Commit | Line | Data |
---|---|---|
61009379 MD |
1 | /* |
2 | * rcuja/rcuja.c | |
3 | * | |
4 | * Userspace RCU library - RCU Judy Array | |
5 | * | |
6 | * Copyright 2012 - Mathieu Desnoyers <mathieu.desnoyers@efficios.com> | |
7 | * | |
8 | * This library is free software; you can redistribute it and/or | |
9 | * modify it under the terms of the GNU Lesser General Public | |
10 | * License as published by the Free Software Foundation; either | |
11 | * version 2.1 of the License, or (at your option) any later version. | |
12 | * | |
13 | * This library is distributed in the hope that it will be useful, | |
14 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
15 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | |
16 | * Lesser General Public License for more details. | |
17 | * | |
18 | * You should have received a copy of the GNU Lesser General Public | |
19 | * License along with this library; if not, write to the Free Software | |
20 | * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA | |
21 | */ | |
22 | ||
195e72d3 | 23 | #define _LGPL_SOURCE |
e5227865 | 24 | #include <stdint.h> |
8e519e3c | 25 | #include <errno.h> |
d68c6810 | 26 | #include <limits.h> |
b1a90ce3 | 27 | #include <string.h> |
61009379 | 28 | #include <urcu/rcuja.h> |
d68c6810 MD |
29 | #include <urcu/compiler.h> |
30 | #include <urcu/arch.h> | |
31 | #include <assert.h> | |
8e519e3c | 32 | #include <urcu-pointer.h> |
f07b240f | 33 | #include <urcu/uatomic.h> |
b4540e8a | 34 | #include <stdint.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, |
b023ba9f MD |
367 | int n, |
368 | enum ja_direction dir) | |
291b2543 MD |
369 | { |
370 | uint8_t nr_child; | |
371 | uint8_t *values; | |
372 | struct cds_ja_inode_flag **pointers; | |
373 | struct cds_ja_inode_flag *ptr; | |
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]; | |
392 | for (i = 0; i < nr_child; i++) { | |
393 | unsigned int v; | |
394 | ||
395 | v = CMM_LOAD_SHARED(values[i]); | |
b023ba9f MD |
396 | if (dir == JA_LEFT) { |
397 | if ((int) v < n && (int) v > match_v) { | |
398 | match_v = v; | |
399 | match_idx = i; | |
400 | } | |
401 | } else { | |
402 | if ((int) v > n && (int) v < match_v) { | |
403 | match_v = v; | |
404 | match_idx = i; | |
405 | } | |
291b2543 MD |
406 | } |
407 | } | |
b023ba9f MD |
408 | |
409 | if (match_idx < 0) { | |
291b2543 MD |
410 | return NULL; |
411 | } | |
b023ba9f MD |
412 | assert(match_v >= 0 && match_v < JA_ENTRY_PER_NODE); |
413 | ||
291b2543 MD |
414 | pointers = (struct cds_ja_inode_flag **) align_ptr_size(&values[type->max_linear_child]); |
415 | ptr = rcu_dereference(pointers[match_idx]); | |
416 | return ptr; | |
417 | } | |
418 | ||
11c5e016 | 419 | static |
5a9a87dd | 420 | void ja_linear_node_get_ith_pos(const struct cds_ja_type *type, |
b4540e8a | 421 | struct cds_ja_inode *node, |
11c5e016 MD |
422 | uint8_t i, |
423 | uint8_t *v, | |
b4540e8a | 424 | struct cds_ja_inode_flag **iter) |
11c5e016 MD |
425 | { |
426 | uint8_t *values; | |
b4540e8a | 427 | struct cds_ja_inode_flag **pointers; |
11c5e016 MD |
428 | |
429 | assert(type->type_class == RCU_JA_LINEAR || type->type_class == RCU_JA_POOL); | |
430 | assert(i < ja_linear_node_get_nr_child(type, node)); | |
431 | ||
432 | values = &node->u.data[1]; | |
433 | *v = values[i]; | |
b4540e8a | 434 | pointers = (struct cds_ja_inode_flag **) align_ptr_size(&values[type->max_linear_child]); |
11c5e016 MD |
435 | *iter = pointers[i]; |
436 | } | |
437 | ||
d68c6810 | 438 | static |
b4540e8a MD |
439 | struct cds_ja_inode_flag *ja_pool_node_get_nth(const struct cds_ja_type *type, |
440 | struct cds_ja_inode *node, | |
b1a90ce3 | 441 | struct cds_ja_inode_flag *node_flag, |
b0ca2d21 | 442 | struct cds_ja_inode_flag ***node_flag_ptr, |
8e519e3c | 443 | uint8_t n) |
d68c6810 | 444 | { |
b4540e8a | 445 | struct cds_ja_inode *linear; |
d68c6810 | 446 | |
fd800776 | 447 | assert(type->type_class == RCU_JA_POOL); |
b1a90ce3 MD |
448 | |
449 | switch (type->nr_pool_order) { | |
450 | case 1: | |
451 | { | |
452 | unsigned long bitsel, index; | |
453 | ||
454 | bitsel = ja_node_pool_1d_bitsel(node_flag); | |
455 | assert(bitsel < CHAR_BIT); | |
19ddcd04 | 456 | index = ((unsigned long) n >> bitsel) & 0x1; |
b1a90ce3 MD |
457 | linear = (struct cds_ja_inode *) &node->u.data[index << type->pool_size_order]; |
458 | break; | |
459 | } | |
460 | case 2: | |
461 | { | |
19ddcd04 MD |
462 | unsigned long bitsel[2], index[2], rindex; |
463 | ||
464 | ja_node_pool_2d_bitsel(node_flag, bitsel); | |
465 | assert(bitsel[0] < CHAR_BIT); | |
466 | assert(bitsel[1] < CHAR_BIT); | |
467 | index[0] = ((unsigned long) n >> bitsel[0]) & 0x1; | |
468 | index[0] <<= 1; | |
469 | index[1] = ((unsigned long) n >> bitsel[1]) & 0x1; | |
470 | rindex = index[0] | index[1]; | |
471 | linear = (struct cds_ja_inode *) &node->u.data[rindex << type->pool_size_order]; | |
b1a90ce3 MD |
472 | break; |
473 | } | |
474 | default: | |
475 | linear = NULL; | |
476 | assert(0); | |
477 | } | |
48cbe001 | 478 | return ja_linear_node_get_nth(type, linear, node_flag_ptr, n); |
d68c6810 MD |
479 | } |
480 | ||
11c5e016 | 481 | static |
b4540e8a MD |
482 | struct cds_ja_inode *ja_pool_node_get_ith_pool(const struct cds_ja_type *type, |
483 | struct cds_ja_inode *node, | |
11c5e016 MD |
484 | uint8_t i) |
485 | { | |
486 | assert(type->type_class == RCU_JA_POOL); | |
b4540e8a | 487 | return (struct cds_ja_inode *) |
11c5e016 MD |
488 | &node->u.data[(unsigned int) i << type->pool_size_order]; |
489 | } | |
490 | ||
291b2543 | 491 | static |
b023ba9f | 492 | struct cds_ja_inode_flag *ja_pool_node_get_direction(const struct cds_ja_type *type, |
291b2543 | 493 | struct cds_ja_inode *node, |
b023ba9f MD |
494 | int n, |
495 | enum ja_direction dir) | |
291b2543 MD |
496 | { |
497 | unsigned int pool_nr; | |
b023ba9f | 498 | int match_v; |
291b2543 MD |
499 | struct cds_ja_inode_flag *match_node_flag = NULL; |
500 | ||
501 | assert(type->type_class == RCU_JA_POOL); | |
b023ba9f MD |
502 | assert(dir == JA_LEFT || dir == JA_RIGHT); |
503 | ||
504 | if (dir == JA_LEFT) { | |
505 | match_v = -1; | |
506 | } else { | |
507 | match_v = JA_ENTRY_PER_NODE; | |
508 | } | |
291b2543 MD |
509 | |
510 | for (pool_nr = 0; pool_nr < (1U << type->nr_pool_order); pool_nr++) { | |
511 | struct cds_ja_inode *pool = | |
512 | ja_pool_node_get_ith_pool(type, | |
513 | node, pool_nr); | |
514 | uint8_t nr_child = | |
515 | ja_linear_node_get_nr_child(type, pool); | |
516 | unsigned int j; | |
517 | ||
518 | for (j = 0; j < nr_child; j++) { | |
519 | struct cds_ja_inode_flag *iter; | |
520 | uint8_t v; | |
521 | ||
522 | ja_linear_node_get_ith_pos(type, pool, | |
