type_index
== NODE_INDEX_NULL
) {
938 new_type_index
= old_type_index
+ 1;
941 case JA_RECOMPACT_DEL
:
942 if (old_type_index
== 0) {
943 new_type_index
= NODE_INDEX_NULL
;
945 new_type_index
= old_type_index
- 1;
952 retry
: /* for fallback */
953 dbg_printf("Recompact from type %d to type %d\n",
954 old_type_index
, new_type_index
);
955 new_type
= &ja_types
[new_type_index
];
956 if (new_type_index
!= NODE_INDEX_NULL
) {
957 new_node
= alloc_cds_ja_node(new_type
);
961 if (new_type
->type_class
== RCU_JA_POOL
) {
962 switch (new_type
->nr_pool_order
) {
965 unsigned int node_distrib_bitsel
= 0;
966 node_distrib_bitsel
=
967 ja_node_sum_distribution_1d(mode
, ja
,
968 old_type_index
, old_type
,
969 old_node
, shadow_node
,
971 nullify_node_flag_ptr
);
972 assert(!((unsigned long) new_node
& JA_POOL_1D_MASK
));
973 new_node_flag
= ja_node_flag_pool_1d(new_node
,
974 new_type_index
, node_distrib_bitsel
);
979 /* TODO: pool order 2 in 2d */
980 assert(!((unsigned long) new_node
& JA_POOL_1D_MASK
));
981 assert(!((unsigned long) new_node
& JA_POOL_2D_MASK
));
982 new_node_flag
= ja_node_flag(new_node
, new_type_index
);
989 new_node_flag
= ja_node_flag(new_node
, new_type_index
);
992 dbg_printf("Recompact inherit lock from %p\n", shadow_node
);
993 new_shadow_node
= rcuja_shadow_set(ja
->ht
, new_node_flag
, shadow_node
, ja
);
994 if (!new_shadow_node
) {
999 new_shadow_node
->fallback_removal_count
=
1000 JA_FALLBACK_REMOVAL_COUNT
;
1003 new_node_flag
= NULL
;
1006 assert(mode
!= JA_RECOMPACT_ADD
|| old_type
->type_class
!= RCU_JA_PIGEON
);
1008 if (new_type_index
== NODE_INDEX_NULL
)
1011 switch (old_type
->type_class
) {
1015 ja_linear_node_get_nr_child(old_type
, old_node
);
1018 for (i
= 0; i
< nr_child
; i
++) {
1019 struct cds_ja_inode_flag
*iter
;
1022 ja_linear_node_get_ith_pos(old_type
, old_node
, i
, &v
, &iter
);
1025 if (mode
== JA_RECOMPACT_DEL
&& *nullify_node_flag_ptr
== iter
)
1027 ret
= _ja_node_set_nth(new_type
, new_node
, new_node_flag
,
1030 if (new_type
->type_class
== RCU_JA_POOL
&& ret
) {
1031 goto fallback_toosmall
;
1039 unsigned int pool_nr
;
1041 for (pool_nr
= 0; pool_nr
< (1U << old_type
->nr_pool_order
); pool_nr
++) {
1042 struct cds_ja_inode
*pool
=
1043 ja_pool_node_get_ith_pool(old_type
,
1046 ja_linear_node_get_nr_child(old_type
, pool
);
1049 for (j
= 0; j
< nr_child
; j
++) {
1050 struct cds_ja_inode_flag
*iter
;
1053 ja_linear_node_get_ith_pos(old_type
, pool
,
1057 if (mode
== JA_RECOMPACT_DEL
&& *nullify_node_flag_ptr
== iter
)
1059 ret
= _ja_node_set_nth(new_type
, new_node
, new_node_flag
,
1062 if (new_type
->type_class
== RCU_JA_POOL
1064 goto fallback_toosmall
;
1072 assert(mode
== JA_RECOMPACT_ADD
);
1079 assert(mode
== JA_RECOMPACT_DEL
);
1080 nr_child
= shadow_node
->nr_child
;
1081 for (i
= 0; i
< nr_child
; i
++) {
1082 struct cds_ja_inode_flag
*iter
;
1084 iter
= ja_pigeon_node_get_ith_pos(old_type
, old_node
, i
);
1087 if (mode
== JA_RECOMPACT_DEL
&& *nullify_node_flag_ptr
== iter
)
1089 ret
= _ja_node_set_nth(new_type
, new_node
, new_node_flag
,
1092 if (new_type
->type_class
== RCU_JA_POOL
&& ret
) {
1093 goto fallback_toosmall
;
1106 if (mode
== JA_RECOMPACT_ADD
) {
1108 ret
= _ja_node_set_nth(new_type
, new_node
, new_node_flag
,
1110 n
, child_node_flag
);
1111 if (new_type
->type_class
== RCU_JA_POOL
&& ret
) {
1112 goto fallback_toosmall
;
1116 /* Return pointer to new recompacted node through old_node_flag_ptr */
1117 *old_node_flag_ptr
= new_node_flag
;
1121 flags
= RCUJA_SHADOW_CLEAR_FREE_NODE
;
1123 * It is OK to free the lock associated with a node
1124 * going to NULL, since we are holding the parent lock.
