Commit | Line | Data |
---|---|---|
61009379 MD |
1 | /* |
2 | * rcuja/rcuja.c | |
3 | * | |
4 | * Userspace RCU library - RCU Judy Array | |
5 | * | |
78ed159a | 6 | * Copyright (C) 2000 - 2002 Hewlett-Packard Company |
170e1186 | 7 | * Copyright 2012-2013 - Mathieu Desnoyers <mathieu.desnoyers@efficios.com> |
61009379 MD |
8 | * |
9 | * This library is free software; you can redistribute it and/or | |
10 | * modify it under the terms of the GNU Lesser General Public | |
11 | * License as published by the Free Software Foundation; either | |
12 | * version 2.1 of the License, or (at your option) any later version. | |
13 | * | |
14 | * This library is distributed in the hope that it will be useful, | |
15 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
16 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | |
17 | * Lesser General Public License for more details. | |
18 | * | |
19 | * You should have received a copy of the GNU Lesser General Public | |
20 | * License along with this library; if not, write to the Free Software | |
21 | * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA | |
22 | */ | |
23 | ||
195e72d3 | 24 | #define _LGPL_SOURCE |
e5227865 | 25 | #include <stdint.h> |
8e519e3c | 26 | #include <errno.h> |
d68c6810 | 27 | #include <limits.h> |
b1a90ce3 | 28 | #include <string.h> |
12f246ac | 29 | #include <assert.h> |
61009379 | 30 | #include <urcu/rcuja.h> |
d68c6810 MD |
31 | #include <urcu/compiler.h> |
32 | #include <urcu/arch.h> | |
8e519e3c | 33 | #include <urcu-pointer.h> |
f07b240f | 34 | #include <urcu/uatomic.h> |
8e519e3c | 35 | |
61009379 MD |
36 | #include "rcuja-internal.h" |
37 | ||
b1a90ce3 MD |
38 | #ifndef abs |
39 | #define abs_int(a) ((int) (a) > 0 ? (int) (a) : -((int) (a))) | |
40 | #endif | |
41 | ||
d96bfb0d | 42 | enum cds_ja_type_class { |
e5227865 | 43 | RCU_JA_LINEAR = 0, /* Type A */ |
fd800776 MD |
44 | /* 32-bit: 1 to 25 children, 8 to 128 bytes */ |
45 | /* 64-bit: 1 to 28 children, 16 to 256 bytes */ | |
46 | RCU_JA_POOL = 1, /* Type B */ | |
47 | /* 32-bit: 26 to 100 children, 256 to 512 bytes */ | |
48 | /* 64-bit: 29 to 112 children, 512 to 1024 bytes */ | |
e5227865 | 49 | RCU_JA_PIGEON = 2, /* Type C */ |
fd800776 MD |
50 | /* 32-bit: 101 to 256 children, 1024 bytes */ |
51 | /* 64-bit: 113 to 256 children, 2048 bytes */ | |
e5227865 | 52 | /* Leaf nodes are implicit from their height in the tree */ |
1db4943c | 53 | RCU_JA_NR_TYPES, |
e1db2db5 MD |
54 | |
55 | RCU_JA_NULL, /* not an encoded type, but keeps code regular */ | |
e5227865 MD |
56 | }; |
57 | ||
d96bfb0d MD |
58 | struct cds_ja_type { |
59 | enum cds_ja_type_class type_class; | |
8e519e3c MD |
60 | uint16_t min_child; /* minimum number of children: 1 to 256 */ |
61 | uint16_t max_child; /* maximum number of children: 1 to 256 */ | |
62 | uint16_t max_linear_child; /* per-pool max nr. children: 1 to 256 */ | |
63 | uint16_t order; /* node size is (1 << order), in bytes */ | |
fd800776 MD |
64 | uint16_t nr_pool_order; /* number of pools */ |
65 | uint16_t pool_size_order; /* pool size */ | |
e5227865 MD |
66 | }; |
67 | ||
68 | /* | |
69 | * Iteration on the array to find the right node size for the number of | |
d68c6810 | 70 | * children stops when it reaches .max_child == 256 (this is the largest |
e5227865 | 71 | * possible node size, which contains 256 children). |
d68c6810 MD |
72 | * The min_child overlaps with the previous max_child to provide an |
73 | * hysteresis loop to reallocation for patterns of cyclic add/removal | |
74 | * within the same node. | |
75 | * The node the index within the following arrays is represented on 3 | |
76 | * bits. It identifies the node type, min/max number of children, and | |
77 | * the size order. | |
3d45251f MD |
78 | * The max_child values for the RCU_JA_POOL below result from |
79 | * statistical approximation: over million populations, the max_child | |
80 | * covers between 97% and 99% of the populations generated. Therefore, a | |
81 | * fallback should exist to cover the rare extreme population unbalance | |
82 | * cases, but it will not have a major impact on speed nor space | |
83 | * consumption, since those are rare cases. | |
e5227865 | 84 | */ |
e5227865 | 85 | |
d68c6810 MD |
86 | #if (CAA_BITS_PER_LONG < 64) |
87 | /* 32-bit pointers */ | |
1db4943c MD |
88 | enum { |
89 | ja_type_0_max_child = 1, | |
90 | ja_type_1_max_child = 3, | |
91 | ja_type_2_max_child = 6, | |
92 | ja_type_3_max_child = 12, | |
93 | ja_type_4_max_child = 25, | |
94 | ja_type_5_max_child = 48, | |
95 | ja_type_6_max_child = 92, | |
96 | ja_type_7_max_child = 256, | |
e1db2db5 | 97 | ja_type_8_max_child = 0, /* NULL */ |
1db4943c MD |
98 | }; |
99 | ||
8e519e3c MD |
100 | enum { |
101 | ja_type_0_max_linear_child = 1, | |
102 | ja_type_1_max_linear_child = 3, | |
103 | ja_type_2_max_linear_child = 6, | |
104 | ja_type_3_max_linear_child = 12, | |
105 | ja_type_4_max_linear_child = 25, | |
106 | ja_type_5_max_linear_child = 24, | |
107 | ja_type_6_max_linear_child = 23, | |
108 | }; | |
109 | ||
1db4943c MD |
110 | enum { |
111 | ja_type_5_nr_pool_order = 1, | |
112 | ja_type_6_nr_pool_order = 2, | |
113 | }; | |
114 | ||
d96bfb0d | 115 | const struct cds_ja_type ja_types[] = { |
8e519e3c MD |
116 | { .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, }, |
117 | { .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, }, | |
118 | { .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, }, | |
119 | { .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, }, | |
120 | { .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 | 121 | |
fd800776 | 122 | /* Pools may fill sooner than max_child */ |
1cee749c | 123 | /* This pool is hardcoded at index 5. See ja_node_ptr(). */ |
8e519e3c | 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, }, |
1cee749c | 125 | /* This pool is hardcoded at index 6. See ja_node_ptr(). */ |
8e519e3c | 126 | { .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 |
127 | |
128 | /* | |
b1a90ce3 MD |
129 | * Upon node removal below min_child, if child pool is filled |
130 | * beyond capacity, we roll back to pigeon. | |
3d45251f | 131 | */ |
58c16c03 | 132 | { .type_class = RCU_JA_PIGEON, .min_child = 83, .max_child = ja_type_7_max_child, .order = 10, }, |
e1db2db5 MD |
133 | |
134 | { .type_class = RCU_JA_NULL, .min_child = 0, .max_child = ja_type_8_max_child, }, | |
d68c6810 | 135 | }; |
d68c6810 MD |
136 | #else /* !(CAA_BITS_PER_LONG < 64) */ |
137 | /* 64-bit pointers */ | |
1db4943c MD |
138 | enum { |
139 | ja_type_0_max_child = 1, | |
140 | ja_type_1_max_child = 3, | |
141 | ja_type_2_max_child = 7, | |
142 | ja_type_3_max_child = 14, | |
143 | ja_type_4_max_child = 28, | |
144 | ja_type_5_max_child = 54, | |
145 | ja_type_6_max_child = 104, | |
146 | ja_type_7_max_child = 256, | |
e1db2db5 | 147 | ja_type_8_max_child = 256, |
1db4943c MD |
148 | }; |
149 | ||
8e519e3c MD |
150 | enum { |
151 | ja_type_0_max_linear_child = 1, | |
152 | ja_type_1_max_linear_child = 3, | |
153 | ja_type_2_max_linear_child = 7, | |
154 | ja_type_3_max_linear_child = 14, | |
155 | ja_type_4_max_linear_child = 28, | |
156 | ja_type_5_max_linear_child = 27, | |
157 | ja_type_6_max_linear_child = 26, | |
158 | }; | |
159 | ||
1db4943c MD |
160 | enum { |
161 | ja_type_5_nr_pool_order = 1, | |
162 | ja_type_6_nr_pool_order = 2, | |
163 | }; | |
164 | ||
d96bfb0d | 165 | const struct cds_ja_type ja_types[] = { |
8e519e3c MD |
166 | { .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, }, |
167 | { .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, }, | |
168 | { .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, }, | |
169 | { .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, }, | |
170 | { .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 | 171 | |
3d45251f | 172 | /* Pools may fill sooner than max_child. */ |
1cee749c | 173 | /* This pool is hardcoded at index 5. See ja_node_ptr(). */ |
8e519e3c | 174 | { .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, }, |
1cee749c | 175 | /* This pool is hardcoded at index 6. See ja_node_ptr(). */ |
8e519e3c | 176 | { .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 | 177 | |
3d45251f | 178 | /* |
b1a90ce3 MD |
179 | * Upon node removal below min_child, if child pool is filled |
180 | * beyond capacity, we roll back to pigeon. | |
3d45251f | 181 | */ |
64457f6c | 182 | { .type_class = RCU_JA_PIGEON, .min_child = 95, .max_child = ja_type_7_max_child, .order = 11, }, |
e1db2db5 MD |
183 | |
184 | { .type_class = RCU_JA_NULL, .min_child = 0, .max_child = ja_type_8_max_child, }, | |
e5227865 | 185 | }; |
d68c6810 | 186 | #endif /* !(BITS_PER_LONG < 64) */ |
e5227865 | 187 | |
1db4943c MD |
188 | static inline __attribute__((unused)) |
189 | void static_array_size_check(void) | |
190 | { | |
e1db2db5 | 191 | CAA_BUILD_BUG_ON(CAA_ARRAY_SIZE(ja_types) < JA_TYPE_MAX_NR); |
1db4943c MD |
192 | } |
193 | ||
e5227865 | 194 | /* |
d96bfb0d | 195 | * The cds_ja_node contains the compressed node data needed for |
1db4943c MD |
196 | * read-side. For linear and pool node configurations, it starts with a |
197 | * byte counting the number of children in the node. Then, the | |
198 | * node-specific data is placed. | |
199 | * The node mutex, if any is needed, protecting concurrent updated of | |
200 | * each node is placed in a separate hash table indexed by node address. | |
201 | * For the pigeon configuration, the number of children is also kept in | |
202 | * a separate hash table, indexed by node address, because it is only | |
203 | * required for updates. | |
e5227865 | 204 | */ |
1db4943c | 205 | |
ff38c745 MD |
206 | #define DECLARE_LINEAR_NODE(index) \ |
207 | struct { \ | |
208 | uint8_t nr_child; \ | |
209 | uint8_t child_value[ja_type_## index ##_max_linear_child]; \ | |
b4540e8a | 210 | struct cds_ja_inode_flag *child_ptr[ja_type_## index ##_max_linear_child]; \ |
ff38c745 MD |
211 | } |
212 | ||
213 | #define DECLARE_POOL_NODE(index) \ | |
214 | struct { \ | |
215 | struct { \ | |
216 | uint8_t nr_child; \ | |
217 | uint8_t child_value[ja_type_## index ##_max_linear_child]; \ | |
b4540e8a | 218 | struct cds_ja_inode_flag *child_ptr[ja_type_## index ##_max_linear_child]; \ |
ff38c745 MD |
219 | } linear[1U << ja_type_## index ##_nr_pool_order]; \ |
220 | } | |
1db4943c | 221 | |
b4540e8a | 222 | struct cds_ja_inode { |
1db4943c MD |
223 | union { |
224 | /* Linear configuration */ | |
225 | DECLARE_LINEAR_NODE(0) conf_0; | |
226 | DECLARE_LINEAR_NODE(1) conf_1; | |
227 | DECLARE_LINEAR_NODE(2) conf_2; | |
228 | DECLARE_LINEAR_NODE(3) conf_3; | |
229 | DECLARE_LINEAR_NODE(4) conf_4; | |
230 | ||
231 | /* Pool configuration */ | |
232 | DECLARE_POOL_NODE(5) conf_5; | |
233 | DECLARE_POOL_NODE(6) conf_6; | |
234 | ||
235 | /* Pigeon configuration */ | |
236 | struct { | |
b4540e8a | 237 | struct cds_ja_inode_flag *child[ja_type_7_max_child]; |
1db4943c MD |
238 | } conf_7; |
239 | /* data aliasing nodes for computed accesses */ | |
b4540e8a | 240 | uint8_t data[sizeof(struct cds_ja_inode_flag *) * ja_type_7_max_child]; |
1db4943c | 241 | } u; |
e5227865 MD |
242 | }; |
243 | ||
2e313670 | 244 | enum ja_recompact { |
19ddcd04 MD |
245 | JA_RECOMPACT_ADD_SAME, |
246 | JA_RECOMPACT_ADD_NEXT, | |
2e313670 MD |
247 | JA_RECOMPACT_DEL, |
248 | }; | |
249 | ||
b023ba9f MD |
250 | enum ja_lookup_inequality { |
251 | JA_LOOKUP_BE, | |
252 | JA_LOOKUP_AE, | |
253 | }; | |
254 | ||
255 | enum ja_direction { | |
256 | JA_LEFT, | |
257 | JA_RIGHT, | |
258 | JA_LEFTMOST, | |
259 | JA_RIGHTMOST, | |
260 | }; | |
261 | ||
b1a90ce3 MD |
262 | static |
263 | struct cds_ja_inode *_ja_node_mask_ptr(struct cds_ja_inode_flag *node) | |
264 | { | |
265 | return (struct cds_ja_inode *) (((unsigned long) node) & JA_PTR_MASK); | |
266 | } | |
267 | ||
268 | unsigned long ja_node_type(struct cds_ja_inode_flag *node) | |
269 | { | |
270 | unsigned long type; | |
271 | ||
272 | if (_ja_node_mask_ptr(node) == NULL) { | |
273 | return NODE_INDEX_NULL; | |
274 | } | |
275 | type = (unsigned int) ((unsigned long) node & JA_TYPE_MASK); | |
276 | assert(type < (1UL << JA_TYPE_BITS)); | |
277 | return type; | |
278 | } | |
279 | ||
354981c2 MD |
280 | static |
281 | struct cds_ja_inode *alloc_cds_ja_node(struct cds_ja *ja, | |
282 | const struct cds_ja_type *ja_type) | |
e5227865 | 283 | { |
b1a90ce3 MD |
284 | size_t len = 1U << ja_type->order; |
285 | void *p; | |
286 | int ret; | |
287 | ||
288 | ret = posix_memalign(&p, len, len); | |
289 | if (ret || !p) { | |
290 | return NULL; | |
291 | } | |
292 | memset(p, 0, len); | |
354981c2 | 293 | uatomic_inc(&ja->nr_nodes_allocated); |
b1a90ce3 | 294 | return p; |
e5227865 MD |
295 | } |
296 | ||
354981c2 | 297 | void free_cds_ja_node(struct cds_ja *ja, struct cds_ja_inode *node) |
e5227865 MD |
298 | { |
299 | free(node); | |
48cbe001 | 300 | if (node) |
354981c2 | 301 | uatomic_inc(&ja->nr_nodes_freed); |
e5227865 MD |
302 | } |
303 | ||
d68c6810 MD |
304 | #define __JA_ALIGN_MASK(v, mask) (((v) + (mask)) & ~(mask)) |
305 | #define JA_ALIGN(v, align) __JA_ALIGN_MASK(v, (typeof(v)) (align) - 1) | |
306 | #define __JA_FLOOR_MASK(v, mask) ((v) & ~(mask)) | |
307 | #define JA_FLOOR(v, align) __JA_FLOOR_MASK(v, (typeof(v)) (align) - 1) | |
308 | ||
309 | static | |
1db4943c | 310 | uint8_t *align_ptr_size(uint8_t *ptr) |
d68c6810 | 311 | { |
1db4943c | 312 | return (uint8_t *) JA_ALIGN((unsigned long) ptr, sizeof(void *)); |
d68c6810 MD |
313 | } |
314 | ||
11c5e016 | 315 | static |
d96bfb0d | 316 | uint8_t ja_linear_node_get_nr_child(const struct cds_ja_type *type, |
b4540e8a | 317 | struct cds_ja_inode *node) |
11c5e016 MD |
318 | { |
319 | assert(type->type_class == RCU_JA_LINEAR || type->type_class == RCU_JA_POOL); | |
2e313670 | 320 | return rcu_dereference(node->u.data[0]); |
11c5e016 MD |