523 | j, &v, &iter); | |
524 | if (!iter) | |
525 | continue; | |
b023ba9f MD |
526 | if (dir == JA_LEFT) { |
527 | if ((int) v < n && (int) v > match_v) { | |
528 | match_v = v; | |
529 | match_node_flag = iter; | |
530 | } | |
531 | } else { | |
532 | if ((int) v > n && (int) v < match_v) { | |
533 | match_v = v; | |
534 | match_node_flag = iter; | |
535 | } | |
291b2543 MD |
536 | } |
537 | } | |
538 | } | |
539 | return match_node_flag; | |
540 | } | |
541 | ||
d68c6810 | 542 | static |
b4540e8a MD |
543 | struct cds_ja_inode_flag *ja_pigeon_node_get_nth(const struct cds_ja_type *type, |
544 | struct cds_ja_inode *node, | |
b0ca2d21 | 545 | struct cds_ja_inode_flag ***node_flag_ptr, |
8e519e3c | 546 | uint8_t n) |
d68c6810 | 547 | { |
48cbe001 MD |
548 | struct cds_ja_inode_flag **child_node_flag_ptr; |
549 | struct cds_ja_inode_flag *child_node_flag; | |
5a9a87dd | 550 | |
d68c6810 | 551 | assert(type->type_class == RCU_JA_PIGEON); |
48cbe001 MD |
552 | child_node_flag_ptr = &((struct cds_ja_inode_flag **) node->u.data)[n]; |
553 | child_node_flag = rcu_dereference(*child_node_flag_ptr); | |
582a6ade | 554 | dbg_printf("ja_pigeon_node_get_nth child_node_flag_ptr %p\n", |
48cbe001 | 555 | child_node_flag_ptr); |
b0ca2d21 | 556 | if (caa_unlikely(node_flag_ptr)) |
48cbe001 MD |
557 | *node_flag_ptr = child_node_flag_ptr; |
558 | return child_node_flag; | |
d68c6810 MD |
559 | } |
560 | ||
291b2543 | 561 | static |
b023ba9f | 562 | struct cds_ja_inode_flag *ja_pigeon_node_get_direction(const struct cds_ja_type *type, |
291b2543 | 563 | struct cds_ja_inode *node, |
b023ba9f MD |
564 | int n, |
565 | enum ja_direction dir) | |
291b2543 MD |
566 | { |
567 | struct cds_ja_inode_flag **child_node_flag_ptr; | |
568 | struct cds_ja_inode_flag *child_node_flag; | |
569 | int i; | |
570 | ||
571 | assert(type->type_class == RCU_JA_PIGEON); | |
b023ba9f MD |
572 | assert(dir == JA_LEFT || dir == JA_RIGHT); |
573 | ||
574 | if (dir == JA_LEFT) { | |
575 | /* n - 1 is first value left of n */ | |
576 | for (i = n - 1; i >= 0; i--) { | |
577 | child_node_flag_ptr = &((struct cds_ja_inode_flag **) node->u.data)[i]; | |
578 | child_node_flag = rcu_dereference(*child_node_flag_ptr); | |
579 | if (child_node_flag) { | |
580 | dbg_printf("ja_pigeon_node_get_left child_node_flag %p\n", | |
581 | child_node_flag); | |
582 | return child_node_flag; | |
583 | } | |
584 | } | |
585 | } else { | |
586 | /* n + 1 is first value right of n */ | |
587 | for (i = n + 1; i < JA_ENTRY_PER_NODE; i++) { | |
588 | child_node_flag_ptr = &((struct cds_ja_inode_flag **) node->u.data)[i]; | |
589 | child_node_flag = rcu_dereference(*child_node_flag_ptr); | |
590 | if (child_node_flag) { | |
591 | dbg_printf("ja_pigeon_node_get_right child_node_flag %p\n", | |
592 | child_node_flag); | |
593 | return child_node_flag; | |
594 | } | |
291b2543 MD |
595 | } |
596 | } | |
597 | return NULL; | |
598 | } | |
599 | ||
2e313670 MD |
600 | static |
601 | struct cds_ja_inode_flag *ja_pigeon_node_get_ith_pos(const struct cds_ja_type *type, | |
602 | struct cds_ja_inode *node, | |
603 | uint8_t i) | |
604 | { | |
48cbe001 | 605 | return ja_pigeon_node_get_nth(type, node, NULL, i); |
2e313670 MD |
606 | } |
607 | ||
13a7f5a6 MD |
608 | /* |
609 | * ja_node_get_nth: get nth item from a node. | |
610 | * node_flag is already rcu_dereference'd. | |
611 | */ | |
d68c6810 | 612 | static |
b62a8d0c | 613 | struct cds_ja_inode_flag *ja_node_get_nth(struct cds_ja_inode_flag *node_flag, |
b0ca2d21 | 614 | struct cds_ja_inode_flag ***node_flag_ptr, |
8e519e3c | 615 | uint8_t n) |
d68c6810 MD |
616 | { |
617 | unsigned int type_index; | |
b4540e8a | 618 | struct cds_ja_inode *node; |
d96bfb0d | 619 | const struct cds_ja_type *type; |
d68c6810 | 620 | |
d68c6810 | 621 | node = ja_node_ptr(node_flag); |
5a9a87dd | 622 | assert(node != NULL); |
d68c6810 MD |
623 | type_index = ja_node_type(node_flag); |
624 | type = &ja_types[type_index]; | |
625 | ||
626 | switch (type->type_class) { | |
627 | case RCU_JA_LINEAR: | |
5a9a87dd | 628 | return ja_linear_node_get_nth(type, node, |
b62a8d0c | 629 | node_flag_ptr, n); |
fd800776 | 630 | case RCU_JA_POOL: |
b1a90ce3 | 631 | return ja_pool_node_get_nth(type, node, node_flag, |
b62a8d0c | 632 | node_flag_ptr, n); |
d68c6810 | 633 | case RCU_JA_PIGEON: |
5a9a87dd | 634 | return ja_pigeon_node_get_nth(type, node, |
b62a8d0c | 635 | node_flag_ptr, n); |
d68c6810 MD |
636 | default: |
637 | assert(0); | |
638 | return (void *) -1UL; | |
639 | } | |
640 | } | |
641 | ||
291b2543 | 642 | static |
b023ba9f MD |
643 | struct cds_ja_inode_flag *ja_node_get_direction(struct cds_ja_inode_flag *node_flag, |
644 | int n, | |
645 | enum ja_direction dir) | |
291b2543 MD |
646 | { |
647 | unsigned int type_index; | |
648 | struct cds_ja_inode *node; | |
649 | const struct cds_ja_type *type; | |
650 | ||
651 | node = ja_node_ptr(node_flag); | |
652 | assert(node != NULL); | |
653 | type_index = ja_node_type(node_flag); | |
654 | type = &ja_types[type_index]; | |
655 | ||
656 | switch (type->type_class) { | |
657 | case RCU_JA_LINEAR: | |
b023ba9f | 658 | return ja_linear_node_get_direction(type, node, n, dir); |
291b2543 | 659 | case RCU_JA_POOL: |
b023ba9f | 660 | return ja_pool_node_get_direction(type, node, n, dir); |
291b2543 | 661 | case RCU_JA_PIGEON: |
b023ba9f | 662 | return ja_pigeon_node_get_direction(type, node, n, dir); |
291b2543 MD |
663 | default: |
664 | assert(0); | |
665 | return (void *) -1UL; | |
666 | } | |
667 | } | |
668 | ||
669 | static | |
b023ba9f MD |
670 | struct cds_ja_inode_flag *ja_node_get_leftright(struct cds_ja_inode_flag *node_flag, |
671 | unsigned int n, | |
672 | enum ja_direction dir) | |
291b2543 | 673 | { |
b023ba9f MD |
674 | return ja_node_get_direction(node_flag, n, dir); |
675 | } | |
676 | ||
677 | static | |
678 | struct cds_ja_inode_flag *ja_node_get_minmax(struct cds_ja_inode_flag *node_flag, | |
679 | enum ja_direction dir) | |
680 | { | |
681 | switch (dir) { | |
682 | case JA_LEFTMOST: | |
683 | return ja_node_get_direction(node_flag, | |
684 | -1, JA_RIGHT); | |
685 | case JA_RIGHTMOST: | |
686 | return ja_node_get_direction(node_flag, | |
687 | JA_ENTRY_PER_NODE, JA_LEFT); | |
688 | default: | |
689 | assert(0); | |
690 | } | |
291b2543 MD |
691 | } |
692 | ||
8e519e3c | 693 | static |
d96bfb0d | 694 | int ja_linear_node_set_nth(const struct cds_ja_type *type, |
b4540e8a | 695 | struct cds_ja_inode *node, |
d96bfb0d | 696 | struct cds_ja_shadow_node *shadow_node, |
8e519e3c | 697 | uint8_t n, |
b4540e8a | 698 | struct cds_ja_inode_flag *child_node_flag) |
8e519e3c MD |
699 | { |
700 | uint8_t nr_child; | |
701 | uint8_t *values, *nr_child_ptr; | |