1125 * This synchronizes removal with re-add of that node.
1127 if (new_type_index
== NODE_INDEX_NULL
)
1128 flags
= RCUJA_SHADOW_CLEAR_FREE_LOCK
;
1129 ret
= rcuja_shadow_clear(ja
->ht
, old_node_flag
, shadow_node
,
1139 /* fallback if next pool is too small */
1140 assert(new_shadow_node
);
1141 ret
= rcuja_shadow_clear(ja
->ht
, new_node_flag
, new_shadow_node
,
1142 RCUJA_SHADOW_CLEAR_FREE_NODE
);
1145 /* Choose fallback type: pigeon */
1146 new_type_index
= (1UL << JA_TYPE_BITS
) - 1;
1147 dbg_printf("Fallback to type %d\n", new_type_index
);
1148 uatomic_inc(&ja
->nr_fallback
);
1154 * Return 0 on success, -EAGAIN if need to retry, or other negative
1155 * error value otherwise.
1158 int ja_node_set_nth(struct cds_ja
*ja
,
1159 struct cds_ja_inode_flag
**node_flag
, uint8_t n
,
1160 struct cds_ja_inode_flag
*child_node_flag
,
1161 struct cds_ja_shadow_node
*shadow_node
)
1164 unsigned int type_index
;
1165 const struct cds_ja_type
*type
;
1166 struct cds_ja_inode
*node
;
1168 dbg_printf("ja_node_set_nth for n=%u, node %p, shadow %p\n",
1169 (unsigned int) n
, ja_node_ptr(*node_flag
), shadow_node
);
1171 node
= ja_node_ptr(*node_flag
);
1172 type_index
= ja_node_type(*node_flag
);
1173 type
= &ja_types
[type_index
];
1174 ret
= _ja_node_set_nth(type
, node
, *node_flag
, shadow_node
,
1175 n
, child_node_flag
);
1178 /* Not enough space in node, need to recompact. */
1179 ret
= ja_node_recompact(JA_RECOMPACT_ADD
, ja
, type_index
, type
, node
,
1180 shadow_node
, node_flag
, n
, child_node_flag
, NULL
);
1183 /* Node needs to be recompacted. */
1184 ret
= ja_node_recompact(JA_RECOMPACT
, ja
, type_index
, type
, node
,
1185 shadow_node
, node_flag
, n
, child_node_flag
, NULL
);
1192 * Return 0 on success, -EAGAIN if need to retry, or other negative
1193 * error value otherwise.