321 | } |
322 | ||
13a7f5a6 MD |
323 | /* |
324 | * The order in which values and pointers are does does not matter: if | |
325 | * a value is missing, we return NULL. If a value is there, but its | |
326 | * associated pointers is still NULL, we return NULL too. | |
327 | */ | |
d68c6810 | 328 | static |
b4540e8a MD |
329 | struct cds_ja_inode_flag *ja_linear_node_get_nth(const struct cds_ja_type *type, |
330 | struct cds_ja_inode *node, | |
b0ca2d21 | 331 | struct cds_ja_inode_flag ***node_flag_ptr, |
8e519e3c | 332 | uint8_t n) |
d68c6810 MD |
333 | { |
334 | uint8_t nr_child; | |
335 | uint8_t *values; | |
b4540e8a MD |
336 | struct cds_ja_inode_flag **pointers; |
337 | struct cds_ja_inode_flag *ptr; | |
d68c6810 MD |
338 | unsigned int i; |
339 | ||
8e519e3c | 340 | assert(type->type_class == RCU_JA_LINEAR || type->type_class == RCU_JA_POOL); |
d68c6810 | 341 | |
11c5e016 | 342 | nr_child = ja_linear_node_get_nr_child(type, node); |
13a7f5a6 | 343 | cmm_smp_rmb(); /* read nr_child before values and pointers */ |
8e519e3c MD |
344 | assert(nr_child <= type->max_linear_child); |
345 | assert(type->type_class != RCU_JA_LINEAR || nr_child >= type->min_child); | |
d68c6810 | 346 | |
1db4943c | 347 | values = &node->u.data[1]; |
d68c6810 | 348 | for (i = 0; i < nr_child; i++) { |
13a7f5a6 | 349 | if (CMM_LOAD_SHARED(values[i]) == n) |
d68c6810 MD |
350 | break; |
351 | } | |
b0ca2d21 MD |
352 | if (i >= nr_child) { |
353 | if (caa_unlikely(node_flag_ptr)) | |
354 | *node_flag_ptr = NULL; | |
d68c6810 | 355 | return NULL; |
b0ca2d21 | 356 | } |
b4540e8a | 357 | pointers = (struct cds_ja_inode_flag **) align_ptr_size(&values[type->max_linear_child]); |
13a7f5a6 | 358 | ptr = rcu_dereference(pointers[i]); |
b0ca2d21 MD |
359 | if (caa_unlikely(node_flag_ptr)) |
360 | *node_flag_ptr = &pointers[i]; | |
d68c6810 MD |
361 | return ptr; |
362 | } | |
363 | ||
291b2543 | 364 | static |
b023ba9f | 365 | struct cds_ja_inode_flag *ja_linear_node_get_direction(const struct cds_ja_type *type, |
291b2543 | 366 | struct cds_ja_inode *node, |
36305a3d | 367 | int n, uint8_t *result_key, |
b023ba9f | 368 | enum ja_direction dir) |
291b2543 MD |
369 | { |
370 | uint8_t nr_child; | |
371 | uint8_t *values; | |
372 | struct cds_ja_inode_flag **pointers; | |
86d700e9 | 373 | struct cds_ja_inode_flag *ptr = NULL; |
b023ba9f MD |
374 | unsigned int i; |
375 | int match_idx = -1, match_v; | |
291b2543 MD |
376 | |
377 | assert(type->type_class == RCU_JA_LINEAR || type->type_class == RCU_JA_POOL); | |
b023ba9f MD |
378 | assert(dir == JA_LEFT || dir == JA_RIGHT); |
379 | ||
380 | if (dir == JA_LEFT) { | |
381 | match_v = -1; | |
382 | } else { | |
383 | match_v = JA_ENTRY_PER_NODE; | |
384 | } | |
291b2543 MD |
385 | |
386 | nr_child = ja_linear_node_get_nr_child(type, node); | |
387 | cmm_smp_rmb(); /* read nr_child before values and pointers */ | |
388 | assert(nr_child <= type->max_linear_child); | |
389 | assert(type->type_class != RCU_JA_LINEAR || nr_child >= type->min_child); | |
390 | ||
391 | values = &node->u.data[1]; | |
86d700e9 | 392 | pointers = (struct cds_ja_inode_flag **) align_ptr_size(&values[type->max_linear_child]); |
291b2543 MD |
393 | for (i = 0; i < nr_child; i++) { |
394 | unsigned int v; | |
395 | ||
396 | v = CMM_LOAD_SHARED(values[i]); | |
86d700e9 MD |
397 | ptr = CMM_LOAD_SHARED(pointers[i]); |
398 | if (!ptr) | |
399 | continue; | |
b023ba9f MD |
400 | if (dir == JA_LEFT) { |
401 | if ((int) v < n && (int) v > match_v) { | |
402 | match_v = v; | |
403 | match_idx = i; | |
404 | } | |
405 | } else { | |
406 | if ((int) v > n && (int) v < match_v) { | |
407 | match_v = v; | |
408 | match_idx = i; | |
409 | } | |
291b2543 MD |
410 | } |
411 | } | |
b023ba9f MD |
412 | |
413 | if (match_idx < 0) { | |
291b2543 MD |
414 | return NULL; |
415 | } | |
b023ba9f MD |
416 | assert(match_v >= 0 && match_v < JA_ENTRY_PER_NODE); |
417 | ||
36305a3d | 418 | *result_key = (uint8_t) match_v; |
291b2543 MD |
419 | ptr = rcu_dereference(pointers[match_idx]); |
420 | return ptr; | |
421 | } | |
422 | ||
11c5e016 | 423 | static |
5a9a87dd | 424 | void ja_linear_node_get_ith_pos(const struct cds_ja_type *type, |
b4540e8a | 425 | struct cds_ja_inode *node, |
11c5e016 MD |
426 | uint8_t i, |
427 | uint8_t *v, | |
b4540e8a | 428 | struct cds_ja_inode_flag **iter) |
11c5e016 MD |
429 | { |
430 | uint8_t *values; | |
b4540e8a | 431 | struct cds_ja_inode_flag **pointers; |
11c5e016 MD |
432 | |
433 | assert(type->type_class == RCU_JA_LINEAR || type->type_class == RCU_JA_POOL); | |
434 | assert(i < ja_linear_node_get_nr_child(type, node)); | |
435 | ||
436 | values = &node->u.data[1]; | |
437 | *v = values[i]; | |
b4540e8a | 438 | pointers = (struct cds_ja_inode_flag **) align_ptr_size(&values[type->max_linear_child]); |
11c5e016 MD |
439 | *iter = pointers[i]; |
440 | } | |
441 | ||
d68c6810 | 442 | static |
b4540e8a MD |
443 | struct cds_ja_inode_flag *ja_pool_node_get_nth(const struct cds_ja_type *type, |
444 | struct cds_ja_inode *node, | |
b1a90ce3 | 445 | struct cds_ja_inode_flag *node_flag, |
b0ca2d21 | 446 | struct cds_ja_inode_flag ***node_flag_ptr, |
8e519e3c | 447 | uint8_t n) |
d68c6810 | 448 | { |
b4540e8a | 449 | struct cds_ja_inode *linear; |
d68c6810 | 450 | |
fd800776 | 451 | assert(type->type_class == RCU_JA_POOL); |
b1a90ce3 MD |
452 | |
453 | switch (type->nr_pool_order) { | |
454 | case 1: | |
455 | { | |
456 | unsigned long bitsel, index; | |
457 | ||
458 | bitsel = ja_node_pool_1d_bitsel(node_flag); | |
459 | assert(bitsel < CHAR_BIT); | |
19ddcd04 | 460 | index = ((unsigned long) n >> bitsel) & 0x1; |
b1a90ce3 MD |
461 | linear = (struct cds_ja_inode *) &node->u.data[index << type->pool_size_order]; |
462 | break; | |
463 | } | |
464 | case 2: | |
465 | { | |
19ddcd04 MD |
466 | unsigned long bitsel[2], index[2], rindex; |
467 | ||
468 | ja_node_pool_2d_bitsel(node_flag, bitsel); | |
469 | assert(bitsel[0] < CHAR_BIT); | |
470 | assert(bitsel[1] < CHAR_BIT); | |
471 | index[0] = ((unsigned long) n >> bitsel[0]) & 0x1; | |
472 | index[0] <<= 1; | |
473 | index[1] = ((unsigned long) n >> bitsel[1]) & 0x1; | |
474 | rindex = index[0] | index[1]; | |
475 | linear = (struct cds_ja_inode *) &node->u.data[rindex << type->pool_size_order]; | |
b1a90ce3 MD |
476 | break; |
477 | } | |
478 | default: | |
479 | linear = NULL; | |
480 | assert(0); | |
481 | } | |
48cbe001 | 482 | return ja_linear_node_get_nth(type, linear, node_flag_ptr, n); |
d68c6810 MD |
483 | } |
484 | ||
11c5e016 | 485 | static |
b4540e8a MD |
486 | struct cds_ja_inode *ja_pool_node_get_ith_pool(const struct cds_ja_type *type, |
487 | struct cds_ja_inode *node, | |
11c5e016 MD |
488 | uint8_t i) |
489 | { | |
490 | assert(type->type_class == RCU_JA_POOL); | |
b4540e8a | 491 | return (struct cds_ja_inode *) |
11c5e016 MD |
492 | &node->u.data[(unsigned int) i << type->pool_size_order]; |
493 | } | |
494 | ||
291b2543 | 495 | static |
b023ba9f | 496 | struct cds_ja_inode_flag *ja_pool_node_get_direction(const struct cds_ja_type *type, |
291b2543 | 497 | struct cds_ja_inode *node, |
36305a3d | 498 | int n, uint8_t *result_key, |
b023ba9f | 499 | enum ja_direction dir) |
291b2543 MD |
500 | { |
501 | unsigned int pool_nr; | |
b023ba9f | 502 | int match_v; |
291b2543 MD |
503 | struct cds_ja_inode_flag *match_node_flag = NULL; |
504 | ||
505 | assert(type->type_class == RCU_JA_POOL); | |
b023ba9f MD |
506 | assert(dir == JA_LEFT || dir == JA_RIGHT); |
507 | ||
508 | if (dir == JA_LEFT) { | |
509 | match_v = -1; | |
510 | } else { | |
511 | match_v = JA_ENTRY_PER_NODE; | |
512 | } | |
291b2543 MD |
513 | |
514 | for (pool_nr = 0; pool_nr < (1U << type->nr_pool_order); pool_nr++) { | |
515 | struct cds_ja_inode *pool = | |
516 | ja_pool_node_get_ith_pool(type, | |
517 | node, pool_nr); | |
518 | uint8_t nr_child = | |
519 | ja_linear_node_get_nr_child(type, pool); | |
520 | unsigned int j; | |
521 | ||
522 | for (j = 0; j < nr_child; j++) { | |
523 | struct cds_ja_inode_flag *iter; | |
524 | uint8_t v; | |
525 | ||
526 | ja_linear_node_get_ith_pos(type, pool, | |
527 | j, &v, &iter); | |
528 | if (!iter) | |
529 | continue; | |
b023ba9f MD |
530 | if (dir == JA_LEFT) { |
531 | if ((int) v < n && (int) v > match_v) { | |
532 | match_v = v; | |
533 | match_node_flag = iter; | |
534 | } | |
535 | } else { | |
536 | if ((int) v > n && (int) v < match_v) { | |
537 | match_v = v; | |
538 | match_node_flag = iter; | |
539 | } | |
291b2543 MD |
540 | } |
541 | } | |
542 | } | |
36305a3d MD |
543 | if (match_node_flag) |
544 | *result_key = (uint8_t) match_v; | |
291b2543 MD |
545 | return match_node_flag; |
546 | } | |
547 | ||
d68c6810 | 548 | static |
b4540e8a MD |
549 | struct cds_ja_inode_flag *ja_pigeon_node_get_nth(const struct cds_ja_type *type, |
550 | struct cds_ja_inode *node, | |
b0ca2d21 | 551 | struct cds_ja_inode_flag ***node_flag_ptr, |
8e519e3c | 552 | uint8_t n) |
d68c6810 | 553 | { |
48cbe001 MD |
554 | struct cds_ja_inode_flag **child_node_flag_ptr; |
555 | struct cds_ja_inode_flag *child_node_flag; | |
5a9a87dd | 556 | |
d68c6810 | 557 | assert(type->type_class == RCU_JA_PIGEON); |
48cbe001 MD |
558 | child_node_flag_ptr = &((struct cds_ja_inode_flag **) node->u.data)[n]; |
559 | child_node_flag = rcu_dereference(*child_node_flag_ptr); | |
582a6ade | 560 | dbg_printf("ja_pigeon_node_get_nth child_node_flag_ptr %p\n", |
48cbe001 | 561 | child_node_flag_ptr); |
b0ca2d21 | 562 | if (caa_unlikely(node_flag_ptr)) |
48cbe001 MD |
563 | *node_flag_ptr = child_node_flag_ptr; |
564 | return child_node_flag; | |
d68c6810 MD |
565 | } |
566 | ||
291b2543 | 567 | static |
b023ba9f | 568 | struct cds_ja_inode_flag *ja_pigeon_node_get_direction(const struct cds_ja_type *type, |
291b2543 | 569 | struct cds_ja_inode *node, |
36305a3d | 570 | int n, uint8_t *result_key, |
b023ba9f | 571 | enum ja_direction dir) |
291b2543 MD |
572 | { |
573 | struct cds_ja_inode_flag **child_node_flag_ptr; | |
574 | struct cds_ja_inode_flag *child_node_flag; | |
575 | int i; | |
576 | ||
577 | assert(type->type_class == RCU_JA_PIGEON); | |
b023ba9f MD |
578 | assert(dir == JA_LEFT || dir == JA_RIGHT); |
579 | ||
580 | if (dir == JA_LEFT) { | |
581 | /* n - 1 is first value left of n */ | |
582 | for (i = n - 1; i >= 0; i--) { | |
583 | child_node_flag_ptr = &((struct cds_ja_inode_flag **) node->u.data)[i]; | |
584 | child_node_flag = rcu_dereference(*child_node_flag_ptr); | |
585 | if (child_node_flag) { | |
586 | dbg_printf("ja_pigeon_node_get_left child_node_flag %p\n", | |
587 | child_node_flag); | |
53f943b0 | 588 | *result_key = (uint8_t) i; |
b023ba9f MD |
589 | return child_node_flag; |
590 | } | |
591 | } | |
592 | } else { | |
593 | /* n + 1 is first value right of n */ | |
594 | for (i = n + 1; i < JA_ENTRY_PER_NODE; i++) { | |
595 | child_node_flag_ptr = &((struct cds_ja_inode_flag **) node->u.data)[i]; | |
596 | child_node_flag = rcu_dereference(*child_node_flag_ptr); | |
597 | if (child_node_flag) { | |
598 | dbg_printf("ja_pigeon_node_get_right child_node_flag %p\n", | |
599 | child_node_flag); | |
53f943b0 | 600 | *result_key = (uint8_t) i; |
b023ba9f MD |
601 | return child_node_flag; |
602 | } | |
291b2543 MD |
603 | } |
604 | } | |
605 | return NULL; | |
606 | } | |
607 | ||
2e313670 MD |
608 | static |
609 | struct cds_ja_inode_flag *ja_pigeon_node_get_ith_pos(const struct cds_ja_type *type, | |
610 | struct cds_ja_inode *node, | |
611 | uint8_t i) | |
612 | { | |
48cbe001 | 613 | return ja_pigeon_node_get_nth(type, node, NULL, i); |
2e313670 MD |
614 | } |
615 | ||
13a7f5a6 MD |
616 | /* |
617 | * ja_node_get_nth: get nth item from a node. | |
618 | * node_flag is already rcu_dereference'd. | |
619 | */ | |
d68c6810 | 620 | static |
b62a8d0c | 621 | struct cds_ja_inode_flag *ja_node_get_nth(struct cds_ja_inode_flag *node_flag, |
b0ca2d21 | 622 | struct cds_ja_inode_flag ***node_flag_ptr, |
8e519e3c | 623 | uint8_t n) |
d68c6810 MD |
624 | { |
625 | unsigned int type_index; | |
b4540e8a | 626 | struct cds_ja_inode *node; |
d96bfb0d | 627 | const struct cds_ja_type *type; |
d68c6810 | 628 | |
d68c6810 | 629 | node = ja_node_ptr(node_flag); |
5a9a87dd | 630 | assert(node != NULL); |
d68c6810 MD |
631 | type_index = ja_node_type(node_flag); |
632 | type = &ja_types[type_index]; | |
633 | ||
634 | switch (type->type_class) { | |
635 | case RCU_JA_LINEAR: | |
5a9a87dd | 636 | return ja_linear_node_get_nth(type, node, |
b62a8d0c | 637 | node_flag_ptr, n); |
fd800776 | 638 | case RCU_JA_POOL: |
b1a90ce3 | 639 | return ja_pool_node_get_nth(type, node, node_flag, |
b62a8d0c | 640 | node_flag_ptr, n); |
d68c6810 | 641 | case RCU_JA_PIGEON: |
5a9a87dd | 642 | return ja_pigeon_node_get_nth(type, node, |
b62a8d0c | 643 | node_flag_ptr, n); |
d68c6810 MD |
644 | default: |
645 | assert(0); | |
646 | return (void *) -1UL; | |
647 | } | |
648 | } | |
649 | ||
291b2543 | 650 | static |
b023ba9f | 651 | struct cds_ja_inode_flag *ja_node_get_direction(struct cds_ja_inode_flag *node_flag, |
36305a3d | 652 | int n, uint8_t *result_key, |
b023ba9f | 653 | enum ja_direction dir) |
291b2543 MD |
654 | { |
655 | unsigned int type_index; | |
656 | struct cds_ja_inode *node; | |
657 | const struct cds_ja_type *type; | |
658 | ||
659 | node = ja_node_ptr(node_flag); | |
660 | assert(node != NULL); | |
661 | type_index = ja_node_type(node_flag); | |
662 | type = &ja_types[type_index]; | |
663 | ||
664 | switch (type->type_class) { | |
665 | case RCU_JA_LINEAR: | |
36305a3d | 666 | return ja_linear_node_get_direction(type, node, n, result_key, dir); |
291b2543 | 667 | case RCU_JA_POOL: |
36305a3d | 668 | return ja_pool_node_get_direction(type, node, n, result_key, dir); |
291b2543 | 669 | case RCU_JA_PIGEON: |
36305a3d | 670 | return ja_pigeon_node_get_direction(type, node, n, result_key, dir); |
291b2543 MD |
671 | default: |