b4540e8a | 702 | struct cds_ja_inode_flag **pointers; |
2e313670 | 703 | unsigned int i, unused = 0; |
8e519e3c MD |
704 | |
705 | assert(type->type_class == RCU_JA_LINEAR || type->type_class == RCU_JA_POOL); | |
706 | ||
707 | nr_child_ptr = &node->u.data[0]; | |
48cbe001 MD |
708 | dbg_printf("linear set nth: n %u, nr_child_ptr %p\n", |
709 | (unsigned int) n, nr_child_ptr); | |
8e519e3c MD |
710 | nr_child = *nr_child_ptr; |
711 | assert(nr_child <= type->max_linear_child); | |
8e519e3c MD |
712 | |
713 | values = &node->u.data[1]; | |
2e313670 MD |
714 | pointers = (struct cds_ja_inode_flag **) align_ptr_size(&values[type->max_linear_child]); |
715 | /* Check if node value is already populated */ | |
8e519e3c | 716 | for (i = 0; i < nr_child; i++) { |
2e313670 MD |
717 | if (values[i] == n) { |
718 | if (pointers[i]) | |
719 | return -EEXIST; | |
720 | else | |
721 | break; | |
722 | } else { | |
723 | if (!pointers[i]) | |
724 | unused++; | |
725 | } | |
8e519e3c | 726 | } |
2e313670 MD |
727 | if (i == nr_child && nr_child >= type->max_linear_child) { |
728 | if (unused) | |
729 | return -ERANGE; /* recompact node */ | |
730 | else | |
731 | return -ENOSPC; /* No space left in this node type */ | |
732 | } | |
733 | ||
734 | assert(pointers[i] == NULL); | |
735 | rcu_assign_pointer(pointers[i], child_node_flag); | |
736 | /* If we expanded the nr_child, increment it */ | |
737 | if (i == nr_child) { | |
738 | CMM_STORE_SHARED(values[nr_child], n); | |
739 | /* write pointer and value before nr_child */ | |
740 | cmm_smp_wmb(); | |
741 | CMM_STORE_SHARED(*nr_child_ptr, nr_child + 1); | |
8e519e3c | 742 | } |
e1db2db5 | 743 | shadow_node->nr_child++; |
a2a7ff59 MD |
744 | dbg_printf("linear set nth: %u child, shadow: %u child, for node %p shadow %p\n", |
745 | (unsigned int) CMM_LOAD_SHARED(*nr_child_ptr), | |
746 | (unsigned int) shadow_node->nr_child, | |
747 | node, shadow_node); | |
748 | ||
8e519e3c MD |
749 | return 0; |
750 | } | |
751 | ||
752 | static | |
d96bfb0d | 753 | int ja_pool_node_set_nth(const struct cds_ja_type *type, |
b4540e8a | 754 | struct cds_ja_inode *node, |
b1a90ce3 | 755 | struct cds_ja_inode_flag *node_flag, |
d96bfb0d | 756 | struct cds_ja_shadow_node *shadow_node, |
8e519e3c | 757 | uint8_t n, |
b4540e8a | 758 | struct cds_ja_inode_flag *child_node_flag) |
8e519e3c | 759 | { |
b4540e8a | 760 | struct cds_ja_inode *linear; |
8e519e3c MD |
761 | |
762 | assert(type->type_class == RCU_JA_POOL); | |
b1a90ce3 MD |
763 | |
764 | switch (type->nr_pool_order) { | |
765 | case 1: | |
766 | { | |
767 | unsigned long bitsel, index; | |
768 | ||
769 | bitsel = ja_node_pool_1d_bitsel(node_flag); | |
770 | assert(bitsel < CHAR_BIT); | |
19ddcd04 | 771 | index = ((unsigned long) n >> bitsel) & 0x1; |
b1a90ce3 MD |
772 | linear = (struct cds_ja_inode *) &node->u.data[index << type->pool_size_order]; |
773 | break; | |
774 | } | |
775 | case 2: | |
776 | { | |
19ddcd04 MD |
777 | unsigned long bitsel[2], index[2], rindex; |
778 | ||
779 | ja_node_pool_2d_bitsel(node_flag, bitsel); | |
780 | assert(bitsel[0] < CHAR_BIT); | |
781 | assert(bitsel[1] < CHAR_BIT); | |
782 | index[0] = ((unsigned long) n >> bitsel[0]) & 0x1; | |
783 | index[0] <<= 1; | |
784 | index[1] = ((unsigned long) n >> bitsel[1]) & 0x1; | |
785 | rindex = index[0] | index[1]; | |
786 | linear = (struct cds_ja_inode *) &node->u.data[rindex << type->pool_size_order]; | |
b1a90ce3 MD |
787 | break; |
788 | } | |
789 | default: | |
790 | linear = NULL; | |
791 | assert(0); | |
792 | } | |
793 | ||
e1db2db5 MD |
794 | return ja_linear_node_set_nth(type, linear, shadow_node, |
795 | n, child_node_flag); | |
8e519e3c MD |
796 | } |
797 | ||
798 | static | |
d96bfb0d | 799 | int ja_pigeon_node_set_nth(const struct cds_ja_type *type, |
b4540e8a | 800 | struct cds_ja_inode *node, |
d96bfb0d | 801 | struct cds_ja_shadow_node *shadow_node, |
8e519e3c | 802 | uint8_t n, |
b4540e8a | 803 | struct cds_ja_inode_flag *child_node_flag) |
8e519e3c | 804 | { |
b4540e8a | 805 | struct cds_ja_inode_flag **ptr; |
8e519e3c MD |
806 | |
807 | assert(type->type_class == RCU_JA_PIGEON); | |
b4540e8a | 808 | ptr = &((struct cds_ja_inode_flag **) node->u.data)[n]; |
5a9a87dd | 809 | if (*ptr) |
8e519e3c MD |
810 | return -EEXIST; |
811 | rcu_assign_pointer(*ptr, child_node_flag); | |
e1db2db5 | 812 | shadow_node->nr_child++; |
8e519e3c MD |
813 | return 0; |
814 | } | |
815 | ||
d68c6810 | 816 | /* |
7a0b2331 | 817 | * _ja_node_set_nth: set nth item within a node. Return an error |
8e519e3c | 818 | * (negative error value) if it is already there. |
d68c6810 | 819 | */ |
8e519e3c | 820 | static |
d96bfb0d | 821 | int _ja_node_set_nth(const struct cds_ja_type *type, |
b4540e8a | 822 | struct cds_ja_inode *node, |
b1a90ce3 | 823 | struct cds_ja_inode_flag *node_flag, |
d96bfb0d | 824 | struct cds_ja_shadow_node *shadow_node, |
e1db2db5 | 825 | uint8_t n, |
b4540e8a | 826 | struct cds_ja_inode_flag *child_node_flag) |
8e519e3c | 827 | { |
8e519e3c MD |
828 | switch (type->type_class) { |
829 | case RCU_JA_LINEAR: | |
e1db2db5 | 830 | return ja_linear_node_set_nth(type, node, shadow_node, n, |
8e519e3c MD |
831 | child_node_flag); |
832 | case RCU_JA_POOL: | |
b1a90ce3 | 833 | return ja_pool_node_set_nth(type, node, node_flag, shadow_node, n, |
8e519e3c MD |
834 | child_node_flag); |
835 | case RCU_JA_PIGEON: | |
e1db2db5 | 836 | return ja_pigeon_node_set_nth(type, node, shadow_node, n, |
8e519e3c | 837 | child_node_flag); |
e1db2db5 MD |
838 | case RCU_JA_NULL: |
839 | return -ENOSPC; | |
8e519e3c MD |
840 | default: |
841 | assert(0); | |
842 | return -EINVAL; | |
843 | } | |
844 | ||
845 | return 0; | |
846 | } | |
7a0b2331 | 847 | |
2e313670 | 848 | static |
af3cbd45 | 849 | int ja_linear_node_clear_ptr(const struct cds_ja_type *type, |
2e313670 MD |
850 | struct cds_ja_inode *node, |
851 | struct cds_ja_shadow_node *shadow_node, | |
af3cbd45 | 852 | struct cds_ja_inode_flag **node_flag_ptr) |
2e313670 MD |
853 | { |
854 | uint8_t nr_child; | |
af3cbd45 | 855 | uint8_t *nr_child_ptr; |
2e313670 MD |
856 | |
857 | assert(type->type_class == RCU_JA_LINEAR || type->type_class == RCU_JA_POOL); | |
858 | ||
859 | nr_child_ptr = &node->u.data[0]; | |
2e313670 MD |
860 | nr_child = *nr_child_ptr; |
861 | assert(nr_child <= type->max_linear_child); | |
862 | ||
48cbe001 MD |
863 | if (type->type_class == RCU_JA_LINEAR) { |
864 | assert(!shadow_node->fallback_removal_count); | |
865 | if (shadow_node->nr_child <= type->min_child) { | |
2e313670 MD |
866 | /* We need to try recompacting the node */ |
867 | return -EFBIG; | |
868 | } | |
869 | } | |