1196 int ja_node_clear_ptr(struct cds_ja
*ja
,
1197 struct cds_ja_inode_flag
**node_flag_ptr
, /* Pointer to location to nullify */
1198 struct cds_ja_inode_flag
**parent_node_flag_ptr
, /* Address of parent ptr in its parent */
1199 struct cds_ja_shadow_node
*shadow_node
, /* of parent */
1203 unsigned int type_index
;
1204 const struct cds_ja_type
*type
;
1205 struct cds_ja_inode
*node
;
1207 dbg_printf("ja_node_clear_ptr for node %p, shadow %p, target ptr %p\n",
1208 ja_node_ptr(*parent_node_flag_ptr
), shadow_node
, node_flag_ptr
);
1210 node
= ja_node_ptr(*parent_node_flag_ptr
);
1211 type_index
= ja_node_type(*parent_node_flag_ptr
);
1212 type
= &ja_types
[type_index
];
1213 ret
= _ja_node_clear_ptr(type
, node
, shadow_node
, node_flag_ptr
, n
);
1214 if (ret
== -EFBIG
) {
1215 /* Should to try recompaction. */
1216 ret
= ja_node_recompact(JA_RECOMPACT_DEL
, ja
, type_index
, type
, node
,
1217 shadow_node
, parent_node_flag_ptr
, n
, NULL
,
1223 struct cds_hlist_head
cds_ja_lookup(struct cds_ja
*ja
, uint64_t key
)
1225 unsigned int tree_depth
, i
;
1226 struct cds_ja_inode_flag
*node_flag
;
1227 struct cds_hlist_head head
= { NULL
};
1229 if (caa_unlikely(key
> ja
->key_max
))
1231 tree_depth
= ja
->tree_depth
;
1232 node_flag
= rcu_dereference(ja
->root
);
1234 /* level 0: root node */
1235 if (!ja_node_ptr(node_flag
))
1238 for (i
= 1; i
< tree_depth
; i
++) {
1241 iter_key
= (uint8_t) (key
>> (JA_BITS_PER_BYTE
* (tree_depth
- i
- 1)));
1242 node_flag
= ja_node_get_nth(node_flag
, NULL
, NULL
, NULL
,
1244 dbg_printf("cds_ja_lookup iter key lookup %u finds node_flag %p\n",
1245 (unsigned int) iter_key
, node_flag
);
1246 if (!ja_node_ptr(node_flag
))
1250 /* Last level lookup succeded. We got an actual match. */
1251 head
.next
= (struct cds_hlist_node
*) node_flag
;
1256 * We reached an unpopulated node. Create it and the children we need,
1257 * and then attach the entire branch to the current node. This may
1258 * trigger recompaction of the current node. Locks needed: node lock
1259 * (for add), and, possibly, parent node lock (to update pointer due to
1260 * node recompaction).
1262 * First take node lock, check if recompaction is needed, then take
1263 * parent lock (if needed). Then we can proceed to create the new
1264 * branch. Publish the new branch, and release locks.
1265 * TODO: we currently always take the parent lock even when not needed.
1268 int ja_attach_node(struct cds_ja
*ja
,
1269 struct cds_ja_inode_flag
**attach_node_flag_ptr
,
1270 struct cds_ja_inode_flag
*attach_node_flag
,
1271 struct cds_ja_inode_flag
**node_flag_ptr
,
1272 struct cds_ja_inode_flag
*node_flag
,
1273 struct cds_ja_inode_flag
*parent_node_flag
,
1276 struct cds_ja_node
*child_node
)
1278 struct cds_ja_shadow_node
*shadow_node
= NULL
,
1279 *parent_shadow_node
= NULL
;
1280 struct cds_ja_inode
*node
= ja_node_ptr(node_flag
);
1281 struct cds_ja_inode
*parent_node
= ja_node_ptr(parent_node_flag
);
1282 struct cds_hlist_head head
;
1283 struct cds_ja_inode_flag
*iter_node_flag
, *iter_dest_node_flag
;
1285 struct cds_ja_inode_flag
*created_nodes
[JA_MAX_DEPTH
];
1286 int nr_created_nodes
= 0;
1288 dbg_printf("Attach node at level %u (node %p, node_flag %p)\n",
1289 level
, node
, node_flag
);
1292 shadow_node
= rcuja_shadow_lookup_lock(ja
->ht
, node_flag
);
1298 parent_shadow_node
= rcuja_shadow_lookup_lock(ja
->ht
,
1300 if (!parent_shadow_node
) {
1306 if (node_flag_ptr
&& ja_node_ptr(*node_flag_ptr
)) {
1308 * Target node has been updated between RCU lookup and
1309 * lock acquisition. We need to re-try lookup and
1316 if (attach_node_flag_ptr
&& ja_node_ptr(*attach_node_flag_ptr
) !=
1317 ja_node_ptr(attach_node_flag
)) {
1319 * Target node has been updated between RCU lookup and
1320 * lock acquisition. We need to re-try lookup and