672 | assert(0); | |
673 | return (void *) -1UL; | |
674 | } | |
675 | } | |
676 | ||
677 | static | |
b023ba9f | 678 | struct cds_ja_inode_flag *ja_node_get_leftright(struct cds_ja_inode_flag *node_flag, |
36305a3d | 679 | unsigned int n, uint8_t *result_key, |
b023ba9f | 680 | enum ja_direction dir) |
291b2543 | 681 | { |
36305a3d | 682 | return ja_node_get_direction(node_flag, n, result_key, dir); |
b023ba9f MD |
683 | } |
684 | ||
685 | static | |
686 | struct cds_ja_inode_flag *ja_node_get_minmax(struct cds_ja_inode_flag *node_flag, | |
36305a3d | 687 | uint8_t *result_key, |
b023ba9f MD |
688 | enum ja_direction dir) |
689 | { | |
690 | switch (dir) { | |
691 | case JA_LEFTMOST: | |
692 | return ja_node_get_direction(node_flag, | |
36305a3d | 693 | -1, result_key, JA_RIGHT); |
b023ba9f MD |
694 | case JA_RIGHTMOST: |
695 | return ja_node_get_direction(node_flag, | |
36305a3d | 696 | JA_ENTRY_PER_NODE, result_key, JA_LEFT); |
b023ba9f MD |
697 | default: |
698 | assert(0); | |
699 | } | |
291b2543 MD |
700 | } |
701 | ||
8e519e3c | 702 | static |
d96bfb0d | 703 | int ja_linear_node_set_nth(const struct cds_ja_type *type, |
b4540e8a | 704 | struct cds_ja_inode *node, |
d96bfb0d | 705 | struct cds_ja_shadow_node *shadow_node, |
8e519e3c | 706 | uint8_t n, |
b4540e8a | 707 | struct cds_ja_inode_flag *child_node_flag) |
8e519e3c MD |
708 | { |
709 | uint8_t nr_child; | |
710 | uint8_t *values, *nr_child_ptr; | |
b4540e8a | 711 | struct cds_ja_inode_flag **pointers; |
2e313670 | 712 | unsigned int i, unused = 0; |
8e519e3c MD |
713 | |
714 | assert(type->type_class == RCU_JA_LINEAR || type->type_class == RCU_JA_POOL); | |
715 | ||
716 | nr_child_ptr = &node->u.data[0]; | |
48cbe001 MD |
717 | dbg_printf("linear set nth: n %u, nr_child_ptr %p\n", |
718 | (unsigned int) n, nr_child_ptr); | |
8e519e3c MD |
719 | nr_child = *nr_child_ptr; |
720 | assert(nr_child <= type->max_linear_child); | |
8e519e3c MD |
721 | |
722 | values = &node->u.data[1]; | |
2e313670 MD |
723 | pointers = (struct cds_ja_inode_flag **) align_ptr_size(&values[type->max_linear_child]); |
724 | /* Check if node value is already populated */ | |
8e519e3c | 725 | for (i = 0; i < nr_child; i++) { |
2e313670 MD |
726 | if (values[i] == n) { |
727 | if (pointers[i]) | |
728 | return -EEXIST; | |
729 | else | |
730 | break; | |
731 | } else { | |
732 | if (!pointers[i]) | |
733 | unused++; | |
734 | } | |
8e519e3c | 735 | } |
2e313670 MD |
736 | if (i == nr_child && nr_child >= type->max_linear_child) { |
737 | if (unused) | |
738 | return -ERANGE; /* recompact node */ | |
739 | else | |
740 | return -ENOSPC; /* No space left in this node type */ | |
741 | } | |
742 | ||
743 | assert(pointers[i] == NULL); | |
744 | rcu_assign_pointer(pointers[i], child_node_flag); | |
745 | /* If we expanded the nr_child, increment it */ | |
746 | if (i == nr_child) { | |
747 | CMM_STORE_SHARED(values[nr_child], n); | |
748 | /* write pointer and value before nr_child */ | |
749 | cmm_smp_wmb(); | |
750 | CMM_STORE_SHARED(*nr_child_ptr, nr_child + 1); | |
8e519e3c | 751 | } |
e1db2db5 | 752 | shadow_node->nr_child++; |
a2a7ff59 MD |
753 | dbg_printf("linear set nth: %u child, shadow: %u child, for node %p shadow %p\n", |
754 | (unsigned int) CMM_LOAD_SHARED(*nr_child_ptr), | |
755 | (unsigned int) shadow_node->nr_child, | |
756 | node, shadow_node); | |
757 | ||
8e519e3c MD |
758 | return 0; |
759 | } | |
760 | ||
761 | static | |
d96bfb0d | 762 | int ja_pool_node_set_nth(const struct cds_ja_type *type, |
b4540e8a | 763 | struct cds_ja_inode *node, |
b1a90ce3 | 764 | struct cds_ja_inode_flag *node_flag, |
d96bfb0d | 765 | struct cds_ja_shadow_node *shadow_node, |
8e519e3c | 766 | uint8_t n, |
b4540e8a | 767 | struct cds_ja_inode_flag *child_node_flag) |
8e519e3c | 768 | { |
b4540e8a | 769 | struct cds_ja_inode *linear; |
8e519e3c MD |
770 | |
771 | assert(type->type_class == RCU_JA_POOL); | |
b1a90ce3 MD |
772 | |
773 | switch (type->nr_pool_order) { | |
774 | case 1: | |
775 | { | |
776 | unsigned long bitsel, index; | |
777 | ||
778 | bitsel = ja_node_pool_1d_bitsel(node_flag); | |
779 | assert(bitsel < CHAR_BIT); | |
19ddcd04 | 780 | index = ((unsigned long) n >> bitsel) & 0x1; |
b1a90ce3 MD |
781 | linear = (struct cds_ja_inode *) &node->u.data[index << type->pool_size_order]; |
782 | break; | |
783 | } | |
784 | case 2: | |
785 | { | |
19ddcd04 MD |
786 | unsigned long bitsel[2], index[2], rindex; |
787 | ||
788 | ja_node_pool_2d_bitsel(node_flag, bitsel); | |
789 | assert(bitsel[0] < CHAR_BIT); | |
790 | assert(bitsel[1] < CHAR_BIT); | |
791 | index[0] = ((unsigned long) n >> bitsel[0]) & 0x1; | |
792 | index[0] <<= 1; | |
793 | index[1] = ((unsigned long) n >> bitsel[1]) & 0x1; | |
794 | rindex = index[0] | index[1]; | |
795 | linear = (struct cds_ja_inode *) &node->u.data[rindex << type->pool_size_order]; | |
b1a90ce3 MD |
796 | break; |
797 | } | |
798 | default: | |
799 | linear = NULL; | |
800 | assert(0); | |
801 | } | |
802 | ||
e1db2db5 MD |
803 | return ja_linear_node_set_nth(type, linear, shadow_node, |
804 | n, child_node_flag); | |
8e519e3c MD |
805 | } |
806 | ||
807 | static | |
d96bfb0d | 808 | int ja_pigeon_node_set_nth(const struct cds_ja_type *type, |
b4540e8a | 809 | struct cds_ja_inode *node, |
d96bfb0d | 810 | struct cds_ja_shadow_node *shadow_node, |
8e519e3c | 811 | uint8_t n, |
b4540e8a | 812 | struct cds_ja_inode_flag *child_node_flag) |
8e519e3c | 813 | { |
b4540e8a | 814 | struct cds_ja_inode_flag **ptr; |
8e519e3c MD |
815 | |
816 | assert(type->type_class == RCU_JA_PIGEON); | |
b4540e8a | 817 | ptr = &((struct cds_ja_inode_flag **) node->u.data)[n]; |
5a9a87dd | 818 | if (*ptr) |
8e519e3c MD |
819 | return -EEXIST; |
820 | rcu_assign_pointer(*ptr, child_node_flag); | |
e1db2db5 | 821 | shadow_node->nr_child++; |
8e519e3c MD |
822 | return 0; |
823 | } | |
824 | ||
d68c6810 | 825 | /* |
7a0b2331 | 826 | * _ja_node_set_nth: set nth item within a node. Return an error |
8e519e3c | 827 | * (negative error value) if it is already there. |
d68c6810 | 828 | */ |
8e519e3c | 829 | static |
d96bfb0d | 830 | int _ja_node_set_nth(const struct cds_ja_type *type, |
b4540e8a | 831 | struct cds_ja_inode *node, |
b1a90ce3 | 832 | struct cds_ja_inode_flag *node_flag, |
d96bfb0d | 833 | struct cds_ja_shadow_node *shadow_node, |
e1db2db5 | 834 | uint8_t n, |
b4540e8a | 835 | struct cds_ja_inode_flag *child_node_flag) |
8e519e3c | 836 | { |
8e519e3c MD |
837 | switch (type->type_class) { |
838 | case RCU_JA_LINEAR: | |
e1db2db5 | 839 | return ja_linear_node_set_nth(type, node, shadow_node, n, |
8e519e3c MD |
840 | child_node_flag); |
841 | case RCU_JA_POOL: | |
b1a90ce3 | 842 | return ja_pool_node_set_nth(type, node, node_flag, shadow_node, n, |
8e519e3c MD |
843 | child_node_flag); |
844 | case RCU_JA_PIGEON: | |
e1db2db5 | 845 | return ja_pigeon_node_set_nth(type, node, shadow_node, n, |
8e519e3c | 846 | child_node_flag); |
e1db2db5 MD |
847 | case RCU_JA_NULL: |
848 | return -ENOSPC; | |
8e519e3c MD |
849 | default: |
850 | assert(0); | |
851 | return -EINVAL; | |
852 | } | |
853 | ||
854 | return 0; | |
855 | } | |
7a0b2331 | 856 | |
2e313670 | 857 | static |
af3cbd45 | 858 | int ja_linear_node_clear_ptr(const struct cds_ja_type *type, |
2e313670 MD |
859 | struct cds_ja_inode *node, |
860 | struct cds_ja_shadow_node *shadow_node, | |
af3cbd45 | 861 | struct cds_ja_inode_flag **node_flag_ptr) |
2e313670 MD |
862 | { |
863 | uint8_t nr_child; | |
af3cbd45 | 864 | uint8_t *nr_child_ptr; |
2e313670 MD |
865 | |
866 | assert(type->type_class == RCU_JA_LINEAR || type->type_class == RCU_JA_POOL); | |
867 | ||
868 | nr_child_ptr = &node->u.data[0]; | |
2e313670 MD |
869 | nr_child = *nr_child_ptr; |
870 | assert(nr_child <= type->max_linear_child); | |
871 | ||
48cbe001 MD |
872 | if (type->type_class == RCU_JA_LINEAR) { |
873 | assert(!shadow_node->fallback_removal_count); | |
874 | if (shadow_node->nr_child <= type->min_child) { | |
2e313670 MD |
875 | /* We need to try recompacting the node */ |
876 | return -EFBIG; | |
877 | } | |
878 | } | |
19ddcd04 | 879 | dbg_printf("linear clear ptr: nr_child_ptr %p\n", nr_child_ptr); |
af3cbd45 MD |
880 | assert(*node_flag_ptr != NULL); |
881 | rcu_assign_pointer(*node_flag_ptr, NULL); | |
2e313670 MD |
882 | /* |
883 | * Value and nr_child are never changed (would cause ABA issue). | |
884 | * Instead, we leave the pointer to NULL and recompact the node | |
885 | * once in a while. It is allowed to set a NULL pointer to a new | |
886 | * value without recompaction though. | |
887 | * Only update the shadow node accounting. | |
888 | */ | |
889 | shadow_node->nr_child--; | |
af3cbd45 | 890 | dbg_printf("linear clear ptr: %u child, shadow: %u child, for node %p shadow %p\n", |
2e313670 MD |
891 | (unsigned int) CMM_LOAD_SHARED(*nr_child_ptr), |
892 | (unsigned int) shadow_node->nr_child, | |
893 | node, shadow_node); | |
2e313670 MD |
894 | return 0; |
895 | } | |
896 | ||
897 | static | |
af3cbd45 | 898 | int ja_pool_node_clear_ptr(const struct cds_ja_type *type, |
2e313670 | 899 | struct cds_ja_inode *node, |
19ddcd04 | 900 | struct cds_ja_inode_flag *node_flag, |
2e313670 | 901 | struct cds_ja_shadow_node *shadow_node, |
af3cbd45 | 902 | struct cds_ja_inode_flag **node_flag_ptr, |
2e313670 MD |
903 | uint8_t n) |
904 | { | |
905 | struct cds_ja_inode *linear; | |
906 | ||
907 | assert(type->type_class == RCU_JA_POOL); | |
19ddcd04 MD |
908 | |
909 | if (shadow_node->fallback_removal_count) { | |
910 | shadow_node->fallback_removal_count--; | |
911 | } else { | |
912 | /* We should try recompacting the node */ | |
913 | if (shadow_node->nr_child <= type->min_child) | |
914 | return -EFBIG; | |
915 | } | |
916 | ||
917 | switch (type->nr_pool_order) { | |
918 | case 1: | |
919 | { | |
920 | unsigned long bitsel, index; | |
921 | ||
922 | bitsel = ja_node_pool_1d_bitsel(node_flag); | |
923 | assert(bitsel < CHAR_BIT); | |
924 | index = ((unsigned long) n >> bitsel) & type->nr_pool_order; | |
925 | linear = (struct cds_ja_inode *) &node->u.data[index << type->pool_size_order]; | |
926 | break; | |
927 | } | |
928 | case 2: | |
929 | { | |
930 | unsigned long bitsel[2], index[2], rindex; | |
931 | ||
932 | ja_node_pool_2d_bitsel(node_flag, bitsel); | |
933 | assert(bitsel[0] < CHAR_BIT); | |
934 | assert(bitsel[1] < CHAR_BIT); | |
935 | index[0] = ((unsigned long) n >> bitsel[0]) & 0x1; | |
936 | index[0] <<= 1; | |
937 | index[1] = ((unsigned long) n >> bitsel[1]) & 0x1; | |
938 | rindex = index[0] | index[1]; | |
939 | linear = (struct cds_ja_inode *) &node->u.data[rindex << type->pool_size_order]; | |
940 | break; | |
941 | } | |
942 | default: | |
943 | linear = NULL; | |
944 | assert(0); | |
945 | } | |
946 | ||
af3cbd45 | 947 | return ja_linear_node_clear_ptr(type, linear, shadow_node, node_flag_ptr); |
2e313670 MD |
948 | } |
949 | ||
950 | static | |
af3cbd45 | 951 | int ja_pigeon_node_clear_ptr(const struct cds_ja_type *type, |
2e313670 MD |
952 | struct cds_ja_inode *node, |
953 | struct cds_ja_shadow_node *shadow_node, | |
af3cbd45 | 954 | struct cds_ja_inode_flag **node_flag_ptr) |
2e313670 | 955 | { |
2e313670 | 956 | assert(type->type_class == RCU_JA_PIGEON); |
19ddcd04 MD |
957 | |
958 | if (shadow_node->fallback_removal_count) { | |
959 | shadow_node->fallback_removal_count--; | |
960 | } else { | |
961 | /* We should try recompacting the node */ | |
962 | if (shadow_node->nr_child <= type->min_child) | |
963 | return -EFBIG; | |
964 | } | |
4d6ef45e | 965 | dbg_printf("ja_pigeon_node_clear_ptr: clearing ptr: %p\n", *node_flag_ptr); |
af3cbd45 | 966 | rcu_assign_pointer(*node_flag_ptr, NULL); |
2e313670 MD |
967 | shadow_node->nr_child--; |
968 | return 0; | |
969 | } | |
970 | ||
971 | /* | |
af3cbd45 | 972 | * _ja_node_clear_ptr: clear ptr item within a node. Return an error |
2e313670 MD |
973 | * (negative error value) if it is not found (-ENOENT). |
974 | */ | |
975 | static | |
af3cbd45 | 976 | int _ja_node_clear_ptr(const struct cds_ja_type *type, |
2e313670 | 977 | struct cds_ja_inode *node, |
19ddcd04 | 978 | struct cds_ja_inode_flag *node_flag, |
2e313670 | 979 | struct cds_ja_shadow_node *shadow_node, |
af3cbd45 | 980 | struct cds_ja_inode_flag **node_flag_ptr, |
2e313670 MD |
981 | uint8_t n) |
982 | { | |
983 | switch (type->type_class) { | |
984 | case RCU_JA_LINEAR: | |
af3cbd45 | 985 | return ja_linear_node_clear_ptr(type, node, shadow_node, node_flag_ptr); |
2e313670 | 986 | case RCU_JA_POOL: |
19ddcd04 | 987 | return ja_pool_node_clear_ptr(type, node, node_flag, shadow_node, node_flag_ptr, n); |
2e313670 | 988 | case RCU_JA_PIGEON: |
af3cbd45 | 989 | return ja_pigeon_node_clear_ptr(type, node, shadow_node, node_flag_ptr); |
2e313670 MD |
990 | case RCU_JA_NULL: |
991 | return -ENOENT; | |
992 | default: | |
993 | assert(0); | |
994 | return -EINVAL; | |
995 | } | |
996 | ||
997 | return 0; | |
998 | } | |
999 | ||
b1a90ce3 MD |
1000 | /* |
1001 | * Calculate bit distribution. Returns the bit (0 to 7) that splits the | |
1002 | * distribution in two sub-distributions containing as much elements one | |
1003 | * compared to the other. | |
1004 | */ | |
1005 | static | |
1006 | unsigned int ja_node_sum_distribution_1d(enum ja_recompact mode, | |
1007 | struct cds_ja *ja, | |
1008 | unsigned int type_index, | |
1009 | const struct cds_ja_type *type, | |
1010 | struct cds_ja_inode *node, | |
1011 | struct cds_ja_shadow_node *shadow_node, | |
1012 | uint8_t n, | |
1013 | struct cds_ja_inode_flag *child_node_flag, | |
1014 | struct cds_ja_inode_flag **nullify_node_flag_ptr) | |
1015 | { | |
1016 | uint8_t nr_one[JA_BITS_PER_BYTE]; | |
1017 | unsigned int bitsel = 0, bit_i, overall_best_distance = UINT_MAX; | |
1018 | unsigned int distrib_nr_child = 0; | |
1019 | ||
1020 | memset(nr_one, 0, sizeof(nr_one)); | |
1021 | ||
