Commit | Line | Data |
---|---|---|
61009379 MD |
1 | /* |
2 | * rcuja/rcuja.c | |
3 | * | |
4 | * Userspace RCU library - RCU Judy Array | |
5 | * | |
6 | * Copyright 2012 - Mathieu Desnoyers <mathieu.desnoyers@efficios.com> | |
7 | * | |
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. | |
12 | * | |
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. | |
17 | * | |
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 | |
21 | */ | |
22 | ||
195e72d3 | 23 | #define _LGPL_SOURCE |
e5227865 | 24 | #include <stdint.h> |
8e519e3c | 25 | #include <errno.h> |
d68c6810 | 26 | #include <limits.h> |
61009379 | 27 | #include <urcu/rcuja.h> |
d68c6810 MD |
28 | #include <urcu/compiler.h> |
29 | #include <urcu/arch.h> | |
30 | #include <assert.h> | |
8e519e3c | 31 | #include <urcu-pointer.h> |
f07b240f | 32 | #include <urcu/uatomic.h> |
b4540e8a | 33 | #include <stdint.h> |
8e519e3c | 34 | |
61009379 | 35 | #include "rcuja-internal.h" |
d68c6810 | 36 | #include "bitfield.h" |
61009379 | 37 | |
d96bfb0d | 38 | enum cds_ja_type_class { |
e5227865 | 39 | RCU_JA_LINEAR = 0, /* Type A */ |
fd800776 MD |
40 | /* 32-bit: 1 to 25 children, 8 to 128 bytes */ |
41 | /* 64-bit: 1 to 28 children, 16 to 256 bytes */ | |
42 | RCU_JA_POOL = 1, /* Type B */ | |
43 | /* 32-bit: 26 to 100 children, 256 to 512 bytes */ | |
44 | /* 64-bit: 29 to 112 children, 512 to 1024 bytes */ | |
e5227865 | 45 | RCU_JA_PIGEON = 2, /* Type C */ |
fd800776 MD |
46 | /* 32-bit: 101 to 256 children, 1024 bytes */ |
47 | /* 64-bit: 113 to 256 children, 2048 bytes */ | |
e5227865 | 48 | /* Leaf nodes are implicit from their height in the tree */ |
1db4943c | 49 | RCU_JA_NR_TYPES, |
e1db2db5 MD |
50 | |
51 | RCU_JA_NULL, /* not an encoded type, but keeps code regular */ | |
e5227865 MD |
52 | }; |
53 | ||
d96bfb0d MD |
54 | struct cds_ja_type { |
55 | enum cds_ja_type_class type_class; | |
8e519e3c MD |
56 | uint16_t min_child; /* minimum number of children: 1 to 256 */ |
57 | uint16_t max_child; /* maximum number of children: 1 to 256 */ | |
58 | uint16_t max_linear_child; /* per-pool max nr. children: 1 to 256 */ | |
59 | uint16_t order; /* node size is (1 << order), in bytes */ | |
fd800776 MD |
60 | uint16_t nr_pool_order; /* number of pools */ |
61 | uint16_t pool_size_order; /* pool size */ | |
e5227865 MD |
62 | }; |
63 | ||
d68c6810 MD |
64 | /* |
65 | * Number of least significant pointer bits reserved to represent the | |
66 | * child type. | |
67 | */ | |
68 | #define JA_TYPE_BITS 3 | |
a2a7ff59 | 69 | #define JA_TYPE_MAX_NR (1UL << JA_TYPE_BITS) |
d68c6810 MD |
70 | #define JA_TYPE_MASK (JA_TYPE_MAX_NR - 1) |
71 | #define JA_PTR_MASK (~JA_TYPE_MASK) | |
72 | ||
73 | #define JA_ENTRY_PER_NODE 256UL | |
582a6ade MD |
74 | #define JA_LOG2_BITS_PER_BYTE 3U |
75 | #define JA_BITS_PER_BYTE (1U << JA_LOG2_BITS_PER_BYTE) | |
d68c6810 | 76 | |
5b0f19e6 | 77 | #define JA_MAX_DEPTH 9 /* Maximum depth, including leafs */ |
5a9a87dd | 78 | |
e1db2db5 MD |
79 | /* |
80 | * Entry for NULL node is at index 8 of the table. It is never encoded | |
81 | * in flags. | |
82 | */ | |
83 | #define NODE_INDEX_NULL 8 | |
84 | ||
f07b240f MD |
85 | /* |
86 | * Number of removals needed on a fallback node before we try to shrink | |
87 | * it. | |
88 | */ | |
89 | #define JA_FALLBACK_REMOVAL_COUNT 8 | |
90 | ||
e5227865 MD |
91 | /* |
92 | * Iteration on the array to find the right node size for the number of | |
d68c6810 | 93 | * children stops when it reaches .max_child == 256 (this is the largest |
e5227865 | 94 | * possible node size, which contains 256 children). |
d68c6810 MD |
95 | * The min_child overlaps with the previous max_child to provide an |
96 | * hysteresis loop to reallocation for patterns of cyclic add/removal | |
97 | * within the same node. | |
98 | * The node the index within the following arrays is represented on 3 | |
99 | * bits. It identifies the node type, min/max number of children, and | |
100 | * the size order. | |
3d45251f MD |
101 | * The max_child values for the RCU_JA_POOL below result from |
102 | * statistical approximation: over million populations, the max_child | |
103 | * covers between 97% and 99% of the populations generated. Therefore, a | |
104 | * fallback should exist to cover the rare extreme population unbalance | |
105 | * cases, but it will not have a major impact on speed nor space | |
106 | * consumption, since those are rare cases. | |
e5227865 | 107 | */ |
e5227865 | 108 | |
d68c6810 MD |
109 | #if (CAA_BITS_PER_LONG < 64) |
110 | /* 32-bit pointers */ | |
1db4943c MD |
111 | enum { |
112 | ja_type_0_max_child = 1, | |
113 | ja_type_1_max_child = 3, | |
114 | ja_type_2_max_child = 6, | |
115 | ja_type_3_max_child = 12, | |
116 | ja_type_4_max_child = 25, | |
117 | ja_type_5_max_child = 48, | |
118 | ja_type_6_max_child = 92, | |
119 | ja_type_7_max_child = 256, | |
e1db2db5 | 120 | ja_type_8_max_child = 0, /* NULL */ |
1db4943c MD |
121 | }; |
122 | ||
8e519e3c MD |
123 | enum { |
124 | ja_type_0_max_linear_child = 1, | |
125 | ja_type_1_max_linear_child = 3, | |
126 | ja_type_2_max_linear_child = 6, | |
127 | ja_type_3_max_linear_child = 12, | |
128 | ja_type_4_max_linear_child = 25, | |
129 | ja_type_5_max_linear_child = 24, | |
130 | ja_type_6_max_linear_child = 23, | |
131 | }; | |
132 | ||
1db4943c MD |
133 | enum { |
134 | ja_type_5_nr_pool_order = 1, | |
135 | ja_type_6_nr_pool_order = 2, | |
136 | }; | |
137 | ||
d96bfb0d | 138 | const struct cds_ja_type ja_types[] = { |
8e519e3c MD |
139 | { .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, }, |
140 | { .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, }, | |
141 | { .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, }, | |
142 | { .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, }, | |
143 | { .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, }, | |
e5227865 | 144 | |
fd800776 | 145 | /* Pools may fill sooner than max_child */ |
8e519e3c MD |
146 | { .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, }, |
147 | { .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, }, | |
3d45251f MD |
148 | |
149 | /* | |
150 | * TODO: Upon node removal below min_child, if child pool is | |
151 | * filled beyond capacity, we need to roll back to pigeon. | |
152 | */ | |
1db4943c | 153 | { .type_class = RCU_JA_PIGEON, .min_child = 89, .max_child = ja_type_7_max_child, .order = 10, }, |
e1db2db5 MD |
154 | |
155 | { .type_class = RCU_JA_NULL, .min_child = 0, .max_child = ja_type_8_max_child, }, | |
d68c6810 | 156 | }; |
d68c6810 MD |
157 | #else /* !(CAA_BITS_PER_LONG < 64) */ |
158 | /* 64-bit pointers */ | |
1db4943c MD |
159 | enum { |
160 | ja_type_0_max_child = 1, | |
161 | ja_type_1_max_child = 3, | |
162 | ja_type_2_max_child = 7, | |
163 | ja_type_3_max_child = 14, | |
164 | ja_type_4_max_child = 28, | |
165 | ja_type_5_max_child = 54, | |
166 | ja_type_6_max_child = 104, | |
167 | ja_type_7_max_child = 256, | |
e1db2db5 | 168 | ja_type_8_max_child = 256, |
1db4943c MD |
169 | }; |
170 | ||
8e519e3c MD |
171 | enum { |
172 | ja_type_0_max_linear_child = 1, | |
173 | ja_type_1_max_linear_child = 3, | |
174 | ja_type_2_max_linear_child = 7, | |
175 | ja_type_3_max_linear_child = 14, | |
176 | ja_type_4_max_linear_child = 28, | |
177 | ja_type_5_max_linear_child = 27, | |
178 | ja_type_6_max_linear_child = 26, | |
179 | }; | |
180 | ||
1db4943c MD |
181 | enum { |
182 | ja_type_5_nr_pool_order = 1, | |
183 | ja_type_6_nr_pool_order = 2, | |
184 | }; | |
185 | ||
d96bfb0d | 186 | const struct cds_ja_type ja_types[] = { |
8e519e3c MD |
187 | { .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, }, |
188 | { .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, }, | |
