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
= 83, .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
= 95, .max_child
= ja_type_7_max_child
, .order
= 11, },
181 { .type_class
= RCU_JA_NULL
, .min_child
= 0, .max_child
= ja_type_8_max_child
, },
183 #endif /* !(BITS_PER_LONG < 64) */
185 static inline __attribute__((unused
))
186 void static_array_size_check(void)
188 CAA_BUILD_BUG_ON(CAA_ARRAY_SIZE(ja_types
) < JA_TYPE_MAX_NR
);
192 * The cds_ja_node contains the compressed node data needed for
193 * read-side. For linear and pool node configurations, it starts with a
194 * byte counting the number of children in the node. Then, the
195 * node-specific data is placed.
196 * The node mutex, if any is needed, protecting concurrent updated of
197 * each node is placed in a separate hash table indexed by node address.
198 * For the pigeon configuration, the number of children is also kept in
199 * a separate hash table, indexed by node address, because it is only
200 * required for updates.
203 #define DECLARE_LINEAR_NODE(index) \
206 uint8_t child_value[ja_type_## index ##_max_linear_child]; \
207 struct cds_ja_inode_flag *child_ptr[ja_type_## index ##_max_linear_child]; \
210 #define DECLARE_POOL_NODE(index) \
214 uint8_t child_value[ja_type_## index ##_max_linear_child]; \
215 struct cds_ja_inode_flag *child_ptr[ja_type_## index ##_max_linear_child]; \
216 } linear[1U << ja_type_## index ##_nr_pool_order]; \
219 struct cds_ja_inode
{
221 /* Linear configuration */
222 DECLARE_LINEAR_NODE(0) conf_0
;
223 DECLARE_LINEAR_NODE(1) conf_1
;
224 DECLARE_LINEAR_NODE(2) conf_2
;
225 DECLARE_LINEAR_NODE(3) conf_3
;
226 DECLARE_LINEAR_NODE(4) conf_4
;
228 /* Pool configuration */
229 DECLARE_POOL_NODE(5) conf_5
;
230 DECLARE_POOL_NODE(6) conf_6
;
232 /* Pigeon configuration */
234 struct cds_ja_inode_flag
*child
[ja_type_7_max_child
];
236 /* data aliasing nodes for computed accesses */
237 uint8_t data
[sizeof(struct cds_ja_inode_flag
*) * ja_type_7_max_child
];
242 JA_RECOMPACT_ADD_SAME
,
243 JA_RECOMPACT_ADD_NEXT
,
248 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
));
290 struct cds_ja_inode
*alloc_cds_ja_node(struct cds_ja
*ja
,
291 const struct cds_ja_type
*ja_type
)
293 size_t len
= 1U << ja_type
->order
;
297 ret
= posix_memalign(&p
, len
, len
);
302 uatomic_inc(&ja
->nr_nodes_allocated
);
306 void free_cds_ja_node(struct cds_ja
*ja
, struct cds_ja_inode
*node
)
310 uatomic_inc(&ja
->nr_nodes_freed
);
313 #define __JA_ALIGN_MASK(v, mask) (((v) + (mask)) & ~(mask))
314 #define JA_ALIGN(v, align) __JA_ALIGN_MASK(v, (typeof(v)) (align) - 1)
315 #define __JA_FLOOR_MASK(v, mask) ((v) & ~(mask))
316 #define JA_FLOOR(v, align) __JA_FLOOR_MASK(v, (typeof(v)) (align) - 1)
319 uint8_t *align_ptr_size(uint8_t *ptr
)
321 return (uint8_t *) JA_ALIGN((unsigned long) ptr
, sizeof(void *));
325 uint8_t ja_linear_node_get_nr_child(const struct cds_ja_type
*type
,
326 struct cds_ja_inode
*node
)
328 assert(type
->type_class
== RCU_JA_LINEAR
|| type
->type_class
== RCU_JA_POOL
);
329 return rcu_dereference(node
->u
.data
[0]);
333 * The order in which values and pointers are does does not matter: if
334 * a value is missing, we return NULL. If a value is there, but its
335 * associated pointers is still NULL, we return NULL too.
338 struct cds_ja_inode_flag
*ja_linear_node_get_nth(const struct cds_ja_type
*type
,
339 struct cds_ja_inode
*node
,
340 struct cds_ja_inode_flag
***node_flag_ptr
,
345 struct cds_ja_inode_flag
**pointers
;
346 struct cds_ja_inode_flag
*ptr
;
349 assert(type
->type_class
== RCU_JA_LINEAR
|| type
->type_class
== RCU_JA_POOL
);
351 nr_child
= ja_linear_node_get_nr_child(type
, node
);
352 cmm_smp_rmb(); /* read nr_child before values and pointers */
353 assert(nr_child
<= type
->max_linear_child
);
354 assert(type
->type_class
!= RCU_JA_LINEAR
|| nr_child
>= type
->min_child
);
356 values
= &node
->u
.data
[1];
357 for (i
= 0; i
< nr_child
; i
++) {
358 if (CMM_LOAD_SHARED(values
[i
]) == n
)
362 if (caa_unlikely(node_flag_ptr
))
363 *node_flag_ptr
= NULL
;
366 pointers
= (struct cds_ja_inode_flag
**) align_ptr_size(&values
[type
->max_linear_child
]);
367 ptr
= rcu_dereference(pointers
[i
]);
368 if (caa_unlikely(node_flag_ptr
))
369 *node_flag_ptr
= &pointers
[i
];
374 void ja_linear_node_get_ith_pos(const struct cds_ja_type
*type
,
375 struct cds_ja_inode
*node
,
378 struct cds_ja_inode_flag
**iter
)
381 struct cds_ja_inode_flag
**pointers
;
383 assert(type
->type_class
== RCU_JA_LINEAR
|| type
->type_class
== RCU_JA_POOL
);
384 assert(i
< ja_linear_node_get_nr_child(type
, node
));
386 values
= &node
->u
.data
[1];
388 pointers
= (struct cds_ja_inode_flag
**) align_ptr_size(&values
[type
->max_linear_child
]);
393 struct cds_ja_inode_flag
*ja_pool_node_get_nth(const struct cds_ja_type
*type
,
394 struct cds_ja_inode
*node
,
395 struct cds_ja_inode_flag
*node_flag
,
396 struct cds_ja_inode_flag
***node_flag_ptr
,
399 struct cds_ja_inode
*linear
;
401 assert(type
->type_class
== RCU_JA_POOL
);
403 switch (type
->nr_pool_order
) {
406 unsigned long bitsel
, index
;
408 bitsel
= ja_node_pool_1d_bitsel(node_flag
);
409 assert(bitsel
< CHAR_BIT
);
410 index
= ((unsigned long) n
>> bitsel
) & 0x1;
411 linear
= (struct cds_ja_inode
*) &node
->u
.data
[index
<< type
->pool_size_order
];
416 unsigned long bitsel
[2], index
[2], rindex
;
418 ja_node_pool_2d_bitsel(node_flag
, bitsel
);
419 assert(bitsel
[0] < CHAR_BIT
);
420 assert(bitsel
[1] < CHAR_BIT
);
421 index
[0] = ((unsigned long) n
>> bitsel
[0]) & 0x1;
423 index
[1] = ((unsigned long) n
>> bitsel
[1]) & 0x1;
424 rindex
= index
[0] | index
[1];
425 linear
= (struct cds_ja_inode
*) &node
->u
.data
[rindex
<< type
->pool_size_order
];
432 return ja_linear_node_get_nth(type
, linear
, 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
***node_flag_ptr
,
451 struct cds_ja_inode_flag
**child_node_flag_ptr
;
452 struct cds_ja_inode_flag
*child_node_flag
;
454 assert(type
->type_class
== RCU_JA_PIGEON
);
455 child_node_flag_ptr
= &((struct cds_ja_inode_flag
**) node
->u
.data
)[n
];
456 child_node_flag
= rcu_dereference(*child_node_flag_ptr
);
457 dbg_printf("ja_pigeon_node_get_nth child_node_flag_ptr %p\n",
458 child_node_flag_ptr
);
459 if (caa_unlikely(node_flag_ptr
))
460 *node_flag_ptr
= child_node_flag_ptr
;
461 return child_node_flag
;
465 struct cds_ja_inode_flag
*ja_pigeon_node_get_ith_pos(const struct cds_ja_type
*type
,
466 struct cds_ja_inode
*node
,
469 return ja_pigeon_node_get_nth(type
, node
, NULL
, i
);
473 * ja_node_get_nth: get nth item from a node.
474 * node_flag is already rcu_dereference'd.