1022 | switch (type->type_class) { | |
1023 | case RCU_JA_LINEAR: | |
1024 | { | |
1025 | uint8_t nr_child = | |
1026 | ja_linear_node_get_nr_child(type, node); | |
1027 | unsigned int i; | |
1028 | ||
1029 | for (i = 0; i < nr_child; i++) { | |
1030 | struct cds_ja_inode_flag *iter; | |
b1a90ce3 MD |
1031 | uint8_t v; |
1032 | ||
1033 | ja_linear_node_get_ith_pos(type, node, i, &v, &iter); | |
1034 | if (!iter) | |
1035 | continue; | |
1036 | if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) | |
1037 | continue; | |
f5531dd9 MD |
1038 | for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) { |
1039 | if (v & (1U << bit_i)) | |
1040 | nr_one[bit_i]++; | |
b1a90ce3 MD |
1041 | } |
1042 | distrib_nr_child++; | |
1043 | } | |
1044 | break; | |
1045 | } | |
1046 | case RCU_JA_POOL: | |
1047 | { | |
1048 | unsigned int pool_nr; | |
1049 | ||
1050 | for (pool_nr = 0; pool_nr < (1U << type->nr_pool_order); pool_nr++) { | |
1051 | struct cds_ja_inode *pool = | |
1052 | ja_pool_node_get_ith_pool(type, | |
1053 | node, pool_nr); | |
1054 | uint8_t nr_child = | |
1055 | ja_linear_node_get_nr_child(type, pool); | |
1056 | unsigned int j; | |
1057 | ||
1058 | for (j = 0; j < nr_child; j++) { | |
1059 | struct cds_ja_inode_flag *iter; | |
b1a90ce3 MD |
1060 | uint8_t v; |
1061 | ||
1062 | ja_linear_node_get_ith_pos(type, pool, | |
1063 | j, &v, &iter); | |
1064 | if (!iter) | |
1065 | continue; | |
1066 | if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) | |
1067 | continue; | |
f5531dd9 MD |
1068 | for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) { |
1069 | if (v & (1U << bit_i)) | |
1070 | nr_one[bit_i]++; | |
b1a90ce3 MD |
1071 | } |
1072 | distrib_nr_child++; | |
1073 | } | |
1074 | } | |
1075 | break; | |
1076 | } | |
1077 | case RCU_JA_PIGEON: | |
1078 | { | |
b1a90ce3 MD |
1079 | unsigned int i; |
1080 | ||
1081 | assert(mode == JA_RECOMPACT_DEL); | |
48cbe001 | 1082 | for (i = 0; i < JA_ENTRY_PER_NODE; i++) { |
b1a90ce3 | 1083 | struct cds_ja_inode_flag *iter; |
b1a90ce3 MD |
1084 | |
1085 | iter = ja_pigeon_node_get_ith_pos(type, node, i); | |
1086 | if (!iter) | |
1087 | continue; | |
1088 | if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) | |
1089 | continue; | |
f5531dd9 MD |
1090 | for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) { |
1091 | if (i & (1U << bit_i)) | |
1092 | nr_one[bit_i]++; | |
b1a90ce3 MD |
1093 | } |
1094 | distrib_nr_child++; | |
1095 | } | |
1096 | break; | |
1097 | } | |
1098 | case RCU_JA_NULL: | |
19ddcd04 | 1099 | assert(mode == JA_RECOMPACT_ADD_NEXT); |
b1a90ce3 MD |
1100 | break; |
1101 | default: | |
1102 | assert(0); | |
1103 | break; | |
1104 | } | |
1105 | ||
19ddcd04 | 1106 | if (mode == JA_RECOMPACT_ADD_NEXT || mode == JA_RECOMPACT_ADD_SAME) { |
f5531dd9 MD |
1107 | for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) { |
1108 | if (n & (1U << bit_i)) | |
1109 | nr_one[bit_i]++; | |
b1a90ce3 MD |
1110 | } |
1111 | distrib_nr_child++; | |
1112 | } | |
1113 | ||
1114 | /* | |
1115 | * The best bit selector is that for which the number of ones is | |
1116 | * closest to half of the number of children in the | |
f5531dd9 MD |
1117 | * distribution. We calculate the distance using the double of |
1118 | * the sub-distribution sizes to eliminate truncation error. | |
b1a90ce3 MD |
1119 | */ |
1120 | for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) { | |
1121 | unsigned int distance_to_best; | |
1122 | ||
1b34283b | 1123 | distance_to_best = abs_int(((unsigned int) nr_one[bit_i] << 1U) - distrib_nr_child); |
b1a90ce3 MD |
1124 | if (distance_to_best < overall_best_distance) { |
1125 | overall_best_distance = distance_to_best; | |
1126 | bitsel = bit_i; | |
1127 | } | |
1128 | } | |
1129 | dbg_printf("1 dimension pool bit selection: (%u)\n", bitsel); | |
1130 | return bitsel; | |
1131 | } | |
1132 | ||
19ddcd04 MD |
1133 | /* |
1134 | * Calculate bit distribution in two dimensions. Returns the two bits | |
1135 | * (each 0 to 7) that splits the distribution in four sub-distributions | |
1136 | * containing as much elements one compared to the other. | |
1137 | */ | |
1138 | static | |
1139 | void ja_node_sum_distribution_2d(enum ja_recompact mode, | |
1140 | struct cds_ja *ja, | |
1141 | unsigned int type_index, | |
1142 | const struct cds_ja_type *type, | |
1143 | struct cds_ja_inode *node, | |
1144 | struct cds_ja_shadow_node *shadow_node, | |
1145 | uint8_t n, | |
1146 | struct cds_ja_inode_flag *child_node_flag, | |
1147 | struct cds_ja_inode_flag **nullify_node_flag_ptr, | |
1148 | unsigned int *_bitsel) | |
1149 | { | |
1150 | uint8_t nr_2d_11[JA_BITS_PER_BYTE][JA_BITS_PER_BYTE], | |
1151 | nr_2d_10[JA_BITS_PER_BYTE][JA_BITS_PER_BYTE], | |
1152 | nr_2d_01[JA_BITS_PER_BYTE][JA_BITS_PER_BYTE], | |
1153 | nr_2d_00[JA_BITS_PER_BYTE][JA_BITS_PER_BYTE]; | |
1154 | unsigned int bitsel[2] = { 0, 1 }; | |
4a073c53 MD |
1155 | unsigned int bit_i, bit_j; |
1156 | int overall_best_distance = INT_MAX; | |
19ddcd04 MD |
1157 | unsigned int distrib_nr_child = 0; |
1158 | ||
1159 | memset(nr_2d_11, 0, sizeof(nr_2d_11)); | |
1160 | memset(nr_2d_10, 0, sizeof(nr_2d_10)); | |
4a073c53 MD |
1161 | memset(nr_2d_01, 0, sizeof(nr_2d_01)); |
1162 | memset(nr_2d_00, 0, sizeof(nr_2d_00)); | |
19ddcd04 MD |
1163 | |
1164 | switch (type->type_class) { | |
1165 | case RCU_JA_LINEAR: | |
1166 | { | |
1167 | uint8_t nr_child = | |
1168 | ja_linear_node_get_nr_child(type, node); | |
1169 | unsigned int i; | |
1170 | ||
1171 | for (i = 0; i < nr_child; i++) { | |
1172 | struct cds_ja_inode_flag *iter; | |
1173 | uint8_t v; | |
1174 | ||
1175 | ja_linear_node_get_ith_pos(type, node, i, &v, &iter); | |
1176 | if (!iter) | |
1177 | continue; | |
1178 | if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) | |
1179 | continue; | |
1180 | for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) { | |
1181 | for (bit_j = 0; bit_j < bit_i; bit_j++) { | |
7f14b43a MD |
1182 | if (v & (1U << bit_i)) { |
1183 | if (v & (1U << bit_j)) { | |
1184 | nr_2d_11[bit_i][bit_j]++; | |
1185 | } else { | |
1186 | nr_2d_10[bit_i][bit_j]++; | |
1187 | } | |
1188 | } else { | |
1189 | if (v & (1U << bit_j)) { | |
1190 | nr_2d_01[bit_i][bit_j]++; | |
1191 | } else { | |
1192 | nr_2d_00[bit_i][bit_j]++; | |
1193 | } | |
19ddcd04 MD |
1194 | } |
1195 | } | |
1196 | } | |
1197 | distrib_nr_child++; | |
1198 | } | |
1199 | break; | |
1200 | } | |
1201 | case RCU_JA_POOL: | |
1202 | { | |
1203 | unsigned int pool_nr; | |
1204 | ||
1205 | for (pool_nr = 0; pool_nr < (1U << type->nr_pool_order); pool_nr++) { | |
1206 | struct cds_ja_inode *pool = | |
1207 | ja_pool_node_get_ith_pool(type, | |
1208 | node, pool_nr); | |
1209 | uint8_t nr_child = | |
1210 | ja_linear_node_get_nr_child(type, pool); | |
1211 | unsigned int j; | |
1212 | ||
1213 | for (j = 0; j < nr_child; j++) { | |
1214 | struct cds_ja_inode_flag *iter; | |
1215 | uint8_t v; | |
1216 | ||
1217 | ja_linear_node_get_ith_pos(type, pool, | |
1218 | j, &v, &iter); | |
1219 | if (!iter) | |
1220 | continue; | |
1221 | if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) | |
1222 | continue; | |
1223 | for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) { | |
1224 | for (bit_j = 0; bit_j < bit_i; bit_j++) { | |
7f14b43a MD |
1225 | if (v & (1U << bit_i)) { |
1226 | if (v & (1U << bit_j)) { | |
1227 | nr_2d_11[bit_i][bit_j]++; | |
1228 | } else { | |
1229 | nr_2d_10[bit_i][bit_j]++; | |
1230 | } | |
1231 | } else { | |
1232 | if (v & (1U << bit_j)) { | |
1233 | nr_2d_01[bit_i][bit_j]++; | |
1234 | } else { | |
1235 | nr_2d_00[bit_i][bit_j]++; | |
1236 | } | |
19ddcd04 MD |
1237 | } |
1238 | } | |
1239 | } | |
1240 | distrib_nr_child++; | |
1241 | } | |
1242 | } | |
1243 | break; | |
1244 | } | |
1245 | case RCU_JA_PIGEON: | |
1246 | { | |
19ddcd04 MD |
1247 | unsigned int i; |
1248 | ||
1249 | assert(mode == JA_RECOMPACT_DEL); | |
48cbe001 | 1250 | for (i = 0; i < JA_ENTRY_PER_NODE; i++) { |
19ddcd04 MD |
1251 | struct cds_ja_inode_flag *iter; |
1252 | ||
1253 | iter = ja_pigeon_node_get_ith_pos(type, node, i); | |
1254 | if (!iter) | |
1255 | continue; | |
1256 | if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) | |
1257 | continue; | |
1258 | for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) { | |
1259 | for (bit_j = 0; bit_j < bit_i; bit_j++) { | |
7f14b43a MD |
1260 | if (i & (1U << bit_i)) { |
1261 | if (i & (1U << bit_j)) { | |
1262 | nr_2d_11[bit_i][bit_j]++; | |
1263 | } else { | |
1264 | nr_2d_10[bit_i][bit_j]++; | |
1265 | } | |
1266 | } else { | |
1267 | if (i & (1U << bit_j)) { | |
1268 | nr_2d_01[bit_i][bit_j]++; | |
1269 | } else { | |
1270 | nr_2d_00[bit_i][bit_j]++; | |
1271 | } | |
19ddcd04 MD |
1272 | } |
1273 | } | |
1274 | } | |
1275 | distrib_nr_child++; | |
1276 | } | |
1277 | break; | |
1278 | } | |
1279 | case RCU_JA_NULL: | |
1280 | assert(mode == JA_RECOMPACT_ADD_NEXT); | |
1281 | break; | |
1282 | default: | |
1283 | assert(0); | |
1284 | break; | |
1285 | } | |
1286 | ||
1287 | if (mode == JA_RECOMPACT_ADD_NEXT || mode == JA_RECOMPACT_ADD_SAME) { | |
1288 | for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) { | |
1289 | for (bit_j = 0; bit_j < bit_i; bit_j++) { | |
7f14b43a MD |
1290 | if (n & (1U << bit_i)) { |
1291 | if (n & (1U << bit_j)) { | |
1292 | nr_2d_11[bit_i][bit_j]++; | |
1293 | } else { | |
1294 | nr_2d_10[bit_i][bit_j]++; | |
1295 | } | |
1296 | } else { | |
1297 | if (n & (1U << bit_j)) { | |
1298 | nr_2d_01[bit_i][bit_j]++; | |
1299 | } else { | |
1300 | nr_2d_00[bit_i][bit_j]++; | |
1301 | } | |
19ddcd04 MD |
1302 | } |
1303 | } | |
1304 | } | |
1305 | distrib_nr_child++; | |
1306 | } | |
1307 | ||
1308 | /* | |
1309 | * The best bit selector is that for which the number of nodes | |
1310 | * in each sub-class is closest to one-fourth of the number of | |
1311 | * children in the distribution. We calculate the distance using | |
1312 | * 4 times the size of the sub-distribution to eliminate | |
1313 | * truncation error. | |
1314 | */ | |
1315 | for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) { | |
1316 | for (bit_j = 0; bit_j < bit_i; bit_j++) { | |
4a073c53 | 1317 | int distance_to_best[4]; |
19ddcd04 | 1318 | |
1b34283b MD |
1319 | distance_to_best[0] = ((unsigned int) nr_2d_11[bit_i][bit_j] << 2U) - distrib_nr_child; |
1320 | distance_to_best[1] = ((unsigned int) nr_2d_10[bit_i][bit_j] << 2U) - distrib_nr_child; | |
1321 | distance_to_best[2] = ((unsigned int) nr_2d_01[bit_i][bit_j] << 2U) - distrib_nr_child; | |
1322 | distance_to_best[3] = ((unsigned int) nr_2d_00[bit_i][bit_j] << 2U) - distrib_nr_child; | |
19ddcd04 | 1323 | |
4a073c53 MD |
1324 | /* Consider worse distance above best */ |
1325 | if (distance_to_best[1] > 0 && distance_to_best[1] > distance_to_best[0]) | |
19ddcd04 | 1326 | distance_to_best[0] = distance_to_best[1]; |
4a073c53 | 1327 | if (distance_to_best[2] > 0 && distance_to_best[2] > distance_to_best[0]) |
19ddcd04 | 1328 | distance_to_best[0] = distance_to_best[2]; |
4a073c53 | 1329 | if (distance_to_best[3] > 0 && distance_to_best[3] > distance_to_best[0]) |
19ddcd04 | 1330 | distance_to_best[0] = distance_to_best[3]; |
4a073c53 | 1331 | |
19ddcd04 MD |
1332 | /* |
1333 | * If our worse distance is better than overall, | |
1334 | * we become new best candidate. | |
1335 | */ | |
1336 | if (distance_to_best[0] < overall_best_distance) { | |
1337 | overall_best_distance = distance_to_best[0]; | |
1338 | bitsel[0] = bit_i; | |
1339 | bitsel[1] = bit_j; | |
1340 | } | |
1341 | } | |
1342 | } | |
1343 | ||
1344 | dbg_printf("2 dimensions pool bit selection: (%u,%u)\n", bitsel[0], bitsel[1]); | |
1345 | ||
1346 | /* Return our bit selection */ | |
1347 | _bitsel[0] = bitsel[0]; | |
1348 | _bitsel[1] = bitsel[1]; | |
1349 | } | |
1350 | ||
48cbe001 MD |
1351 | static |
1352 | unsigned int find_nearest_type_index(unsigned int type_index, | |
1353 | unsigned int nr_nodes) | |
1354 | { | |
1355 | const struct cds_ja_type *type; | |
1356 | ||
1357 | assert(type_index != NODE_INDEX_NULL); | |
1358 | if (nr_nodes == 0) | |
1359 | return NODE_INDEX_NULL; | |
1360 | for (;;) { | |
1361 | type = &ja_types[type_index]; | |
1362 | if (nr_nodes < type->min_child) | |
1363 | type_index--; | |
1364 | else if (nr_nodes > type->max_child) | |
1365 | type_index++; | |
1366 | else | |
1367 | break; | |
1368 | } | |
1369 | return type_index; | |
1370 | } | |
1371 | ||
7a0b2331 MD |
1372 | /* |
1373 | * ja_node_recompact_add: recompact a node, adding a new child. | |
2e313670 | 1374 | * Return 0 on success, -EAGAIN if need to retry, or other negative |
5a9a87dd | 1375 | * error value otherwise. |
7a0b2331 MD |
1376 | */ |
1377 | static | |
2e313670 MD |
1378 | int ja_node_recompact(enum ja_recompact mode, |
1379 | struct cds_ja *ja, | |
e1db2db5 | 1380 | unsigned int old_type_index, |
d96bfb0d | 1381 | const struct cds_ja_type *old_type, |
b4540e8a | 1382 | struct cds_ja_inode *old_node, |
5a9a87dd | 1383 | struct cds_ja_shadow_node *shadow_node, |
3d8fe307 | 1384 | struct cds_ja_inode_flag **old_node_flag_ptr, uint8_t n, |
af3cbd45 | 1385 | struct cds_ja_inode_flag *child_node_flag, |
48cbe001 MD |
1386 | struct cds_ja_inode_flag **nullify_node_flag_ptr, |
1387 | int level) | |
7a0b2331 | 1388 | { |
e1db2db5 | 1389 | unsigned int new_type_index; |
b4540e8a | 1390 | struct cds_ja_inode *new_node; |
af3cbd45 | 1391 | struct cds_ja_shadow_node *new_shadow_node = NULL; |
d96bfb0d | 1392 | const struct cds_ja_type *new_type; |
3d8fe307 | 1393 | struct cds_ja_inode_flag *new_node_flag, *old_node_flag; |
7a0b2331 | 1394 | int ret; |
f07b240f | 1395 | int fallback = 0; |
7a0b2331 | 1396 | |
3d8fe307 MD |
1397 | old_node_flag = *old_node_flag_ptr; |
1398 | ||
48cbe001 MD |
1399 | /* |
1400 | * Need to find nearest type index even for ADD_SAME, because | |
1401 | * this recompaction, when applied to linear nodes, will garbage | |