19ddcd04 | 870 | dbg_printf("linear clear ptr: nr_child_ptr %p\n", nr_child_ptr); |
af3cbd45 MD |
871 | assert(*node_flag_ptr != NULL); |
872 | rcu_assign_pointer(*node_flag_ptr, NULL); | |
2e313670 MD |
873 | /* |
874 | * Value and nr_child are never changed (would cause ABA issue). | |
875 | * Instead, we leave the pointer to NULL and recompact the node | |
876 | * once in a while. It is allowed to set a NULL pointer to a new | |
877 | * value without recompaction though. | |
878 | * Only update the shadow node accounting. | |
879 | */ | |
880 | shadow_node->nr_child--; | |
af3cbd45 | 881 | dbg_printf("linear clear ptr: %u child, shadow: %u child, for node %p shadow %p\n", |
2e313670 MD |
882 | (unsigned int) CMM_LOAD_SHARED(*nr_child_ptr), |
883 | (unsigned int) shadow_node->nr_child, | |
884 | node, shadow_node); | |
2e313670 MD |
885 | return 0; |
886 | } | |
887 | ||
888 | static | |
af3cbd45 | 889 | int ja_pool_node_clear_ptr(const struct cds_ja_type *type, |
2e313670 | 890 | struct cds_ja_inode *node, |
19ddcd04 | 891 | struct cds_ja_inode_flag *node_flag, |
2e313670 | 892 | struct cds_ja_shadow_node *shadow_node, |
af3cbd45 | 893 | struct cds_ja_inode_flag **node_flag_ptr, |
2e313670 MD |
894 | uint8_t n) |
895 | { | |
896 | struct cds_ja_inode *linear; | |
897 | ||
898 | assert(type->type_class == RCU_JA_POOL); | |
19ddcd04 MD |
899 | |
900 | if (shadow_node->fallback_removal_count) { | |
901 | shadow_node->fallback_removal_count--; | |
902 | } else { | |
903 | /* We should try recompacting the node */ | |
904 | if (shadow_node->nr_child <= type->min_child) | |
905 | return -EFBIG; | |
906 | } | |
907 | ||
908 | switch (type->nr_pool_order) { | |
909 | case 1: | |
910 | { | |
911 | unsigned long bitsel, index; | |
912 | ||
913 | bitsel = ja_node_pool_1d_bitsel(node_flag); | |
914 | assert(bitsel < CHAR_BIT); | |
915 | index = ((unsigned long) n >> bitsel) & type->nr_pool_order; | |
916 | linear = (struct cds_ja_inode *) &node->u.data[index << type->pool_size_order]; | |
917 | break; | |
918 | } | |
919 | case 2: | |
920 | { | |
921 | unsigned long bitsel[2], index[2], rindex; | |
922 | ||
923 | ja_node_pool_2d_bitsel(node_flag, bitsel); | |
924 | assert(bitsel[0] < CHAR_BIT); | |
925 | assert(bitsel[1] < CHAR_BIT); | |
926 | index[0] = ((unsigned long) n >> bitsel[0]) & 0x1; | |
927 | index[0] <<= 1; | |
928 | index[1] = ((unsigned long) n >> bitsel[1]) & 0x1; | |
929 | rindex = index[0] | index[1]; | |
930 | linear = (struct cds_ja_inode *) &node->u.data[rindex << type->pool_size_order]; | |
931 | break; | |
932 | } | |
933 | default: | |
934 | linear = NULL; | |
935 | assert(0); | |
936 | } | |
937 | ||
af3cbd45 | 938 | return ja_linear_node_clear_ptr(type, linear, shadow_node, node_flag_ptr); |
2e313670 MD |
939 | } |
940 | ||
941 | static | |
af3cbd45 | 942 | int ja_pigeon_node_clear_ptr(const struct cds_ja_type *type, |
2e313670 MD |
943 | struct cds_ja_inode *node, |
944 | struct cds_ja_shadow_node *shadow_node, | |
af3cbd45 | 945 | struct cds_ja_inode_flag **node_flag_ptr) |
2e313670 | 946 | { |
2e313670 | 947 | assert(type->type_class == RCU_JA_PIGEON); |
19ddcd04 MD |
948 | |
949 | if (shadow_node->fallback_removal_count) { | |
950 | shadow_node->fallback_removal_count--; | |
951 | } else { | |
952 | /* We should try recompacting the node */ | |
953 | if (shadow_node->nr_child <= type->min_child) | |
954 | return -EFBIG; | |
955 | } | |
4d6ef45e | 956 | dbg_printf("ja_pigeon_node_clear_ptr: clearing ptr: %p\n", *node_flag_ptr); |
af3cbd45 | 957 | rcu_assign_pointer(*node_flag_ptr, NULL); |
2e313670 MD |
958 | shadow_node->nr_child--; |
959 | return 0; | |
960 | } | |
961 | ||
962 | /* | |
af3cbd45 | 963 | * _ja_node_clear_ptr: clear ptr item within a node. Return an error |
2e313670 MD |
964 | * (negative error value) if it is not found (-ENOENT). |
965 | */ | |
966 | static | |
af3cbd45 | 967 | int _ja_node_clear_ptr(const struct cds_ja_type *type, |
2e313670 | 968 | struct cds_ja_inode *node, |
19ddcd04 | 969 | struct cds_ja_inode_flag *node_flag, |
2e313670 | 970 | struct cds_ja_shadow_node *shadow_node, |
af3cbd45 | 971 | struct cds_ja_inode_flag **node_flag_ptr, |
2e313670 MD |
972 | uint8_t n) |
973 | { | |
974 | switch (type->type_class) { | |
975 | case RCU_JA_LINEAR: | |
af3cbd45 | 976 | return ja_linear_node_clear_ptr(type, node, shadow_node, node_flag_ptr); |
2e313670 | 977 | case RCU_JA_POOL: |
19ddcd04 | 978 | return ja_pool_node_clear_ptr(type, node, node_flag, shadow_node, node_flag_ptr, n); |
2e313670 | 979 | case RCU_JA_PIGEON: |
af3cbd45 | 980 | return ja_pigeon_node_clear_ptr(type, node, shadow_node, node_flag_ptr); |
2e313670 MD |
981 | case RCU_JA_NULL: |
982 | return -ENOENT; | |
983 | default: | |
984 | assert(0); | |
985 | return -EINVAL; | |
986 | } | |
987 | ||
988 | return 0; | |
989 | } | |
990 | ||
b1a90ce3 MD |
991 | /* |
992 | * Calculate bit distribution. Returns the bit (0 to 7) that splits the | |
993 | * distribution in two sub-distributions containing as much elements one | |
994 | * compared to the other. | |
995 | */ | |
996 | static | |
997 | unsigned int ja_node_sum_distribution_1d(enum ja_recompact mode, | |
998 | struct cds_ja *ja, | |
999 | unsigned int type_index, | |
1000 | const struct cds_ja_type *type, | |
1001 | struct cds_ja_inode *node, | |
1002 | struct cds_ja_shadow_node *shadow_node, | |
1003 | uint8_t n, | |
1004 | struct cds_ja_inode_flag *child_node_flag, | |
1005 | struct cds_ja_inode_flag **nullify_node_flag_ptr) | |
1006 | { | |
1007 | uint8_t nr_one[JA_BITS_PER_BYTE]; | |
1008 | unsigned int bitsel = 0, bit_i, overall_best_distance = UINT_MAX; | |
1009 | unsigned int distrib_nr_child = 0; | |
1010 | ||
1011 | memset(nr_one, 0, sizeof(nr_one)); | |
1012 | ||
1013 | switch (type->type_class) { | |
1014 | case RCU_JA_LINEAR: | |
1015 | { | |
1016 | uint8_t nr_child = | |
1017 | ja_linear_node_get_nr_child(type, node); | |
1018 | unsigned int i; | |
1019 | ||
1020 | for (i = 0; i < nr_child; i++) { | |
1021 | struct cds_ja_inode_flag *iter; | |
b1a90ce3 MD |
1022 | uint8_t v; |
1023 | ||
1024 | ja_linear_node_get_ith_pos(type, node, i, &v, &iter); | |
1025 | if (!iter) | |
1026 | continue; | |
1027 | if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) | |
1028 | continue; | |
f5531dd9 MD |
1029 | for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) { |
1030 | if (v & (1U << bit_i)) | |
1031 | nr_one[bit_i]++; | |
b1a90ce3 MD |
1032 | } |
1033 | distrib_nr_child++; | |
1034 | } | |
1035 | break; | |
1036 | } | |
1037 | case RCU_JA_POOL: | |
1038 | { | |
1039 | unsigned int pool_nr; | |
1040 | ||
1041 | for (pool_nr = 0; pool_nr < (1U << type->nr_pool_order); pool_nr++) { | |