477 struct cds_ja_inode_flag
*ja_node_get_nth(struct cds_ja_inode_flag
*node_flag
,
478 struct cds_ja_inode_flag
***node_flag_ptr
,
481 unsigned int type_index
;
482 struct cds_ja_inode
*node
;
483 const struct cds_ja_type
*type
;
485 node
= ja_node_ptr(node_flag
);
486 assert(node
!= NULL
);
487 type_index
= ja_node_type(node_flag
);
488 type
= &ja_types
[type_index
];
490 switch (type
->type_class
) {
492 return ja_linear_node_get_nth(type
, node
,
495 return ja_pool_node_get_nth(type
, node
, node_flag
,
498 return ja_pigeon_node_get_nth(type
, node
,
502 return (void *) -1UL;
507 int ja_linear_node_set_nth(const struct cds_ja_type
*type
,
508 struct cds_ja_inode
*node
,
509 struct cds_ja_shadow_node
*shadow_node
,
511 struct cds_ja_inode_flag
*child_node_flag
)
514 uint8_t *values
, *nr_child_ptr
;
515 struct cds_ja_inode_flag
**pointers
;
516 unsigned int i
, unused
= 0;
518 assert(type
->type_class
== RCU_JA_LINEAR
|| type
->type_class
== RCU_JA_POOL
);
520 nr_child_ptr
= &node
->u
.data
[0];
521 dbg_printf("linear set nth: n %u, nr_child_ptr %p\n",
522 (unsigned int) n
, nr_child_ptr
);
523 nr_child
= *nr_child_ptr
;
524 assert(nr_child
<= type
->max_linear_child
);
526 values
= &node
->u
.data
[1];
527 pointers
= (struct cds_ja_inode_flag
**) align_ptr_size(&values
[type
->max_linear_child
]);
528 /* Check if node value is already populated */
529 for (i
= 0; i
< nr_child
; i
++) {
530 if (values
[i
] == n
) {
540 if (i
== nr_child
&& nr_child
>= type
->max_linear_child
) {
542 return -ERANGE
; /* recompact node */
544 return -ENOSPC
; /* No space left in this node type */
547 assert(pointers
[i
] == NULL
);
548 rcu_assign_pointer(pointers
[i
], child_node_flag
);
549 /* If we expanded the nr_child, increment it */
551 CMM_STORE_SHARED(values
[nr_child
], n
);
552 /* write pointer and value before nr_child */
554 CMM_STORE_SHARED(*nr_child_ptr
, nr_child
+ 1);
556 shadow_node
->nr_child
++;
557 dbg_printf("linear set nth: %u child, shadow: %u child, for node %p shadow %p\n",
558 (unsigned int) CMM_LOAD_SHARED(*nr_child_ptr
),
559 (unsigned int) shadow_node
->nr_child
,
566 int ja_pool_node_set_nth(const struct cds_ja_type
*type
,
567 struct cds_ja_inode
*node
,
568 struct cds_ja_inode_flag
*node_flag
,
569 struct cds_ja_shadow_node
*shadow_node
,
571 struct cds_ja_inode_flag
*child_node_flag
)
573 struct cds_ja_inode
*linear
;
575 assert(type
->type_class
== RCU_JA_POOL
);
577 switch (type
->nr_pool_order
) {
580 unsigned long bitsel
, index
;
582 bitsel
= ja_node_pool_1d_bitsel(node_flag
);
583 assert(bitsel
< CHAR_BIT
);
584 index
= ((unsigned long) n
>> bitsel
) & 0x1;
585 linear
= (struct cds_ja_inode
*) &node
->u
.data
[index
<< type
->pool_size_order
];
590 unsigned long bitsel
[2], index
[2], rindex
;
592 ja_node_pool_2d_bitsel(node_flag
, bitsel
);
593 assert(bitsel
[0] < CHAR_BIT
);
594 assert(bitsel
[1] < CHAR_BIT
);
595 index
[0] = ((unsigned long) n
>> bitsel
[0]) & 0x1;
597 index
[1] = ((unsigned long) n
>> bitsel
[1]) & 0x1;
598 rindex
= index
[0] | index
[1];
599 linear
= (struct cds_ja_inode
*) &node
->u
.data
[rindex
<< type
->pool_size_order
];
607 return ja_linear_node_set_nth(type
, linear
, shadow_node
,
612 int ja_pigeon_node_set_nth(const struct cds_ja_type
*type
,
613 struct cds_ja_inode
*node
,
614 struct cds_ja_shadow_node
*shadow_node
,
616 struct cds_ja_inode_flag
*child_node_flag
)
618 struct cds_ja_inode_flag
**ptr
;
620 assert(type
->type_class
== RCU_JA_PIGEON
);
621 ptr
= &((struct cds_ja_inode_flag
**) node
->u
.data
)[n
];
624 rcu_assign_pointer(*ptr
, child_node_flag
);
625 shadow_node
->nr_child
++;
630 * _ja_node_set_nth: set nth item within a node. Return an error
631 * (negative error value) if it is already there.
634 int _ja_node_set_nth(const struct cds_ja_type
*type
,
635 struct cds_ja_inode
*node
,
636 struct cds_ja_inode_flag
*node_flag
,
637 struct cds_ja_shadow_node
*shadow_node
,
639 struct cds_ja_inode_flag
*child_node_flag
)
641 switch (type
->type_class
) {
643 return ja_linear_node_set_nth(type
, node
, shadow_node
, n
,
646 return ja_pool_node_set_nth(type
, node
, node_flag
, shadow_node
, n
,
649 return ja_pigeon_node_set_nth(type
, node
, shadow_node
, n
,
662 int ja_linear_node_clear_ptr(const struct cds_ja_type
*type
,
663 struct cds_ja_inode
*node
,
664 struct cds_ja_shadow_node
*shadow_node
,
665 struct cds_ja_inode_flag
**node_flag_ptr
)
668 uint8_t *nr_child_ptr
;
670 assert(type
->type_class
== RCU_JA_LINEAR
|| type
->type_class
== RCU_JA_POOL
);
672 nr_child_ptr
= &node
->u
.data
[0];
673 nr_child
= *nr_child_ptr
;
674 assert(nr_child
<= type
->max_linear_child
);
676 if (type
->type_class
== RCU_JA_LINEAR
) {
677 assert(!shadow_node
->fallback_removal_count
);
678 if (shadow_node
->nr_child
<= type
->min_child
) {
679 /* We need to try recompacting the node */
683 dbg_printf("linear clear ptr: nr_child_ptr %p\n", nr_child_ptr
);
684 assert(*node_flag_ptr
!= NULL
);
685 rcu_assign_pointer(*node_flag_ptr
, NULL
);
687 * Value and nr_child are never changed (would cause ABA issue).
688 * Instead, we leave the pointer to NULL and recompact the node
689 * once in a while. It is allowed to set a NULL pointer to a new
690 * value without recompaction though.
691 * Only update the shadow node accounting.
693 shadow_node
->nr_child
--;
694 dbg_printf("linear clear ptr: %u child, shadow: %u child, for node %p shadow %p\n",
695 (unsigned int) CMM_LOAD_SHARED(*nr_child_ptr
),
696 (unsigned int) shadow_node
->nr_child
,
702 int ja_pool_node_clear_ptr(const struct cds_ja_type
*type
,
703 struct cds_ja_inode
*node
,
704 struct cds_ja_inode_flag
*node_flag
,
705 struct cds_ja_shadow_node
*shadow_node
,
706 struct cds_ja_inode_flag
**node_flag_ptr
,
709 struct cds_ja_inode
*linear
;
711 assert(type
->type_class
== RCU_JA_POOL
);
713 if (shadow_node
->fallback_removal_count
) {
714 shadow_node
->fallback_removal_count
--;
716 /* We should try recompacting the node */
717 if (shadow_node
->nr_child
<= type
->min_child
)
721 switch (type
->nr_pool_order
) {
724 unsigned long bitsel
, index
;
726 bitsel
= ja_node_pool_1d_bitsel(node_flag
);
727 assert(bitsel
< CHAR_BIT
);
728 index
= ((unsigned long) n
>> bitsel
) & type
->nr_pool_order
;
729 linear
= (struct cds_ja_inode
*) &node
->u
.data
[index
<< type
->pool_size_order
];
734 unsigned long bitsel
[2], index
[2], rindex
;
736 ja_node_pool_2d_bitsel(node_flag
, bitsel
);
737 assert(bitsel
[0] < CHAR_BIT
);
738 assert(bitsel
[1] < CHAR_BIT
);
739 index
[0] = ((unsigned long) n
>> bitsel
[0]) & 0x1;
741 index
[1] = ((unsigned long) n
>> bitsel
[1]) & 0x1;
742 rindex
= index
[0] | index
[1];
743 linear
= (struct cds_ja_inode
*) &node
->u
.data
[rindex
<< type
->pool_size_order
];
751 return ja_linear_node_clear_ptr(type
, linear
, shadow_node
, node_flag_ptr
);
755 int ja_pigeon_node_clear_ptr(const struct cds_ja_type
*type
,
756 struct cds_ja_inode
*node
,
757 struct cds_ja_shadow_node
*shadow_node
,
758 struct cds_ja_inode_flag
**node_flag_ptr
)
760 assert(type
->type_class
== RCU_JA_PIGEON
);
762 if (shadow_node
->fallback_removal_count
) {
763 shadow_node
->fallback_removal_count
--;
765 /* We should try recompacting the node */
766 if (shadow_node
->nr_child
<= type
->min_child
)
769 dbg_printf("ja_pigeon_node_clear_ptr: clearing ptr: %p\n", *node_flag_ptr
);
770 rcu_assign_pointer(*node_flag_ptr
, NULL
);
771 shadow_node
->nr_child
--;
776 * _ja_node_clear_ptr: clear ptr item within a node. Return an error
777 * (negative error value) if it is not found (-ENOENT).