189 | { .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, }, | |
190 | { .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, }, | |
191 | { .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, }, | |
e5227865 | 192 | |
3d45251f | 193 | /* Pools may fill sooner than max_child. */ |
8e519e3c MD |
194 | { .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, }, |
195 | { .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, }, | |
e5227865 | 196 | |
3d45251f MD |
197 | /* |
198 | * TODO: Upon node removal below min_child, if child pool is | |
199 | * filled beyond capacity, we need to roll back to pigeon. | |
200 | */ | |
1db4943c | 201 | { .type_class = RCU_JA_PIGEON, .min_child = 101, .max_child = ja_type_7_max_child, .order = 11, }, |
e1db2db5 MD |
202 | |
203 | { .type_class = RCU_JA_NULL, .min_child = 0, .max_child = ja_type_8_max_child, }, | |
e5227865 | 204 | }; |
d68c6810 | 205 | #endif /* !(BITS_PER_LONG < 64) */ |
e5227865 | 206 | |
1db4943c MD |
207 | static inline __attribute__((unused)) |
208 | void static_array_size_check(void) | |
209 | { | |
e1db2db5 | 210 | CAA_BUILD_BUG_ON(CAA_ARRAY_SIZE(ja_types) < JA_TYPE_MAX_NR); |
1db4943c MD |
211 | } |
212 | ||
e5227865 | 213 | /* |
d96bfb0d | 214 | * The cds_ja_node contains the compressed node data needed for |
1db4943c MD |
215 | * read-side. For linear and pool node configurations, it starts with a |
216 | * byte counting the number of children in the node. Then, the | |
217 | * node-specific data is placed. | |
218 | * The node mutex, if any is needed, protecting concurrent updated of | |
219 | * each node is placed in a separate hash table indexed by node address. | |
220 | * For the pigeon configuration, the number of children is also kept in | |
221 | * a separate hash table, indexed by node address, because it is only | |
222 | * required for updates. | |
e5227865 | 223 | */ |
1db4943c | 224 | |
ff38c745 MD |
225 | #define DECLARE_LINEAR_NODE(index) \ |
226 | struct { \ | |
227 | uint8_t nr_child; \ | |
228 | uint8_t child_value[ja_type_## index ##_max_linear_child]; \ | |
b4540e8a | 229 | struct cds_ja_inode_flag *child_ptr[ja_type_## index ##_max_linear_child]; \ |
ff38c745 MD |
230 | } |
231 | ||
232 | #define DECLARE_POOL_NODE(index) \ | |
233 | struct { \ | |
234 | struct { \ | |
235 | uint8_t nr_child; \ | |
236 | uint8_t child_value[ja_type_## index ##_max_linear_child]; \ | |
b4540e8a | 237 | struct cds_ja_inode_flag *child_ptr[ja_type_## index ##_max_linear_child]; \ |
ff38c745 MD |
238 | } linear[1U << ja_type_## index ##_nr_pool_order]; \ |
239 | } | |
1db4943c | 240 | |
b4540e8a | 241 | struct cds_ja_inode { |
1db4943c MD |
242 | union { |
243 | /* Linear configuration */ | |
244 | DECLARE_LINEAR_NODE(0) conf_0; | |
245 | DECLARE_LINEAR_NODE(1) conf_1; | |
246 | DECLARE_LINEAR_NODE(2) conf_2; | |
247 | DECLARE_LINEAR_NODE(3) conf_3; | |
248 | DECLARE_LINEAR_NODE(4) conf_4; | |
249 | ||
250 | /* Pool configuration */ | |
251 | DECLARE_POOL_NODE(5) conf_5; | |
252 | DECLARE_POOL_NODE(6) conf_6; | |
253 | ||
254 | /* Pigeon configuration */ | |
255 | struct { | |
b4540e8a | 256 | struct cds_ja_inode_flag *child[ja_type_7_max_child]; |
1db4943c MD |
257 | } conf_7; |
258 | /* data aliasing nodes for computed accesses */ | |
b4540e8a | 259 | uint8_t data[sizeof(struct cds_ja_inode_flag *) * ja_type_7_max_child]; |
1db4943c | 260 | } u; |
e5227865 MD |
261 | }; |
262 | ||
2e313670 MD |
263 | enum ja_recompact { |
264 | JA_RECOMPACT, | |
265 | JA_RECOMPACT_ADD, | |
266 | JA_RECOMPACT_DEL, | |
267 | }; | |
268 | ||
d68c6810 | 269 | static |
b4540e8a | 270 | struct cds_ja_inode_flag *ja_node_flag(struct cds_ja_inode *node, |
a2a7ff59 | 271 | unsigned long type) |
d68c6810 | 272 | { |
a2a7ff59 | 273 | assert(type < (1UL << JA_TYPE_BITS)); |
b4540e8a | 274 | return (struct cds_ja_inode_flag *) (((unsigned long) node) | type); |
d68c6810 MD |
275 | } |
276 | ||
e1db2db5 | 277 | static |
b4540e8a | 278 | struct cds_ja_inode *ja_node_ptr(struct cds_ja_inode_flag *node) |
e1db2db5 | 279 | { |
a2a7ff59 | 280 | return (struct cds_ja_inode *) (((unsigned long) node) & JA_PTR_MASK); |
e1db2db5 MD |
281 | } |
282 | ||
d68c6810 | 283 | static |
a2a7ff59 | 284 | unsigned long ja_node_type(struct cds_ja_inode_flag *node) |
d68c6810 | 285 | { |
a2a7ff59 | 286 | unsigned long type; |
d68c6810 | 287 | |
e1db2db5 MD |
288 | if (ja_node_ptr(node) == NULL) { |
289 | return NODE_INDEX_NULL; | |
290 | } | |
d68c6810 | 291 | type = (unsigned int) ((unsigned long) node & JA_TYPE_MASK); |
a2a7ff59 | 292 | assert(type < (1UL << JA_TYPE_BITS)); |
d68c6810 MD |
293 | return type; |
294 | } | |
295 | ||
b4540e8a | 296 | struct cds_ja_inode *alloc_cds_ja_node(const struct cds_ja_type *ja_type) |
e5227865 | 297 | { |
1db4943c | 298 | return calloc(1U << ja_type->order, sizeof(char)); |
e5227865 MD |
299 | } |
300 | ||
b4540e8a | 301 | void free_cds_ja_node(struct cds_ja_inode *node) |
e5227865 MD |
302 | { |
303 | free(node); | |
304 | } | |
305 | ||
d68c6810 MD |
306 | #define __JA_ALIGN_MASK(v, mask) (((v) + (mask)) & ~(mask)) |
307 | #define JA_ALIGN(v, align) __JA_ALIGN_MASK(v, (typeof(v)) (align) - 1) | |
308 | #define __JA_FLOOR_MASK(v, mask) ((v) & ~(mask)) | |
309 | #define JA_FLOOR(v, align) __JA_FLOOR_MASK(v, (typeof(v)) (align) - 1) | |
310 | ||
311 | static | |
1db4943c | 312 | uint8_t *align_ptr_size(uint8_t *ptr) |
d68c6810 | 313 | { |
1db4943c | 314 | return (uint8_t *) JA_ALIGN((unsigned long) ptr, sizeof(void *)); |
d68c6810 MD |
315 | } |
316 | ||
11c5e016 | 317 | static |
d96bfb0d | 318 | uint8_t ja_linear_node_get_nr_child(const struct cds_ja_type *type, |
b4540e8a | 319 | struct cds_ja_inode *node) |
11c5e016 MD |
320 | { |
321 | assert(type->type_class == RCU_JA_LINEAR || type->type_class == RCU_JA_POOL); | |
2e313670 | 322 | return rcu_dereference(node->u.data[0]); |
11c5e016 MD |
323 | } |
324 | ||
13a7f5a6 MD |
325 | /* |
326 | * The order in which values and pointers are does does not matter: if | |
327 | * a value is missing, we return NULL. If a value is there, but its | |
328 | * associated pointers is still NULL, we return NULL too. | |
329 | */ | |
d68c6810 | 330 | static |
b4540e8a MD |
331 | struct cds_ja_inode_flag *ja_linear_node_get_nth(const struct cds_ja_type *type, |
332 | struct cds_ja_inode *node, | |
5a9a87dd | 333 | struct cds_ja_inode_flag ***child_node_flag_ptr, |
8e519e3c | 334 | uint8_t n) |
d68c6810 MD |
335 | { |
336 | uint8_t nr_child; | |
337 | uint8_t *values; | |
b4540e8a MD |
338 | struct cds_ja_inode_flag **pointers; |
339 | struct cds_ja_inode_flag *ptr; | |
d68c6810 MD |
340 | unsigned int i; |
341 | ||
8e519e3c | 342 | assert(type->type_class == RCU_JA_LINEAR || type->type_class == RCU_JA_POOL); |
d68c6810 | 343 | |
11c5e016 | 344 | nr_child = ja_linear_node_get_nr_child(type, node); |
13a7f5a6 | 345 | cmm_smp_rmb(); /* read nr_child before values and pointers */ |
8e519e3c MD |
346 | assert(nr_child <= type->max_linear_child); |
347 | assert(type->type_class != RCU_JA_LINEAR || nr_child >= type->min_child); | |
d68c6810 | 348 | |
1db4943c | 349 | values = &node->u.data[1]; |
d68c6810 | 350 | for (i = 0; i < nr_child; i++) { |
13a7f5a6 | 351 | if (CMM_LOAD_SHARED(values[i]) == n) |
d68c6810 MD |
352 | break; |
353 | } | |
354 | if (i >= nr_child) | |
355 | return NULL; | |
b4540e8a | 356 | pointers = (struct cds_ja_inode_flag **) align_ptr_size(&values[type->max_linear_child]); |
13a7f5a6 | 357 | ptr = rcu_dereference(pointers[i]); |
2e313670 MD |
358 | if (caa_unlikely(child_node_flag_ptr) && ptr) |
359 | *child_node_flag_ptr = &pointers[i]; | |
d68c6810 MD |
360 | return ptr; |
361 | } | |
362 | ||
11c5e016 | 363 | static |