1402 | * collect dummy (NULL) entries, and can therefore cause a few | |
1403 | * linear representations to be skipped. | |
1404 | */ | |
2e313670 | 1405 | switch (mode) { |
19ddcd04 | 1406 | case JA_RECOMPACT_ADD_SAME: |
48cbe001 MD |
1407 | new_type_index = find_nearest_type_index(old_type_index, |
1408 | shadow_node->nr_child + 1); | |
1409 | dbg_printf("Recompact for node with %u children\n", | |
1410 | shadow_node->nr_child + 1); | |
2e313670 | 1411 | break; |
19ddcd04 | 1412 | case JA_RECOMPACT_ADD_NEXT: |
2e313670 MD |
1413 | if (!shadow_node || old_type_index == NODE_INDEX_NULL) { |
1414 | new_type_index = 0; | |
48cbe001 | 1415 | dbg_printf("Recompact for NULL\n"); |
2e313670 | 1416 | } else { |
48cbe001 MD |
1417 | new_type_index = find_nearest_type_index(old_type_index, |
1418 | shadow_node->nr_child + 1); | |
1419 | dbg_printf("Recompact for node with %u children\n", | |
1420 | shadow_node->nr_child + 1); | |
2e313670 MD |
1421 | } |
1422 | break; | |
1423 | case JA_RECOMPACT_DEL: | |
48cbe001 MD |
1424 | new_type_index = find_nearest_type_index(old_type_index, |
1425 | shadow_node->nr_child - 1); | |
1426 | dbg_printf("Recompact for node with %u children\n", | |
1427 | shadow_node->nr_child - 1); | |
2e313670 MD |
1428 | break; |
1429 | default: | |
1430 | assert(0); | |
7a0b2331 | 1431 | } |
a2a7ff59 | 1432 | |
f07b240f | 1433 | retry: /* for fallback */ |
582a6ade MD |
1434 | dbg_printf("Recompact from type %d to type %d\n", |
1435 | old_type_index, new_type_index); | |
7a0b2331 | 1436 | new_type = &ja_types[new_type_index]; |
2e313670 | 1437 | if (new_type_index != NODE_INDEX_NULL) { |
354981c2 | 1438 | new_node = alloc_cds_ja_node(ja, new_type); |
2e313670 MD |
1439 | if (!new_node) |
1440 | return -ENOMEM; | |
b1a90ce3 MD |
1441 | |
1442 | if (new_type->type_class == RCU_JA_POOL) { | |
1443 | switch (new_type->nr_pool_order) { | |
1444 | case 1: | |
1445 | { | |
19ddcd04 MD |
1446 | unsigned int node_distrib_bitsel; |
1447 | ||
b1a90ce3 MD |
1448 | node_distrib_bitsel = |
1449 | ja_node_sum_distribution_1d(mode, ja, | |
1450 | old_type_index, old_type, | |
1451 | old_node, shadow_node, | |
1452 | n, child_node_flag, | |
1453 | nullify_node_flag_ptr); | |
1454 | assert(!((unsigned long) new_node & JA_POOL_1D_MASK)); | |
1455 | new_node_flag = ja_node_flag_pool_1d(new_node, | |
1456 | new_type_index, node_distrib_bitsel); | |
1457 | break; | |
1458 | } | |
1459 | case 2: | |
1460 | { | |
19ddcd04 MD |
1461 | unsigned int node_distrib_bitsel[2]; |
1462 | ||
1463 | ja_node_sum_distribution_2d(mode, ja, | |
1464 | old_type_index, old_type, | |
1465 | old_node, shadow_node, | |
1466 | n, child_node_flag, | |
1467 | nullify_node_flag_ptr, | |
1468 | node_distrib_bitsel); | |
b1a90ce3 MD |
1469 | assert(!((unsigned long) new_node & JA_POOL_1D_MASK)); |
1470 | assert(!((unsigned long) new_node & JA_POOL_2D_MASK)); | |
19ddcd04 MD |
1471 | new_node_flag = ja_node_flag_pool_2d(new_node, |
1472 | new_type_index, node_distrib_bitsel); | |
b1a90ce3 MD |
1473 | break; |
1474 | } | |
1475 | default: | |
1476 | assert(0); | |
1477 | } | |
1478 | } else { | |
1479 | new_node_flag = ja_node_flag(new_node, new_type_index); | |
1480 | } | |
1481 | ||
2e313670 | 1482 | dbg_printf("Recompact inherit lock from %p\n", shadow_node); |
48cbe001 | 1483 | new_shadow_node = rcuja_shadow_set(ja->ht, new_node_flag, shadow_node, ja, level); |
2e313670 | 1484 | if (!new_shadow_node) { |
354981c2 | 1485 | free_cds_ja_node(ja, new_node); |
2e313670 MD |
1486 | return -ENOMEM; |
1487 | } | |
1488 | if (fallback) | |
1489 | new_shadow_node->fallback_removal_count = | |
1490 | JA_FALLBACK_REMOVAL_COUNT; | |
1491 | } else { | |
1492 | new_node = NULL; | |
1493 | new_node_flag = NULL; | |
e1db2db5 | 1494 | } |
11c5e016 | 1495 | |
19ddcd04 | 1496 | assert(mode != JA_RECOMPACT_ADD_NEXT || old_type->type_class != RCU_JA_PIGEON); |
2e313670 MD |
1497 | |
1498 | if (new_type_index == NODE_INDEX_NULL) | |
1499 | goto skip_copy; | |
1500 | ||
11c5e016 MD |
1501 | switch (old_type->type_class) { |
1502 | case RCU_JA_LINEAR: | |
1503 | { | |
1504 | uint8_t nr_child = | |
1505 | ja_linear_node_get_nr_child(old_type, old_node); | |
1506 | unsigned int i; | |
1507 | ||
1508 | for (i = 0; i < nr_child; i++) { | |
b4540e8a | 1509 | struct cds_ja_inode_flag *iter; |
11c5e016 MD |
1510 | uint8_t v; |
1511 | ||
1512 | ja_linear_node_get_ith_pos(old_type, old_node, i, &v, &iter); | |
1513 | if (!iter) | |
1514 | continue; | |
af3cbd45 | 1515 | if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) |
2e313670 | 1516 | continue; |
b1a90ce3 | 1517 | ret = _ja_node_set_nth(new_type, new_node, new_node_flag, |
f07b240f | 1518 | new_shadow_node, |
11c5e016 | 1519 | v, iter); |
f07b240f MD |
1520 | if (new_type->type_class == RCU_JA_POOL && ret) { |
1521 | goto fallback_toosmall; | |
1522 | } | |
11c5e016 MD |
1523 | assert(!ret); |
1524 | } | |
1525 | break; | |
1526 | } | |
1527 | case RCU_JA_POOL: | |
1528 | { | |
1529 | unsigned int pool_nr; | |
1530 | ||
1531 | for (pool_nr = 0; pool_nr < (1U << old_type->nr_pool_order); pool_nr++) { | |
b4540e8a | 1532 | struct cds_ja_inode *pool = |
11c5e016 MD |
1533 | ja_pool_node_get_ith_pool(old_type, |
1534 | old_node, pool_nr); | |
1535 | uint8_t nr_child = | |
1536 | ja_linear_node_get_nr_child(old_type, pool); | |
1537 | unsigned int j; | |
1538 | ||
1539 | for (j = 0; j < nr_child; j++) { | |
b4540e8a | 1540 | struct cds_ja_inode_flag *iter; |
11c5e016 MD |
1541 | uint8_t v; |
1542 | ||
1543 | ja_linear_node_get_ith_pos(old_type, pool, | |
1544 | j, &v, &iter); | |
1545 | if (!iter) | |
1546 | continue; | |
af3cbd45 | 1547 | if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) |
2e313670 | 1548 | continue; |
b1a90ce3 | 1549 | ret = _ja_node_set_nth(new_type, new_node, new_node_flag, |
f07b240f | 1550 | new_shadow_node, |
11c5e016 | 1551 | v, iter); |
f07b240f MD |
1552 | if (new_type->type_class == RCU_JA_POOL |
1553 | && ret) { | |
1554 | goto fallback_toosmall; | |
1555 | } | |
11c5e016 MD |
1556 | assert(!ret); |
1557 | } | |
1558 | } | |
1559 | break; | |
7a0b2331 | 1560 | } |
a2a7ff59 | 1561 | case RCU_JA_NULL: |
19ddcd04 | 1562 | assert(mode == JA_RECOMPACT_ADD_NEXT); |
a2a7ff59 | 1563 | break; |
11c5e016 | 1564 | case RCU_JA_PIGEON: |
2e313670 | 1565 | { |
2e313670 MD |
1566 | unsigned int i; |
1567 | ||
1568 | assert(mode == JA_RECOMPACT_DEL); | |
48cbe001 | 1569 | for (i = 0; i < JA_ENTRY_PER_NODE; i++) { |
2e313670 MD |
1570 | struct cds_ja_inode_flag *iter; |
1571 | ||
1572 | iter = ja_pigeon_node_get_ith_pos(old_type, old_node, i); | |
1573 | if (!iter) | |
1574 | continue; | |
af3cbd45 | 1575 | if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) |
2e313670 | 1576 | continue; |
b1a90ce3 | 1577 | ret = _ja_node_set_nth(new_type, new_node, new_node_flag, |
2e313670 MD |
1578 | new_shadow_node, |
1579 | i, iter); | |
1580 | if (new_type->type_class == RCU_JA_POOL && ret) { | |
1581 | goto fallback_toosmall; | |
1582 | } | |
1583 | assert(!ret); | |
1584 | } | |
1585 | break; | |
1586 | } | |
11c5e016 MD |
1587 | default: |
1588 | assert(0); | |
5a9a87dd | 1589 | ret = -EINVAL; |
f07b240f | 1590 | goto end; |
11c5e016 | 1591 | } |
2e313670 | 1592 | skip_copy: |
11c5e016 | 1593 | |
19ddcd04 | 1594 | if (mode == JA_RECOMPACT_ADD_NEXT || mode == JA_RECOMPACT_ADD_SAME) { |
2e313670 | 1595 | /* add node */ |
b1a90ce3 | 1596 | ret = _ja_node_set_nth(new_type, new_node, new_node_flag, |
2e313670 MD |
1597 | new_shadow_node, |
1598 | n, child_node_flag); | |
7b413155 MD |
1599 | if (new_type->type_class == RCU_JA_POOL && ret) { |
1600 | goto fallback_toosmall; | |
1601 | } | |
2e313670 MD |
1602 | assert(!ret); |
1603 | } | |
19ddcd04 MD |
1604 | |
1605 | if (fallback) { | |
1606 | dbg_printf("Using fallback for %u children, node type index: %u, mode %s\n", | |
1607 | new_shadow_node->nr_child, old_type_index, mode == JA_RECOMPACT_ADD_NEXT ? "add_next" : | |
1608 | (mode == JA_RECOMPACT_DEL ? "del" : "add_same")); | |
354981c2 | 1609 | uatomic_inc(&ja->node_fallback_count_distribution[new_shadow_node->nr_child]); |
19ddcd04 MD |
1610 | } |
1611 | ||
3d8fe307 MD |
1612 | /* Return pointer to new recompacted node through old_node_flag_ptr */ |
1613 | *old_node_flag_ptr = new_node_flag; | |
a2a7ff59 | 1614 | if (old_node) { |
2e313670 MD |
1615 | int flags; |
1616 | ||
1617 | flags = RCUJA_SHADOW_CLEAR_FREE_NODE; | |
1618 | /* | |
1619 | * It is OK to free the lock associated with a node | |
1620 | * going to NULL, since we are holding the parent lock. | |
1621 | * This synchronizes removal with re-add of that node. | |
1622 | */ | |
1623 | if (new_type_index == NODE_INDEX_NULL) | |
48cbe001 | 1624 | flags |= RCUJA_SHADOW_CLEAR_FREE_LOCK; |
3d8fe307 | 1625 | ret = rcuja_shadow_clear(ja->ht, old_node_flag, shadow_node, |
2e313670 | 1626 | flags); |
a2a7ff59 MD |
1627 | assert(!ret); |
1628 | } | |
5a9a87dd MD |
1629 | |
1630 | ret = 0; | |
f07b240f | 1631 | end: |
5a9a87dd | 1632 | return ret; |
f07b240f MD |
1633 | |
1634 | fallback_toosmall: | |
1635 | /* fallback if next pool is too small */ | |
af3cbd45 | 1636 | assert(new_shadow_node); |
3d8fe307 | 1637 | ret = rcuja_shadow_clear(ja->ht, new_node_flag, new_shadow_node, |
f07b240f MD |
1638 | RCUJA_SHADOW_CLEAR_FREE_NODE); |
1639 | assert(!ret); | |
1640 | ||
19ddcd04 MD |
1641 | switch (mode) { |
1642 | case JA_RECOMPACT_ADD_SAME: | |
1643 | /* | |
1644 | * JA_RECOMPACT_ADD_SAME is only triggered if a linear | |
1645 | * node within a pool has unused entries. It should | |
1646 | * therefore _never_ be too small. | |
1647 | */ | |
4a073c53 | 1648 | assert(0); |
4cde8267 MD |
1649 | |
1650 | /* Fall-through */ | |
19ddcd04 MD |
1651 | case JA_RECOMPACT_ADD_NEXT: |
1652 | { | |
1653 | const struct cds_ja_type *next_type; | |
1654 | ||
1655 | /* | |
1656 | * Recompaction attempt on add failed. Should only | |
1657 | * happen if target node type is pool. Caused by | |
1658 | * hard-to-split distribution. Recompact using the next | |
1659 | * distribution size. | |
1660 | */ | |
1661 | assert(new_type->type_class == RCU_JA_POOL); | |
1662 | next_type = &ja_types[new_type_index + 1]; | |
1663 | /* | |
1664 | * Try going to the next pool size if our population | |
1665 | * fits within its range. This is not flagged as a | |
1666 | * fallback. | |
1667 | */ | |
1668 | if (shadow_node->nr_child + 1 >= next_type->min_child | |
1669 | && shadow_node->nr_child + 1 <= next_type->max_child) { | |
1670 | new_type_index++; | |
1671 | goto retry; | |
1672 | } else { | |
1673 | new_type_index++; | |
1674 | dbg_printf("Add fallback to type %d\n", new_type_index); | |
1675 | uatomic_inc(&ja->nr_fallback); | |
1676 | fallback = 1; | |
1677 | goto retry; | |
1678 | } | |
1679 | break; | |
1680 | } | |
1681 | case JA_RECOMPACT_DEL: | |
1682 | /* | |
1683 | * Recompaction attempt on delete failed. Should only | |
1684 | * happen if target node type is pool. This is caused by | |
1685 | * a hard-to-split distribution. Recompact on same node | |
1686 | * size, but flag current node as "fallback" to ensure | |
1687 | * we don't attempt recompaction before some activity | |
1688 | * has reshuffled our node. | |
1689 | */ | |
1690 | assert(new_type->type_class == RCU_JA_POOL); | |
1691 | new_type_index = old_type_index; | |
1692 | dbg_printf("Delete fallback keeping type %d\n", new_type_index); | |
1693 | uatomic_inc(&ja->nr_fallback); | |
1694 | fallback = 1; | |
1695 | goto retry; | |
1696 | default: | |
1697 | assert(0); | |
1698 | return -EINVAL; | |
1699 | } | |
1700 | ||
1701 | /* | |
1702 | * Last resort fallback: pigeon. | |
1703 | */ | |
f07b240f MD |
1704 | new_type_index = (1UL << JA_TYPE_BITS) - 1; |
1705 | dbg_printf("Fallback to type %d\n", new_type_index); | |
1706 | uatomic_inc(&ja->nr_fallback); | |
1707 | fallback = 1; | |
1708 | goto retry; | |
7a0b2331 MD |
1709 | } |
1710 | ||
5a9a87dd | 1711 | /* |
2e313670 | 1712 | * Return 0 on success, -EAGAIN if need to retry, or other negative |
5a9a87dd MD |
1713 | * error value otherwise. |
1714 | */ | |
7a0b2331 | 1715 | static |
d96bfb0d | 1716 | int ja_node_set_nth(struct cds_ja *ja, |
b4540e8a | 1717 | struct cds_ja_inode_flag **node_flag, uint8_t n, |
5a9a87dd | 1718 | struct cds_ja_inode_flag *child_node_flag, |
48cbe001 MD |
1719 | struct cds_ja_shadow_node *shadow_node, |
1720 | int level) | |
7a0b2331 MD |
1721 | { |
1722 | int ret; | |
e1db2db5 | 1723 | unsigned int type_index; |
d96bfb0d | 1724 | const struct cds_ja_type *type; |
b4540e8a | 1725 | struct cds_ja_inode *node; |
7a0b2331 | 1726 | |
a2a7ff59 MD |
1727 | dbg_printf("ja_node_set_nth for n=%u, node %p, shadow %p\n", |
1728 | (unsigned int) n, ja_node_ptr(*node_flag), shadow_node); | |
1729 | ||
e1db2db5 MD |
1730 | node = ja_node_ptr(*node_flag); |
1731 | type_index = ja_node_type(*node_flag); | |
1732 | type = &ja_types[type_index]; | |
b1a90ce3 | 1733 | ret = _ja_node_set_nth(type, node, *node_flag, shadow_node, |
e1db2db5 | 1734 | n, child_node_flag); |
2e313670 MD |
1735 | switch (ret) { |
1736 | case -ENOSPC: | |
19ddcd04 MD |
1737 | /* Not enough space in node, need to recompact to next type. */ |
1738 | ret = ja_node_recompact(JA_RECOMPACT_ADD_NEXT, ja, type_index, type, node, | |
48cbe001 | 1739 | shadow_node, node_flag, n, child_node_flag, NULL, level); |
2e313670 MD |
1740 | break; |
1741 | case -ERANGE: | |
1742 | /* Node needs to be recompacted. */ | |
19ddcd04 | 1743 | ret = ja_node_recompact(JA_RECOMPACT_ADD_SAME, ja, type_index, type, node, |
48cbe001 | 1744 | shadow_node, node_flag, n, child_node_flag, NULL, level); |
2e313670 MD |
1745 | break; |
1746 | } | |
1747 | return ret; | |
1748 | } | |
1749 | ||
1750 | /* | |
1751 | * Return 0 on success, -EAGAIN if need to retry, or other negative | |
1752 | * error value otherwise. | |
1753 | */ | |