1327 /* Create new branch, starting from bottom */
1328 CDS_INIT_HLIST_HEAD(&head
);
1329 cds_hlist_add_head_rcu(&child_node
->list
, &head
);
1330 iter_node_flag
= (struct cds_ja_inode_flag
*) head
.next
;
1332 for (i
= ja
->tree_depth
; i
> (int) level
; i
--) {
1335 iter_key
= (uint8_t) (key
>> (JA_BITS_PER_BYTE
* (ja
->tree_depth
- i
)));
1336 dbg_printf("branch creation level %d, key %u\n",
1337 i
- 1, (unsigned int) iter_key
);
1338 iter_dest_node_flag
= NULL
;
1339 ret
= ja_node_set_nth(ja
, &iter_dest_node_flag
,
1345 created_nodes
[nr_created_nodes
++] = iter_dest_node_flag
;
1346 iter_node_flag
= iter_dest_node_flag
;
1352 iter_key
= (uint8_t) (key
>> (JA_BITS_PER_BYTE
* (ja
->tree_depth
- level
)));
1353 /* We need to use set_nth on the previous level. */
1354 iter_dest_node_flag
= node_flag
;
1355 ret
= ja_node_set_nth(ja
, &iter_dest_node_flag
,
1361 created_nodes
[nr_created_nodes
++] = iter_dest_node_flag
;
1362 iter_node_flag
= iter_dest_node_flag
;
1365 /* Publish new branch */
1366 dbg_printf("Publish branch %p, replacing %p\n",
1367 iter_node_flag
, *attach_node_flag_ptr
);
1368 rcu_assign_pointer(*attach_node_flag_ptr
, iter_node_flag
);
1375 for (i
= 0; i
< nr_created_nodes
; i
++) {
1379 flags
= RCUJA_SHADOW_CLEAR_FREE_LOCK
;
1381 flags
|= RCUJA_SHADOW_CLEAR_FREE_NODE
;
1382 tmpret
= rcuja_shadow_clear(ja
->ht
,
1390 if (parent_shadow_node
)
1391 rcuja_shadow_unlock(parent_shadow_node
);
1394 rcuja_shadow_unlock(shadow_node
);
1400 * Lock the parent containing the hlist head pointer, and add node to list of
1401 * duplicates. Failure can happen if concurrent update changes the
1402 * parent before we get the lock. We return -EAGAIN in that case.
1403 * Return 0 on success, negative error value on failure.
1406 int ja_chain_node(struct cds_ja
*ja
,
1407 struct cds_ja_inode_flag
*parent_node_flag
,
1408 struct cds_ja_inode_flag
**node_flag_ptr
,
1409 struct cds_ja_inode_flag
*node_flag
,
1410 struct cds_hlist_head
*head
,
1411 struct cds_ja_node
*node
)
1413 struct cds_ja_shadow_node
*shadow_node
;
1416 shadow_node
= rcuja_shadow_lookup_lock(ja
->ht
, parent_node_flag
);
1420 if (ja_node_ptr(*node_flag_ptr
) != ja_node_ptr(node_flag
)) {
1424 cds_hlist_add_head_rcu(&node
->list
, head
);
1426 rcuja_shadow_unlock(shadow_node
);
1430 int cds_ja_add(struct cds_ja
*ja
, uint64_t key
,
1431 struct cds_ja_node
*new_node
)
1433 unsigned int tree_depth
, i
;
1434 struct cds_ja_inode_flag
**attach_node_flag_ptr
,
1436 struct cds_ja_inode_flag
*node_flag
,
1442 if (caa_unlikely(key
> ja
->key_max
)) {
1445 tree_depth
= ja
->tree_depth
;
1448 dbg_printf("cds_ja_add attempt: key %" PRIu64
", node %p\n",
1450 parent2_node_flag
= NULL
;
1452 (struct cds_ja_inode_flag
*) &ja
->root
; /* Use root ptr address as key for mutex */
1453 attach_node_flag_ptr
= &ja
->root
;
1454 attach_node_flag
= rcu_dereference(ja
->root
);
1455 node_flag_ptr
= &ja
->root
;
1456 node_flag
= rcu_dereference(ja
->root
);
1458 /* Iterate on all internal levels */
1459 for (i
= 1; i
< tree_depth
; i
++) {
1462 dbg_printf("cds_ja_add iter attach_node_flag_ptr %p node_flag_ptr %p node_flag %p\n",
1463 attach_node_flag_ptr
, node_flag_ptr
, node_flag
);
1464 if (!ja_node_ptr(node_flag
)) {
1465 ret
= ja_attach_node(ja
, attach_node_flag_ptr
,
1471 if (ret
== -EAGAIN
|| ret
== -EEXIST
)
1476 iter_key
= (uint8_t) (key
>> (JA_BITS_PER_BYTE
* (tree_depth
- i
- 1)));
1477 parent2_node_flag
= parent_node_flag
;
1478 parent_node_flag
= node_flag
;
1479 node_flag
= ja_node_get_nth(node_flag
,
1480 &attach_node_flag_ptr
,
1484 dbg_printf("cds_ja_add iter key lookup %u finds node_flag %p attach_node_flag_ptr %p node_flag_ptr %p\n",
1485 (unsigned int) iter_key
, node_flag
,
1486 attach_node_flag_ptr
,
1491 * We reached bottom of tree, simply add node to last internal
1492 * level, or chain it if key is already present.