5a9a87dd | 364 | void ja_linear_node_get_ith_pos(const struct cds_ja_type *type, |
b4540e8a | 365 | struct cds_ja_inode *node, |
11c5e016 MD |
366 | uint8_t i, |
367 | uint8_t *v, | |
b4540e8a | 368 | struct cds_ja_inode_flag **iter) |
11c5e016 MD |
369 | { |
370 | uint8_t *values; | |
b4540e8a | 371 | struct cds_ja_inode_flag **pointers; |
11c5e016 MD |
372 | |
373 | assert(type->type_class == RCU_JA_LINEAR || type->type_class == RCU_JA_POOL); | |
374 | assert(i < ja_linear_node_get_nr_child(type, node)); | |
375 | ||
376 | values = &node->u.data[1]; | |
377 | *v = values[i]; | |
b4540e8a | 378 | pointers = (struct cds_ja_inode_flag **) align_ptr_size(&values[type->max_linear_child]); |
11c5e016 MD |
379 | *iter = pointers[i]; |
380 | } | |
381 | ||
d68c6810 | 382 | static |
b4540e8a MD |
383 | struct cds_ja_inode_flag *ja_pool_node_get_nth(const struct cds_ja_type *type, |
384 | struct cds_ja_inode *node, | |
5a9a87dd | 385 | struct cds_ja_inode_flag ***child_node_flag_ptr, |
8e519e3c | 386 | uint8_t n) |
d68c6810 | 387 | { |
b4540e8a | 388 | struct cds_ja_inode *linear; |
d68c6810 | 389 | |
fd800776 | 390 | assert(type->type_class == RCU_JA_POOL); |
e1db2db5 MD |
391 | /* |
392 | * TODO: currently, we select the pool by highest bits. We | |
393 | * should support various encodings. | |
394 | */ | |
b4540e8a | 395 | linear = (struct cds_ja_inode *) |
1db4943c | 396 | &node->u.data[((unsigned long) n >> (CHAR_BIT - type->nr_pool_order)) << type->pool_size_order]; |
5a9a87dd | 397 | return ja_linear_node_get_nth(type, linear, child_node_flag_ptr, n); |
d68c6810 MD |
398 | } |
399 | ||
11c5e016 | 400 | static |
b4540e8a MD |
401 | struct cds_ja_inode *ja_pool_node_get_ith_pool(const struct cds_ja_type *type, |
402 | struct cds_ja_inode *node, | |
11c5e016 MD |
403 | uint8_t i) |
404 | { | |
405 | assert(type->type_class == RCU_JA_POOL); | |
b4540e8a | 406 | return (struct cds_ja_inode *) |
11c5e016 MD |
407 | &node->u.data[(unsigned int) i << type->pool_size_order]; |
408 | } | |
409 | ||
d68c6810 | 410 | static |
b4540e8a MD |
411 | struct cds_ja_inode_flag *ja_pigeon_node_get_nth(const struct cds_ja_type *type, |
412 | struct cds_ja_inode *node, | |
5a9a87dd | 413 | struct cds_ja_inode_flag ***child_node_flag_ptr, |
8e519e3c | 414 | uint8_t n) |
d68c6810 | 415 | { |
5a9a87dd MD |
416 | struct cds_ja_inode_flag **child_node_flag; |
417 | ||
d68c6810 | 418 | assert(type->type_class == RCU_JA_PIGEON); |
5a9a87dd | 419 | child_node_flag = &((struct cds_ja_inode_flag **) node->u.data)[n]; |
582a6ade MD |
420 | dbg_printf("ja_pigeon_node_get_nth child_node_flag_ptr %p\n", |
421 | child_node_flag); | |
422 | if (caa_unlikely(child_node_flag_ptr) && *child_node_flag) | |
5a9a87dd MD |
423 | *child_node_flag_ptr = child_node_flag; |
424 | return rcu_dereference(*child_node_flag); | |
d68c6810 MD |
425 | } |
426 | ||
2e313670 MD |
427 | static |
428 | struct cds_ja_inode_flag *ja_pigeon_node_get_ith_pos(const struct cds_ja_type *type, | |
429 | struct cds_ja_inode *node, | |
430 | uint8_t i) | |
431 | { | |
432 | return ja_pigeon_node_get_nth(type, node, NULL, i); | |
433 | } | |
434 | ||
13a7f5a6 MD |
435 | /* |
436 | * ja_node_get_nth: get nth item from a node. | |
437 | * node_flag is already rcu_dereference'd. | |
438 | */ | |
d68c6810 | 439 | static |
41975c12 | 440 | struct cds_ja_inode_flag * ja_node_get_nth(struct cds_ja_inode_flag *node_flag, |
5a9a87dd | 441 | struct cds_ja_inode_flag ***child_node_flag_ptr, |
8e519e3c | 442 | uint8_t n) |
d68c6810 MD |
443 | { |
444 | unsigned int type_index; | |
b4540e8a | 445 | struct cds_ja_inode *node; |
d96bfb0d | 446 | const struct cds_ja_type *type; |
d68c6810 | 447 | |
d68c6810 | 448 | node = ja_node_ptr(node_flag); |
5a9a87dd | 449 | assert(node != NULL); |
d68c6810 MD |
450 | type_index = ja_node_type(node_flag); |
451 | type = &ja_types[type_index]; | |
452 | ||
453 | switch (type->type_class) { | |
454 | case RCU_JA_LINEAR: | |
5a9a87dd MD |
455 | return ja_linear_node_get_nth(type, node, |
456 | child_node_flag_ptr, n); | |
fd800776 | 457 | case RCU_JA_POOL: |
5a9a87dd MD |
458 | return ja_pool_node_get_nth(type, node, |
459 | child_node_flag_ptr, n); | |
d68c6810 | 460 | case RCU_JA_PIGEON: |
5a9a87dd MD |
461 | return ja_pigeon_node_get_nth(type, node, |
462 | child_node_flag_ptr, n); | |
d68c6810 MD |
463 | default: |
464 | assert(0); | |
465 | return (void *) -1UL; | |
466 | } | |
467 | } | |
468 | ||
8e519e3c | 469 | static |
d96bfb0d | 470 | int ja_linear_node_set_nth(const struct cds_ja_type *type, |
b4540e8a | 471 | struct cds_ja_inode *node, |
d96bfb0d | 472 | struct cds_ja_shadow_node *shadow_node, |
8e519e3c | 473 | uint8_t n, |
b4540e8a | 474 | struct cds_ja_inode_flag *child_node_flag) |
8e519e3c MD |
475 | { |
476 | uint8_t nr_child; | |
477 | uint8_t *values, *nr_child_ptr; | |
b4540e8a | 478 | struct cds_ja_inode_flag **pointers; |
2e313670 | 479 | unsigned int i, unused = 0; |
8e519e3c MD |
480 | |
481 | assert(type->type_class == RCU_JA_LINEAR || type->type_class == RCU_JA_POOL); | |
482 | ||
483 | nr_child_ptr = &node->u.data[0]; | |
a2a7ff59 | 484 | dbg_printf("linear set nth: nr_child_ptr %p\n", nr_child_ptr); |
8e519e3c MD |
485 | nr_child = *nr_child_ptr; |
486 | assert(nr_child <= type->max_linear_child); | |
8e519e3c MD |
487 | |
488 | values = &node->u.data[1]; | |
2e313670 MD |
489 | pointers = (struct cds_ja_inode_flag **) align_ptr_size(&values[type->max_linear_child]); |
490 | /* Check if node value is already populated */ | |
8e519e3c | 491 | for (i = 0; i < nr_child; i++) { |
2e313670 MD |
492 | if (values[i] == n) { |
493 | if (pointers[i]) | |
494 | return -EEXIST; | |
495 | else | |
496 | break; | |
497 | } else { | |
498 | if (!pointers[i]) | |
499 | unused++; | |
500 | } | |
8e519e3c | 501 | } |
2e313670 MD |
502 | if (i == nr_child && nr_child >= type->max_linear_child) { |
503 | if (unused) | |
504 | return -ERANGE; /* recompact node */ | |
505 | else | |
506 | return -ENOSPC; /* No space left in this node type */ | |
507 | } | |
508 | ||
509 | assert(pointers[i] == NULL); | |
510 | rcu_assign_pointer(pointers[i], child_node_flag); | |
511 | /* If we expanded the nr_child, increment it */ | |
512 | if (i == nr_child) { | |
513 | CMM_STORE_SHARED(values[nr_child], n); | |
514 | /* write pointer and value before nr_child */ | |
515 | cmm_smp_wmb(); | |
516 | CMM_STORE_SHARED(*nr_child_ptr, nr_child + 1); | |
8e519e3c | 517 | } |
e1db2db5 | 518 | shadow_node->nr_child++; |
a2a7ff59 MD |
519 | dbg_printf("linear set nth: %u child, shadow: %u child, for node %p shadow %p\n", |
520 | (unsigned int) CMM_LOAD_SHARED(*nr_child_ptr), | |
521 | (unsigned int) shadow_node->nr_child, | |
522 | node, shadow_node); | |
523 | ||
8e519e3c MD |
524 | return 0; |
525 | } | |
526 | ||
527 | static | |
d96bfb0d | 528 | int ja_pool_node_set_nth(const struct cds_ja_type *type, |
b4540e8a | 529 | struct cds_ja_inode *node, |
d96bfb0d | 530 | struct cds_ja_shadow_node *shadow_node, |
8e519e3c | 531 | uint8_t n, |
b4540e8a | 532 | struct cds_ja_inode_flag *child_node_flag) |
8e519e3c | 533 | { |
b4540e8a | 534 | struct cds_ja_inode *linear; |
8e519e3c MD |
535 | |
536 | assert(type->type_class == RCU_JA_POOL); | |
b4540e8a | 537 | linear = (struct cds_ja_inode *) |
8e519e3c | 538 | &node->u.data[((unsigned long) n >> (CHAR_BIT - type->nr_pool_order)) << type->pool_size_order]; |
e1db2db5 MD |
539 | return ja_linear_node_set_nth(type, linear, shadow_node, |
540 | n, child_node_flag); | |
8e519e3c MD |
541 | } |
542 | ||
543 | static | |
d96bfb0d | 544 | int ja_pigeon_node_set_nth(const struct cds_ja_type *type, |
b4540e8a | 545 | struct cds_ja_inode *node, |
d96bfb0d | 546 | struct cds_ja_shadow_node *shadow_node, |