1754 | static | |
af3cbd45 MD |
1755 | int ja_node_clear_ptr(struct cds_ja *ja, |
1756 | struct cds_ja_inode_flag **node_flag_ptr, /* Pointer to location to nullify */ | |
1757 | struct cds_ja_inode_flag **parent_node_flag_ptr, /* Address of parent ptr in its parent */ | |
1758 | struct cds_ja_shadow_node *shadow_node, /* of parent */ | |
48cbe001 | 1759 | uint8_t n, int level) |
2e313670 MD |
1760 | { |
1761 | int ret; | |
1762 | unsigned int type_index; | |
1763 | const struct cds_ja_type *type; | |
1764 | struct cds_ja_inode *node; | |
1765 | ||
af3cbd45 MD |
1766 | dbg_printf("ja_node_clear_ptr for node %p, shadow %p, target ptr %p\n", |
1767 | ja_node_ptr(*parent_node_flag_ptr), shadow_node, node_flag_ptr); | |
2e313670 | 1768 | |
af3cbd45 MD |
1769 | node = ja_node_ptr(*parent_node_flag_ptr); |
1770 | type_index = ja_node_type(*parent_node_flag_ptr); | |
2e313670 | 1771 | type = &ja_types[type_index]; |
19ddcd04 | 1772 | ret = _ja_node_clear_ptr(type, node, *parent_node_flag_ptr, shadow_node, node_flag_ptr, n); |
2e313670 | 1773 | if (ret == -EFBIG) { |
19ddcd04 | 1774 | /* Should try recompaction. */ |
2e313670 | 1775 | ret = ja_node_recompact(JA_RECOMPACT_DEL, ja, type_index, type, node, |
af3cbd45 | 1776 | shadow_node, parent_node_flag_ptr, n, NULL, |
48cbe001 | 1777 | node_flag_ptr, level); |
7a0b2331 MD |
1778 | } |
1779 | return ret; | |
1780 | } | |
be9a7474 | 1781 | |
03ec1aeb | 1782 | struct cds_ja_node *cds_ja_lookup(struct cds_ja *ja, uint64_t key) |
b4540e8a | 1783 | { |
41975c12 MD |
1784 | unsigned int tree_depth, i; |
1785 | struct cds_ja_inode_flag *node_flag; | |
1786 | ||
1787 | if (caa_unlikely(key > ja->key_max)) | |
03ec1aeb | 1788 | return NULL; |
41975c12 | 1789 | tree_depth = ja->tree_depth; |
5a9a87dd | 1790 | node_flag = rcu_dereference(ja->root); |
41975c12 | 1791 | |
5a9a87dd MD |
1792 | /* level 0: root node */ |
1793 | if (!ja_node_ptr(node_flag)) | |
03ec1aeb | 1794 | return NULL; |
5a9a87dd MD |
1795 | |
1796 | for (i = 1; i < tree_depth; i++) { | |
79b41067 MD |
1797 | uint8_t iter_key; |
1798 | ||
1799 | iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (tree_depth - i - 1))); | |
48cbe001 | 1800 | node_flag = ja_node_get_nth(node_flag, NULL, iter_key); |
582a6ade MD |
1801 | dbg_printf("cds_ja_lookup iter key lookup %u finds node_flag %p\n", |
1802 | (unsigned int) iter_key, node_flag); | |
41975c12 | 1803 | if (!ja_node_ptr(node_flag)) |
03ec1aeb | 1804 | return NULL; |
41975c12 MD |
1805 | } |
1806 | ||
5a9a87dd | 1807 | /* Last level lookup succeded. We got an actual match. */ |
03ec1aeb | 1808 | return (struct cds_ja_node *) node_flag; |
5a9a87dd MD |
1809 | } |
1810 | ||
b023ba9f MD |
1811 | static |
1812 | struct cds_ja_node *cds_ja_lookup_inequality(struct cds_ja *ja, uint64_t key, | |
36305a3d | 1813 | uint64_t *result_key, enum ja_lookup_inequality mode) |
291b2543 MD |
1814 | { |
1815 | int tree_depth, level; | |
1816 | struct cds_ja_inode_flag *node_flag, *cur_node_depth[JA_MAX_DEPTH]; | |
36305a3d MD |
1817 | uint8_t cur_key[JA_MAX_DEPTH]; |
1818 | uint64_t _result_key = 0; | |
b023ba9f | 1819 | enum ja_direction dir; |
291b2543 | 1820 | |
b023ba9f MD |
1821 | switch (mode) { |
1822 | case JA_LOOKUP_BE: | |
1823 | if (caa_unlikely(key > ja->key_max || key == 0)) | |
1824 | return NULL; | |
1825 | break; | |
1826 | case JA_LOOKUP_AE: | |
1827 | if (caa_unlikely(key >= ja->key_max)) | |
1828 | return NULL; | |
1829 | break; | |
1830 | default: | |
03ec1aeb | 1831 | return NULL; |
b023ba9f | 1832 | } |
291b2543 MD |
1833 | |
1834 | memset(cur_node_depth, 0, sizeof(cur_node_depth)); | |
36305a3d | 1835 | memset(cur_key, 0, sizeof(cur_key)); |
291b2543 MD |
1836 | tree_depth = ja->tree_depth; |
1837 | node_flag = rcu_dereference(ja->root); | |
1838 | cur_node_depth[0] = node_flag; | |
1839 | ||
1840 | /* level 0: root node */ | |
1841 | if (!ja_node_ptr(node_flag)) | |
03ec1aeb | 1842 | return NULL; |
291b2543 MD |
1843 | |
1844 | for (level = 1; level < tree_depth; level++) { | |
1845 | uint8_t iter_key; | |
1846 | ||
1847 | iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (tree_depth - level - 1))); | |
1848 | node_flag = ja_node_get_nth(node_flag, NULL, iter_key); | |
1849 | if (!ja_node_ptr(node_flag)) | |
1850 | break; | |
36305a3d | 1851 | cur_key[level - 1] = iter_key; |
291b2543 | 1852 | cur_node_depth[level] = node_flag; |
b023ba9f | 1853 | dbg_printf("cds_ja_lookup_inequality iter key lookup %u finds node_flag %p\n", |
291b2543 MD |
1854 | (unsigned int) iter_key, node_flag); |
1855 | } | |
1856 | ||
1857 | if (level == tree_depth) { | |
1858 | /* Last level lookup succeded. We got an equal match. */ | |
36305a3d MD |
1859 | if (result_key) |
1860 | *result_key = key; | |
03ec1aeb | 1861 | return (struct cds_ja_node *) node_flag; |
291b2543 MD |
1862 | } |
1863 | ||
1864 | /* | |
b023ba9f | 1865 | * Find highest value left/right of current node. |
291b2543 | 1866 | * Current node is cur_node_depth[level]. |
b023ba9f MD |
1867 | * Start at current level. If we cannot find any key left/right |
1868 | * of ours, go one level up, seek highest value left/right of | |
1869 | * current (recursively), and when we find one, get the | |
1870 | * rightmost/leftmost child of its rightmost/leftmost child | |
1871 | * (recursively). | |
291b2543 | 1872 | */ |
b023ba9f MD |
1873 | switch (mode) { |
1874 | case JA_LOOKUP_BE: | |
1875 | dir = JA_LEFT; | |
1876 | break; | |
1877 | case JA_LOOKUP_AE: | |
1878 | dir = JA_RIGHT; | |
1879 | break; | |
1880 | default: | |
1881 | assert(0); | |
1882 | } | |
291b2543 MD |
1883 | for (; level > 0; level--) { |
1884 | uint8_t iter_key; | |
1885 | ||
1886 | iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (tree_depth - level - 1))); | |
b023ba9f | 1887 | node_flag = ja_node_get_leftright(cur_node_depth[level - 1], |
36305a3d | 1888 | iter_key, &cur_key[level - 1], dir); |
86d700e9 MD |
1889 | dbg_printf("cds_ja_lookup_inequality find sibling from %u at %u finds node_flag %p\n", |
1890 | (unsigned int) iter_key, (unsigned int) cur_key[level - 1], | |
1891 | node_flag); | |
36305a3d | 1892 | /* If found left/right sibling, find rightmost/leftmost child. */ |
291b2543 MD |
1893 | if (ja_node_ptr(node_flag)) |
1894 | break; | |
1895 | } | |
1896 | ||
1897 | if (!level) { | |
b023ba9f | 1898 | /* Reached the root and could not find a left/right sibling. */ |
03ec1aeb | 1899 | return NULL; |
291b2543 MD |
1900 | } |
1901 | ||
1902 | level++; | |
3c52f0f9 MD |
1903 | |
1904 | /* | |
4cef6f97 | 1905 | * From this point, we are guaranteed to be able to find a |
b023ba9f MD |
1906 | * "below than"/"above than" match. ja_attach_node() and |
1907 | * ja_detach_node() both guarantee that it is not possible for a | |
1908 | * lookup to reach a dead-end. | |
3c52f0f9 MD |
1909 | */ |
1910 | ||
b023ba9f MD |
1911 | /* |
1912 | * Find rightmost/leftmost child of rightmost/leftmost child | |
1913 | * (recursively). | |
1914 | */ | |
1915 | switch (mode) { | |
1916 | case JA_LOOKUP_BE: | |
1917 | dir = JA_RIGHTMOST; | |
1918 | break; | |
1919 | case JA_LOOKUP_AE: | |
1920 | dir = JA_LEFTMOST; | |
1921 | break; | |
1922 | default: | |
1923 | assert(0); | |
1924 | } | |
291b2543 | 1925 | for (; level < tree_depth; level++) { |
36305a3d | 1926 | node_flag = ja_node_get_minmax(node_flag, &cur_key[level - 1], dir); |
86d700e9 MD |
1927 | dbg_printf("cds_ja_lookup_inequality find minmax at %u finds node_flag %p\n", |
1928 | (unsigned int) cur_key[level - 1], | |
1929 | node_flag); | |
291b2543 MD |
1930 | if (!ja_node_ptr(node_flag)) |
1931 | break; | |
1932 | } | |
1933 | ||
4cef6f97 | 1934 | assert(level == tree_depth); |
291b2543 | 1935 | |
36305a3d MD |
1936 | if (result_key) { |
1937 | for (level = 1; level < tree_depth; level++) { | |
1938 | _result_key |= ((uint64_t) cur_key[level - 1]) | |
1939 | << (JA_BITS_PER_BYTE * (tree_depth - level - 1)); | |
1940 | } | |
1941 | *result_key = _result_key; | |
1942 | } | |
03ec1aeb | 1943 | return (struct cds_ja_node *) node_flag; |
291b2543 MD |
1944 | } |
1945 | ||
36305a3d MD |
1946 | struct cds_ja_node *cds_ja_lookup_below_equal(struct cds_ja *ja, |
1947 | uint64_t key, uint64_t *result_key) | |
b023ba9f | 1948 | { |
86d700e9 | 1949 | dbg_printf("cds_ja_lookup_below_equal key %" PRIu64 "\n", key); |
36305a3d | 1950 | return cds_ja_lookup_inequality(ja, key, result_key, JA_LOOKUP_BE); |
b023ba9f MD |
1951 | } |
1952 | ||
36305a3d MD |
1953 | struct cds_ja_node *cds_ja_lookup_above_equal(struct cds_ja *ja, |
1954 | uint64_t key, uint64_t *result_key) | |
b023ba9f | 1955 | { |
86d700e9 | 1956 | dbg_printf("cds_ja_lookup_above_equal key %" PRIu64 "\n", key); |
36305a3d | 1957 | return cds_ja_lookup_inequality(ja, key, result_key, JA_LOOKUP_AE); |
b023ba9f MD |
1958 | } |
1959 | ||
5a9a87dd MD |
1960 | /* |
1961 | * We reached an unpopulated node. Create it and the children we need, | |
1962 | * and then attach the entire branch to the current node. This may | |
1963 | * trigger recompaction of the current node. Locks needed: node lock | |
1964 | * (for add), and, possibly, parent node lock (to update pointer due to | |
1965 | * node recompaction). | |
1966 | * | |
1967 | * First take node lock, check if recompaction is needed, then take | |
1968 | * parent lock (if needed). Then we can proceed to create the new | |
1969 | * branch. Publish the new branch, and release locks. | |
1970 | * TODO: we currently always take the parent lock even when not needed. | |
47d2eab3 MD |
1971 | * |
1972 | * ja_attach_node() ensures that a lookup will _never_ see a branch that | |
1973 | * leads to a dead-end: before attaching a branch, the entire content of | |
1974 | * the new branch is populated, thus creating a cluster, before | |
1975 | * attaching the cluster to the rest of the tree, thus making it visible | |
1976 | * to lookups. | |
5a9a87dd MD |
1977 | */ |
1978 | static | |
1979 | int ja_attach_node(struct cds_ja *ja, | |
b0ca2d21 | 1980 | struct cds_ja_inode_flag **attach_node_flag_ptr, |
b62a8d0c | 1981 | struct cds_ja_inode_flag *attach_node_flag, |
48cbe001 MD |
1982 | struct cds_ja_inode_flag *parent_attach_node_flag, |
1983 | struct cds_ja_inode_flag **old_node_flag_ptr, | |
1984 | struct cds_ja_inode_flag *old_node_flag, | |
5a9a87dd | 1985 | uint64_t key, |
79b41067 | 1986 | unsigned int level, |
5a9a87dd MD |
1987 | struct cds_ja_node *child_node) |
1988 | { | |
1989 | struct cds_ja_shadow_node *shadow_node = NULL, | |
af3cbd45 | 1990 | *parent_shadow_node = NULL; |
5a9a87dd MD |
1991 | struct cds_ja_inode_flag *iter_node_flag, *iter_dest_node_flag; |
1992 | int ret, i; | |
a2a7ff59 | 1993 | struct cds_ja_inode_flag *created_nodes[JA_MAX_DEPTH]; |
5a9a87dd MD |
1994 | int nr_created_nodes = 0; |
1995 | ||
48cbe001 MD |
1996 | 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", |
1997 | level, old_node_flag, attach_node_flag_ptr, attach_node_flag, parent_attach_node_flag); | |
a2a7ff59 | 1998 | |
48cbe001 MD |
1999 | assert(!old_node_flag); |
2000 | if (attach_node_flag) { | |
2001 | shadow_node = rcuja_shadow_lookup_lock(ja->ht, attach_node_flag); | |
2002 | if (!shadow_node) { | |
2003 | ret = -EAGAIN; | |
2004 | goto end; | |
2005 | } | |
5a9a87dd | 2006 | } |
48cbe001 | 2007 | if (parent_attach_node_flag) { |
5a9a87dd | 2008 | parent_shadow_node = rcuja_shadow_lookup_lock(ja->ht, |
48cbe001 | 2009 | parent_attach_node_flag); |
5a9a87dd | 2010 | if (!parent_shadow_node) { |
2e313670 | 2011 | ret = -EAGAIN; |
5a9a87dd MD |
2012 | goto unlock_shadow; |
2013 | } | |
2014 | } | |
2015 | ||
48cbe001 | 2016 | if (old_node_flag_ptr && ja_node_ptr(*old_node_flag_ptr)) { |
b306a0fe | 2017 | /* |
c112acaa MD |
2018 | * Target node has been updated between RCU lookup and |
2019 | * lock acquisition. We need to re-try lookup and | |
2020 | * attach. | |
2021 | */ | |
2022 | ret = -EAGAIN; | |
2023 | goto unlock_parent; | |
2024 | } | |
2025 | ||
9be99d4a MD |
2026 | /* |
2027 | * Perform a lookup query to handle the case where | |
2028 | * old_node_flag_ptr is NULL. We cannot use it to check if the | |
2029 | * node has been populated between RCU lookup and mutex | |
2030 | * acquisition. | |
2031 | */ | |
2032 | if (!old_node_flag_ptr) { | |
2033 | uint8_t iter_key; | |
2034 | struct cds_ja_inode_flag *lookup_node_flag; | |
2035 | struct cds_ja_inode_flag **lookup_node_flag_ptr; | |
2036 | ||
2037 | iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (ja->tree_depth - level))); | |
2038 | lookup_node_flag = ja_node_get_nth(attach_node_flag, | |
2039 | &lookup_node_flag_ptr, | |
2040 | iter_key); | |
2041 | if (lookup_node_flag) { | |
2042 | ret = -EEXIST; | |
2043 | goto unlock_parent; | |
2044 | } | |
2045 | } | |
2046 | ||
c112acaa | 2047 | if (attach_node_flag_ptr && ja_node_ptr(*attach_node_flag_ptr) != |
b62a8d0c | 2048 | ja_node_ptr(attach_node_flag)) { |
c112acaa MD |
2049 | /* |
2050 | * Target node has been updated between RCU lookup and | |
2051 | * lock acquisition. We need to re-try lookup and | |
2052 | * attach. | |
b306a0fe MD |
2053 | */ |
2054 | ret = -EAGAIN; | |
2055 | goto unlock_parent; | |
2056 | } | |
2057 | ||
a2a7ff59 | 2058 | /* Create new branch, starting from bottom */ |
03ec1aeb | 2059 | iter_node_flag = (struct cds_ja_inode_flag *) child_node; |
5a9a87dd | 2060 | |
48cbe001 | 2061 | for (i = ja->tree_depth - 1; i >= (int) level; i--) { |
79b41067 MD |
2062 | uint8_t iter_key; |
2063 | ||
48cbe001 | 2064 | iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (ja->tree_depth - i - 1))); |
79b41067 | 2065 | dbg_printf("branch creation level %d, key %u\n", |
48cbe001 | 2066 | i, (unsigned int) iter_key); |
5a9a87dd MD |
2067 | iter_dest_node_flag = NULL; |
2068 | ret = ja_node_set_nth(ja, &iter_dest_node_flag, | |
79b41067 | 2069 | iter_key, |
5a9a87dd | 2070 | iter_node_flag, |
48cbe001 | 2071 | NULL, i); |