1494 if (!ja_node_ptr(node_flag
)) {
1495 dbg_printf("cds_ja_add attach_node_flag_ptr %p node_flag_ptr %p node_flag %p\n",
1496 attach_node_flag_ptr
, node_flag_ptr
, node_flag
);
1497 ret
= ja_attach_node(ja
, attach_node_flag_ptr
,
1499 node_flag_ptr
, parent_node_flag
,
1500 parent2_node_flag
, key
, i
, new_node
);
1502 ret
= ja_chain_node(ja
,
1506 (struct cds_hlist_head
*) attach_node_flag_ptr
,
1509 if (ret
== -EAGAIN
|| ret
== -EEXIST
)
1516 * Note: there is no need to lookup the pointer address associated with
1517 * each node's nth item after taking the lock: it's already been done by
1518 * cds_ja_del while holding the rcu read-side lock, and our node rules
1519 * ensure that when a match value -> pointer is found in a node, it is
1520 * _NEVER_ changed for that node without recompaction, and recompaction
1521 * reallocates the node.
1522 * However, when a child is removed from "linear" nodes, its pointer
1523 * is set to NULL. We therefore check, while holding the locks, if this
1524 * pointer is NULL, and return -ENOENT to the caller if it is the case.
1527 int ja_detach_node(struct cds_ja
*ja
,
1528 struct cds_ja_inode_flag
**snapshot
,
1529 struct cds_ja_inode_flag
***snapshot_ptr
,
1530 uint8_t *snapshot_n
,
1533 struct cds_ja_node
*node
)
1535 struct cds_ja_shadow_node
*shadow_nodes
[JA_MAX_DEPTH
];
1536 struct cds_ja_inode_flag
**node_flag_ptr
= NULL
,
1537 *parent_node_flag
= NULL
,
1538 **parent_node_flag_ptr
= NULL
;
1539 struct cds_ja_inode_flag
*iter_node_flag
;
1540 int ret
, i
, nr_shadow
= 0, nr_clear
= 0, nr_branch
= 0;
1543 assert(nr_snapshot
== ja
->tree_depth
+ 1);
1546 * From the last internal level node going up, get the node
1547 * lock, check if the node has only one child left. If it is the
1548 * case, we continue iterating upward. When we reach a node
1549 * which has more that one child left, we lock the parent, and
1550 * proceed to the node deletion (removing its children too).
1552 for (i
= nr_snapshot
- 2; i
>= 1; i
--) {
1553 struct cds_ja_shadow_node
*shadow_node
;
1555 shadow_node
= rcuja_shadow_lookup_lock(ja
->ht
,
1561 shadow_nodes
[nr_shadow
++] = shadow_node
;
1564 * Check if node has been removed between RCU
1565 * lookup and lock acquisition.
1567 assert(snapshot_ptr
[i
+ 1]);
1568 if (ja_node_ptr(*snapshot_ptr
[i
+ 1])
1569 != ja_node_ptr(snapshot
[i
+ 1])) {
1574 assert(shadow_node
->nr_child
> 0);
1575 if (shadow_node
->nr_child
== 1 && i
> 1)
1578 if (shadow_node
->nr_child
> 1 || i
== 1) {
1579 /* Lock parent and break */
1580 shadow_node
= rcuja_shadow_lookup_lock(ja
->ht
,
1586 shadow_nodes
[nr_shadow
++] = shadow_node
;
1589 * Check if node has been removed between RCU
1590 * lookup and lock acquisition.