8e519e3c | 547 | uint8_t n, |
b4540e8a | 548 | struct cds_ja_inode_flag *child_node_flag) |
8e519e3c | 549 | { |
b4540e8a | 550 | struct cds_ja_inode_flag **ptr; |
8e519e3c MD |
551 | |
552 | assert(type->type_class == RCU_JA_PIGEON); | |
b4540e8a | 553 | ptr = &((struct cds_ja_inode_flag **) node->u.data)[n]; |
5a9a87dd | 554 | if (*ptr) |
8e519e3c MD |
555 | return -EEXIST; |
556 | rcu_assign_pointer(*ptr, child_node_flag); | |
e1db2db5 | 557 | shadow_node->nr_child++; |
8e519e3c MD |
558 | return 0; |
559 | } | |
560 | ||
d68c6810 | 561 | /* |
7a0b2331 | 562 | * _ja_node_set_nth: set nth item within a node. Return an error |
8e519e3c | 563 | * (negative error value) if it is already there. |
d68c6810 | 564 | */ |
8e519e3c | 565 | static |
d96bfb0d | 566 | int _ja_node_set_nth(const struct cds_ja_type *type, |
b4540e8a | 567 | struct cds_ja_inode *node, |
d96bfb0d | 568 | struct cds_ja_shadow_node *shadow_node, |
e1db2db5 | 569 | uint8_t n, |
b4540e8a | 570 | struct cds_ja_inode_flag *child_node_flag) |
8e519e3c | 571 | { |
8e519e3c MD |
572 | switch (type->type_class) { |
573 | case RCU_JA_LINEAR: | |
e1db2db5 | 574 | return ja_linear_node_set_nth(type, node, shadow_node, n, |
8e519e3c MD |
575 | child_node_flag); |
576 | case RCU_JA_POOL: | |
e1db2db5 | 577 | return ja_pool_node_set_nth(type, node, shadow_node, n, |
8e519e3c MD |
578 | child_node_flag); |
579 | case RCU_JA_PIGEON: | |
e1db2db5 | 580 | return ja_pigeon_node_set_nth(type, node, shadow_node, n, |
8e519e3c | 581 | child_node_flag); |
e1db2db5 MD |
582 | case RCU_JA_NULL: |
583 | return -ENOSPC; | |
8e519e3c MD |
584 | default: |
585 | assert(0); | |
586 | return -EINVAL; | |
587 | } | |
588 | ||
589 | return 0; | |
590 | } | |
7a0b2331 | 591 | |
2e313670 | 592 | static |
af3cbd45 | 593 | int ja_linear_node_clear_ptr(const struct cds_ja_type *type, |
2e313670 MD |
594 | struct cds_ja_inode *node, |
595 | struct cds_ja_shadow_node *shadow_node, | |
af3cbd45 | 596 | struct cds_ja_inode_flag **node_flag_ptr) |
2e313670 MD |
597 | { |
598 | uint8_t nr_child; | |
af3cbd45 | 599 | uint8_t *nr_child_ptr; |
2e313670 MD |
600 | |
601 | assert(type->type_class == RCU_JA_LINEAR || type->type_class == RCU_JA_POOL); | |
602 | ||
603 | nr_child_ptr = &node->u.data[0]; | |
af3cbd45 | 604 | dbg_printf("linear clear ptr: nr_child_ptr %p\n", nr_child_ptr); |
2e313670 MD |
605 | nr_child = *nr_child_ptr; |
606 | assert(nr_child <= type->max_linear_child); | |
607 | ||
2e313670 MD |
608 | if (shadow_node->fallback_removal_count) { |
609 | shadow_node->fallback_removal_count--; | |
610 | } else { | |
611 | if (shadow_node->nr_child <= type->min_child) { | |
612 | /* We need to try recompacting the node */ | |
613 | return -EFBIG; | |
614 | } | |
615 | } | |
af3cbd45 MD |
616 | assert(*node_flag_ptr != NULL); |
617 | rcu_assign_pointer(*node_flag_ptr, NULL); | |
2e313670 MD |
618 | /* |
619 | * Value and nr_child are never changed (would cause ABA issue). | |
620 | * Instead, we leave the pointer to NULL and recompact the node | |
621 | * once in a while. It is allowed to set a NULL pointer to a new | |
622 | * value without recompaction though. | |
623 | * Only update the shadow node accounting. | |
624 | */ | |
625 | shadow_node->nr_child--; | |
af3cbd45 | 626 | dbg_printf("linear clear ptr: %u child, shadow: %u child, for node %p shadow %p\n", |
2e313670 MD |
627 | (unsigned int) CMM_LOAD_SHARED(*nr_child_ptr), |
628 | (unsigned int) shadow_node->nr_child, | |
629 | node, shadow_node); | |
630 | ||
631 | return 0; | |
632 | } | |
633 | ||
634 | static | |
af3cbd45 | 635 | int ja_pool_node_clear_ptr(const struct cds_ja_type *type, |
2e313670 MD |
636 | struct cds_ja_inode *node, |
637 | struct cds_ja_shadow_node *shadow_node, | |
af3cbd45 | 638 | struct cds_ja_inode_flag **node_flag_ptr, |
2e313670 MD |
639 | uint8_t n) |
640 | { | |
641 | struct cds_ja_inode *linear; | |
642 | ||
643 | assert(type->type_class == RCU_JA_POOL); | |
644 | linear = (struct cds_ja_inode *) | |
645 | &node->u.data[((unsigned long) n >> (CHAR_BIT - type->nr_pool_order)) << type->pool_size_order]; | |
af3cbd45 | 646 | return ja_linear_node_clear_ptr(type, linear, shadow_node, node_flag_ptr); |
2e313670 MD |
647 | } |
648 | ||
649 | static | |
af3cbd45 | 650 | int ja_pigeon_node_clear_ptr(const struct cds_ja_type *type, |
2e313670 MD |
651 | struct cds_ja_inode *node, |
652 | struct cds_ja_shadow_node *shadow_node, | |
af3cbd45 | 653 | struct cds_ja_inode_flag **node_flag_ptr) |
2e313670 | 654 | { |
2e313670 | 655 | assert(type->type_class == RCU_JA_PIGEON); |
af3cbd45 | 656 | rcu_assign_pointer(*node_flag_ptr, NULL); |
2e313670 MD |
657 | shadow_node->nr_child--; |
658 | return 0; | |
659 | } | |
660 | ||
661 | /* | |
af3cbd45 | 662 | * _ja_node_clear_ptr: clear ptr item within a node. Return an error |
2e313670 MD |
663 | * (negative error value) if it is not found (-ENOENT). |
664 | */ | |
665 | static | |
af3cbd45 | 666 | int _ja_node_clear_ptr(const struct cds_ja_type *type, |
2e313670 MD |
667 | struct cds_ja_inode *node, |
668 | struct cds_ja_shadow_node *shadow_node, | |
af3cbd45 | 669 | struct cds_ja_inode_flag **node_flag_ptr, |
2e313670 MD |
670 | uint8_t n) |
671 | { | |
672 | switch (type->type_class) { | |
673 | case RCU_JA_LINEAR: | |
af3cbd45 | 674 | return ja_linear_node_clear_ptr(type, node, shadow_node, node_flag_ptr); |
2e313670 | 675 | case RCU_JA_POOL: |
af3cbd45 | 676 | return ja_pool_node_clear_ptr(type, node, shadow_node, node_flag_ptr, n); |
2e313670 | 677 | case RCU_JA_PIGEON: |
af3cbd45 | 678 | return ja_pigeon_node_clear_ptr(type, node, shadow_node, node_flag_ptr); |
2e313670 MD |
679 | case RCU_JA_NULL: |
680 | return -ENOENT; | |
681 | default: | |
682 | assert(0); | |
683 | return -EINVAL; | |
684 | } | |
685 | ||
686 | return 0; | |
687 | } | |
688 | ||
7a0b2331 MD |
689 | /* |
690 | * ja_node_recompact_add: recompact a node, adding a new child. | |
e1db2db5 | 691 | * TODO: for pool type, take selection bit(s) into account. |
2e313670 | 692 | * Return 0 on success, -EAGAIN if need to retry, or other negative |
5a9a87dd | 693 | * error value otherwise. |
7a0b2331 MD |
694 | */ |
695 | static | |
2e313670 MD |
696 | int ja_node_recompact(enum ja_recompact mode, |
697 | struct cds_ja *ja, | |
e1db2db5 | 698 | unsigned int old_type_index, |
d96bfb0d | 699 | const struct cds_ja_type *old_type, |
b4540e8a | 700 | struct cds_ja_inode *old_node, |
5a9a87dd | 701 | struct cds_ja_shadow_node *shadow_node, |
b4540e8a | 702 | struct cds_ja_inode_flag **old_node_flag, uint8_t n, |
af3cbd45 MD |
703 | struct cds_ja_inode_flag *child_node_flag, |
704 | struct cds_ja_inode_flag **nullify_node_flag_ptr) | |
7a0b2331 | 705 | { |
e1db2db5 | 706 | unsigned int new_type_index; |
b4540e8a | 707 | struct cds_ja_inode *new_node; |
af3cbd45 | 708 | struct cds_ja_shadow_node *new_shadow_node = NULL; |
d96bfb0d | 709 | const struct cds_ja_type *new_type; |
b4540e8a | 710 | struct cds_ja_inode_flag *new_node_flag; |
7a0b2331 | 711 | int ret; |
f07b240f | 712 | int fallback = 0; |
7a0b2331 | 713 | |
2e313670 MD |
714 | switch (mode) { |
715 | case JA_RECOMPACT: | |
716 | new_type_index = old_type_index; | |
717 | break; | |
718 | case JA_RECOMPACT_ADD: | |
719 | if (!shadow_node || old_type_index == NODE_INDEX_NULL) { | |
720 | new_type_index = 0; | |
721 | } else { | |
722 | new_type_index = old_type_index + 1; | |
723 | } | |
724 | break; | |
725 | case JA_RECOMPACT_DEL: | |
726 | if (old_type_index == 0) { | |
727 | new_type_index = NODE_INDEX_NULL; | |
728 | } else { | |
729 | new_type_index = old_type_index - 1; | |
730 | } | |
731 | break; | |
732 | default: | |
733 | assert(0); | |
7a0b2331 | 734 | } |
a2a7ff59 | 735 | |