780 int _ja_node_clear_ptr(const struct cds_ja_type
*type
,
781 struct cds_ja_inode
*node
,
782 struct cds_ja_inode_flag
*node_flag
,
783 struct cds_ja_shadow_node
*shadow_node
,
784 struct cds_ja_inode_flag
**node_flag_ptr
,
787 switch (type
->type_class
) {
789 return ja_linear_node_clear_ptr(type
, node
, shadow_node
, node_flag_ptr
);
791 return ja_pool_node_clear_ptr(type
, node
, node_flag
, shadow_node
, node_flag_ptr
, n
);
793 return ja_pigeon_node_clear_ptr(type
, node
, shadow_node
, node_flag_ptr
);
805 * Calculate bit distribution. Returns the bit (0 to 7) that splits the
806 * distribution in two sub-distributions containing as much elements one
807 * compared to the other.
810 unsigned int ja_node_sum_distribution_1d(enum ja_recompact mode
,
812 unsigned int type_index
,
813 const struct cds_ja_type
*type
,
814 struct cds_ja_inode
*node
,
815 struct cds_ja_shadow_node
*shadow_node
,
817 struct cds_ja_inode_flag
*child_node_flag
,
818 struct cds_ja_inode_flag
**nullify_node_flag_ptr
)
820 uint8_t nr_one
[JA_BITS_PER_BYTE
];
821 unsigned int bitsel
= 0, bit_i
, overall_best_distance
= UINT_MAX
;
822 unsigned int distrib_nr_child
= 0;
824 memset(nr_one
, 0, sizeof(nr_one
));
826 switch (type
->type_class
) {
830 ja_linear_node_get_nr_child(type
, node
);
833 for (i
= 0; i
< nr_child
; i
++) {
834 struct cds_ja_inode_flag
*iter
;
837 ja_linear_node_get_ith_pos(type
, node
, i
, &v
, &iter
);
840 if (mode
== JA_RECOMPACT_DEL
&& *nullify_node_flag_ptr
== iter
)
842 for (bit_i
= 0; bit_i
< JA_BITS_PER_BYTE
; bit_i
++) {
843 if (v
& (1U << bit_i
))
852 unsigned int pool_nr
;
854 for (pool_nr
= 0; pool_nr
< (1U << type
->nr_pool_order
); pool_nr
++) {
855 struct cds_ja_inode
*pool
=
856 ja_pool_node_get_ith_pool(type
,
859 ja_linear_node_get_nr_child(type
, pool
);
862 for (j
= 0; j
< nr_child
; j
++) {
863 struct cds_ja_inode_flag
*iter
;
866 ja_linear_node_get_ith_pos(type
, pool
,
870 if (mode
== JA_RECOMPACT_DEL
&& *nullify_node_flag_ptr
== iter
)
872 for (bit_i
= 0; bit_i
< JA_BITS_PER_BYTE
; bit_i
++) {
873 if (v
& (1U << bit_i
))
885 assert(mode
== JA_RECOMPACT_DEL
);
886 for (i
= 0; i
< JA_ENTRY_PER_NODE
; i
++) {
887 struct cds_ja_inode_flag
*iter
;
889 iter
= ja_pigeon_node_get_ith_pos(type
, node
, i
);
892 if (mode
== JA_RECOMPACT_DEL
&& *nullify_node_flag_ptr
== iter
)
894 for (bit_i
= 0; bit_i
< JA_BITS_PER_BYTE
; bit_i
++) {
895 if (i
& (1U << bit_i
))
903 assert(mode
== JA_RECOMPACT_ADD_NEXT
);
910 if (mode
== JA_RECOMPACT_ADD_NEXT
|| mode
== JA_RECOMPACT_ADD_SAME
) {
911 for (bit_i
= 0; bit_i
< JA_BITS_PER_BYTE
; bit_i
++) {
912 if (n
& (1U << bit_i
))
919 * The best bit selector is that for which the number of ones is
920 * closest to half of the number of children in the
921 * distribution. We calculate the distance using the double of
922 * the sub-distribution sizes to eliminate truncation error.
924 for (bit_i
= 0; bit_i
< JA_BITS_PER_BYTE
; bit_i
++) {
925 unsigned int distance_to_best
;
927 distance_to_best
= abs_int((nr_one
[bit_i
] << 1U) - distrib_nr_child
);
928 if (distance_to_best
< overall_best_distance
) {
929 overall_best_distance
= distance_to_best
;
933 dbg_printf("1 dimension pool bit selection: (%u)\n", bitsel
);
938 * Calculate bit distribution in two dimensions. Returns the two bits
939 * (each 0 to 7) that splits the distribution in four sub-distributions
940 * containing as much elements one compared to the other.
943 void ja_node_sum_distribution_2d(enum ja_recompact mode
,
945 unsigned int type_index
,
946 const struct cds_ja_type
*type
,
947 struct cds_ja_inode
*node
,
948 struct cds_ja_shadow_node
*shadow_node
,
950 struct cds_ja_inode_flag
*child_node_flag
,
951 struct cds_ja_inode_flag
**nullify_node_flag_ptr
,
952 unsigned int *_bitsel
)
954 uint8_t nr_2d_11
[JA_BITS_PER_BYTE
][JA_BITS_PER_BYTE
],
955 nr_2d_10
[JA_BITS_PER_BYTE
][JA_BITS_PER_BYTE
],
956 nr_2d_01
[JA_BITS_PER_BYTE
][JA_BITS_PER_BYTE
],
957 nr_2d_00
[JA_BITS_PER_BYTE
][JA_BITS_PER_BYTE
];
958 unsigned int bitsel
[2] = { 0, 1 };
959 unsigned int bit_i
, bit_j
;
960 int overall_best_distance
= INT_MAX
;
961 unsigned int distrib_nr_child
= 0;
963 memset(nr_2d_11
, 0, sizeof(nr_2d_11
));
964 memset(nr_2d_10
, 0, sizeof(nr_2d_10
));
965 memset(nr_2d_01
, 0, sizeof(nr_2d_01
));
966 memset(nr_2d_00
, 0, sizeof(nr_2d_00
));
968 switch (type
->type_class
) {
972 ja_linear_node_get_nr_child(type
, node
);
975 for (i
= 0; i
< nr_child
; i
++) {
976 struct cds_ja_inode_flag
*iter
;
979 ja_linear_node_get_ith_pos(type
, node
, i
, &v
, &iter
);
982 if (mode
== JA_RECOMPACT_DEL
&& *nullify_node_flag_ptr
== iter
)
984 for (bit_i
= 0; bit_i
< JA_BITS_PER_BYTE
; bit_i
++) {
985 for (bit_j
= 0; bit_j
< bit_i
; bit_j
++) {
986 if ((v
& (1U << bit_i
)) && (v
& (1U << bit_j
))) {
987 nr_2d_11
[bit_i
][bit_j
]++;
989 if ((v
& (1U << bit_i
)) && !(v
& (1U << bit_j
))) {
990 nr_2d_10
[bit_i
][bit_j
]++;
992 if (!(v
& (1U << bit_i
)) && (v
& (1U << bit_j
))) {
993 nr_2d_01
[bit_i
][bit_j
]++;
995 if (!(v
& (1U << bit_i
)) && !(v
& (1U << bit_j
))) {
996 nr_2d_00
[bit_i
][bit_j
]++;
1006 unsigned int pool_nr
;
1008 for (pool_nr
= 0; pool_nr
< (1U << type
->nr_pool_order
); pool_nr
++) {
1009 struct cds_ja_inode
*pool
=
1010 ja_pool_node_get_ith_pool(type
,
1013 ja_linear_node_get_nr_child(type
, pool
);
1016 for (j
= 0; j
< nr_child
; j
++) {
1017 struct cds_ja_inode_flag
*iter
;
1020 ja_linear_node_get_ith_pos(type
, pool
,
1024 if (mode
== JA_RECOMPACT_DEL
&& *nullify_node_flag_ptr
== iter
)
1026 for (bit_i
= 0; bit_i
< JA_BITS_PER_BYTE
; bit_i
++) {
1027 for (bit_j
= 0; bit_j
< bit_i
; bit_j
++) {
1028 if ((v
& (1U << bit_i
)) && (v
& (1U << bit_j
))) {
1029 nr_2d_11
[bit_i
][bit_j
]++;
1031 if ((v
& (1U << bit_i
)) && !(v
& (1U << bit_j
))) {
1032 nr_2d_10
[bit_i
][bit_j
]++;
1034 if (!(v
& (1U << bit_i
)) && (v
& (1U << bit_j
))) {
1035 nr_2d_01
[bit_i
][bit_j
]++;
1037 if (!(v
& (1U << bit_i
)) && !(v
& (1U << bit_j
))) {
1038 nr_2d_00
[bit_i
][bit_j
]++;
1051 assert(mode
== JA_RECOMPACT_DEL
);
1052 for (i
= 0; i
< JA_ENTRY_PER_NODE
; i
++) {
1053 struct cds_ja_inode_flag
*iter
;
1055 iter
= ja_pigeon_node_get_ith_pos(type
, node
, i
);
1058 if (mode
== JA_RECOMPACT_DEL
&& *nullify_node_flag_ptr
== iter
)
1060 for (bit_i
= 0; bit_i
< JA_BITS_PER_BYTE
; bit_i
++) {
1061 for (bit_j
= 0; bit_j
< bit_i
; bit_j
++) {
1062 if ((i
& (1U << bit_i
)) && (i
& (1U << bit_j
))) {
1063 nr_2d_11
[bit_i
][bit_j
]++;
1065 if ((i
& (1U << bit_i
)) && !(i
& (1U << bit_j
))) {
1066 nr_2d_10
[bit_i
][bit_j
]++;
1068 if (!(i
& (1U << bit_i
)) && (i
& (1U << bit_j
))) {
1069 nr_2d_01
[bit_i
][bit_j
]++;
1071 if (!(i
& (1U << bit_i
)) && !(i
& (1U << bit_j
))) {
1072 nr_2d_00
[bit_i
][bit_j
]++;
1081 assert(mode
== JA_RECOMPACT_ADD_NEXT
);
1088 if (mode
== JA_RECOMPACT_ADD_NEXT
|| mode
== JA_RECOMPACT_ADD_SAME
) {
1089 for (bit_i
= 0; bit_i
< JA_BITS_PER_BYTE