1592 assert(snapshot_ptr
[i
]);
1593 if (ja_node_ptr(*snapshot_ptr
[i
])
1594 != ja_node_ptr(snapshot
[i
])) {
1599 node_flag_ptr
= snapshot_ptr
[i
+ 1];
1600 n
= snapshot_n
[i
+ 1];
1601 parent_node_flag_ptr
= snapshot_ptr
[i
];
1602 parent_node_flag
= snapshot
[i
];
1606 * Lock parent's parent, in case we need
1607 * to recompact parent.
1609 shadow_node
= rcuja_shadow_lookup_lock(ja
->ht
,
1615 shadow_nodes
[nr_shadow
++] = shadow_node
;
1618 * Check if node has been removed between RCU
1619 * lookup and lock acquisition.
1621 assert(snapshot_ptr
[i
- 1]);
1622 if (ja_node_ptr(*snapshot_ptr
[i
- 1])
1623 != ja_node_ptr(snapshot
[i
- 1])) {
1634 * At this point, we want to delete all nodes that are about to
1635 * be removed from shadow_nodes (except the last one, which is
1636 * either the root or the parent of the upmost node with 1
1637 * child). OK to free lock here, because RCU read lock is held,
1638 * and free only performed in call_rcu.
1641 for (i
= 0; i
< nr_clear
; i
++) {
1642 ret
= rcuja_shadow_clear(ja
->ht
,
1643 shadow_nodes
[i
]->node_flag
,
1645 RCUJA_SHADOW_CLEAR_FREE_NODE
1646 | RCUJA_SHADOW_CLEAR_FREE_LOCK
);
1650 iter_node_flag
= parent_node_flag
;
1651 /* Remove from parent */
1652 ret
= ja_node_clear_ptr(ja
,
1653 node_flag_ptr
, /* Pointer to location to nullify */
1654 &iter_node_flag
, /* Old new parent ptr in its parent */
1655 shadow_nodes
[nr_branch
- 1], /* of parent */
1660 dbg_printf("ja_detach_node: publish %p instead of %p\n",
1661 iter_node_flag
, *parent_node_flag_ptr
);
1662 /* Update address of parent ptr in its parent */
1663 rcu_assign_pointer(*parent_node_flag_ptr
, iter_node_flag
);
1666 for (i
= 0; i
< nr_shadow
; i
++)
1667 rcuja_shadow_unlock(shadow_nodes
[i
]);
1672 int ja_unchain_node(struct cds_ja
*ja
,
1673 struct cds_ja_inode_flag
*parent_node_flag
,
1674 struct cds_ja_inode_flag
**node_flag_ptr
,
1675 struct cds_ja_inode_flag
*node_flag
,
1676 struct cds_ja_node
*node
)
1678 struct cds_ja_shadow_node
*shadow_node
;
1679 struct cds_hlist_node
*hlist_node
;
1680 struct cds_hlist_head hlist_head
;
1681 int ret
= 0, count
= 0, found
= 0;
1683 shadow_node
= rcuja_shadow_lookup_lock(ja
->ht
, parent_node_flag
);
1686 if (ja_node_ptr(*node_flag_ptr
) != ja_node_ptr(node_flag
)) {
1690 hlist_head
.next
= (struct cds_hlist_node
*) ja_node_ptr(node_flag
);
1692 * Retry if another thread removed all but one of duplicates
1693 * since check (this check was performed without lock).
1694 * Ensure that the node we are about to remove is still in the
1695 * list (while holding lock).
1697 cds_hlist_for_each_rcu(hlist_node
, &hlist_head
) {
1699 /* FIXME: currently a work-around */
1700 hlist_node
->prev
= (struct cds_hlist_node
*) node_flag_ptr
;
1703 if (hlist_node
== &node
->list
)
1707 if (!found
|| count
== 1) {
1711 cds_hlist_del_rcu(&node
->list
);
1713 * Validate that we indeed removed the node from linked list.
1715 assert(ja_node_ptr(*node_flag_ptr
) != (struct cds_ja_inode
*) node
);
1717 rcuja_shadow_unlock(shadow_node
);
1722 * Called with RCU read lock held.