1042 | struct cds_ja_inode *pool = | |
1043 | ja_pool_node_get_ith_pool(type, | |
1044 | node, pool_nr); | |
1045 | uint8_t nr_child = | |
1046 | ja_linear_node_get_nr_child(type, pool); | |
1047 | unsigned int j; | |
1048 | ||
1049 | for (j = 0; j < nr_child; j++) { | |
1050 | struct cds_ja_inode_flag *iter; | |
b1a90ce3 MD |
1051 | uint8_t v; |
1052 | ||
1053 | ja_linear_node_get_ith_pos(type, pool, | |
1054 | j, &v, &iter); | |
1055 | if (!iter) | |
1056 | continue; | |
1057 | if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) | |
1058 | continue; | |
f5531dd9 MD |
1059 | for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) { |
1060 | if (v & (1U << bit_i)) | |
1061 | nr_one[bit_i]++; | |
b1a90ce3 MD |
1062 | } |
1063 | distrib_nr_child++; | |
1064 | } | |
1065 | } | |
1066 | break; | |
1067 | } | |
1068 | case RCU_JA_PIGEON: | |
1069 | { | |
b1a90ce3 MD |
1070 | unsigned int i; |
1071 | ||
1072 | assert(mode == JA_RECOMPACT_DEL); | |
48cbe001 | 1073 | for (i = 0; i < JA_ENTRY_PER_NODE; i++) { |
b1a90ce3 | 1074 | struct cds_ja_inode_flag *iter; |
b1a90ce3 MD |
1075 | |
1076 | iter = ja_pigeon_node_get_ith_pos(type, node, i); | |
1077 | if (!iter) | |
1078 | continue; | |
1079 | if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) | |
1080 | continue; | |
f5531dd9 MD |
1081 | for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) { |
1082 | if (i & (1U << bit_i)) | |
1083 | nr_one[bit_i]++; | |
b1a90ce3 MD |
1084 | } |
1085 | distrib_nr_child++; | |
1086 | } | |
1087 | break; | |
1088 | } | |
1089 | case RCU_JA_NULL: | |
19ddcd04 | 1090 | assert(mode == JA_RECOMPACT_ADD_NEXT); |
b1a90ce3 MD |
1091 | break; |
1092 | default: | |
1093 | assert(0); | |
1094 | break; | |
1095 | } | |
1096 | ||
19ddcd04 | 1097 | if (mode == JA_RECOMPACT_ADD_NEXT || mode == JA_RECOMPACT_ADD_SAME) { |
f5531dd9 MD |
1098 | for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) { |
1099 | if (n & (1U << bit_i)) | |
1100 | nr_one[bit_i]++; | |
b1a90ce3 MD |
1101 | } |
1102 | distrib_nr_child++; | |
1103 | } | |
1104 | ||
1105 | /* | |
1106 | * The best bit selector is that for which the number of ones is | |
1107 | * closest to half of the number of children in the | |
f5531dd9 MD |
1108 | * distribution. We calculate the distance using the double of |
1109 | * the sub-distribution sizes to eliminate truncation error. | |
b1a90ce3 MD |
1110 | */ |
1111 | for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) { | |
1112 | unsigned int distance_to_best; | |
1113 | ||
1b34283b | 1114 | distance_to_best = abs_int(((unsigned int) nr_one[bit_i] << 1U) - distrib_nr_child); |
b1a90ce3 MD |
1115 | if (distance_to_best < overall_best_distance) { |
1116 | overall_best_distance = distance_to_best; | |
1117 | bitsel = bit_i; | |
1118 | } | |
1119 | } | |
1120 | dbg_printf("1 dimension pool bit selection: (%u)\n", bitsel); | |
1121 | return bitsel; | |
1122 | } | |
1123 | ||
19ddcd04 MD |
1124 | /* |
1125 | * Calculate bit distribution in two dimensions. Returns the two bits | |
1126 | * (each 0 to 7) that splits the distribution in four sub-distributions | |
1127 | * containing as much elements one compared to the other. | |
1128 | */ | |
1129 | static | |
1130 | void ja_node_sum_distribution_2d(enum ja_recompact mode, | |
1131 | struct cds_ja *ja, | |
1132 | unsigned int type_index, | |
1133 | const struct cds_ja_type *type, | |
1134 | struct cds_ja_inode *node, | |
1135 | struct cds_ja_shadow_node *shadow_node, | |
1136 | uint8_t n, | |
1137 | struct cds_ja_inode_flag *child_node_flag, | |
1138 | struct cds_ja_inode_flag **nullify_node_flag_ptr, | |
1139 | unsigned int *_bitsel) | |
1140 | { | |
1141 | uint8_t nr_2d_11[JA_BITS_PER_BYTE][JA_BITS_PER_BYTE], | |
1142 | nr_2d_10[JA_BITS_PER_BYTE][JA_BITS_PER_BYTE], | |
1143 | nr_2d_01[JA_BITS_PER_BYTE][JA_BITS_PER_BYTE], | |
1144 | nr_2d_00[JA_BITS_PER_BYTE][JA_BITS_PER_BYTE]; | |
1145 | unsigned int bitsel[2] = { 0, 1 }; | |
4a073c53 MD |
1146 | unsigned int bit_i, bit_j; |
1147 | int overall_best_distance = INT_MAX; | |
19ddcd04 MD |
1148 | unsigned int distrib_nr_child = 0; |
1149 | ||
1150 | memset(nr_2d_11, 0, sizeof(nr_2d_11)); | |
1151 | memset(nr_2d_10, 0, sizeof(nr_2d_10)); | |
4a073c53 MD |
1152 | memset(nr_2d_01, 0, sizeof(nr_2d_01)); |
1153 | memset(nr_2d_00, 0, sizeof(nr_2d_00)); | |
19ddcd04 MD |
1154 | |
1155 | switch (type->type_class) { | |
1156 | case RCU_JA_LINEAR: | |
1157 | { | |
1158 | uint8_t nr_child = | |
1159 | ja_linear_node_get_nr_child(type, node); | |
1160 | unsigned int i; | |
1161 | ||
1162 | for (i = 0; i < nr_child; i++) { | |
1163 | struct cds_ja_inode_flag *iter; | |
1164 | uint8_t v; | |
1165 | ||
1166 | ja_linear_node_get_ith_pos(type, node, i, &v, &iter); | |
1167 | if (!iter) | |
1168 | continue; | |
1169 | if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) | |
1170 | continue; | |
1171 | for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) { | |
1172 | for (bit_j = 0; bit_j < bit_i; bit_j++) { | |
7f14b43a MD |
1173 | if (v & (1U << bit_i)) { |
1174 | if (v & (1U << bit_j)) { | |
1175 | nr_2d_11[bit_i][bit_j]++; | |
1176 | } else { | |
1177 | nr_2d_10[bit_i][bit_j]++; | |
1178 | } | |
1179 | } else { | |
1180 | if (v & (1U << bit_j)) { | |
1181 | nr_2d_01[bit_i][bit_j]++; | |
1182 | } else { | |
1183 | nr_2d_00[bit_i][bit_j]++; | |
1184 | } | |
19ddcd04 MD |
1185 | } |
1186 | } | |
1187 | } | |
1188 | distrib_nr_child++; | |
1189 | } | |
1190 | break; | |
1191 | } | |
1192 | case RCU_JA_POOL: | |
1193 | { | |
1194 | unsigned int pool_nr; | |
1195 | ||
1196 | for (pool_nr = 0; pool_nr < (1U << type->nr_pool_order); pool_nr++) { | |
1197 | struct cds_ja_inode *pool = | |
1198 | ja_pool_node_get_ith_pool(type, | |
1199 | node, pool_nr); | |
1200 | uint8_t nr_child = | |
1201 | ja_linear_node_get_nr_child(type, pool); | |
1202 | unsigned int j; | |
1203 | ||
1204 | for (j = 0; j < nr_child; j++) { | |
1205 | struct cds_ja_inode_flag *iter; | |
1206 | uint8_t v; | |
1207 | ||
1208 | ja_linear_node_get_ith_pos(type, pool, | |
1209 | j, &v, &iter); | |
1210 | if (!iter) | |
1211 | continue; | |
1212 | if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) | |
1213 | continue; | |
1214 | for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) { | |
1215 | for (bit_j = 0; bit_j < bit_i; bit_j++) { | |
7f14b43a MD |
1216 | if (v & (1U << bit_i)) { |
1217 | if (v & (1U << bit_j)) { | |
1218 | nr_2d_11[bit_i][bit_j]++; | |
1219 | } else { | |
1220 | nr_2d_10[bit_i][bit_j]++; | |
1221 | } | |
1222 | } else { | |
1223 | if (v & (1U << bit_j)) { | |
1224 | nr_2d_01[bit_i][bit_j]++; | |
1225 | } else { | |
1226 | nr_2d_00[bit_i][bit_j]++; | |
1227 | } | |
19ddcd04 MD |
1228 | } |