f07b240f | 736 | retry: /* for fallback */ |
582a6ade MD |
737 | dbg_printf("Recompact from type %d to type %d\n", |
738 | old_type_index, new_type_index); | |
7a0b2331 | 739 | new_type = &ja_types[new_type_index]; |
2e313670 MD |
740 | if (new_type_index != NODE_INDEX_NULL) { |
741 | new_node = alloc_cds_ja_node(new_type); | |
742 | if (!new_node) | |
743 | return -ENOMEM; | |
744 | new_node_flag = ja_node_flag(new_node, new_type_index); | |
745 | dbg_printf("Recompact inherit lock from %p\n", shadow_node); | |
746 | new_shadow_node = rcuja_shadow_set(ja->ht, new_node, shadow_node); | |
747 | if (!new_shadow_node) { | |
748 | free(new_node); | |
749 | return -ENOMEM; | |
750 | } | |
751 | if (fallback) | |
752 | new_shadow_node->fallback_removal_count = | |
753 | JA_FALLBACK_REMOVAL_COUNT; | |
754 | } else { | |
755 | new_node = NULL; | |
756 | new_node_flag = NULL; | |
e1db2db5 | 757 | } |
11c5e016 | 758 | |
2e313670 MD |
759 | assert(mode != JA_RECOMPACT_ADD || old_type->type_class != RCU_JA_PIGEON); |
760 | ||
761 | if (new_type_index == NODE_INDEX_NULL) | |
762 | goto skip_copy; | |
763 | ||
11c5e016 MD |
764 | switch (old_type->type_class) { |
765 | case RCU_JA_LINEAR: | |
766 | { | |
767 | uint8_t nr_child = | |
768 | ja_linear_node_get_nr_child(old_type, old_node); | |
769 | unsigned int i; | |
770 | ||
771 | for (i = 0; i < nr_child; i++) { | |
b4540e8a | 772 | struct cds_ja_inode_flag *iter; |
11c5e016 MD |
773 | uint8_t v; |
774 | ||
775 | ja_linear_node_get_ith_pos(old_type, old_node, i, &v, &iter); | |
776 | if (!iter) | |
777 | continue; | |
af3cbd45 | 778 | if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) |
2e313670 | 779 | continue; |
f07b240f MD |
780 | ret = _ja_node_set_nth(new_type, new_node, |
781 | new_shadow_node, | |
11c5e016 | 782 | v, iter); |
f07b240f MD |
783 | if (new_type->type_class == RCU_JA_POOL && ret) { |
784 | goto fallback_toosmall; | |
785 | } | |
11c5e016 MD |
786 | assert(!ret); |
787 | } | |
788 | break; | |
789 | } | |
790 | case RCU_JA_POOL: | |
791 | { | |
792 | unsigned int pool_nr; | |
793 | ||
794 | for (pool_nr = 0; pool_nr < (1U << old_type->nr_pool_order); pool_nr++) { | |
b4540e8a | 795 | struct cds_ja_inode *pool = |
11c5e016 MD |
796 | ja_pool_node_get_ith_pool(old_type, |
797 | old_node, pool_nr); | |
798 | uint8_t nr_child = | |
799 | ja_linear_node_get_nr_child(old_type, pool); | |
800 | unsigned int j; | |
801 | ||
802 | for (j = 0; j < nr_child; j++) { | |
b4540e8a | 803 | struct cds_ja_inode_flag *iter; |
11c5e016 MD |
804 | uint8_t v; |
805 | ||
806 | ja_linear_node_get_ith_pos(old_type, pool, | |
807 | j, &v, &iter); | |
808 | if (!iter) | |
809 | continue; | |
af3cbd45 | 810 | if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) |
2e313670 | 811 | continue; |
f07b240f MD |
812 | ret = _ja_node_set_nth(new_type, new_node, |
813 | new_shadow_node, | |
11c5e016 | 814 | v, iter); |
f07b240f MD |
815 | if (new_type->type_class == RCU_JA_POOL |
816 | && ret) { | |
817 | goto fallback_toosmall; | |
818 | } | |
11c5e016 MD |
819 | assert(!ret); |
820 | } | |
821 | } | |
822 | break; | |
7a0b2331 | 823 | } |
a2a7ff59 | 824 | case RCU_JA_NULL: |
2e313670 | 825 | assert(mode == JA_RECOMPACT_ADD); |
a2a7ff59 | 826 | break; |
11c5e016 | 827 | case RCU_JA_PIGEON: |
2e313670 MD |
828 | { |
829 | uint8_t nr_child; | |
830 | unsigned int i; | |
831 | ||
832 | assert(mode == JA_RECOMPACT_DEL); | |
833 | nr_child = shadow_node->nr_child; | |
834 | for (i = 0; i < nr_child; i++) { | |
835 | struct cds_ja_inode_flag *iter; | |
836 | ||
837 | iter = ja_pigeon_node_get_ith_pos(old_type, old_node, i); | |
838 | if (!iter) | |
839 | continue; | |
af3cbd45 | 840 | if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) |
2e313670 MD |
841 | continue; |
842 | ret = _ja_node_set_nth(new_type, new_node, | |
843 | new_shadow_node, | |
844 | i, iter); | |
845 | if (new_type->type_class == RCU_JA_POOL && ret) { | |
846 | goto fallback_toosmall; | |
847 | } | |
848 | assert(!ret); | |
849 | } | |
850 | break; | |
851 | } | |
11c5e016 MD |
852 | default: |
853 | assert(0); | |
5a9a87dd | 854 | ret = -EINVAL; |
f07b240f | 855 | goto end; |
11c5e016 | 856 | } |
2e313670 | 857 | skip_copy: |
11c5e016 | 858 | |
2e313670 MD |
859 | if (JA_RECOMPACT_ADD) { |
860 | /* add node */ | |
861 | ret = _ja_node_set_nth(new_type, new_node, | |
862 | new_shadow_node, | |
863 | n, child_node_flag); | |
864 | assert(!ret); | |
865 | } | |
866 | /* Return pointer to new recompacted node through old_node_flag */ | |
5a9a87dd | 867 | *old_node_flag = new_node_flag; |
a2a7ff59 | 868 | if (old_node) { |
2e313670 MD |
869 | int flags; |
870 | ||
871 | flags = RCUJA_SHADOW_CLEAR_FREE_NODE; | |
872 | /* | |
873 | * It is OK to free the lock associated with a node | |
874 | * going to NULL, since we are holding the parent lock. | |
875 | * This synchronizes removal with re-add of that node. | |
876 | */ | |
877 | if (new_type_index == NODE_INDEX_NULL) | |
878 | flags = RCUJA_SHADOW_CLEAR_FREE_LOCK; | |
a2a7ff59 | 879 | ret = rcuja_shadow_clear(ja->ht, old_node, shadow_node, |
2e313670 | 880 | flags); |
a2a7ff59 MD |
881 | assert(!ret); |
882 | } | |
5a9a87dd MD |
883 | |
884 | ret = 0; | |
f07b240f | 885 | end: |
5a9a87dd | 886 | return ret; |
f07b240f MD |
887 | |
888 | fallback_toosmall: | |
889 | /* fallback if next pool is too small */ | |
af3cbd45 | 890 | assert(new_shadow_node); |
f07b240f MD |
891 | ret = rcuja_shadow_clear(ja->ht, new_node, new_shadow_node, |
892 | RCUJA_SHADOW_CLEAR_FREE_NODE); | |
893 | assert(!ret); | |
894 | ||
2e313670 | 895 | /* Choose fallback type: pigeon */ |
f07b240f MD |
896 | new_type_index = (1UL << JA_TYPE_BITS) - 1; |
897 | dbg_printf("Fallback to type %d\n", new_type_index); | |
898 | uatomic_inc(&ja->nr_fallback); | |
899 | fallback = 1; | |
900 | goto retry; | |
7a0b2331 MD |
901 | } |
902 | ||
5a9a87dd | 903 | /* |
2e313670 | 904 | * Return 0 on success, -EAGAIN if need to retry, or other negative |
5a9a87dd MD |
905 | * error value otherwise. |
906 | */ | |
7a0b2331 | 907 | static |
d96bfb0d | 908 | int ja_node_set_nth(struct cds_ja *ja, |
b4540e8a | 909 | struct cds_ja_inode_flag **node_flag, uint8_t n, |
5a9a87dd MD |
910 | struct cds_ja_inode_flag *child_node_flag, |
911 | struct cds_ja_shadow_node *shadow_node) | |
7a0b2331 MD |
912 | { |
913 | int ret; | |
e1db2db5 | 914 | unsigned int type_index; |
d96bfb0d | 915 | const struct cds_ja_type *type; |
b4540e8a | 916 | struct cds_ja_inode *node; |
7a0b2331 | 917 | |
a2a7ff59 MD |
918 | dbg_printf("ja_node_set_nth for n=%u, node %p, shadow %p\n", |
919 | (unsigned int) n, ja_node_ptr(*node_flag), shadow_node); | |
920 | ||
e1db2db5 MD |
921 | node = ja_node_ptr(*node_flag); |
922 | type_index = ja_node_type(*node_flag); | |
923 | type = &ja_types[type_index]; | |
e1db2db5 MD |
924 | ret = _ja_node_set_nth(type, node, shadow_node, |
925 | n, child_node_flag); | |
2e313670 MD |
926 | switch (ret) { |
927 | case -ENOSPC: | |
e1db2db5 | 928 | /* Not enough space in node, need to recompact. */ |
2e313670 | 929 | ret = ja_node_recompact(JA_RECOMPACT_ADD, ja, type_index, type, node, |
af3cbd45 | 930 | shadow_node, node_flag, n, child_node_flag, NULL); |
2e313670 MD |
931 | break; |
932 | case -ERANGE: | |
933 | /* Node needs to be recompacted. */ | |
934 | ret = ja_node_recompact(JA_RECOMPACT, ja, type_index, type, node, | |
af3cbd45 | 935 | shadow_node, node_flag, n, child_node_flag, NULL); |
2e313670 MD |
936 | break; |
937 | } | |
938 | return ret; | |
939 | } | |
940 | ||
941 | /* | |
942 | * Return 0 on success, -EAGAIN if need to retry, or other negative | |
943 | * error value otherwise. | |
944 | */ | |
945 | static | |
af3cbd45 MD |
946 | int ja_node_clear_ptr(struct cds_ja *ja, |