9be99d4a MD |
2072 | if (ret) { |
2073 | dbg_printf("branch creation error %d\n", ret); | |
5a9a87dd | 2074 | goto check_error; |
9be99d4a | 2075 | } |
5a9a87dd MD |
2076 | created_nodes[nr_created_nodes++] = iter_dest_node_flag; |
2077 | iter_node_flag = iter_dest_node_flag; | |
2078 | } | |
48cbe001 | 2079 | assert(level > 0); |
5a9a87dd | 2080 | |
48cbe001 MD |
2081 | /* Publish branch */ |
2082 | if (level == 1) { | |
2083 | /* | |
2084 | * Attaching to root node. | |
2085 | */ | |
2086 | rcu_assign_pointer(ja->root, iter_node_flag); | |
2087 | } else { | |
79b41067 MD |
2088 | uint8_t iter_key; |
2089 | ||
2090 | iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (ja->tree_depth - level))); | |
48cbe001 MD |
2091 | dbg_printf("publish branch at level %d, key %u\n", |
2092 | level - 1, (unsigned int) iter_key); | |
a2a7ff59 | 2093 | /* We need to use set_nth on the previous level. */ |
48cbe001 | 2094 | iter_dest_node_flag = attach_node_flag; |
a2a7ff59 | 2095 | ret = ja_node_set_nth(ja, &iter_dest_node_flag, |
79b41067 | 2096 | iter_key, |
a2a7ff59 | 2097 | iter_node_flag, |
48cbe001 | 2098 | shadow_node, level - 1); |
9be99d4a MD |
2099 | if (ret) { |
2100 | dbg_printf("branch publish error %d\n", ret); | |
a2a7ff59 | 2101 | goto check_error; |
9be99d4a | 2102 | } |
48cbe001 MD |
2103 | /* |
2104 | * Attach branch | |
2105 | */ | |
2106 | rcu_assign_pointer(*attach_node_flag_ptr, iter_dest_node_flag); | |
a2a7ff59 MD |
2107 | } |
2108 | ||
5a9a87dd MD |
2109 | /* Success */ |
2110 | ret = 0; | |
2111 | ||
2112 | check_error: | |
2113 | if (ret) { | |
2114 | for (i = 0; i < nr_created_nodes; i++) { | |
2115 | int tmpret; | |
a2a7ff59 MD |
2116 | int flags; |
2117 | ||
2118 | flags = RCUJA_SHADOW_CLEAR_FREE_LOCK; | |
2119 | if (i) | |
2120 | flags |= RCUJA_SHADOW_CLEAR_FREE_NODE; | |
5a9a87dd | 2121 | tmpret = rcuja_shadow_clear(ja->ht, |
3d8fe307 | 2122 | created_nodes[i], |
a2a7ff59 MD |
2123 | NULL, |
2124 | flags); | |
5a9a87dd MD |
2125 | assert(!tmpret); |
2126 | } | |
2127 | } | |
b306a0fe | 2128 | unlock_parent: |
5a9a87dd MD |
2129 | if (parent_shadow_node) |
2130 | rcuja_shadow_unlock(parent_shadow_node); | |
2131 | unlock_shadow: | |
2132 | if (shadow_node) | |
2133 | rcuja_shadow_unlock(shadow_node); | |
2134 | end: | |
2135 | return ret; | |
2136 | } | |
2137 | ||
2138 | /* | |
03ec1aeb MD |
2139 | * Lock the parent containing the pointer to list of duplicates, and add |
2140 | * node to this list. Failure can happen if concurrent update changes | |
2141 | * the parent before we get the lock. We return -EAGAIN in that case. | |
5a9a87dd MD |
2142 | * Return 0 on success, negative error value on failure. |
2143 | */ | |
2144 | static | |
2145 | int ja_chain_node(struct cds_ja *ja, | |
af3cbd45 | 2146 | struct cds_ja_inode_flag *parent_node_flag, |
fa112799 | 2147 | struct cds_ja_inode_flag **node_flag_ptr, |
c112acaa | 2148 | struct cds_ja_inode_flag *node_flag, |
28fd1038 | 2149 | struct cds_ja_node *last_node, |
5a9a87dd MD |
2150 | struct cds_ja_node *node) |
2151 | { | |
2152 | struct cds_ja_shadow_node *shadow_node; | |
28fd1038 MD |
2153 | struct cds_ja_node *iter_node; |
2154 | int ret = 0, found = 0; | |
5a9a87dd | 2155 | |
3d8fe307 | 2156 | shadow_node = rcuja_shadow_lookup_lock(ja->ht, parent_node_flag); |
b306a0fe | 2157 | if (!shadow_node) { |
2e313670 | 2158 | return -EAGAIN; |
b306a0fe | 2159 | } |
28fd1038 MD |
2160 | /* |
2161 | * Ensure that previous node is still there at end of list. | |
2162 | */ | |
2163 | iter_node = (struct cds_ja_node *) ja_node_ptr(node_flag); | |
2164 | if ((struct cds_ja_node *) ja_node_ptr(*node_flag_ptr) != iter_node) { | |
2165 | ret = -EAGAIN; | |
2166 | goto end; | |
2167 | } | |
2168 | cds_ja_for_each_duplicate(iter_node) { | |
2169 | if (iter_node == last_node) | |
2170 | found = 1; | |
2171 | } | |
2172 | if (!found) { | |
fa112799 MD |
2173 | ret = -EAGAIN; |
2174 | goto end; | |
2175 | } | |
03ec1aeb | 2176 | /* |
28fd1038 MD |
2177 | * Add node to tail of list to ensure that RCU traversals will |
2178 | * always see either the prior node or the newly added if | |
2179 | * executed concurrently with a sequence of add followed by del | |
2180 | * on the same key. Safe against concurrent RCU read traversals. | |
03ec1aeb | 2181 | */ |
28fd1038 MD |
2182 | node->next = NULL; |
2183 | rcu_assign_pointer(last_node->next, node); | |
fa112799 | 2184 | end: |
5a9a87dd | 2185 | rcuja_shadow_unlock(shadow_node); |
fa112799 | 2186 | return ret; |
5a9a87dd MD |
2187 | } |
2188 | ||
75d573aa MD |
2189 | static |
2190 | int _cds_ja_add(struct cds_ja *ja, uint64_t key, | |
6475613c | 2191 | struct cds_ja_node *node, |
75d573aa | 2192 | struct cds_ja_node **unique_node_ret) |
5a9a87dd MD |
2193 | { |
2194 | unsigned int tree_depth, i; | |
48cbe001 | 2195 | struct cds_ja_inode_flag *attach_node_flag, |
5a9a87dd | 2196 | *parent_node_flag, |
b62a8d0c | 2197 | *parent2_node_flag, |
48cbe001 MD |
2198 | *node_flag, |
2199 | *parent_attach_node_flag; | |
2200 | struct cds_ja_inode_flag **attach_node_flag_ptr, | |
2201 | **parent_node_flag_ptr, | |
2202 | **node_flag_ptr; | |
5a9a87dd MD |
2203 | int ret; |
2204 | ||
b306a0fe | 2205 | if (caa_unlikely(key > ja->key_max)) { |
5a9a87dd | 2206 | return -EINVAL; |
b306a0fe | 2207 | } |
5a9a87dd MD |
2208 | tree_depth = ja->tree_depth; |
2209 | ||
2210 | retry: | |
a2a7ff59 | 2211 | dbg_printf("cds_ja_add attempt: key %" PRIu64 ", node %p\n", |
6475613c | 2212 | key, node); |
5a9a87dd | 2213 | parent2_node_flag = NULL; |
b0f74e47 MD |
2214 | parent_node_flag = |
2215 | (struct cds_ja_inode_flag *) &ja->root; /* Use root ptr address as key for mutex */ | |
48cbe001 | 2216 | parent_node_flag_ptr = NULL; |
35170a44 | 2217 | node_flag = rcu_dereference(ja->root); |
48cbe001 | 2218 | node_flag_ptr = &ja->root; |
5a9a87dd MD |
2219 | |
2220 | /* Iterate on all internal levels */ | |
a2a7ff59 | 2221 | for (i = 1; i < tree_depth; i++) { |
79b41067 MD |
2222 | uint8_t iter_key; |
2223 | ||
48cbe001 MD |
2224 | if (!ja_node_ptr(node_flag)) |
2225 | break; | |
2226 | dbg_printf("cds_ja_add iter parent2_node_flag %p parent_node_flag %p node_flag_ptr %p node_flag %p\n", | |
2227 | parent2_node_flag, parent_node_flag, node_flag_ptr, node_flag); | |
79b41067 | 2228 | iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (tree_depth - i - 1))); |
5a9a87dd MD |
2229 | parent2_node_flag = parent_node_flag; |
2230 | parent_node_flag = node_flag; | |
48cbe001 | 2231 | parent_node_flag_ptr = node_flag_ptr; |
5a9a87dd MD |
2232 | node_flag = ja_node_get_nth(node_flag, |
2233 | &node_flag_ptr, | |
79b41067 | 2234 | iter_key); |
5a9a87dd MD |
2235 | } |
2236 | ||
2237 | /* | |
48cbe001 MD |
2238 | * We reached either bottom of tree or internal NULL node, |
2239 | * simply add node to last internal level, or chain it if key is | |
2240 | * already present. | |
5a9a87dd MD |
2241 | */ |
2242 | if (!ja_node_ptr(node_flag)) { | |
48cbe001 MD |
2243 | dbg_printf("cds_ja_add NULL parent2_node_flag %p parent_node_flag %p node_flag_ptr %p node_flag %p\n", |
2244 | parent2_node_flag, parent_node_flag, node_flag_ptr, node_flag); | |
75d573aa | 2245 | |
48cbe001 MD |
2246 | attach_node_flag = parent_node_flag; |
2247 | attach_node_flag_ptr = parent_node_flag_ptr; | |
2248 | parent_attach_node_flag = parent2_node_flag; | |
2249 | ||
b0ca2d21 | 2250 | ret = ja_attach_node(ja, attach_node_flag_ptr, |
b62a8d0c | 2251 | attach_node_flag, |
48cbe001 MD |
2252 | parent_attach_node_flag, |
2253 | node_flag_ptr, | |
2254 | node_flag, | |
6475613c | 2255 | key, i, node); |
5a9a87dd | 2256 | } else { |
28fd1038 MD |
2257 | struct cds_ja_node *iter_node, *last_node = NULL; |
2258 | ||
75d573aa MD |
2259 | if (unique_node_ret) { |
2260 | *unique_node_ret = (struct cds_ja_node *) ja_node_ptr(node_flag); | |
2261 | return -EEXIST; | |
2262 | } | |
2263 | ||
28fd1038 MD |
2264 | /* Find last duplicate */ |
2265 | iter_node = (struct cds_ja_node *) ja_node_ptr(node_flag); | |
2266 | cds_ja_for_each_duplicate_rcu(iter_node) | |
2267 | last_node = iter_node; | |
2268 | ||
48cbe001 MD |
2269 | dbg_printf("cds_ja_add duplicate parent2_node_flag %p parent_node_flag %p node_flag_ptr %p node_flag %p\n", |
2270 | parent2_node_flag, parent_node_flag, node_flag_ptr, node_flag); | |
75d573aa | 2271 | |
48cbe001 MD |
2272 | attach_node_flag = node_flag; |
2273 | attach_node_flag_ptr = node_flag_ptr; | |
2274 | parent_attach_node_flag = parent_node_flag; | |
2275 | ||
5a9a87dd | 2276 | ret = ja_chain_node(ja, |
48cbe001 MD |
2277 | parent_attach_node_flag, |
2278 | attach_node_flag_ptr, | |
2279 | attach_node_flag, | |
28fd1038 | 2280 | last_node, |
6475613c | 2281 | node); |
5a9a87dd | 2282 | } |
b306a0fe | 2283 | if (ret == -EAGAIN || ret == -EEXIST) |
5a9a87dd | 2284 | goto retry; |
48cbe001 | 2285 | |
5a9a87dd | 2286 | return ret; |
b4540e8a MD |
2287 | } |
2288 | ||
75d573aa | 2289 | int cds_ja_add(struct cds_ja *ja, uint64_t key, |
6475613c | 2290 | struct cds_ja_node *node) |
75d573aa | 2291 | { |
6475613c | 2292 | return _cds_ja_add(ja, key, node, NULL); |
75d573aa MD |
2293 | } |
2294 | ||
2295 | struct cds_ja_node *cds_ja_add_unique(struct cds_ja *ja, uint64_t key, | |
6475613c | 2296 | struct cds_ja_node *node) |
75d573aa MD |
2297 | { |
2298 | int ret; | |
2299 | struct cds_ja_node *ret_node; | |
2300 | ||
6475613c | 2301 | ret = _cds_ja_add(ja, key, node, &ret_node); |
75d573aa MD |
2302 | if (ret == -EEXIST) |
2303 | return ret_node; | |
2304 | else | |
6475613c | 2305 | return node; |
75d573aa MD |
2306 | } |
2307 | ||
af3cbd45 MD |
2308 | /* |
2309 | * Note: there is no need to lookup the pointer address associated with | |
2310 | * each node's nth item after taking the lock: it's already been done by | |
2311 | * cds_ja_del while holding the rcu read-side lock, and our node rules | |
2312 | * ensure that when a match value -> pointer is found in a node, it is | |
2313 | * _NEVER_ changed for that node without recompaction, and recompaction | |
2314 | * reallocates the node. | |
b306a0fe MD |
2315 | * However, when a child is removed from "linear" nodes, its pointer |
2316 | * is set to NULL. We therefore check, while holding the locks, if this | |
2317 | * pointer is NULL, and return -ENOENT to the caller if it is the case. | |
4cef6f97 MD |
2318 | * |
2319 | * ja_detach_node() ensures that a lookup will _never_ see a branch that | |
2320 | * leads to a dead-end: when removing branch, it makes sure to perform | |
2321 | * the "cut" at the highest node that has only one child, effectively | |
2322 | * replacing it with a NULL pointer. | |
af3cbd45 | 2323 | */ |
35170a44 MD |
2324 | static |
2325 | int ja_detach_node(struct cds_ja *ja, | |
2326 | struct cds_ja_inode_flag **snapshot, | |
af3cbd45 MD |
2327 | struct cds_ja_inode_flag ***snapshot_ptr, |
2328 | uint8_t *snapshot_n, | |
35170a44 MD |
2329 | int nr_snapshot, |
2330 | uint64_t key, | |
2331 | struct cds_ja_node *node) | |
2332 | { | |
af3cbd45 MD |
2333 | struct cds_ja_shadow_node *shadow_nodes[JA_MAX_DEPTH]; |
2334 | struct cds_ja_inode_flag **node_flag_ptr = NULL, | |
2335 | *parent_node_flag = NULL, | |
2336 | **parent_node_flag_ptr = NULL; | |
b62a8d0c | 2337 | struct cds_ja_inode_flag *iter_node_flag; |
4d6ef45e MD |
2338 | int ret, i, nr_shadow = 0, nr_clear = 0, nr_branch = 0; |
2339 | uint8_t n = 0; | |
35170a44 | 2340 | |
4d6ef45e | 2341 | assert(nr_snapshot == ja->tree_depth + 1); |
35170a44 | 2342 | |
af3cbd45 MD |
2343 | /* |
2344 | * From the last internal level node going up, get the node | |
2345 | * lock, check if the node has only one child left. If it is the | |
2346 | * case, we continue iterating upward. When we reach a node | |
2347 | * which has more that one child left, we lock the parent, and | |
2348 | * proceed to the node deletion (removing its children too). | |
2349 | */ | |
4d6ef45e | 2350 | for (i = nr_snapshot - 2; i >= 1; i--) { |
af3cbd45 MD |
2351 | struct cds_ja_shadow_node *shadow_node; |
2352 | ||
2353 | shadow_node = rcuja_shadow_lookup_lock(ja->ht, | |
3d8fe307 | 2354 | snapshot[i]); |
af3cbd45 MD |
2355 | if (!shadow_node) { |
2356 | ret = -EAGAIN; | |
2357 | goto end; | |
2358 | } | |
af3cbd45 | 2359 | shadow_nodes[nr_shadow++] = shadow_node; |
b62a8d0c MD |
2360 | |
2361 | /* | |
2362 | * Check if node has been removed between RCU | |
2363 | * lookup and lock acquisition. | |
2364 | */ | |
2365 | assert(snapshot_ptr[i + 1]); | |
2366 | if (ja_node_ptr(*snapshot_ptr[i + 1]) | |
2367 | != ja_node_ptr(snapshot[i + 1])) { | |
2368 | ret = -ENOENT; | |
2369 | goto end; | |
2370 | } | |
2371 | ||
2372 | assert(shadow_node->nr_child > 0); | |
d810c97f | 2373 | if (shadow_node->nr_child == 1 && i > 1) |
4d6ef45e MD |
2374 | nr_clear++; |
2375 | nr_branch++; | |
af3cbd45 MD |
2376 | if (shadow_node->nr_child > 1 || i == 1) { |
2377 | /* Lock parent and break */ | |
2378 | shadow_node = rcuja_shadow_lookup_lock(ja->ht, | |
3d8fe307 | 2379 | snapshot[i - 1]); |
af3cbd45 MD |
2380 | if (!shadow_node) { |
2381 | ret = -EAGAIN; | |
2382 | goto end; | |
2383 | } | |
2384 | shadow_nodes[nr_shadow++] = shadow_node; | |
b62a8d0c | 2385 | |
c112acaa MD |
2386 | /* |
2387 | * Check if node has been removed between RCU | |
2388 | * lookup and lock acquisition. | |
2389 | */ | |
b62a8d0c MD |
2390 | assert(snapshot_ptr[i]); |
2391 | if (ja_node_ptr(*snapshot_ptr[i]) | |
2392 | != ja_node_ptr(snapshot[i])) { | |
c112acaa MD |
2393 | ret = -ENOENT; |
2394 | goto end; | |
2395 | } | |
2396 | ||
b62a8d0c | 2397 | node_flag_ptr = snapshot_ptr[i + 1]; |
4d6ef45e MD |
2398 | n = snapshot_n[i + 1]; |
2399 | parent_node_flag_ptr = snapshot_ptr[i]; | |
2400 | parent_node_flag = snapshot[i]; | |
c112acaa | 2401 | |
af3cbd45 MD |
2402 | if (i > 1) { |
2403 | /* | |
2404 | * Lock parent's parent, in case we need | |
2405 | * to recompact parent. | |