1229 | } | |
1230 | } | |
1231 | distrib_nr_child++; | |
1232 | } | |
1233 | } | |
1234 | break; | |
1235 | } | |
1236 | case RCU_JA_PIGEON: | |
1237 | { | |
19ddcd04 MD |
1238 | unsigned int i; |
1239 | ||
1240 | assert(mode == JA_RECOMPACT_DEL); | |
48cbe001 | 1241 | for (i = 0; i < JA_ENTRY_PER_NODE; i++) { |
19ddcd04 MD |
1242 | struct cds_ja_inode_flag *iter; |
1243 | ||
1244 | iter = ja_pigeon_node_get_ith_pos(type, node, i); | |
1245 | if (!iter) | |
1246 | continue; | |
1247 | if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) | |
1248 | continue; | |
1249 | for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) { | |
1250 | for (bit_j = 0; bit_j < bit_i; bit_j++) { | |
7f14b43a MD |
1251 | if (i & (1U << bit_i)) { |
1252 | if (i & (1U << bit_j)) { | |
1253 | nr_2d_11[bit_i][bit_j]++; | |
1254 | } else { | |
1255 | nr_2d_10[bit_i][bit_j]++; | |
1256 | } | |
1257 | } else { | |
1258 | if (i & (1U << bit_j)) { | |
1259 | nr_2d_01[bit_i][bit_j]++; | |
1260 | } else { | |
1261 | nr_2d_00[bit_i][bit_j]++; | |
1262 | } | |
19ddcd04 MD |
1263 | } |
1264 | } | |
1265 | } | |
1266 | distrib_nr_child++; | |
1267 | } | |
1268 | break; | |
1269 | } | |
1270 | case RCU_JA_NULL: | |
1271 | assert(mode == JA_RECOMPACT_ADD_NEXT); | |
1272 | break; | |
1273 | default: | |
1274 | assert(0); | |
1275 | break; | |
1276 | } | |
1277 | ||
1278 | if (mode == JA_RECOMPACT_ADD_NEXT || mode == JA_RECOMPACT_ADD_SAME) { | |
1279 | for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) { | |
1280 | for (bit_j = 0; bit_j < bit_i; bit_j++) { | |
7f14b43a MD |
1281 | if (n & (1U << bit_i)) { |
1282 | if (n & (1U << bit_j)) { | |
1283 | nr_2d_11[bit_i][bit_j]++; | |
1284 | } else { | |
1285 | nr_2d_10[bit_i][bit_j]++; | |
1286 | } | |
1287 | } else { | |
1288 | if (n & (1U << bit_j)) { | |
1289 | nr_2d_01[bit_i][bit_j]++; | |
1290 | } else { | |
1291 | nr_2d_00[bit_i][bit_j]++; | |
1292 | } | |
19ddcd04 MD |
1293 | } |
1294 | } | |
1295 | } | |
1296 | distrib_nr_child++; | |
1297 | } | |
1298 | ||
1299 | /* | |
1300 | * The best bit selector is that for which the number of nodes | |
1301 | * in each sub-class is closest to one-fourth of the number of | |
1302 | * children in the distribution. We calculate the distance using | |
1303 | * 4 times the size of the sub-distribution to eliminate | |
1304 | * truncation error. | |
1305 | */ | |
1306 | for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) { | |
1307 | for (bit_j = 0; bit_j < bit_i; bit_j++) { | |
4a073c53 | 1308 | int distance_to_best[4]; |
19ddcd04 | 1309 | |
1b34283b MD |
1310 | distance_to_best[0] = ((unsigned int) nr_2d_11[bit_i][bit_j] << 2U) - distrib_nr_child; |
1311 | distance_to_best[1] = ((unsigned int) nr_2d_10[bit_i][bit_j] << 2U) - distrib_nr_child; | |
1312 | distance_to_best[2] = ((unsigned int) nr_2d_01[bit_i][bit_j] << 2U) - distrib_nr_child; | |
1313 | distance_to_best[3] = ((unsigned int) nr_2d_00[bit_i][bit_j] << 2U) - distrib_nr_child; | |
19ddcd04 | 1314 | |
4a073c53 MD |
1315 | /* Consider worse distance above best */ |
1316 | if (distance_to_best[1] > 0 && distance_to_best[1] > distance_to_best[0]) | |
19ddcd04 | 1317 | distance_to_best[0] = distance_to_best[1]; |
4a073c53 | 1318 | if (distance_to_best[2] > 0 && distance_to_best[2] > distance_to_best[0]) |
19ddcd04 | 1319 | distance_to_best[0] = distance_to_best[2]; |
4a073c53 | 1320 | if (distance_to_best[3] > 0 && distance_to_best[3] > distance_to_best[0]) |
19ddcd04 | 1321 | distance_to_best[0] = distance_to_best[3]; |
4a073c53 | 1322 | |
19ddcd04 MD |
1323 | /* |
1324 | * If our worse distance is better than overall, | |
1325 | * we become new best candidate. | |
1326 | */ | |
1327 | if (distance_to_best[0] < overall_best_distance) { | |
1328 | overall_best_distance = distance_to_best[0]; | |
1329 | bitsel[0] = bit_i; | |
1330 | bitsel[1] = bit_j; | |
1331 | } | |
1332 | } | |
1333 | } | |
1334 | ||
1335 | dbg_printf("2 dimensions pool bit selection: (%u,%u)\n", bitsel[0], bitsel[1]); | |
1336 | ||
1337 | /* Return our bit selection */ | |
1338 | _bitsel[0] = bitsel[0]; | |
1339 | _bitsel[1] = bitsel[1]; | |
1340 | } | |
1341 | ||
48cbe001 MD |
1342 | static |
1343 | unsigned int find_nearest_type_index(unsigned int type_index, | |
1344 | unsigned int nr_nodes) | |
1345 | { | |
1346 | const struct cds_ja_type *type; | |
1347 | ||
1348 | assert(type_index != NODE_INDEX_NULL); | |
1349 | if (nr_nodes == 0) | |
1350 | return NODE_INDEX_NULL; | |
1351 | for (;;) { | |
1352 | type = &ja_types[type_index]; | |
1353 | if (nr_nodes < type->min_child) | |
1354 | type_index--; | |
1355 | else if (nr_nodes > type->max_child) | |
1356 | type_index++; | |
1357 | else | |
1358 | break; | |
1359 | } | |
1360 | return type_index; | |
1361 | } | |
1362 | ||
7a0b2331 MD |
1363 | /* |
1364 | * ja_node_recompact_add: recompact a node, adding a new child. | |
2e313670 | 1365 | * Return 0 on success, -EAGAIN if need to retry, or other negative |
5a9a87dd | 1366 | * error value otherwise. |
7a0b2331 MD |
1367 | */ |
1368 | static | |
2e313670 MD |
1369 | int ja_node_recompact(enum ja_recompact mode, |
1370 | struct cds_ja *ja, | |
e1db2db5 | 1371 | unsigned int old_type_index, |
d96bfb0d | 1372 | const struct cds_ja_type *old_type, |
b4540e8a | 1373 | struct cds_ja_inode *old_node, |
5a9a87dd | 1374 | struct cds_ja_shadow_node *shadow_node, |
3d8fe307 | 1375 | struct cds_ja_inode_flag **old_node_flag_ptr, uint8_t n, |
af3cbd45 | 1376 | struct cds_ja_inode_flag *child_node_flag, |
48cbe001 MD |
1377 | struct cds_ja_inode_flag **nullify_node_flag_ptr, |
1378 | int level) | |
7a0b2331 | 1379 | { |
e1db2db5 | 1380 | unsigned int new_type_index; |
b4540e8a | 1381 | struct cds_ja_inode *new_node; |
af3cbd45 | 1382 | struct cds_ja_shadow_node *new_shadow_node = NULL; |
d96bfb0d | 1383 | const struct cds_ja_type *new_type; |
3d8fe307 | 1384 | struct cds_ja_inode_flag *new_node_flag, *old_node_flag; |
7a0b2331 | 1385 | int ret; |
f07b240f | 1386 | int fallback = 0; |
7a0b2331 | 1387 | |
3d8fe307 MD |
1388 | old_node_flag = *old_node_flag_ptr; |
1389 | ||
48cbe001 MD |
1390 | /* |
1391 | * Need to find nearest type index even for ADD_SAME, because | |
1392 | * this recompaction, when applied to linear nodes, will garbage | |
1393 | * collect dummy (NULL) entries, and can therefore cause a few | |
1394 | * linear representations to be skipped. | |
1395 | */ | |
2e313670 | 1396 | switch (mode) { |
19ddcd04 | 1397 | case JA_RECOMPACT_ADD_SAME: |
48cbe001 MD |
1398 | new_type_index = find_nearest_type_index(old_type_index, |
1399 | shadow_node->nr_child + 1); | |
1400 | dbg_printf("Recompact for node with %u children\n", | |
1401 | shadow_node->nr_child + 1); | |