; bit_i
++) {
1090 for (bit_j
= 0; bit_j
< bit_i
; bit_j
++) {
1091 if ((n
& (1U << bit_i
)) && (n
& (1U << bit_j
))) {
1092 nr_2d_11
[bit_i
][bit_j
]++;
1094 if ((n
& (1U << bit_i
)) && !(n
& (1U << bit_j
))) {
1095 nr_2d_10
[bit_i
][bit_j
]++;
1097 if (!(n
& (1U << bit_i
)) && (n
& (1U << bit_j
))) {
1098 nr_2d_01
[bit_i
][bit_j
]++;
1100 if (!(n
& (1U << bit_i
)) && !(n
& (1U << bit_j
))) {
1101 nr_2d_00
[bit_i
][bit_j
]++;
1109 * The best bit selector is that for which the number of nodes
1110 * in each sub-class is closest to one-fourth of the number of
1111 * children in the distribution. We calculate the distance using
1112 * 4 times the size of the sub-distribution to eliminate
1115 for (bit_i
= 0; bit_i
< JA_BITS_PER_BYTE
; bit_i
++) {
1116 for (bit_j
= 0; bit_j
< bit_i
; bit_j
++) {
1117 int distance_to_best
[4];
1119 distance_to_best
[0] = (nr_2d_11
[bit_i
][bit_j
] << 2U) - distrib_nr_child
;
1120 distance_to_best
[1] = (nr_2d_10
[bit_i
][bit_j
] << 2U) - distrib_nr_child
;
1121 distance_to_best
[2] = (nr_2d_01
[bit_i
][bit_j
] << 2U) - distrib_nr_child
;
1122 distance_to_best
[3] = (nr_2d_00
[bit_i
][bit_j
] << 2U) - distrib_nr_child
;
1124 /* Consider worse distance above best */
1125 if (distance_to_best
[1] > 0 && distance_to_best
[1] > distance_to_best
[0])
1126 distance_to_best
[0] = distance_to_best
[1];
1127 if (distance_to_best
[2] > 0 && distance_to_best
[2] > distance_to_best
[0])
1128 distance_to_best
[0] = distance_to_best
[2];
1129 if (distance_to_best
[3] > 0 && distance_to_best
[3] > distance_to_best
[0])
1130 distance_to_best
[0] = distance_to_best
[3];
1133 * If our worse distance is better than overall,
1134 * we become new best candidate.
1136 if (distance_to_best
[0] < overall_best_distance
) {
1137 overall_best_distance
= distance_to_best
[0];
1144 dbg_printf("2 dimensions pool bit selection: (%u,%u)\n", bitsel
[0], bitsel
[1]);
1146 /* Return our bit selection */
1147 _bitsel
[0] = bitsel
[0];
1148 _bitsel
[1] = bitsel
[1];
1152 unsigned int find_nearest_type_index(unsigned int type_index
,
1153 unsigned int nr_nodes
)
1155 const struct cds_ja_type
*type
;
1157 assert(type_index
!= NODE_INDEX_NULL
);
1159 return NODE_INDEX_NULL
;
1161 type
= &ja_types
[type_index
];
1162 if (nr_nodes
< type
->min_child
)
1164 else if (nr_nodes
> type
->max_child
)
1173 * ja_node_recompact_add: recompact a node, adding a new child.
1174 * Return 0 on success, -EAGAIN if need to retry, or other negative
1175 * error value otherwise.
1178 int ja_node_recompact(enum ja_recompact mode
,
1180 unsigned int old_type_index
,
1181 const struct cds_ja_type
*old_type
,
1182 struct cds_ja_inode
*old_node
,
1183 struct cds_ja_shadow_node
*shadow_node
,
1184 struct cds_ja_inode_flag
**old_node_flag_ptr
, uint8_t n
,
1185 struct cds_ja_inode_flag
*child_node_flag
,
1186 struct cds_ja_inode_flag
**nullify_node_flag_ptr
,
1189 unsigned int new_type_index
;
1190 struct cds_ja_inode
*new_node
;
1191 struct cds_ja_shadow_node
*new_shadow_node
= NULL
;
1192 const struct cds_ja_type
*new_type
;
1193 struct cds_ja_inode_flag
*new_node_flag
, *old_node_flag
;
1197 old_node_flag
= *old_node_flag_ptr
;
1200 * Need to find nearest type index even for ADD_SAME, because
1201 * this recompaction, when applied to linear nodes, will garbage
1202 * collect dummy (NULL) entries, and can therefore cause a few
1203 * linear representations to be skipped.
1206 case JA_RECOMPACT_ADD_SAME
:
1207 new_type_index
= find_nearest_type_index(old_type_index
,
1208 shadow_node
->nr_child
+ 1);
1209 dbg_printf("Recompact for node with %u children\n",
1210 shadow_node
->nr_child
+ 1);
1212 case JA_RECOMPACT_ADD_NEXT
:
1213 if (!shadow_node
|| old_type_index
== NODE_INDEX_NULL
) {
1215 dbg_printf("Recompact for NULL\n");
1217 new_type_index
= find_nearest_type_index(old_type_index
,
1218 shadow_node
->nr_child
+ 1);
1219 dbg_printf("Recompact for node with %u children\n",
1220 shadow_node
->nr_child
+ 1);
1223 case JA_RECOMPACT_DEL
:
1224 new_type_index
= find_nearest_type_index(old_type_index
,
1225 shadow_node
->nr_child
- 1);
1226 dbg_printf("Recompact for node with %u children\n",
1227 shadow_node
->nr_child
- 1);
1233 retry
: /* for fallback */
1234 dbg_printf("Recompact from type %d to type %d\n",
1235 old_type_index
, new_type_index
);
1236 new_type
= &ja_types
[new_type_index
];
1237 if (new_type_index
!= NODE_INDEX_NULL
) {
1238 new_node
= alloc_cds_ja_node(ja
, new_type
);
1242 if (new_type
->type_class
== RCU_JA_POOL
) {
1243 switch (new_type
->nr_pool_order
) {
1246 unsigned int node_distrib_bitsel
;
1248 node_distrib_bitsel
=
1249 ja_node_sum_distribution_1d(mode
, ja
,
1250 old_type_index
, old_type
,
1251 old_node
, shadow_node
,
1253 nullify_node_flag_ptr
);
1254 assert(!((unsigned long) new_node
& JA_POOL_1D_MASK
));
1255 new_node_flag
= ja_node_flag_pool_1d(new_node
,
1256 new_type_index
, node_distrib_bitsel
);
1261 unsigned int node_distrib_bitsel
[2];
1263 ja_node_sum_distribution_2d(mode
, ja
,
1264 old_type_index
, old_type
,
1265 old_node
, shadow_node
,
1267 nullify_node_flag_ptr
,
1268 node_distrib_bitsel
);
1269 assert(!((unsigned long) new_node
& JA_POOL_1D_MASK
));
1270 assert(!((unsigned long) new_node
& JA_POOL_2D_MASK
));
1271 new_node_flag
= ja_node_flag_pool_2d(new_node
,
1272 new_type_index
, node_distrib_bitsel
);
1279 new_node_flag
= ja_node_flag(new_node
, new_type_index
);
1282 dbg_printf("Recompact inherit lock from %p\n", shadow_node
);
1283 new_shadow_node
= rcuja_shadow_set(ja
->ht
, new_node_flag
, shadow_node
, ja
, level
);
1284 if (!new_shadow_node
) {
1285 free_cds_ja_node(ja
, new_node
);
1289 new_shadow_node
->fallback_removal_count
=
1290 JA_FALLBACK_REMOVAL_COUNT
;
1293 new_node_flag
= NULL
;
1296 assert(mode
!= JA_RECOMPACT_ADD_NEXT
|| old_type
->type_class
!= RCU_JA_PIGEON
);
1298 if (new_type_index
== NODE_INDEX_NULL
)
1301 switch (old_type
->type_class
) {
1305 ja_linear_node_get_nr_child(old_type
, old_node
);
1308 for (i
= 0; i
< nr_child
; i
++) {
1309 struct cds_ja_inode_flag
*iter
;
1312 ja_linear_node_get_ith_pos(old_type
, old_node
, i
, &v
, &iter
);
1315 if (mode
== JA_RECOMPACT_DEL
&& *nullify_node_flag_ptr
== iter
)
1317 ret
= _ja_node_set_nth(new_type
, new_node
, new_node_flag
,
1320 if (new_type
->type_class
== RCU_JA_POOL
&& ret
) {
1321 goto fallback_toosmall
;
1329 unsigned int pool_nr
;
1331 for (pool_nr
= 0; pool_nr
< (1U << old_type
->nr_pool_order
); pool_nr
++) {
1332 struct cds_ja_inode
*pool
=
1333 ja_pool_node_get_ith_pool(old_type
,
1336 ja_linear_node_get_nr_child(old_type
, pool
);
1339 for (j
= 0; j
< nr_child
; j
++) {
1340 struct cds_ja_inode_flag
*iter
;
1343 ja_linear_node_get_ith_pos(old_type
, pool
,
1347 if (mode
== JA_RECOMPACT_DEL
&& *nullify_node_flag_ptr
== iter
)
1349 ret
= _ja_node_set_nth(new_type
, new_node
, new_node_flag
,
1352 if (new_type
->type_class
== RCU_JA_POOL