947 | struct cds_ja_inode_flag **node_flag_ptr, /* Pointer to location to nullify */ | |
948 | struct cds_ja_inode_flag **parent_node_flag_ptr, /* Address of parent ptr in its parent */ | |
949 | struct cds_ja_shadow_node *shadow_node, /* of parent */ | |
950 | uint8_t n) | |
2e313670 MD |
951 | { |
952 | int ret; | |
953 | unsigned int type_index; | |
954 | const struct cds_ja_type *type; | |
955 | struct cds_ja_inode *node; | |
956 | ||
af3cbd45 MD |
957 | dbg_printf("ja_node_clear_ptr for node %p, shadow %p, target ptr %p\n", |
958 | ja_node_ptr(*parent_node_flag_ptr), shadow_node, node_flag_ptr); | |
2e313670 | 959 | |
af3cbd45 MD |
960 | node = ja_node_ptr(*parent_node_flag_ptr); |
961 | type_index = ja_node_type(*parent_node_flag_ptr); | |
2e313670 | 962 | type = &ja_types[type_index]; |
af3cbd45 | 963 | ret = _ja_node_clear_ptr(type, node, shadow_node, node_flag_ptr, n); |
2e313670 MD |
964 | if (ret == -EFBIG) { |
965 | /* Should to try recompaction. */ | |
966 | ret = ja_node_recompact(JA_RECOMPACT_DEL, ja, type_index, type, node, | |
af3cbd45 MD |
967 | shadow_node, parent_node_flag_ptr, n, NULL, |
968 | node_flag_ptr); | |
7a0b2331 MD |
969 | } |
970 | return ret; | |
971 | } | |
be9a7474 | 972 | |
af3cbd45 | 973 | struct cds_hlist_head cds_ja_lookup(struct cds_ja *ja, uint64_t key) |
b4540e8a | 974 | { |
41975c12 MD |
975 | unsigned int tree_depth, i; |
976 | struct cds_ja_inode_flag *node_flag; | |
af3cbd45 | 977 | struct cds_hlist_head head = { NULL }; |
41975c12 MD |
978 | |
979 | if (caa_unlikely(key > ja->key_max)) | |
af3cbd45 | 980 | return head; |
41975c12 | 981 | tree_depth = ja->tree_depth; |
5a9a87dd | 982 | node_flag = rcu_dereference(ja->root); |
41975c12 | 983 | |
5a9a87dd MD |
984 | /* level 0: root node */ |
985 | if (!ja_node_ptr(node_flag)) | |
af3cbd45 | 986 | return head; |
5a9a87dd MD |
987 | |
988 | for (i = 1; i < tree_depth; i++) { | |
79b41067 MD |
989 | uint8_t iter_key; |
990 | ||
991 | iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (tree_depth - i - 1))); | |
5a9a87dd | 992 | node_flag = ja_node_get_nth(node_flag, NULL, |
79b41067 | 993 | iter_key); |
582a6ade MD |
994 | dbg_printf("cds_ja_lookup iter key lookup %u finds node_flag %p\n", |
995 | (unsigned int) iter_key, node_flag); | |
41975c12 | 996 | if (!ja_node_ptr(node_flag)) |
af3cbd45 | 997 | return head; |
41975c12 MD |
998 | } |
999 | ||
5a9a87dd | 1000 | /* Last level lookup succeded. We got an actual match. */ |
af3cbd45 MD |
1001 | head.next = (struct cds_hlist_node *) node_flag; |
1002 | return head; | |
5a9a87dd MD |
1003 | } |
1004 | ||
1005 | /* | |
1006 | * We reached an unpopulated node. Create it and the children we need, | |
1007 | * and then attach the entire branch to the current node. This may | |
1008 | * trigger recompaction of the current node. Locks needed: node lock | |
1009 | * (for add), and, possibly, parent node lock (to update pointer due to | |
1010 | * node recompaction). | |
1011 | * | |
1012 | * First take node lock, check if recompaction is needed, then take | |
1013 | * parent lock (if needed). Then we can proceed to create the new | |
1014 | * branch. Publish the new branch, and release locks. | |
1015 | * TODO: we currently always take the parent lock even when not needed. | |
1016 | */ | |
1017 | static | |
1018 | int ja_attach_node(struct cds_ja *ja, | |
1019 | struct cds_ja_inode_flag **node_flag_ptr, | |
1020 | struct cds_ja_inode_flag *node_flag, | |
1021 | struct cds_ja_inode_flag *parent_node_flag, | |
1022 | uint64_t key, | |
79b41067 | 1023 | unsigned int level, |
5a9a87dd MD |
1024 | struct cds_ja_node *child_node) |
1025 | { | |
1026 | struct cds_ja_shadow_node *shadow_node = NULL, | |
af3cbd45 | 1027 | *parent_shadow_node = NULL; |
5a9a87dd MD |
1028 | struct cds_ja_inode *node = ja_node_ptr(node_flag); |
1029 | struct cds_ja_inode *parent_node = ja_node_ptr(parent_node_flag); | |
1030 | struct cds_hlist_head head; | |
1031 | struct cds_ja_inode_flag *iter_node_flag, *iter_dest_node_flag; | |
1032 | int ret, i; | |
a2a7ff59 | 1033 | struct cds_ja_inode_flag *created_nodes[JA_MAX_DEPTH]; |
5a9a87dd MD |
1034 | int nr_created_nodes = 0; |
1035 | ||
582a6ade MD |
1036 | dbg_printf("Attach node at level %u (node %p, node_flag %p)\n", |
1037 | level, node, node_flag); | |
a2a7ff59 | 1038 | |
5a9a87dd MD |
1039 | assert(node); |
1040 | shadow_node = rcuja_shadow_lookup_lock(ja->ht, node); | |
1041 | if (!shadow_node) { | |
2e313670 | 1042 | ret = -EAGAIN; |
5a9a87dd MD |
1043 | goto end; |
1044 | } | |
1045 | if (parent_node) { | |
1046 | parent_shadow_node = rcuja_shadow_lookup_lock(ja->ht, | |
1047 | parent_node); | |
1048 | if (!parent_shadow_node) { | |
2e313670 | 1049 | ret = -EAGAIN; |
5a9a87dd MD |
1050 | goto unlock_shadow; |
1051 | } | |
1052 | } | |
1053 | ||
a2a7ff59 | 1054 | /* Create new branch, starting from bottom */ |
5a9a87dd MD |
1055 | CDS_INIT_HLIST_HEAD(&head); |
1056 | cds_hlist_add_head_rcu(&child_node->list, &head); | |
a2a7ff59 | 1057 | iter_node_flag = (struct cds_ja_inode_flag *) head.next; |
5a9a87dd | 1058 | |
79b41067 MD |
1059 | for (i = ja->tree_depth; i > (int) level; i--) { |
1060 | uint8_t iter_key; | |
1061 | ||
1062 | iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (ja->tree_depth - i))); | |
1063 | dbg_printf("branch creation level %d, key %u\n", | |
1064 | i - 1, (unsigned int) iter_key); | |
5a9a87dd MD |
1065 | iter_dest_node_flag = NULL; |
1066 | ret = ja_node_set_nth(ja, &iter_dest_node_flag, | |
79b41067 | 1067 | iter_key, |
5a9a87dd MD |
1068 | iter_node_flag, |
1069 | NULL); | |
1070 | if (ret) | |
1071 | goto check_error; | |
1072 | created_nodes[nr_created_nodes++] = iter_dest_node_flag; | |
1073 | iter_node_flag = iter_dest_node_flag; | |
1074 | } | |
1075 | ||
79b41067 MD |
1076 | if (level > 1) { |
1077 | uint8_t iter_key; | |
1078 | ||
1079 | iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (ja->tree_depth - level))); | |
a2a7ff59 MD |
1080 | /* We need to use set_nth on the previous level. */ |
1081 | iter_dest_node_flag = node_flag; | |
1082 | ret = ja_node_set_nth(ja, &iter_dest_node_flag, | |
79b41067 | 1083 | iter_key, |
a2a7ff59 MD |
1084 | iter_node_flag, |
1085 | shadow_node); | |
1086 | if (ret) | |
1087 | goto check_error; | |
1088 | created_nodes[nr_created_nodes++] = iter_dest_node_flag; | |
1089 | iter_node_flag = iter_dest_node_flag; | |
1090 | } | |
1091 | ||
5a9a87dd | 1092 | /* Publish new branch */ |
a2a7ff59 MD |
1093 | dbg_printf("Publish branch %p, replacing %p\n", |
1094 | iter_node_flag, *node_flag_ptr); | |
5a9a87dd MD |
1095 | rcu_assign_pointer(*node_flag_ptr, iter_node_flag); |
1096 | ||
1097 | /* Success */ | |
1098 | ret = 0; | |
1099 | ||
1100 | check_error: | |
1101 | if (ret) { | |
1102 | for (i = 0; i < nr_created_nodes; i++) { | |
1103 | int tmpret; | |
a2a7ff59 MD |
1104 | int flags; |
1105 | ||
1106 | flags = RCUJA_SHADOW_CLEAR_FREE_LOCK; | |
1107 | if (i) | |
1108 | flags |= RCUJA_SHADOW_CLEAR_FREE_NODE; | |
5a9a87dd MD |
1109 | tmpret = rcuja_shadow_clear(ja->ht, |
1110 | ja_node_ptr(created_nodes[i]), | |
a2a7ff59 MD |
1111 | NULL, |
1112 | flags); | |
5a9a87dd MD |
1113 | assert(!tmpret); |
1114 | } | |
1115 | } | |
5a9a87dd MD |
1116 | if (parent_shadow_node) |
1117 | rcuja_shadow_unlock(parent_shadow_node); | |
1118 | unlock_shadow: | |
1119 | if (shadow_node) | |
1120 | rcuja_shadow_unlock(shadow_node); | |
1121 | end: | |
1122 | return ret; | |
1123 | } | |
1124 | ||
1125 | /* | |
af3cbd45 MD |
1126 | * Lock the parent containing the hlist head pointer, and add node to list of |
1127 | * duplicates. Failure can happen if concurrent update changes the | |
1128 | * parent before we get the lock. We return -EAGAIN in that case. | |