1724 int cds_ja_del(struct cds_ja
*ja
, uint64_t key
,
1725 struct cds_ja_node
*node
)
1727 unsigned int tree_depth
, i
;
1728 struct cds_ja_inode_flag
*snapshot
[JA_MAX_DEPTH
];
1729 struct cds_ja_inode_flag
**snapshot_ptr
[JA_MAX_DEPTH
];
1730 uint8_t snapshot_n
[JA_MAX_DEPTH
];
1731 struct cds_ja_inode_flag
*node_flag
;
1732 struct cds_ja_inode_flag
**prev_node_flag_ptr
,
1737 if (caa_unlikely(key
> ja
->key_max
))
1739 tree_depth
= ja
->tree_depth
;
1743 dbg_printf("cds_ja_del attempt: key %" PRIu64
", node %p\n",
1746 /* snapshot for level 0 is only for shadow node lookup */
1749 snapshot_ptr
[nr_snapshot
] = NULL
;
1750 snapshot
[nr_snapshot
++] = (struct cds_ja_inode_flag
*) &ja
->root
;
1751 node_flag
= rcu_dereference(ja
->root
);
1752 prev_node_flag_ptr
= &ja
->root
;
1753 node_flag_ptr
= &ja
->root
;
1755 /* Iterate on all internal levels */
1756 for (i
= 1; i
< tree_depth
; i
++) {
1759 dbg_printf("cds_ja_del iter node_flag %p\n",
1761 if (!ja_node_ptr(node_flag
)) {
1764 iter_key
= (uint8_t) (key
>> (JA_BITS_PER_BYTE
* (tree_depth
- i
- 1)));
1765 snapshot_n
[nr_snapshot
+ 1] = iter_key
;
1766 snapshot_ptr
[nr_snapshot
] = prev_node_flag_ptr
;
1767 snapshot
[nr_snapshot
++] = node_flag
;
1768 node_flag
= ja_node_get_nth(node_flag
,
1769 &prev_node_flag_ptr
,
1773 dbg_printf("cds_ja_del iter key lookup %u finds node_flag %p, prev_node_flag_ptr %p\n",
1774 (unsigned int) iter_key
, node_flag
,
1775 prev_node_flag_ptr
);
1778 * We reached bottom of tree, try to find the node we are trying
1779 * to remove. Fail if we cannot find it.
1781 if (!ja_node_ptr(node_flag
)) {
1782 dbg_printf("cds_ja_del: no node found for key %" PRIu64
"\n",
1786 struct cds_hlist_head hlist_head
;
1787 struct cds_hlist_node
*hlist_node
;
1788 struct cds_ja_node
*entry
, *match
= NULL
;
1792 (struct cds_hlist_node
*) ja_node_ptr(node_flag
);
1793 cds_hlist_for_each_entry_rcu(entry
,
1797 dbg_printf("cds_ja_del: compare %p with entry %p\n", node
, entry
);
1803 dbg_printf("cds_ja_del: no node match for node %p key %" PRIu64
"\n", node
, key
);
1809 * Removing last of duplicates. Last snapshot
1810 * does not have a shadow node (external leafs).
1812 snapshot_ptr
[nr_snapshot
] = prev_node_flag_ptr
;
1813 snapshot
[nr_snapshot
++] = node_flag
;
1814 ret
= ja_detach_node(ja
, snapshot
, snapshot_ptr
,
1815 snapshot_n
, nr_snapshot
, key
, node
);
1817 ret
= ja_unchain_node(ja
, snapshot
[nr_snapshot
- 1],
1818 node_flag_ptr
, node_flag
, match
);
1822 * Explanation of -ENOENT handling: caused by concurrent delete
1823 * between RCU lookup and actual removal. Need to re-do the
1824 * lookup and removal attempt.
1826 if (ret
== -EAGAIN
|| ret
== -ENOENT
)
1831 struct cds_ja
*_cds_ja_new(unsigned int key_bits
,
1832 const struct rcu_flavor_struct
*flavor
)
1836 struct cds_ja_shadow_node
*root_shadow_node
;
1838 ja
= calloc(sizeof(*ja
), 1);
1850 ja
->key_max
= (1ULL << key_bits
) - 1;
1853 ja
->key_max
= UINT64_MAX
;
1859 /* ja->root is NULL */
1860 /* tree_depth 0 is for pointer to root node */
1861 ja
->tree_depth
= (key_bits
>> JA_LOG2_BITS_PER_BYTE
) + 1;
1862 assert(ja
->tree_depth
<= JA_MAX_DEPTH
);
1863 ja
->ht
= rcuja_create_ht(flavor
);
1868 * Note: we should not free this node until judy array destroy.