1354 goto fallback_toosmall
;
1362 assert(mode
== JA_RECOMPACT_ADD_NEXT
);
1368 assert(mode
== JA_RECOMPACT_DEL
);
1369 for (i
= 0; i
< JA_ENTRY_PER_NODE
; i
++) {
1370 struct cds_ja_inode_flag
*iter
;
1372 iter
= ja_pigeon_node_get_ith_pos(old_type
, old_node
, i
);
1375 if (mode
== JA_RECOMPACT_DEL
&& *nullify_node_flag_ptr
== iter
)
1377 ret
= _ja_node_set_nth(new_type
, new_node
, new_node_flag
,
1380 if (new_type
->type_class
== RCU_JA_POOL
&& ret
) {
1381 goto fallback_toosmall
;
1394 if (mode
== JA_RECOMPACT_ADD_NEXT
|| mode
== JA_RECOMPACT_ADD_SAME
) {
1396 ret
= _ja_node_set_nth(new_type
, new_node
, new_node_flag
,
1398 n
, child_node_flag
);
1399 if (new_type
->type_class
== RCU_JA_POOL
&& ret
) {
1400 goto fallback_toosmall
;
1406 dbg_printf("Using fallback for %u children, node type index: %u, mode %s\n",
1407 new_shadow_node
->nr_child
, old_type_index
, mode
== JA_RECOMPACT_ADD_NEXT
? "add_next" :
1408 (mode
== JA_RECOMPACT_DEL
? "del" : "add_same"));
1409 uatomic_inc(&ja
->node_fallback_count_distribution
[new_shadow_node
->nr_child
]);
1412 /* Return pointer to new recompacted node through old_node_flag_ptr */
1413 *old_node_flag_ptr
= new_node_flag
;
1417 flags
= RCUJA_SHADOW_CLEAR_FREE_NODE
;
1419 * It is OK to free the lock associated with a node
1420 * going to NULL, since we are holding the parent lock.
1421 * This synchronizes removal with re-add of that node.
1423 if (new_type_index
== NODE_INDEX_NULL
)
1424 flags
|= RCUJA_SHADOW_CLEAR_FREE_LOCK
;
1425 ret
= rcuja_shadow_clear(ja
->ht
, old_node_flag
, shadow_node
,
1435 /* fallback if next pool is too small */
1436 assert(new_shadow_node
);
1437 ret
= rcuja_shadow_clear(ja
->ht
, new_node_flag
, new_shadow_node
,
1438 RCUJA_SHADOW_CLEAR_FREE_NODE
);
1442 case JA_RECOMPACT_ADD_SAME
:
1444 * JA_RECOMPACT_ADD_SAME is only triggered if a linear
1445 * node within a pool has unused entries. It should
1446 * therefore _never_ be too small.
1451 case JA_RECOMPACT_ADD_NEXT
:
1453 const struct cds_ja_type
*next_type
;
1456 * Recompaction attempt on add failed. Should only
1457 * happen if target node type is pool. Caused by
1458 * hard-to-split distribution. Recompact using the next
1459 * distribution size.
1461 assert(new_type
->type_class
== RCU_JA_POOL
);
1462 next_type
= &ja_types
[new_type_index
+ 1];
1464 * Try going to the next pool size if our population
1465 * fits within its range. This is not flagged as a
1468 if (shadow_node
->nr_child
+ 1 >= next_type
->min_child
1469 && shadow_node
->nr_child
+ 1 <= next_type
->max_child
) {
1474 dbg_printf("Add fallback to type %d\n", new_type_index
);
1475 uatomic_inc(&ja
->nr_fallback
);
1481 case JA_RECOMPACT_DEL
:
1483 * Recompaction attempt on delete failed. Should only
1484 * happen if target node type is pool. This is caused by
1485 * a hard-to-split distribution. Recompact on same node
1486 * size, but flag current node as "fallback" to ensure
1487 * we don't attempt recompaction before some activity
1488 * has reshuffled our node.
1490 assert(new_type
->type_class
== RCU_JA_POOL
);
1491 new_type_index
= old_type_index
;
1492 dbg_printf("Delete fallback keeping type %d\n", new_type_index
);
1493 uatomic_inc(&ja
->nr_fallback
);
1502 * Last resort fallback: pigeon.
1504 new_type_index
= (1UL << JA_TYPE_BITS
) - 1;
1505 dbg_printf("Fallback to type %d\n", new_type_index
);
1506 uatomic_inc(&ja
->nr_fallback
);
1512 * Return 0 on success, -EAGAIN if need to retry, or other negative
1513 * error value otherwise.
1516 int ja_node_set_nth(struct cds_ja
*ja
,
1517 struct cds_ja_inode_flag
**node_flag
, uint8_t n
,
1518 struct cds_ja_inode_flag
*child_node_flag
,
1519 struct cds_ja_shadow_node
*shadow_node
,
1523 unsigned int type_index
;
1524 const struct cds_ja_type
*type
;
1525 struct cds_ja_inode
*node
;
1527 dbg_printf("ja_node_set_nth for n=%u, node %p, shadow %p\n",
1528 (unsigned int) n
, ja_node_ptr(*node_flag
), shadow_node
);
1530 node
= ja_node_ptr(*node_flag
);
1531 type_index
= ja_node_type(*node_flag
);
1532 type
= &ja_types
[type_index
];
1533 ret
= _ja_node_set_nth(type
, node
, *node_flag
, shadow_node
,
1534 n
, child_node_flag
);
1537 /* Not enough space in node, need to recompact to next type. */
1538 ret
= ja_node_recompact(JA_RECOMPACT_ADD_NEXT
, ja
, type_index
, type
, node
,
1539 shadow_node
, node_flag
, n
, child_node_flag
, NULL
, level
);
1542 /* Node needs to be recompacted. */
1543 ret
= ja_node_recompact(JA_RECOMPACT_ADD_SAME
, ja
, type_index
, type
, node
,
1544 shadow_node
, node_flag
, n
, child_node_flag
, NULL
, level
);
1551 * Return 0 on success, -EAGAIN if need to retry, or other negative
1552 * error value otherwise.
1555 int ja_node_clear_ptr(struct cds_ja
*ja
,
1556 struct cds_ja_inode_flag
**node_flag_ptr
, /* Pointer to location to nullify */
1557 struct cds_ja_inode_flag
**parent_node_flag_ptr
, /* Address of parent ptr in its parent */
1558 struct cds_ja_shadow_node
*shadow_node
, /* of parent */
1559 uint8_t n
, int level
)
1562 unsigned int type_index
;
1563 const struct cds_ja_type
*type
;
1564 struct cds_ja_inode
*node
;
1566 dbg_printf("ja_node_clear_ptr for node %p, shadow %p, target ptr %p\n",
1567 ja_node_ptr(*parent_node_flag_ptr
), shadow_node
, node_flag_ptr
);
1569 node
= ja_node_ptr(*parent_node_flag_ptr
);
1570 type_index
= ja_node_type(*parent_node_flag_ptr
);
1571 type
= &ja_types
[type_index
];
1572 ret
= _ja_node_clear_ptr(type
, node
, *parent_node_flag_ptr
, shadow_node
, node_flag_ptr
, n
);
1573 if (ret
== -EFBIG
) {
1574 /* Should try recompaction. */
1575 ret
= ja_node_recompact(JA_RECOMPACT_DEL
, ja
, type_index
, type
, node
,
1576 shadow_node
, parent_node_flag_ptr
, n
, NULL
,
1577 node_flag_ptr
, level
);
1582 struct cds_hlist_head
cds_ja_lookup(struct cds_ja
*ja
, uint64_t key
)
1584 unsigned int tree_depth
, i
;
1585 struct cds_ja_inode_flag
*node_flag
;
1586 struct cds_hlist_head head
= { NULL
};
1588 if (caa_unlikely(key
> ja
->key_max
))
1590 tree_depth
= ja
->tree_depth
;
1591 node_flag
= rcu_dereference(ja
->root
);
1593 /* level 0: root node */
1594 if (!ja_node_ptr(node_flag
))
1597 for (i
= 1; i
< tree_depth
; i
++) {
1600 iter_key
= (uint8_t) (key
>> (JA_BITS_PER_BYTE
* (tree_depth
- i
- 1)));
1601 node_flag
= ja_node_get_nth(node_flag
, NULL
, iter_key
);
1602 dbg_printf("cds_ja_lookup iter key lookup %u finds node_flag %p\n",
1603 (unsigned int) iter_key
, node_flag
);
1604 if (!ja_node_ptr(node_flag
))
1608 /* Last level lookup succeded. We got an actual match. */
1609 head
.next
= (struct cds_hlist_node
*) node_flag
;
1614 * We reached an unpopulated node. Create it and the children we need,
1615 * and then attach the entire branch to the current node. This may
1616 * trigger recompaction of the current node. Locks needed: node lock
1617 * (for add), and, possibly, parent node lock (to update pointer due to
1618 * node recompaction).