2e313670 | 1402 | break; |
19ddcd04 | 1403 | case JA_RECOMPACT_ADD_NEXT: |
2e313670 MD |
1404 | if (!shadow_node || old_type_index == NODE_INDEX_NULL) { |
1405 | new_type_index = 0; | |
48cbe001 | 1406 | dbg_printf("Recompact for NULL\n"); |
2e313670 | 1407 | } else { |
48cbe001 MD |
1408 | new_type_index = find_nearest_type_index(old_type_index, |
1409 | shadow_node->nr_child + 1); | |
1410 | dbg_printf("Recompact for node with %u children\n", | |
1411 | shadow_node->nr_child + 1); | |
2e313670 MD |
1412 | } |
1413 | break; | |
1414 | case JA_RECOMPACT_DEL: | |
48cbe001 MD |
1415 | new_type_index = find_nearest_type_index(old_type_index, |
1416 | shadow_node->nr_child - 1); | |
1417 | dbg_printf("Recompact for node with %u children\n", | |
1418 | shadow_node->nr_child - 1); | |
2e313670 MD |
1419 | break; |
1420 | default: | |
1421 | assert(0); | |
7a0b2331 | 1422 | } |
a2a7ff59 | 1423 | |
f07b240f | 1424 | retry: /* for fallback */ |
582a6ade MD |
1425 | dbg_printf("Recompact from type %d to type %d\n", |
1426 | old_type_index, new_type_index); | |
7a0b2331 | 1427 | new_type = &ja_types[new_type_index]; |
2e313670 | 1428 | if (new_type_index != NODE_INDEX_NULL) { |
354981c2 | 1429 | new_node = alloc_cds_ja_node(ja, new_type); |
2e313670 MD |
1430 | if (!new_node) |
1431 | return -ENOMEM; | |
b1a90ce3 MD |
1432 | |
1433 | if (new_type->type_class == RCU_JA_POOL) { | |
1434 | switch (new_type->nr_pool_order) { | |
1435 | case 1: | |
1436 | { | |
19ddcd04 MD |
1437 | unsigned int node_distrib_bitsel; |
1438 | ||
b1a90ce3 MD |
1439 | node_distrib_bitsel = |
1440 | ja_node_sum_distribution_1d(mode, ja, | |
1441 | old_type_index, old_type, | |
1442 | old_node, shadow_node, | |
1443 | n, child_node_flag, | |
1444 | nullify_node_flag_ptr); | |
1445 | assert(!((unsigned long) new_node & JA_POOL_1D_MASK)); | |
1446 | new_node_flag = ja_node_flag_pool_1d(new_node, | |
1447 | new_type_index, node_distrib_bitsel); | |
1448 | break; | |
1449 | } | |
1450 | case 2: | |
1451 | { | |
19ddcd04 MD |
1452 | unsigned int node_distrib_bitsel[2]; |
1453 | ||
1454 | ja_node_sum_distribution_2d(mode, ja, | |
1455 | old_type_index, old_type, | |
1456 | old_node, shadow_node, | |
1457 | n, child_node_flag, | |
1458 | nullify_node_flag_ptr, | |
1459 | node_distrib_bitsel); | |
b1a90ce3 MD |
1460 | assert(!((unsigned long) new_node & JA_POOL_1D_MASK)); |
1461 | assert(!((unsigned long) new_node & JA_POOL_2D_MASK)); | |
19ddcd04 MD |
1462 | new_node_flag = ja_node_flag_pool_2d(new_node, |
1463 | new_type_index, node_distrib_bitsel); | |
b1a90ce3 MD |
1464 | break; |
1465 | } | |
1466 | default: | |
1467 | assert(0); | |
1468 | } | |
1469 | } else { | |
1470 | new_node_flag = ja_node_flag(new_node, new_type_index); | |
1471 | } | |
1472 | ||
2e313670 | 1473 | dbg_printf("Recompact inherit lock from %p\n", shadow_node); |
48cbe001 | 1474 | new_shadow_node = rcuja_shadow_set(ja->ht, new_node_flag, shadow_node, ja, level); |
2e313670 | 1475 | if (!new_shadow_node) { |
354981c2 | 1476 | free_cds_ja_node(ja, new_node); |
2e313670 MD |
1477 | return -ENOMEM; |
1478 | } | |
1479 | if (fallback) | |
1480 | new_shadow_node->fallback_removal_count = | |
1481 | JA_FALLBACK_REMOVAL_COUNT; | |
1482 | } else { | |
1483 | new_node = NULL; | |
1484 | new_node_flag = NULL; | |
e1db2db5 | 1485 | } |
11c5e016 | 1486 | |
19ddcd04 | 1487 | assert(mode != JA_RECOMPACT_ADD_NEXT || old_type->type_class != RCU_JA_PIGEON); |
2e313670 MD |
1488 | |
1489 | if (new_type_index == NODE_INDEX_NULL) | |
1490 | goto skip_copy; | |
1491 | ||
11c5e016 MD |
1492 | switch (old_type->type_class) { |
1493 | case RCU_JA_LINEAR: | |
1494 | { | |
1495 | uint8_t nr_child = | |
1496 | ja_linear_node_get_nr_child(old_type, old_node); | |
1497 | unsigned int i; | |
1498 | ||
1499 | for (i = 0; i < nr_child; i++) { | |
b4540e8a | 1500 | struct cds_ja_inode_flag *iter; |
11c5e016 MD |
1501 | uint8_t v; |
1502 | ||
1503 | ja_linear_node_get_ith_pos(old_type, old_node, i, &v, &iter); | |
1504 | if (!iter) | |
1505 | continue; | |
af3cbd45 | 1506 | if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) |
2e313670 | 1507 | continue; |
b1a90ce3 | 1508 | ret = _ja_node_set_nth(new_type, new_node, new_node_flag, |
f07b240f | 1509 | new_shadow_node, |
11c5e016 | 1510 | v, iter); |
f07b240f MD |
1511 | if (new_type->type_class == RCU_JA_POOL && ret) { |
1512 | goto fallback_toosmall; | |
1513 | } | |
11c5e016 MD |
1514 | assert(!ret); |
1515 | } | |
1516 | break; | |
1517 | } | |
1518 | case RCU_JA_POOL: | |
1519 | { | |
1520 | unsigned int pool_nr; | |
1521 | ||
1522 | for (pool_nr = 0; pool_nr < (1U << old_type->nr_pool_order); pool_nr++) { | |
b4540e8a | 1523 | struct cds_ja_inode *pool = |
11c5e016 MD |
1524 | ja_pool_node_get_ith_pool(old_type, |
1525 | old_node, pool_nr); | |
1526 | uint8_t nr_child = | |
1527 | ja_linear_node_get_nr_child(old_type, pool); | |
1528 | unsigned int j; | |
1529 | ||
1530 | for (j = 0; j < nr_child; j++) { | |
b4540e8a | 1531 | struct cds_ja_inode_flag *iter; |
11c5e016 MD |
1532 | uint8_t v; |
1533 | ||
1534 | ja_linear_node_get_ith_pos(old_type, pool, | |
1535 | j, &v, &iter); | |
1536 | if (!iter) | |
1537 | continue; | |
af3cbd45 | 1538 | if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) |
2e313670 | 1539 | continue; |
b1a90ce3 | 1540 | ret = _ja_node_set_nth(new_type, new_node, new_node_flag, |
f07b240f | 1541 | new_shadow_node, |
11c5e016 | 1542 | v, iter); |
f07b240f MD |
1543 | if (new_type->type_class == RCU_JA_POOL |
1544 | && ret) { | |
1545 | goto fallback_toosmall; | |
1546 | } | |
11c5e016 MD |
1547 | assert(!ret); |
1548 | } | |
1549 | } | |
1550 | break; | |
7a0b2331 | 1551 | } |
a2a7ff59 | 1552 | case RCU_JA_NULL: |
19ddcd04 | 1553 | assert(mode == JA_RECOMPACT_ADD_NEXT); |
a2a7ff59 | 1554 | break; |
11c5e016 | 1555 | case RCU_JA_PIGEON: |
2e313670 | 1556 | { |
2e313670 MD |
1557 | unsigned int i; |
1558 | ||
1559 | assert(mode == JA_RECOMPACT_DEL); | |
48cbe001 | 1560 | for (i = 0; i < JA_ENTRY_PER_NODE; i++) { |
2e313670 MD |
1561 | struct cds_ja_inode_flag *iter; |
1562 | ||
1563 | iter = ja_pigeon_node_get_ith_pos(old_type, old_node, i); | |
1564 | if (!iter) | |
1565 | continue; | |
af3cbd45 | 1566 | if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) |
2e313670 | 1567 | continue; |
b1a90ce3 | 1568 | ret = _ja_node_set_nth(new_type, new_node, new_node_flag, |
2e313670 MD |
1569 | new_shadow_node, |
1570 | i, iter); | |
1571 | if (new_type->type_class == RCU_JA_POOL && ret) { | |
1572 | goto fallback_toosmall; | |
1573 | } | |
1574 | assert(!ret); | |
1575 | } | |
1576 | break; | |
1577 | } | |
11c5e016 MD |
1578 | default: |
1579 | assert(0); | |