1870 root_shadow_node
= rcuja_shadow_set(ja
->ht
,
1871 (struct cds_ja_inode_flag
*) &ja
->root
,
1873 if (!root_shadow_node
) {
1877 root_shadow_node
->level
= 0;
1882 ret
= rcuja_delete_ht(ja
->ht
);
1892 * Called from RCU read-side CS.
1894 __attribute__((visibility("protected")))
1895 void rcuja_free_all_children(struct cds_ja_shadow_node
*shadow_node
,
1896 struct cds_ja_inode_flag
*node_flag
,
1897 void (*free_node_cb
)(struct rcu_head
*head
))
1899 const struct rcu_flavor_struct
*flavor
;
1900 unsigned int type_index
;
1901 struct cds_ja_inode
*node
;
1902 const struct cds_ja_type
*type
;
1904 flavor
= cds_lfht_rcu_flavor(shadow_node
->ja
->ht
);
1905 node
= ja_node_ptr(node_flag
);
1906 assert(node
!= NULL
);
1907 type_index
= ja_node_type(node_flag
);
1908 type
= &ja_types
[type_index
];
1910 switch (type
->type_class
) {
1914 ja_linear_node_get_nr_child(type
, node
);
1917 for (i
= 0; i
< nr_child
; i
++) {
1918 struct cds_ja_inode_flag
*iter
;
1919 struct cds_hlist_head head
;
1920 struct cds_ja_node
*entry
;
1921 struct cds_hlist_node
*pos
;
1924 ja_linear_node_get_ith_pos(type
, node
, i
, &v
, &iter
);
1927 head
.next
= (struct cds_hlist_node
*) iter
;
1928 cds_hlist_for_each_entry_rcu(entry
, pos
, &head
, list
) {
1929 flavor
->update_call_rcu(&entry
->head
, free_node_cb
);
1936 unsigned int pool_nr
;
1938 for (pool_nr
= 0; pool_nr
< (1U << type
->nr_pool_order
); pool_nr
++) {
1939 struct cds_ja_inode
*pool
=
1940 ja_pool_node_get_ith_pool(type
, node
, pool_nr
);
1942 ja_linear_node_get_nr_child(type
, pool
);
1945 for (j
= 0; j
< nr_child
; j
++) {
1946 struct cds_ja_inode_flag
*iter
;
1947 struct cds_hlist_head head
;
1948 struct cds_ja_node
*entry
;
1949 struct cds_hlist_node
*pos
;
1952 ja_linear_node_get_ith_pos(type
, node
, j
, &v
, &iter
);
1955 head
.next
= (struct cds_hlist_node
*) iter
;
1956 cds_hlist_for_each_entry_rcu(entry
, pos
, &head
, list
) {
1957 flavor
->update_call_rcu(&entry
->head
, free_node_cb
);
1970 nr_child
= shadow_node
->nr_child
;
1971 for (i
= 0; i
< nr_child
; i
++) {
1972 struct cds_ja_inode_flag
*iter
;
1973 struct cds_hlist_head head
;
1974 struct cds_ja_node
*entry
;
1975 struct cds_hlist_node
*pos
;
1977 iter
= ja_pigeon_node_get_ith_pos(type
, node
, i
);
1980 head
.next
= (struct cds_hlist_node
*) iter
;
1981 cds_hlist_for_each_entry_rcu(entry
, pos
, &head
, list
) {
1982 flavor
->update_call_rcu(&entry
->head
, free_node_cb
);
1993 * There should be no more concurrent add to the judy array while it is
1994 * being destroyed (ensured by the caller).
1996 int cds_ja_destroy(struct cds_ja
*ja
,
1997 void (*free_node_cb
)(struct rcu_head
*head
))
2001 rcuja_shadow_prune(ja
->ht
,
2002 RCUJA_SHADOW_CLEAR_FREE_NODE
| RCUJA_SHADOW_CLEAR_FREE_LOCK
,
2004 ret
= rcuja_delete_ht(ja
->ht
);
2007 if (uatomic_read(&ja
->nr_fallback
))
2009 "[warning] RCU Judy Array used %lu fallback node(s)\n",
2010 uatomic_read(&ja
->nr_fallback
));