1620 * First take node lock, check if recompaction is needed, then take
1621 * parent lock (if needed). Then we can proceed to create the new
1622 * branch. Publish the new branch, and release locks.
1623 * TODO: we currently always take the parent lock even when not needed.
1626 int ja_attach_node(struct cds_ja
*ja
,
1627 struct cds_ja_inode_flag
**attach_node_flag_ptr
,
1628 struct cds_ja_inode_flag
*attach_node_flag
,
1629 struct cds_ja_inode_flag
*parent_attach_node_flag
,
1630 struct cds_ja_inode_flag
**old_node_flag_ptr
,
1631 struct cds_ja_inode_flag
*old_node_flag
,
1634 struct cds_ja_node
*child_node
)
1636 struct cds_ja_shadow_node
*shadow_node
= NULL
,
1637 *parent_shadow_node
= NULL
;
1638 struct cds_hlist_head head
;
1639 struct cds_ja_inode_flag
*iter_node_flag
, *iter_dest_node_flag
;
1641 struct cds_ja_inode_flag
*created_nodes
[JA_MAX_DEPTH
];
1642 int nr_created_nodes
= 0;
1644 dbg_printf("Attach node at level %u (old_node_flag %p, attach_node_flag_ptr %p attach_node_flag %p, parent_attach_node_flag %p)\n",
1645 level
, old_node_flag
, attach_node_flag_ptr
, attach_node_flag
, parent_attach_node_flag
);
1647 assert(!old_node_flag
);
1648 if (attach_node_flag
) {
1649 shadow_node
= rcuja_shadow_lookup_lock(ja
->ht
, attach_node_flag
);
1655 if (parent_attach_node_flag
) {
1656 parent_shadow_node
= rcuja_shadow_lookup_lock(ja
->ht
,
1657 parent_attach_node_flag
);
1658 if (!parent_shadow_node
) {
1664 if (old_node_flag_ptr
&& ja_node_ptr(*old_node_flag_ptr
)) {
1666 * Target node has been updated between RCU lookup and
1667 * lock acquisition. We need to re-try lookup and
1675 * Perform a lookup query to handle the case where
1676 * old_node_flag_ptr is NULL. We cannot use it to check if the
1677 * node has been populated between RCU lookup and mutex
1680 if (!old_node_flag_ptr
) {
1682 struct cds_ja_inode_flag
*lookup_node_flag
;
1683 struct cds_ja_inode_flag
**lookup_node_flag_ptr
;
1685 iter_key
= (uint8_t) (key
>> (JA_BITS_PER_BYTE
* (ja
->tree_depth
- level
)));
1686 lookup_node_flag
= ja_node_get_nth(attach_node_flag
,
1687 &lookup_node_flag_ptr
,
1689 if (lookup_node_flag
) {
1695 if (attach_node_flag_ptr
&& ja_node_ptr(*attach_node_flag_ptr
) !=
1696 ja_node_ptr(attach_node_flag
)) {
1698 * Target node has been updated between RCU lookup and
1699 * lock acquisition. We need to re-try lookup and
1706 /* Create new branch, starting from bottom */
1707 CDS_INIT_HLIST_HEAD(&head
);
1708 cds_hlist_add_head_rcu(&child_node
->list
, &head
);
1709 iter_node_flag
= (struct cds_ja_inode_flag
*) head
.next
;
1711 for (i
= ja
->tree_depth
- 1; i
>= (int) level
; i
--) {
1714 iter_key
= (uint8_t) (key
>> (JA_BITS_PER_BYTE
* (ja
->tree_depth
- i
- 1)));
1715 dbg_printf("branch creation level %d, key %u\n",
1716 i
, (unsigned int) iter_key
);
1717 iter_dest_node_flag
= NULL
;
1718 ret
= ja_node_set_nth(ja
, &iter_dest_node_flag
,
1723 dbg_printf("branch creation error %d\n", ret
);
1726 created_nodes
[nr_created_nodes
++] = iter_dest_node_flag
;
1727 iter_node_flag
= iter_dest_node_flag
;
1731 /* Publish branch */
1734 * Attaching to root node.
1736 rcu_assign_pointer(ja
->root
, iter_node_flag
);
1740 iter_key
= (uint8_t) (key
>> (JA_BITS_PER_BYTE
* (ja
->tree_depth
- level
)));
1741 dbg_printf("publish branch at level %d, key %u\n",
1742 level
- 1, (unsigned int) iter_key
);
1743 /* We need to use set_nth on the previous level. */
1744 iter_dest_node_flag
= attach_node_flag
;
1745 ret
= ja_node_set_nth(ja
, &iter_dest_node_flag
,
1748 shadow_node
, level
- 1);
1750 dbg_printf("branch publish error %d\n", ret
);
1756 rcu_assign_pointer(*attach_node_flag_ptr
, iter_dest_node_flag
);
1764 for (i
= 0; i
< nr_created_nodes
; i
++) {
1768 flags
= RCUJA_SHADOW_CLEAR_FREE_LOCK
;
1770 flags
|= RCUJA_SHADOW_CLEAR_FREE_NODE
;
1771 tmpret
= rcuja_shadow_clear(ja
->ht
,
1779 if (parent_shadow_node
)
1780 rcuja_shadow_unlock(parent_shadow_node
);
1783 rcuja_shadow_unlock(shadow_node
);
1789 * Lock the parent containing the hlist head pointer, and add node to list of
1790 * duplicates. Failure can happen if concurrent update changes the
1791 * parent before we get the lock. We return -EAGAIN in that case.
1792 * Return 0 on success, negative error value on failure.
1795 int ja_chain_node(struct cds_ja
*ja
,
1796 struct cds_ja_inode_flag
*parent_node_flag
,
1797 struct cds_ja_inode_flag
**node_flag_ptr
,
1798 struct cds_ja_inode_flag
*node_flag
,
1799 struct cds_ja_node
*node
)
1801 struct cds_ja_shadow_node
*shadow_node
;
1804 shadow_node
= rcuja_shadow_lookup_lock(ja
->ht
, parent_node_flag
);
1808 if (ja_node_ptr(*node_flag_ptr
) != ja_node_ptr(node_flag
)) {
1812 cds_hlist_add_head_rcu(&node
->list
, (struct cds_hlist_head
*) node_flag_ptr
);
1814 rcuja_shadow_unlock(shadow_node
);
1819 int _cds_ja_add(struct cds_ja
*ja
, uint64_t key
,
1820 struct cds_ja_node
*new_node
,
1821 struct cds_ja_node
**unique_node_ret
)
1823 unsigned int tree_depth
, i
;
1824 struct cds_ja_inode_flag
*attach_node_flag
,
1828 *parent_attach_node_flag
;
1829 struct cds_ja_inode_flag
**attach_node_flag_ptr
,
1830 **parent_node_flag_ptr
,
1834 if (caa_unlikely(key
> ja
->key_max
)) {
1837 tree_depth
= ja
->tree_depth
;
1840 dbg_printf("cds_ja_add attempt: key %" PRIu64
", node %p\n",
1842 parent2_node_flag
= NULL
;
1844 (struct cds_ja_inode_flag
*) &ja
->root
; /* Use root ptr address as key for mutex */
1845 parent_node_flag_ptr
= NULL
;
1846 node_flag
= rcu_dereference(ja
->root
);
1847 node_flag_ptr
= &ja
->root
;
1849 /* Iterate on all internal levels */
1850 for (i
= 1; i
< tree_depth
; i
++) {
1853 if (!ja_node_ptr(node_flag
))
1855 dbg_printf("cds_ja_add iter parent2_node_flag %p parent_node_flag %p node_flag_ptr %p node_flag %p\n",
1856 parent2_node_flag
, parent_node_flag
, node_flag_ptr
, node_flag
);
1857 iter_key
= (uint8_t) (key
>> (JA_BITS_PER_BYTE
* (tree_depth
- i
- 1)));
1858 parent2_node_flag
= parent_node_flag
;
1859 parent_node_flag
= node_flag
;
1860 parent_node_flag_ptr
= node_flag_ptr
;
1861 node_flag
= ja_node_get_nth(node_flag
,
1867 * We reached either bottom of tree or internal NULL node,
1868 * simply add node to last internal level, or chain it if key is
1871 if (!ja_node_ptr(node_flag
)) {
1872 dbg_printf("cds_ja_add NULL parent2_node_flag %p parent_node_flag %p node_flag_ptr %p node_flag %p\n",
1873 parent2_node_flag
, parent_node_flag
, node_flag_ptr
, node_flag
);
1875 attach_node_flag
= parent_node_flag
;
1876 attach_node_flag_ptr
= parent_node_flag_ptr
;
1877 parent_attach_node_flag
= parent2_node_flag
;
1879 ret
= ja_attach_node(ja
, attach_node_flag_ptr
,
1881 parent_attach_node_flag
,
1886 if (unique_node_ret
) {
1887 *unique_node_ret
= (struct cds_ja_node
*) ja_node_ptr(node_flag
);
1891 dbg_printf("cds_ja_add duplicate parent2_node_flag %p parent_node_flag %p node_flag_ptr %p node_flag %p\n",
1892 parent2_node_flag
, parent_node_flag
, node_flag_ptr
, node_flag
);
1894 attach_node_flag
= node_flag
;
1895 attach_node_flag_ptr
= node_flag_ptr
;
1896 parent_attach_node_flag
= parent_node_flag
;
1898 ret
= ja_chain_node(ja
,
1899 parent_attach_node_flag
,
1900 attach_node_flag_ptr
,
1904 if (ret
== -EAGAIN
|| ret
== -EEXIST
)
1910 int cds_ja_add(struct cds_ja
*ja
, uint64_t key
,
1911 struct cds_ja_node
*new_node
)
1913 return _cds_ja_add(ja
, key
, new_node
, NULL
);
1916 struct cds_ja_node
*cds_ja_add_unique(struct cds_ja
*ja
, uint64_t key
,
1917 struct cds_ja_node
*new_node
)
1920 struct cds_ja_node
*ret_node
;
1922 ret
= _cds_ja_add(ja
, key
, new_node
, &ret_node
);
1930 * Note: there is no need to lookup the pointer address associated with
1931 * each node's nth item after taking the lock: it's already been done by
1932 * cds_ja_del while holding the rcu read-side lock, and our node rules
1933 * ensure that when a match value -> pointer is found in a node, it is
1934 * _NEVER_ changed for that node without recompaction, and recompaction
1935 * reallocates the node.