5a9a87dd MD |
1129 | * Return 0 on success, negative error value on failure. |
1130 | */ | |
1131 | static | |
1132 | int ja_chain_node(struct cds_ja *ja, | |
af3cbd45 | 1133 | struct cds_ja_inode_flag *parent_node_flag, |
5a9a87dd MD |
1134 | struct cds_hlist_head *head, |
1135 | struct cds_ja_node *node) | |
1136 | { | |
1137 | struct cds_ja_shadow_node *shadow_node; | |
1138 | ||
1139 | shadow_node = rcuja_shadow_lookup_lock(ja->ht, | |
af3cbd45 | 1140 | ja_node_ptr(parent_node_flag)); |
5a9a87dd | 1141 | if (!shadow_node) |
2e313670 | 1142 | return -EAGAIN; |
5a9a87dd MD |
1143 | cds_hlist_add_head_rcu(&node->list, head); |
1144 | rcuja_shadow_unlock(shadow_node); | |
1145 | return 0; | |
1146 | } | |
1147 | ||
1148 | int cds_ja_add(struct cds_ja *ja, uint64_t key, | |
1149 | struct cds_ja_node *new_node) | |
1150 | { | |
1151 | unsigned int tree_depth, i; | |
5a9a87dd MD |
1152 | struct cds_ja_inode_flag **node_flag_ptr; /* in parent */ |
1153 | struct cds_ja_inode_flag *node_flag, | |
1154 | *parent_node_flag, | |
1155 | *parent2_node_flag; | |
1156 | int ret; | |
1157 | ||
1158 | if (caa_unlikely(key > ja->key_max)) | |
1159 | return -EINVAL; | |
1160 | tree_depth = ja->tree_depth; | |
1161 | ||
1162 | retry: | |
a2a7ff59 MD |
1163 | dbg_printf("cds_ja_add attempt: key %" PRIu64 ", node %p\n", |
1164 | key, new_node); | |
5a9a87dd | 1165 | parent2_node_flag = NULL; |
b0f74e47 MD |
1166 | parent_node_flag = |
1167 | (struct cds_ja_inode_flag *) &ja->root; /* Use root ptr address as key for mutex */ | |
5a9a87dd | 1168 | node_flag_ptr = &ja->root; |
35170a44 | 1169 | node_flag = rcu_dereference(ja->root); |
5a9a87dd MD |
1170 | |
1171 | /* Iterate on all internal levels */ | |
a2a7ff59 | 1172 | for (i = 1; i < tree_depth; i++) { |
79b41067 MD |
1173 | uint8_t iter_key; |
1174 | ||
582a6ade MD |
1175 | dbg_printf("cds_ja_add iter node_flag_ptr %p node_flag %p\n", |
1176 | *node_flag_ptr, node_flag); | |
5a9a87dd MD |
1177 | if (!ja_node_ptr(node_flag)) { |
1178 | ret = ja_attach_node(ja, node_flag_ptr, | |
1179 | parent_node_flag, parent2_node_flag, | |
1180 | key, i, new_node); | |
2e313670 | 1181 | if (ret == -EAGAIN || ret == -EEXIST) |
5a9a87dd MD |
1182 | goto retry; |
1183 | else | |
1184 | goto end; | |
1185 | } | |
79b41067 | 1186 | iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (tree_depth - i - 1))); |
5a9a87dd MD |
1187 | parent2_node_flag = parent_node_flag; |
1188 | parent_node_flag = node_flag; | |
1189 | node_flag = ja_node_get_nth(node_flag, | |
1190 | &node_flag_ptr, | |
79b41067 | 1191 | iter_key); |
582a6ade MD |
1192 | dbg_printf("cds_ja_add iter key lookup %u finds node_flag %p node_flag_ptr %p\n", |
1193 | (unsigned int) iter_key, node_flag, *node_flag_ptr); | |
5a9a87dd MD |
1194 | } |
1195 | ||
1196 | /* | |
1197 | * We reached bottom of tree, simply add node to last internal | |
1198 | * level, or chain it if key is already present. | |
1199 | */ | |
1200 | if (!ja_node_ptr(node_flag)) { | |
582a6ade MD |
1201 | dbg_printf("cds_ja_add last node_flag_ptr %p node_flag %p\n", |
1202 | *node_flag_ptr, node_flag); | |
5a9a87dd MD |
1203 | ret = ja_attach_node(ja, node_flag_ptr, parent_node_flag, |
1204 | parent2_node_flag, key, i, new_node); | |
1205 | } else { | |
1206 | ret = ja_chain_node(ja, | |
af3cbd45 MD |
1207 | parent_node_flag, |
1208 | (struct cds_hlist_head *) node_flag_ptr, | |
5a9a87dd MD |
1209 | new_node); |
1210 | } | |
2e313670 | 1211 | if (ret == -EAGAIN) |
5a9a87dd MD |
1212 | goto retry; |
1213 | end: | |
1214 | return ret; | |
b4540e8a MD |
1215 | } |
1216 | ||
af3cbd45 MD |
1217 | /* |
1218 | * Note: there is no need to lookup the pointer address associated with | |
1219 | * each node's nth item after taking the lock: it's already been done by | |
1220 | * cds_ja_del while holding the rcu read-side lock, and our node rules | |
1221 | * ensure that when a match value -> pointer is found in a node, it is | |
1222 | * _NEVER_ changed for that node without recompaction, and recompaction | |
1223 | * reallocates the node. | |
1224 | */ | |
35170a44 MD |
1225 | static |
1226 | int ja_detach_node(struct cds_ja *ja, | |
1227 | struct cds_ja_inode_flag **snapshot, | |
af3cbd45 MD |
1228 | struct cds_ja_inode_flag ***snapshot_ptr, |
1229 | uint8_t *snapshot_n, | |
35170a44 MD |
1230 | int nr_snapshot, |
1231 | uint64_t key, | |
1232 | struct cds_ja_node *node) | |
1233 | { | |
af3cbd45 MD |
1234 | struct cds_ja_shadow_node *shadow_nodes[JA_MAX_DEPTH]; |
1235 | struct cds_ja_inode_flag **node_flag_ptr = NULL, | |
1236 | *parent_node_flag = NULL, | |
1237 | **parent_node_flag_ptr = NULL; | |
1238 | struct cds_ja_inode_flag *iter_node_flag; | |
1239 | int ret, i, nr_shadow = 0, nr_clear = 0; | |
1240 | uint8_t n; | |
35170a44 MD |
1241 | |
1242 | assert(nr_snapshot == ja->tree_depth - 1); | |
1243 | ||
af3cbd45 MD |
1244 | /* |
1245 | * From the last internal level node going up, get the node | |
1246 | * lock, check if the node has only one child left. If it is the | |
1247 | * case, we continue iterating upward. When we reach a node | |
1248 | * which has more that one child left, we lock the parent, and | |
1249 | * proceed to the node deletion (removing its children too). | |
1250 | */ | |
1251 | for (i = nr_snapshot - 1; i >= 1; i--) { | |
1252 | struct cds_ja_shadow_node *shadow_node; | |
1253 | ||
1254 | shadow_node = rcuja_shadow_lookup_lock(ja->ht, | |
1255 | ja_node_ptr(snapshot[i])); | |
1256 | if (!shadow_node) { | |
1257 | ret = -EAGAIN; | |
1258 | goto end; | |
1259 | } | |
1260 | assert(shadow_node->nr_child > 0); | |
1261 | shadow_nodes[nr_shadow++] = shadow_node; | |
1262 | nr_clear++; | |
1263 | if (i == nr_snapshot - 1) { | |
1264 | /* | |
1265 | * Re-check that last internal node level has | |
1266 | * only one child, else trigger a retry. | |
1267 | */ | |
1268 | if (shadow_node->nr_child != 1) { | |
1269 | ret = -EAGAIN; | |
1270 | goto end; | |
1271 | } | |
1272 | } | |
1273 | if (shadow_node->nr_child > 1 || i == 1) { | |
1274 | /* Lock parent and break */ | |
1275 | shadow_node = rcuja_shadow_lookup_lock(ja->ht, | |
1276 | ja_node_ptr(snapshot[i - 1])); | |
1277 | if (!shadow_node) { | |
1278 | ret = -EAGAIN; | |
1279 | goto end; | |
1280 | } | |
1281 | shadow_nodes[nr_shadow++] = shadow_node; | |
1282 | node_flag_ptr = snapshot_ptr[i]; | |
1283 | n = snapshot_n[i]; | |
1284 | parent_node_flag_ptr = snapshot_ptr[i - 1]; | |
1285 | parent_node_flag = snapshot[i - 1]; | |
1286 | if (i > 1) { | |
1287 | /* | |
1288 | * Lock parent's parent, in case we need | |
1289 | * to recompact parent. | |
1290 | */ | |
1291 | shadow_node = rcuja_shadow_lookup_lock(ja->ht, | |
1292 | ja_node_ptr(snapshot[i - 2])); | |
1293 | if (!shadow_node) { | |
1294 | ret = -EAGAIN; | |
1295 | goto end; | |
1296 | } | |
1297 | shadow_nodes[nr_shadow++] = shadow_node; | |
1298 | } | |
1299 | break; | |
1300 | } | |
1301 | } | |
1302 | ||
1303 | /* | |
1304 | * At this point, we want to delete all nodes in shadow_nodes | |
1305 | * (except the last one, which is either the root or the parent | |
1306 | * of the upmost node with 1 child). OK to as to free lock here, | |
1307 | * because RCU read lock is held, and free only performed in | |
1308 | * call_rcu. | |
1309 | */ | |
1310 | ||
1311 | for (i = 0; i < nr_clear; i++) { | |
1312 | ret = rcuja_shadow_clear(ja->ht, | |
1313 | shadow_nodes[i]->node, | |
1314 | shadow_nodes[i], | |
1315 | RCUJA_SHADOW_CLEAR_FREE_NODE | |
1316 | | RCUJA_SHADOW_CLEAR_FREE_LOCK); | |
1317 | assert(!ret); | |
1318 | } | |
1319 | ||
1320 | iter_node_flag = parent_node_flag; | |
1321 | /* Remove from parent */ | |
1322 | ret = ja_node_clear_ptr(ja, | |