2406 | */ | |
2407 | shadow_node = rcuja_shadow_lookup_lock(ja->ht, | |
3d8fe307 | 2408 | snapshot[i - 2]); |
af3cbd45 MD |
2409 | if (!shadow_node) { |
2410 | ret = -EAGAIN; | |
2411 | goto end; | |
2412 | } | |
2413 | shadow_nodes[nr_shadow++] = shadow_node; | |
b62a8d0c MD |
2414 | |
2415 | /* | |
2416 | * Check if node has been removed between RCU | |
2417 | * lookup and lock acquisition. | |
2418 | */ | |
2419 | assert(snapshot_ptr[i - 1]); | |
2420 | if (ja_node_ptr(*snapshot_ptr[i - 1]) | |
2421 | != ja_node_ptr(snapshot[i - 1])) { | |
2422 | ret = -ENOENT; | |
2423 | goto end; | |
2424 | } | |
af3cbd45 | 2425 | } |
b62a8d0c | 2426 | |
af3cbd45 MD |
2427 | break; |
2428 | } | |
2429 | } | |
2430 | ||
2431 | /* | |
4d6ef45e MD |
2432 | * At this point, we want to delete all nodes that are about to |
2433 | * be removed from shadow_nodes (except the last one, which is | |
2434 | * either the root or the parent of the upmost node with 1 | |
b62a8d0c MD |
2435 | * child). OK to free lock here, because RCU read lock is held, |
2436 | * and free only performed in call_rcu. | |
af3cbd45 MD |
2437 | */ |
2438 | ||
2439 | for (i = 0; i < nr_clear; i++) { | |
2440 | ret = rcuja_shadow_clear(ja->ht, | |
3d8fe307 | 2441 | shadow_nodes[i]->node_flag, |
af3cbd45 MD |
2442 | shadow_nodes[i], |
2443 | RCUJA_SHADOW_CLEAR_FREE_NODE | |
2444 | | RCUJA_SHADOW_CLEAR_FREE_LOCK); | |
2445 | assert(!ret); | |
2446 | } | |
2447 | ||
2448 | iter_node_flag = parent_node_flag; | |
2449 | /* Remove from parent */ | |
2450 | ret = ja_node_clear_ptr(ja, | |
2451 | node_flag_ptr, /* Pointer to location to nullify */ | |
2452 | &iter_node_flag, /* Old new parent ptr in its parent */ | |
4d6ef45e | 2453 | shadow_nodes[nr_branch - 1], /* of parent */ |
48cbe001 | 2454 | n, nr_branch - 1); |
b306a0fe MD |
2455 | if (ret) |
2456 | goto end; | |
af3cbd45 | 2457 | |
4d6ef45e MD |
2458 | dbg_printf("ja_detach_node: publish %p instead of %p\n", |
2459 | iter_node_flag, *parent_node_flag_ptr); | |
af3cbd45 MD |
2460 | /* Update address of parent ptr in its parent */ |
2461 | rcu_assign_pointer(*parent_node_flag_ptr, iter_node_flag); | |
2462 | ||
2463 | end: | |
2464 | for (i = 0; i < nr_shadow; i++) | |
2465 | rcuja_shadow_unlock(shadow_nodes[i]); | |
35170a44 MD |
2466 | return ret; |
2467 | } | |
2468 | ||
af3cbd45 MD |
2469 | static |
2470 | int ja_unchain_node(struct cds_ja *ja, | |
2471 | struct cds_ja_inode_flag *parent_node_flag, | |
fa112799 | 2472 | struct cds_ja_inode_flag **node_flag_ptr, |
013a6083 | 2473 | struct cds_ja_inode_flag *node_flag, |
af3cbd45 MD |
2474 | struct cds_ja_node *node) |
2475 | { | |
2476 | struct cds_ja_shadow_node *shadow_node; | |
03ec1aeb | 2477 | struct cds_ja_node *iter_node, **iter_node_ptr, **prev_node_ptr = NULL; |
013a6083 | 2478 | int ret = 0, count = 0, found = 0; |
af3cbd45 | 2479 | |
3d8fe307 | 2480 | shadow_node = rcuja_shadow_lookup_lock(ja->ht, parent_node_flag); |
af3cbd45 MD |
2481 | if (!shadow_node) |
2482 | return -EAGAIN; | |
013a6083 | 2483 | if (ja_node_ptr(*node_flag_ptr) != ja_node_ptr(node_flag)) { |
fa112799 MD |
2484 | ret = -EAGAIN; |
2485 | goto end; | |
2486 | } | |
af3cbd45 | 2487 | /* |
03ec1aeb MD |
2488 | * Find the previous node's next pointer pointing to our node, |
2489 | * so we can update it. Retry if another thread removed all but | |
2490 | * one of duplicates since check (this check was performed | |
2491 | * without lock). Ensure that the node we are about to remove is | |
2492 | * still in the list (while holding lock). No need for RCU | |
2493 | * traversal here since we hold the lock on the parent. | |
af3cbd45 | 2494 | */ |
03ec1aeb MD |
2495 | iter_node_ptr = (struct cds_ja_node **) node_flag_ptr; |
2496 | iter_node = (struct cds_ja_node *) ja_node_ptr(node_flag); | |
2497 | cds_ja_for_each_duplicate(iter_node) { | |
f2758d14 | 2498 | count++; |
03ec1aeb MD |
2499 | if (iter_node == node) { |
2500 | prev_node_ptr = iter_node_ptr; | |
013a6083 | 2501 | found++; |
03ec1aeb MD |
2502 | } |
2503 | iter_node_ptr = &iter_node->next; | |
f2758d14 | 2504 | } |
013a6083 MD |
2505 | assert(found <= 1); |
2506 | if (!found || count == 1) { | |
af3cbd45 MD |
2507 | ret = -EAGAIN; |
2508 | goto end; | |
2509 | } | |
03ec1aeb | 2510 | CMM_STORE_SHARED(*prev_node_ptr, node->next); |
ade342cb MD |
2511 | /* |
2512 | * Validate that we indeed removed the node from linked list. | |
2513 | */ | |
2514 | assert(ja_node_ptr(*node_flag_ptr) != (struct cds_ja_inode *) node); | |
af3cbd45 MD |
2515 | end: |
2516 | rcuja_shadow_unlock(shadow_node); | |
2517 | return ret; | |
2518 | } | |
2519 | ||
2520 | /* | |
2521 | * Called with RCU read lock held. | |
2522 | */ | |
35170a44 MD |
2523 | int cds_ja_del(struct cds_ja *ja, uint64_t key, |
2524 | struct cds_ja_node *node) | |
2525 | { | |
2526 | unsigned int tree_depth, i; | |
2527 | struct cds_ja_inode_flag *snapshot[JA_MAX_DEPTH]; | |
af3cbd45 MD |
2528 | struct cds_ja_inode_flag **snapshot_ptr[JA_MAX_DEPTH]; |
2529 | uint8_t snapshot_n[JA_MAX_DEPTH]; | |
35170a44 | 2530 | struct cds_ja_inode_flag *node_flag; |
fa112799 MD |
2531 | struct cds_ja_inode_flag **prev_node_flag_ptr, |
2532 | **node_flag_ptr; | |
4d6ef45e | 2533 | int nr_snapshot; |
35170a44 MD |
2534 | int ret; |
2535 | ||
2536 | if (caa_unlikely(key > ja->key_max)) | |
2537 | return -EINVAL; | |
2538 | tree_depth = ja->tree_depth; | |
2539 | ||
2540 | retry: | |
4d6ef45e | 2541 | nr_snapshot = 0; |
35170a44 MD |
2542 | dbg_printf("cds_ja_del attempt: key %" PRIu64 ", node %p\n", |
2543 | key, node); | |
2544 | ||
2545 | /* snapshot for level 0 is only for shadow node lookup */ | |
4d6ef45e MD |
2546 | snapshot_n[0] = 0; |
2547 | snapshot_n[1] = 0; | |
af3cbd45 | 2548 | snapshot_ptr[nr_snapshot] = NULL; |
35170a44 MD |
2549 | snapshot[nr_snapshot++] = (struct cds_ja_inode_flag *) &ja->root; |
2550 | node_flag = rcu_dereference(ja->root); | |
af3cbd45 | 2551 | prev_node_flag_ptr = &ja->root; |
fa112799 | 2552 | node_flag_ptr = &ja->root; |
35170a44 MD |
2553 | |
2554 | /* Iterate on all internal levels */ | |
2555 | for (i = 1; i < tree_depth; i++) { | |
2556 | uint8_t iter_key; | |
2557 | ||
2558 | dbg_printf("cds_ja_del iter node_flag %p\n", | |
2559 | node_flag); | |
2560 | if (!ja_node_ptr(node_flag)) { | |
2561 | return -ENOENT; | |
2562 | } | |
35170a44 | 2563 | iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (tree_depth - i - 1))); |
4d6ef45e | 2564 | snapshot_n[nr_snapshot + 1] = iter_key; |
af3cbd45 MD |
2565 | snapshot_ptr[nr_snapshot] = prev_node_flag_ptr; |
2566 | snapshot[nr_snapshot++] = node_flag; | |
35170a44 | 2567 | node_flag = ja_node_get_nth(node_flag, |
fa112799 | 2568 | &node_flag_ptr, |
35170a44 | 2569 | iter_key); |
48cbe001 MD |
2570 | if (node_flag) |
2571 | prev_node_flag_ptr = node_flag_ptr; | |
af3cbd45 MD |
2572 | dbg_printf("cds_ja_del iter key lookup %u finds node_flag %p, prev_node_flag_ptr %p\n", |
2573 | (unsigned int) iter_key, node_flag, | |
2574 | prev_node_flag_ptr); | |
35170a44 | 2575 | } |
35170a44 MD |
2576 | /* |
2577 | * We reached bottom of tree, try to find the node we are trying | |
2578 | * to remove. Fail if we cannot find it. | |
2579 | */ | |
2580 | if (!ja_node_ptr(node_flag)) { | |
4d6ef45e MD |
2581 | dbg_printf("cds_ja_del: no node found for key %" PRIu64 "\n", |
2582 | key); | |
35170a44 MD |
2583 | return -ENOENT; |
2584 | } else { | |
03ec1aeb | 2585 | struct cds_ja_node *iter_node, *match = NULL; |
af3cbd45 | 2586 | int count = 0; |
35170a44 | 2587 | |
03ec1aeb MD |
2588 | iter_node = (struct cds_ja_node *) ja_node_ptr(node_flag); |
2589 | cds_ja_for_each_duplicate_rcu(iter_node) { | |
2590 | dbg_printf("cds_ja_del: compare %p with iter_node %p\n", node, iter_node); | |
2591 | if (iter_node == node) | |
2592 | match = iter_node; | |
af3cbd45 | 2593 | count++; |
35170a44 | 2594 | } |
03ec1aeb | 2595 | |
4d6ef45e MD |
2596 | if (!match) { |
2597 | dbg_printf("cds_ja_del: no node match for node %p key %" PRIu64 "\n", node, key); | |
35170a44 | 2598 | return -ENOENT; |
4d6ef45e | 2599 | } |
af3cbd45 MD |
2600 | assert(count > 0); |
2601 | if (count == 1) { | |
2602 | /* | |
4d6ef45e MD |
2603 | * Removing last of duplicates. Last snapshot |
2604 | * does not have a shadow node (external leafs). | |
af3cbd45 MD |
2605 | */ |
2606 | snapshot_ptr[nr_snapshot] = prev_node_flag_ptr; | |
2607 | snapshot[nr_snapshot++] = node_flag; | |
2608 | ret = ja_detach_node(ja, snapshot, snapshot_ptr, | |
2609 | snapshot_n, nr_snapshot, key, node); | |
2610 | } else { | |
f2758d14 | 2611 | ret = ja_unchain_node(ja, snapshot[nr_snapshot - 1], |
013a6083 | 2612 | node_flag_ptr, node_flag, match); |
af3cbd45 | 2613 | } |
35170a44 | 2614 | } |
b306a0fe MD |
2615 | /* |
2616 | * Explanation of -ENOENT handling: caused by concurrent delete | |
2617 | * between RCU lookup and actual removal. Need to re-do the | |
2618 | * lookup and removal attempt. | |
2619 | */ | |
2620 | if (ret == -EAGAIN || ret == -ENOENT) | |
35170a44 MD |
2621 | goto retry; |
2622 | return ret; | |
2623 | } | |
2624 | ||
b4540e8a MD |
2625 | struct cds_ja *_cds_ja_new(unsigned int key_bits, |
2626 | const struct rcu_flavor_struct *flavor) | |
be9a7474 MD |
2627 | { |
2628 | struct cds_ja *ja; | |
b0f74e47 | 2629 | int ret; |
f07b240f | 2630 | struct cds_ja_shadow_node *root_shadow_node; |
be9a7474 MD |
2631 | |
2632 | ja = calloc(sizeof(*ja), 1); | |
2633 | if (!ja) | |
2634 | goto ja_error; | |
b4540e8a MD |
2635 | |
2636 | switch (key_bits) { | |
2637 | case 8: | |
b4540e8a | 2638 | case 16: |
1216b3d2 | 2639 | case 24: |
b4540e8a | 2640 | case 32: |
1216b3d2 MD |
2641 | case 40: |
2642 | case 48: | |
2643 | case 56: | |
2644 | ja->key_max = (1ULL << key_bits) - 1; | |
b4540e8a MD |
2645 | break; |
2646 | case 64: | |
2647 | ja->key_max = UINT64_MAX; | |
2648 | break; | |
2649 | default: | |
2650 | goto check_error; | |
2651 | } | |
2652 | ||
be9a7474 | 2653 | /* ja->root is NULL */ |
5a9a87dd | 2654 | /* tree_depth 0 is for pointer to root node */ |
582a6ade | 2655 | ja->tree_depth = (key_bits >> JA_LOG2_BITS_PER_BYTE) + 1; |
a2a7ff59 | 2656 | assert(ja->tree_depth <= JA_MAX_DEPTH); |
be9a7474 MD |
2657 | ja->ht = rcuja_create_ht(flavor); |
2658 | if (!ja->ht) | |
2659 | goto ht_error; | |
b0f74e47 MD |
2660 | |
2661 | /* | |
2662 | * Note: we should not free this node until judy array destroy. | |
2663 | */ | |
f07b240f | 2664 | root_shadow_node = rcuja_shadow_set(ja->ht, |
3d8fe307 | 2665 | (struct cds_ja_inode_flag *) &ja->root, |
48cbe001 | 2666 | NULL, ja, 0); |
f07b240f MD |
2667 | if (!root_shadow_node) { |
2668 | ret = -ENOMEM; | |
b0f74e47 | 2669 | goto ht_node_error; |
f07b240f | 2670 | } |
b0f74e47 | 2671 | |
be9a7474 MD |
2672 | return ja; |
2673 | ||
b0f74e47 MD |
2674 | ht_node_error: |
2675 | ret = rcuja_delete_ht(ja->ht); | |
2676 | assert(!ret); | |
be9a7474 | 2677 | ht_error: |
b4540e8a | 2678 | check_error: |
be9a7474 MD |
2679 | free(ja); |
2680 | ja_error: | |
2681 | return NULL; | |
2682 | } | |
2683 | ||
19ddcd04 | 2684 | static |
354981c2 | 2685 | void print_debug_fallback_distribution(struct cds_ja *ja) |
19ddcd04 MD |
2686 | { |
2687 | int i; | |
2688 | ||
2689 | fprintf(stderr, "Fallback node distribution:\n"); | |
2690 | for (i = 0; i < JA_ENTRY_PER_NODE; i++) { | |
354981c2 | 2691 | if (!ja->node_fallback_count_distribution[i]) |
19ddcd04 MD |
2692 | continue; |
2693 | fprintf(stderr, " %3u: %4lu\n", | |
354981c2 | 2694 | i, ja->node_fallback_count_distribution[i]); |
19ddcd04 MD |
2695 | } |
2696 | } | |
2697 | ||
021c72c0 | 2698 | static |
19a748d9 | 2699 | int ja_final_checks(struct cds_ja *ja) |
021c72c0 MD |
2700 | { |
2701 | double fallback_ratio; | |
2702 | unsigned long na, nf, nr_fallback; | |
19a748d9 | 2703 | int ret = 0; |
021c72c0 MD |
2704 | |
2705 | fallback_ratio = (double) uatomic_read(&ja->nr_fallback); | |
2706 | fallback_ratio /= (double) uatomic_read(&ja->nr_nodes_allocated); | |
2707 | nr_fallback = uatomic_read(&ja->nr_fallback); | |
2708 | if (nr_fallback) | |
2709 | fprintf(stderr, | |
2710 | "[warning] RCU Judy Array used %lu fallback node(s) (ratio: %g)\n", | |
2711 | uatomic_read(&ja->nr_fallback), | |
2712 | fallback_ratio); | |
2713 | ||
2714 | na = uatomic_read(&ja->nr_nodes_allocated); | |
2715 | nf = uatomic_read(&ja->nr_nodes_freed); | |
19a748d9 MD |
2716 | dbg_printf("Nodes allocated: %lu, Nodes freed: %lu.\n", na, nf); |
2717 | if (nr_fallback) | |
2718 | print_debug_fallback_distribution(ja); | |
2719 | ||
021c72c0 MD |
2720 | if (na != nf) { |
2721 | fprintf(stderr, "[error] Judy array leaked %ld nodes. Allocated: %lu, freed: %lu.\n", | |
2722 | (long) na - nf, na, nf); | |
19a748d9 | 2723 | ret = -1; |
021c72c0 | 2724 | } |
19a748d9 | 2725 | return ret; |
021c72c0 MD |
2726 | } |
2727 | ||
be9a7474 | 2728 | /* |
dc0e9798 MD |
2729 | * There should be no more concurrent add, delete, nor look-up performed |
2730 | * on the Judy array while it is being destroyed (ensured by the | |
2731 | * caller). | |
be9a7474 | 2732 | */ |
99e6e3dc | 2733 | int cds_ja_destroy(struct cds_ja *ja) |
be9a7474 | 2734 | { |
48cbe001 | 2735 | const struct rcu_flavor_struct *flavor; |
b4540e8a MD |
2736 | int ret; |
2737 | ||
48cbe001 | 2738 | flavor = cds_lfht_rcu_flavor(ja->ht); |
be9a7474 | 2739 | rcuja_shadow_prune(ja->ht, |
99e6e3dc | 2740 | RCUJA_SHADOW_CLEAR_FREE_NODE | RCUJA_SHADOW_CLEAR_FREE_LOCK); |
48cbe001 | 2741 | flavor->thread_offline(); |
b4540e8a MD |
2742 | ret = rcuja_delete_ht(ja->ht); |
2743 | if (ret) | |
2744 | return ret; | |
f2ae7af7 MD |
2745 | |
2746 | /* Wait for in-flight call_rcu free to complete. */ | |
2747 | flavor->barrier(); | |
2748 | ||
48cbe001 | 2749 | flavor->thread_online(); |
19a748d9 | 2750 | ret = ja_final_checks(ja); |
b4540e8a | 2751 | free(ja); |
19a748d9 | 2752 | return ret; |
be9a7474 | 2753 | } |