1936 * However, when a child is removed from "linear" nodes, its pointer
1937 * is set to NULL. We therefore check, while holding the locks, if this
1938 * pointer is NULL, and return -ENOENT to the caller if it is the case.
1941 int ja_detach_node(struct cds_ja
*ja
,
1942 struct cds_ja_inode_flag
**snapshot
,
1943 struct cds_ja_inode_flag
***snapshot_ptr
,
1944 uint8_t *snapshot_n
,
1947 struct cds_ja_node
*node
)
1949 struct cds_ja_shadow_node
*shadow_nodes
[JA_MAX_DEPTH
];
1950 struct cds_ja_inode_flag
**node_flag_ptr
= NULL
,
1951 *parent_node_flag
= NULL
,
1952 **parent_node_flag_ptr
= NULL
;
1953 struct cds_ja_inode_flag
*iter_node_flag
;
1954 int ret
, i
, nr_shadow
= 0, nr_clear
= 0, nr_branch
= 0;
1957 assert(nr_snapshot
== ja
->tree_depth
+ 1);
1960 * From the last internal level node going up, get the node
1961 * lock, check if the node has only one child left. If it is the
1962 * case, we continue iterating upward. When we reach a node
1963 * which has more that one child left, we lock the parent, and
1964 * proceed to the node deletion (removing its children too).
1966 for (i
= nr_snapshot
- 2; i
>= 1; i
--) {
1967 struct cds_ja_shadow_node
*shadow_node
;
1969 shadow_node
= rcuja_shadow_lookup_lock(ja
->ht
,
1975 shadow_nodes
[nr_shadow
++] = shadow_node
;
1978 * Check if node has been removed between RCU
1979 * lookup and lock acquisition.
1981 assert(snapshot_ptr
[i
+ 1]);
1982 if (ja_node_ptr(*snapshot_ptr
[i
+ 1])
1983 != ja_node_ptr(snapshot
[i
+ 1])) {
1988 assert(shadow_node
->nr_child
> 0);
1989 if (shadow_node
->nr_child
== 1 && i
> 1)
1992 if (shadow_node
->nr_child
> 1 || i
== 1) {
1993 /* Lock parent and break */
1994 shadow_node
= rcuja_shadow_lookup_lock(ja
->ht
,
2000 shadow_nodes
[nr_shadow
++] = shadow_node
;
2003 * Check if node has been removed between RCU
2004 * lookup and lock acquisition.
2006 assert(snapshot_ptr
[i
]);
2007 if (ja_node_ptr(*snapshot_ptr
[i
])
2008 != ja_node_ptr(snapshot
[i
])) {
2013 node_flag_ptr
= snapshot_ptr
[i
+ 1];
2014 n
= snapshot_n
[i
+ 1];
2015 parent_node_flag_ptr
= snapshot_ptr
[i
];
2016 parent_node_flag
= snapshot
[i
];
2020 * Lock parent's parent, in case we need
2021 * to recompact parent.
2023 shadow_node
= rcuja_shadow_lookup_lock(ja
->ht
,
2029 shadow_nodes
[nr_shadow
++] = shadow_node
;
2032 * Check if node has been removed between RCU
2033 * lookup and lock acquisition.
2035 assert(snapshot_ptr
[i
- 1]);
2036 if (ja_node_ptr(*snapshot_ptr
[i
- 1])
2037 != ja_node_ptr(snapshot
[i
- 1])) {
2048 * At this point, we want to delete all nodes that are about to
2049 * be removed from shadow_nodes (except the last one, which is
2050 * either the root or the parent of the upmost node with 1
2051 * child). OK to free lock here, because RCU read lock is held,
2052 * and free only performed in call_rcu.
2055 for (i
= 0; i
< nr_clear
; i
++) {
2056 ret
= rcuja_shadow_clear(ja
->ht
,
2057 shadow_nodes
[i
]->node_flag
,
2059 RCUJA_SHADOW_CLEAR_FREE_NODE
2060 | RCUJA_SHADOW_CLEAR_FREE_LOCK
);
2064 iter_node_flag
= parent_node_flag
;
2065 /* Remove from parent */
2066 ret
= ja_node_clear_ptr(ja
,
2067 node_flag_ptr
, /* Pointer to location to nullify */
2068 &iter_node_flag
, /* Old new parent ptr in its parent */
2069 shadow_nodes
[nr_branch
- 1], /* of parent */
2074 dbg_printf("ja_detach_node: publish %p instead of %p\n",
2075 iter_node_flag
, *parent_node_flag_ptr
);
2076 /* Update address of parent ptr in its parent */
2077 rcu_assign_pointer(*parent_node_flag_ptr
, iter_node_flag
);
2080 for (i
= 0; i
< nr_shadow
; i
++)
2081 rcuja_shadow_unlock(shadow_nodes
[i
]);
2086 int ja_unchain_node(struct cds_ja
*ja
,
2087 struct cds_ja_inode_flag
*parent_node_flag
,
2088 struct cds_ja_inode_flag
**node_flag_ptr
,
2089 struct cds_ja_inode_flag
*node_flag
,
2090 struct cds_ja_node
*node
)
2092 struct cds_ja_shadow_node
*shadow_node
;
2093 struct cds_hlist_node
*hlist_node
;
2094 struct cds_hlist_head hlist_head
;
2095 int ret
= 0, count
= 0, found
= 0;
2097 shadow_node
= rcuja_shadow_lookup_lock(ja
->ht
, parent_node_flag
);
2100 if (ja_node_ptr(*node_flag_ptr
) != ja_node_ptr(node_flag
)) {
2104 hlist_head
.next
= (struct cds_hlist_node
*) ja_node_ptr(node_flag
);
2106 * Retry if another thread removed all but one of duplicates
2107 * since check (this check was performed without lock).
2108 * Ensure that the node we are about to remove is still in the
2109 * list (while holding lock).
2111 cds_hlist_for_each_rcu(hlist_node
, &hlist_head
) {
2113 /* FIXME: currently a work-around */
2114 hlist_node
->prev
= (struct cds_hlist_node
*) node_flag_ptr
;
2117 if (hlist_node
== &node
->list
)
2121 if (!found
|| count
== 1) {
2125 cds_hlist_del_rcu(&node
->list
);
2127 * Validate that we indeed removed the node from linked list.
2129 assert(ja_node_ptr(*node_flag_ptr
) != (struct cds_ja_inode
*) node
);
2131 rcuja_shadow_unlock(shadow_node
);
2136 * Called with RCU read lock held.