1323 | node_flag_ptr, /* Pointer to location to nullify */ | |
1324 | &iter_node_flag, /* Old new parent ptr in its parent */ | |
1325 | shadow_nodes[nr_clear], /* of parent */ | |
1326 | n); | |
1327 | ||
1328 | /* Update address of parent ptr in its parent */ | |
1329 | rcu_assign_pointer(*parent_node_flag_ptr, iter_node_flag); | |
1330 | ||
1331 | end: | |
1332 | for (i = 0; i < nr_shadow; i++) | |
1333 | rcuja_shadow_unlock(shadow_nodes[i]); | |
35170a44 MD |
1334 | return ret; |
1335 | } | |
1336 | ||
af3cbd45 MD |
1337 | static |
1338 | int ja_unchain_node(struct cds_ja *ja, | |
1339 | struct cds_ja_inode_flag *parent_node_flag, | |
1340 | struct cds_ja_node *node) | |
1341 | { | |
1342 | struct cds_ja_shadow_node *shadow_node; | |
1343 | int ret = 0; | |
1344 | ||
1345 | shadow_node = rcuja_shadow_lookup_lock(ja->ht, | |
1346 | ja_node_ptr(parent_node_flag)); | |
1347 | if (!shadow_node) | |
1348 | return -EAGAIN; | |
1349 | /* | |
1350 | * Retry if another thread removed all but one of duplicates | |
1351 | * since check. | |
1352 | */ | |
1353 | if (shadow_node->nr_child == 1) { | |
1354 | ret = -EAGAIN; | |
1355 | goto end; | |
1356 | } | |
1357 | cds_hlist_del_rcu(&node->list); | |
1358 | end: | |
1359 | rcuja_shadow_unlock(shadow_node); | |
1360 | return ret; | |
1361 | } | |
1362 | ||
1363 | /* | |
1364 | * Called with RCU read lock held. | |
1365 | */ | |
35170a44 MD |
1366 | int cds_ja_del(struct cds_ja *ja, uint64_t key, |
1367 | struct cds_ja_node *node) | |
1368 | { | |
1369 | unsigned int tree_depth, i; | |
1370 | struct cds_ja_inode_flag *snapshot[JA_MAX_DEPTH]; | |
af3cbd45 MD |
1371 | struct cds_ja_inode_flag **snapshot_ptr[JA_MAX_DEPTH]; |
1372 | uint8_t snapshot_n[JA_MAX_DEPTH]; | |
35170a44 | 1373 | struct cds_ja_inode_flag *node_flag; |
af3cbd45 | 1374 | struct cds_ja_inode_flag **prev_node_flag_ptr; |
35170a44 MD |
1375 | int nr_snapshot = 0; |
1376 | int ret; | |
1377 | ||
1378 | if (caa_unlikely(key > ja->key_max)) | |
1379 | return -EINVAL; | |
1380 | tree_depth = ja->tree_depth; | |
1381 | ||
1382 | retry: | |
1383 | dbg_printf("cds_ja_del attempt: key %" PRIu64 ", node %p\n", | |
1384 | key, node); | |
1385 | ||
1386 | /* snapshot for level 0 is only for shadow node lookup */ | |
af3cbd45 MD |
1387 | snapshot_n[nr_snapshot] = 0; |
1388 | snapshot_ptr[nr_snapshot] = NULL; | |
35170a44 MD |
1389 | snapshot[nr_snapshot++] = (struct cds_ja_inode_flag *) &ja->root; |
1390 | node_flag = rcu_dereference(ja->root); | |
af3cbd45 | 1391 | prev_node_flag_ptr = &ja->root; |
35170a44 MD |
1392 | |
1393 | /* Iterate on all internal levels */ | |
1394 | for (i = 1; i < tree_depth; i++) { | |
1395 | uint8_t iter_key; | |
1396 | ||
1397 | dbg_printf("cds_ja_del iter node_flag %p\n", | |
1398 | node_flag); | |
1399 | if (!ja_node_ptr(node_flag)) { | |
1400 | return -ENOENT; | |
1401 | } | |
35170a44 | 1402 | iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (tree_depth - i - 1))); |
af3cbd45 MD |
1403 | if (nr_snapshot <= 1) |
1404 | snapshot_n[nr_snapshot] = 0; | |
1405 | else | |
1406 | snapshot_n[nr_snapshot - 1] = iter_key; | |
1407 | ||
1408 | snapshot_ptr[nr_snapshot] = prev_node_flag_ptr; | |
1409 | snapshot[nr_snapshot++] = node_flag; | |
35170a44 | 1410 | node_flag = ja_node_get_nth(node_flag, |
af3cbd45 | 1411 | &prev_node_flag_ptr, |
35170a44 | 1412 | iter_key); |
af3cbd45 MD |
1413 | dbg_printf("cds_ja_del iter key lookup %u finds node_flag %p, prev_node_flag_ptr %p\n", |
1414 | (unsigned int) iter_key, node_flag, | |
1415 | prev_node_flag_ptr); | |
35170a44 MD |
1416 | } |
1417 | ||
1418 | /* | |
1419 | * We reached bottom of tree, try to find the node we are trying | |
1420 | * to remove. Fail if we cannot find it. | |
1421 | */ | |
1422 | if (!ja_node_ptr(node_flag)) { | |
1423 | return -ENOENT; | |
1424 | } else { | |
1425 | struct cds_hlist_head *hlist_head; | |
1426 | struct cds_hlist_node *hlist_node; | |
af3cbd45 MD |
1427 | struct cds_ja_node *entry, *match = NULL; |
1428 | int count = 0; | |
35170a44 MD |
1429 | |
1430 | hlist_head = (struct cds_hlist_head *) ja_node_ptr(node_flag); | |
af3cbd45 | 1431 | cds_hlist_for_each_entry_rcu(entry, |
35170a44 MD |
1432 | hlist_node, |
1433 | hlist_head, | |
1434 | list) { | |
af3cbd45 MD |
1435 | if (entry == node) |
1436 | match = entry; | |
1437 | count++; | |
35170a44 | 1438 | } |
af3cbd45 | 1439 | if (!match) |
35170a44 | 1440 | return -ENOENT; |
af3cbd45 MD |
1441 | assert(count > 0); |
1442 | if (count == 1) { | |
1443 | /* | |
1444 | * Removing last of duplicates. | |
1445 | */ | |
1446 | snapshot_ptr[nr_snapshot] = prev_node_flag_ptr; | |
1447 | snapshot[nr_snapshot++] = node_flag; | |
1448 | ret = ja_detach_node(ja, snapshot, snapshot_ptr, | |
1449 | snapshot_n, nr_snapshot, key, node); | |
1450 | } else { | |
1451 | ret = ja_unchain_node(ja, node_flag, entry); | |
1452 | } | |
35170a44 MD |
1453 | } |
1454 | if (ret == -EAGAIN) | |
1455 | goto retry; | |
1456 | return ret; | |
1457 | } | |
1458 | ||
b4540e8a MD |
1459 | struct cds_ja *_cds_ja_new(unsigned int key_bits, |
1460 | const struct rcu_flavor_struct *flavor) | |
be9a7474 MD |
1461 | { |
1462 | struct cds_ja *ja; | |
b0f74e47 | 1463 | int ret; |
f07b240f | 1464 | struct cds_ja_shadow_node *root_shadow_node; |
be9a7474 MD |
1465 | |
1466 | ja = calloc(sizeof(*ja), 1); | |
1467 | if (!ja) | |
1468 | goto ja_error; | |
b4540e8a MD |
1469 | |
1470 | switch (key_bits) { | |
1471 | case 8: | |
1472 | ja->key_max = UINT8_MAX; | |
1473 | break; | |
1474 | case 16: | |
1475 | ja->key_max = UINT16_MAX; | |
1476 | break; | |
1477 | case 32: | |
1478 | ja->key_max = UINT32_MAX; | |
1479 | break; | |
1480 | case 64: | |
1481 | ja->key_max = UINT64_MAX; | |
1482 | break; | |
1483 | default: | |
1484 | goto check_error; | |
1485 | } | |
1486 | ||
be9a7474 | 1487 | /* ja->root is NULL */ |
5a9a87dd | 1488 | /* tree_depth 0 is for pointer to root node */ |
582a6ade | 1489 | ja->tree_depth = (key_bits >> JA_LOG2_BITS_PER_BYTE) + 1; |
a2a7ff59 | 1490 | assert(ja->tree_depth <= JA_MAX_DEPTH); |
be9a7474 MD |
1491 | ja->ht = rcuja_create_ht(flavor); |
1492 | if (!ja->ht) | |
1493 | goto ht_error; | |
b0f74e47 MD |
1494 | |
1495 | /* | |
1496 | * Note: we should not free this node until judy array destroy. | |
1497 | */ | |
f07b240f | 1498 | root_shadow_node = rcuja_shadow_set(ja->ht, |
b0f74e47 MD |
1499 | ja_node_ptr((struct cds_ja_inode_flag *) &ja->root), |
1500 | NULL); | |
f07b240f MD |
1501 | if (!root_shadow_node) { |
1502 | ret = -ENOMEM; | |
b0f74e47 | 1503 | goto ht_node_error; |
f07b240f | 1504 | } |
582a6ade | 1505 | root_shadow_node->is_root = 1; |
b0f74e47 | 1506 | |
be9a7474 MD |
1507 | return ja; |
1508 | ||
b0f74e47 MD |
1509 | ht_node_error: |
1510 | ret = rcuja_delete_ht(ja->ht); | |
1511 | assert(!ret); | |
be9a7474 | 1512 | ht_error: |
b4540e8a | 1513 | check_error: |
be9a7474 MD |
1514 | free(ja); |
1515 | ja_error: | |
1516 | return NULL; | |
1517 | } | |
1518 | ||
1519 | /* | |
1520 | * There should be no more concurrent add to the judy array while it is | |
1521 | * being destroyed (ensured by the caller). | |
1522 | */ | |
1523 | int cds_ja_destroy(struct cds_ja *ja) | |
1524 | { | |
b4540e8a MD |
1525 | int ret; |
1526 | ||
be9a7474 MD |
1527 | rcuja_shadow_prune(ja->ht, |
1528 | RCUJA_SHADOW_CLEAR_FREE_NODE | RCUJA_SHADOW_CLEAR_FREE_LOCK); | |
b4540e8a MD |
1529 | ret = rcuja_delete_ht(ja->ht); |
1530 | if (ret) | |
1531 | return ret; | |
f07b240f MD |
1532 | if (uatomic_read(&ja->nr_fallback)) |
1533 | fprintf(stderr, | |
1534 | "[warning] RCU Judy Array used %lu fallback node(s)\n", | |
1535 | uatomic_read(&ja->nr_fallback)); | |
b4540e8a | 1536 | free(ja); |
41975c12 | 1537 | return 0; |
be9a7474 | 1538 | } |