2138 int cds_ja_del(struct cds_ja
*ja
, uint64_t key
,
2139 struct cds_ja_node
*node
)
2141 unsigned int tree_depth
, i
;
2142 struct cds_ja_inode_flag
*snapshot
[JA_MAX_DEPTH
];
2143 struct cds_ja_inode_flag
**snapshot_ptr
[JA_MAX_DEPTH
];
2144 uint8_t snapshot_n
[JA_MAX_DEPTH
];
2145 struct cds_ja_inode_flag
*node_flag
;
2146 struct cds_ja_inode_flag
**prev_node_flag_ptr
,
2151 if (caa_unlikely(key
> ja
->key_max
))
2153 tree_depth
= ja
->tree_depth
;
2157 dbg_printf("cds_ja_del attempt: key %" PRIu64
", node %p\n",
2160 /* snapshot for level 0 is only for shadow node lookup */
2163 snapshot_ptr
[nr_snapshot
] = NULL
;
2164 snapshot
[nr_snapshot
++] = (struct cds_ja_inode_flag
*) &ja
->root
;
2165 node_flag
= rcu_dereference(ja
->root
);
2166 prev_node_flag_ptr
= &ja
->root
;
2167 node_flag_ptr
= &ja
->root
;
2169 /* Iterate on all internal levels */
2170 for (i
= 1; i
< tree_depth
; i
++) {
2173 dbg_printf("cds_ja_del iter node_flag %p\n",
2175 if (!ja_node_ptr(node_flag
)) {
2178 iter_key
= (uint8_t) (key
>> (JA_BITS_PER_BYTE
* (tree_depth
- i
- 1)));
2179 snapshot_n
[nr_snapshot
+ 1] = iter_key
;
2180 snapshot_ptr
[nr_snapshot
] = prev_node_flag_ptr
;
2181 snapshot
[nr_snapshot
++] = node_flag
;
2182 node_flag
= ja_node_get_nth(node_flag
,
2186 prev_node_flag_ptr
= node_flag_ptr
;
2187 dbg_printf("cds_ja_del iter key lookup %u finds node_flag %p, prev_node_flag_ptr %p\n",
2188 (unsigned int) iter_key
, node_flag
,
2189 prev_node_flag_ptr
);
2192 * We reached bottom of tree, try to find the node we are trying
2193 * to remove. Fail if we cannot find it.
2195 if (!ja_node_ptr(node_flag
)) {
2196 dbg_printf("cds_ja_del: no node found for key %" PRIu64
"\n",
2200 struct cds_hlist_head hlist_head
;
2201 struct cds_hlist_node
*hlist_node
;
2202 struct cds_ja_node
*entry
, *match
= NULL
;
2206 (struct cds_hlist_node
*) ja_node_ptr(node_flag
);
2207 cds_hlist_for_each_entry_rcu(entry
,
2211 dbg_printf("cds_ja_del: compare %p with entry %p\n", node
, entry
);
2217 dbg_printf("cds_ja_del: no node match for node %p key %" PRIu64
"\n", node
, key
);
2223 * Removing last of duplicates. Last snapshot
2224 * does not have a shadow node (external leafs).
2226 snapshot_ptr
[nr_snapshot
] = prev_node_flag_ptr
;
2227 snapshot
[nr_snapshot
++] = node_flag
;
2228 ret
= ja_detach_node(ja
, snapshot
, snapshot_ptr
,
2229 snapshot_n
, nr_snapshot
, key
, node
);
2231 ret
= ja_unchain_node(ja
, snapshot
[nr_snapshot
- 1],
2232 node_flag_ptr
, node_flag
, match
);
2236 * Explanation of -ENOENT handling: caused by concurrent delete
2237 * between RCU lookup and actual removal. Need to re-do the
2238 * lookup and removal attempt.
2240 if (ret
== -EAGAIN
|| ret
== -ENOENT
)
2245 struct cds_ja
*_cds_ja_new(unsigned int key_bits
,
2246 const struct rcu_flavor_struct
*flavor
)
2250 struct cds_ja_shadow_node
*root_shadow_node
;
2252 ja
= calloc(sizeof(*ja
), 1);
2264 ja
->key_max
= (1ULL << key_bits
) - 1;
2267 ja
->key_max
= UINT64_MAX
;
2273 /* ja->root is NULL */
2274 /* tree_depth 0 is for pointer to root node */
2275 ja
->tree_depth
= (key_bits
>> JA_LOG2_BITS_PER_BYTE
) + 1;
2276 assert(ja
->tree_depth
<= JA_MAX_DEPTH
);
2277 ja
->ht
= rcuja_create_ht(flavor
);
2282 * Note: we should not free this node until judy array destroy.
2284 root_shadow_node
= rcuja_shadow_set(ja
->ht
,
2285 (struct cds_ja_inode_flag
*) &ja
->root
,
2287 if (!root_shadow_node
) {
2295 ret
= rcuja_delete_ht(ja
->ht
);
2305 * Called from RCU read-side CS.
2307 __attribute__((visibility("protected")))
2308 void rcuja_free_all_children(struct cds_ja_shadow_node
*shadow_node
,
2309 struct cds_ja_inode_flag
*node_flag
,
2310 void (*free_node_cb
)(struct rcu_head
*head
))
2312 const struct rcu_flavor_struct
*flavor
;
2313 unsigned int type_index
;
2314 struct cds_ja_inode
*node
;
2315 const struct cds_ja_type
*type
;
2317 flavor
= cds_lfht_rcu_flavor(shadow_node
->ja
->ht
);
2318 node
= ja_node_ptr(node_flag
);
2319 assert(node
!= NULL
);
2320 type_index
= ja_node_type(node_flag
);
2321 type
= &ja_types
[type_index
];
2323 switch (type
->type_class
) {
2327 ja_linear_node_get_nr_child(type
, node
);
2330 for (i
= 0; i
< nr_child
; i
++) {
2331 struct cds_ja_inode_flag
*iter
;
2332 struct cds_hlist_head head
;
2333 struct cds_ja_node
*entry
;
2334 struct cds_hlist_node
*pos
, *tmp
;
2337 ja_linear_node_get_ith_pos(type
, node
, i
, &v
, &iter
);
2340 head
.next
= (struct cds_hlist_node
*) iter
;
2341 cds_hlist_for_each_entry_safe(entry
, pos
, tmp
, &head
, list
) {
2342 flavor
->update_call_rcu(&entry
->head
, free_node_cb
);
2349 unsigned int pool_nr
;
2351 for (pool_nr
= 0; pool_nr
< (1U << type
->nr_pool_order
); pool_nr
++) {
2352 struct cds_ja_inode
*pool
=
2353 ja_pool_node_get_ith_pool(type
, node
, pool_nr
);
2355 ja_linear_node_get_nr_child(type
, pool
);
2358 for (j
= 0; j
< nr_child
; j
++) {
2359 struct cds_ja_inode_flag
*iter
;
2360 struct cds_hlist_head head
;
2361 struct cds_ja_node
*entry
;
2362 struct cds_hlist_node
*pos
, *tmp
;
2365 ja_linear_node_get_ith_pos(type
, pool
, j
, &v
, &iter
);
2368 head
.next
= (struct cds_hlist_node
*) iter
;
2369 cds_hlist_for_each_entry_safe(entry
, pos
, tmp
, &head
, list
) {
2370 flavor
->update_call_rcu(&entry
->head
, free_node_cb
);
2382 for (i
= 0; i
< JA_ENTRY_PER_NODE
; i
++) {
2383 struct cds_ja_inode_flag
*iter
;
2384 struct cds_hlist_head head
;
2385 struct cds_ja_node
*entry
;
2386 struct cds_hlist_node
*pos
, *tmp
;
2388 iter
= ja_pigeon_node_get_ith_pos(type
, node
, i
);
2391 head
.next
= (struct cds_hlist_node
*) iter
;
2392 cds_hlist_for_each_entry_safe(entry
, pos
, tmp
, &head
, list
) {
2393 flavor
->update_call_rcu(&entry
->head
, free_node_cb
);
2404 void print_debug_fallback_distribution(struct cds_ja
*ja
)
2408 fprintf(stderr
, "Fallback node distribution:\n");
2409 for (i
= 0; i
< JA_ENTRY_PER_NODE
; i
++) {
2410 if (!ja
->node_fallback_count_distribution
[i
])
2412 fprintf(stderr
, " %3u: %4lu\n",
2413 i
, ja
->node_fallback_count_distribution
[i
]);
2418 * There should be no more concurrent add to the judy array while it is
2419 * being destroyed (ensured by the caller).
2421 int cds_ja_destroy(struct cds_ja
*ja
,
2422 void (*free_node_cb
)(struct rcu_head
*head
))
2424 const struct rcu_flavor_struct
*flavor
;
2427 flavor
= cds_lfht_rcu_flavor(ja
->ht
);
2428 rcuja_shadow_prune(ja
->ht
,
2429 RCUJA_SHADOW_CLEAR_FREE_NODE
| RCUJA_SHADOW_CLEAR_FREE_LOCK
,
2431 flavor
->thread_offline();
2432 ret
= rcuja_delete_ht(ja
->ht
);
2436 /* Wait for in-flight call_rcu free to complete. */
2439 flavor
->thread_online();
2440 if (uatomic_read(&ja
->nr_fallback
))
2442 "[warning] RCU Judy Array used %lu fallback node(s)\n",
2443 uatomic_read(&ja
->nr_fallback
));
2444 fprintf(stderr
, "Nodes allocated: %lu, Nodes freed: %lu. Fallback ratio: %g\n",
2445 uatomic_read(&ja
->nr_nodes_allocated
),
2446 uatomic_read(&ja
->nr_nodes_freed
),
2447 (double) uatomic_read(&ja
->nr_fallback
) / (double) uatomic_read(&ja
->nr_nodes_allocated
));
2448 print_debug_fallback_distribution(ja
);