5a9a87dd | 1580 | ret = -EINVAL; |
f07b240f | 1581 | goto end; |
11c5e016 | 1582 | } |
2e313670 | 1583 | skip_copy: |
11c5e016 | 1584 | |
19ddcd04 | 1585 | if (mode == JA_RECOMPACT_ADD_NEXT || mode == JA_RECOMPACT_ADD_SAME) { |
2e313670 | 1586 | /* add node */ |
b1a90ce3 | 1587 | ret = _ja_node_set_nth(new_type, new_node, new_node_flag, |
2e313670 MD |
1588 | new_shadow_node, |
1589 | n, child_node_flag); | |
7b413155 MD |
1590 | if (new_type->type_class == RCU_JA_POOL && ret) { |
1591 | goto fallback_toosmall; | |
1592 | } | |
2e313670 MD |
1593 | assert(!ret); |
1594 | } | |
19ddcd04 MD |
1595 | |
1596 | if (fallback) { | |
1597 | dbg_printf("Using fallback for %u children, node type index: %u, mode %s\n", | |
1598 | new_shadow_node->nr_child, old_type_index, mode == JA_RECOMPACT_ADD_NEXT ? "add_next" : | |
1599 | (mode == JA_RECOMPACT_DEL ? "del" : "add_same")); | |
354981c2 | 1600 | uatomic_inc(&ja->node_fallback_count_distribution[new_shadow_node->nr_child]); |
19ddcd04 MD |
1601 | } |
1602 | ||
3d8fe307 MD |
1603 | /* Return pointer to new recompacted node through old_node_flag_ptr */ |
1604 | *old_node_flag_ptr = new_node_flag; | |
a2a7ff59 | 1605 | if (old_node) { |
2e313670 MD |
1606 | int flags; |
1607 | ||
1608 | flags = RCUJA_SHADOW_CLEAR_FREE_NODE; | |
1609 | /* | |
1610 | * It is OK to free the lock associated with a node | |
1611 | * going to NULL, since we are holding the parent lock. | |
1612 | * This synchronizes removal with re-add of that node. | |
1613 | */ | |
1614 | if (new_type_index == NODE_INDEX_NULL) | |
48cbe001 | 1615 | flags |= RCUJA_SHADOW_CLEAR_FREE_LOCK; |
3d8fe307 | 1616 | ret = rcuja_shadow_clear(ja->ht, old_node_flag, shadow_node, |
2e313670 | 1617 | flags); |
a2a7ff59 MD |
1618 | assert(!ret); |
1619 | } | |
5a9a87dd MD |
1620 | |
1621 | ret = 0; | |
f07b240f | 1622 | end: |
5a9a87dd | 1623 | return ret; |
f07b240f MD |
1624 | |
1625 | fallback_toosmall: | |
1626 | /* fallback if next pool is too small */ | |
af3cbd45 | 1627 | assert(new_shadow_node); |
3d8fe307 | 1628 | ret = rcuja_shadow_clear(ja->ht, new_node_flag, new_shadow_node, |
f07b240f MD |
1629 | RCUJA_SHADOW_CLEAR_FREE_NODE); |
1630 | assert(!ret); | |
1631 | ||
19ddcd04 MD |
1632 | switch (mode) { |
1633 | case JA_RECOMPACT_ADD_SAME: | |
1634 | /* | |
1635 | * JA_RECOMPACT_ADD_SAME is only triggered if a linear | |
1636 | * node within a pool has unused entries. It should | |
1637 | * therefore _never_ be too small. | |
1638 | */ | |
4a073c53 | 1639 | assert(0); |
4cde8267 MD |
1640 | |
1641 | /* Fall-through */ | |
19ddcd04 MD |
1642 | case JA_RECOMPACT_ADD_NEXT: |
1643 | { | |
1644 | const struct cds_ja_type *next_type; | |
1645 | ||
1646 | /* | |
1647 | * Recompaction attempt on add failed. Should only | |
1648 | * happen if target node type is pool. Caused by | |
1649 | * hard-to-split distribution. Recompact using the next | |
1650 | * distribution size. | |
1651 | */ | |
1652 | assert(new_type->type_class == RCU_JA_POOL); | |
1653 | next_type = &ja_types[new_type_index + 1]; | |
1654 | /* | |
1655 | * Try going to the next pool size if our population | |
1656 | * fits within its range. This is not flagged as a | |
1657 | * fallback. | |
1658 | */ | |
1659 | if (shadow_node->nr_child + 1 >= next_type->min_child | |
1660 | && shadow_node->nr_child + 1 <= next_type->max_child) { | |
1661 | new_type_index++; | |
1662 | goto retry; | |
1663 | } else { | |
1664 | new_type_index++; | |
1665 | dbg_printf("Add fallback to type %d\n", new_type_index); | |
1666 | uatomic_inc(&ja->nr_fallback); | |
1667 | fallback = 1; | |
1668 | goto retry; | |
1669 | } | |
1670 | break; | |
1671 | } | |
1672 | case JA_RECOMPACT_DEL: | |
1673 | /* | |
1674 | * Recompaction attempt on delete failed. Should only | |
1675 | * happen if target node type is pool. This is caused by | |
1676 | * a hard-to-split distribution. Recompact on same node | |
1677 | * size, but flag current node as "fallback" to ensure | |
1678 | * we don't attempt recompaction before some activity | |
1679 | * has reshuffled our node. | |
1680 | */ | |
1681 | assert(new_type->type_class == RCU_JA_POOL); | |
1682 | new_type_index = old_type_index; | |
1683 | dbg_printf("Delete fallback keeping type %d\n", new_type_index); | |
1684 | uatomic_inc(&ja->nr_fallback); | |
1685 | fallback = 1; | |
1686 | goto retry; | |
1687 | default: | |
1688 | assert(0); | |
1689 | return -EINVAL; | |
1690 | } | |
1691 | ||
1692 | /* | |
1693 | * Last resort fallback: pigeon. | |
1694 | */ | |
f07b240f MD |
1695 | new_type_index = (1UL << JA_TYPE_BITS) - 1; |
1696 | dbg_printf("Fallback to type %d\n", new_type_index); | |
1697 | uatomic_inc(&ja->nr_fallback); | |
1698 | fallback = 1; | |
1699 | goto retry; | |
7a0b2331 MD |
1700 | } |
1701 | ||
5a9a87dd | 1702 | /* |
2e313670 | 1703 | * Return 0 on success, -EAGAIN if need to retry, or other negative |
5a9a87dd MD |
1704 | * error value otherwise. |
1705 | */ | |
7a0b2331 | 1706 | static |
d96bfb0d | 1707 | int ja_node_set_nth(struct cds_ja *ja, |
b4540e8a | 1708 | struct cds_ja_inode_flag **node_flag, uint8_t n, |
5a9a87dd | 1709 | struct cds_ja_inode_flag *child_node_flag, |
48cbe001 MD |
1710 | struct cds_ja_shadow_node *shadow_node, |
1711 | int level) | |
7a0b2331 MD |
1712 | { |
1713 | int ret; | |
e1db2db5 | 1714 | unsigned int type_index; |
d96bfb0d | 1715 | const struct cds_ja_type *type; |
b4540e8a | 1716 | struct cds_ja_inode *node; |
7a0b2331 | 1717 | |
a2a7ff59 MD |
1718 | dbg_printf("ja_node_set_nth for n=%u, node %p, shadow %p\n", |
1719 | (unsigned int) n, ja_node_ptr(*node_flag), shadow_node); | |
1720 | ||
e1db2db5 MD |
1721 | node = ja_node_ptr(*node_flag); |
1722 | type_index = ja_node_type(*node_flag); | |
1723 | type = &ja_types[type_index]; | |
b1a90ce3 | 1724 | ret = _ja_node_set_nth(type, node, *node_flag, shadow_node, |
e1db2db5 | 1725 | n, child_node_flag); |
2e313670 MD |
1726 | switch (ret) { |
1727 | case -ENOSPC: | |
19ddcd04 MD |
1728 | /* Not enough space in node, need to recompact to next type. */ |
1729 | ret = ja_node_recompact(JA_RECOMPACT_ADD_NEXT, ja, type_index, type, node, | |
48cbe001 | 1730 | shadow_node, node_flag, n, child_node_flag, NULL, level); |
2e313670 MD |
1731 | break; |
1732 | case -ERANGE: | |
1733 | /* Node needs to be recompacted. */ | |
19ddcd04 | 1734 | ret = ja_node_recompact(JA_RECOMPACT_ADD_SAME, ja, type_index, type, node, |
48cbe001 | 1735 | shadow_node, node_flag, n, child_node_flag, NULL, level); |
2e313670 MD |
1736 | break; |
1737 | } | |
1738 | return ret; | |
1739 | } | |
1740 | ||
1741 | /* | |
1742 | * Return 0 on success, -EAGAIN if need to retry, or other negative | |
1743 | * error value otherwise. | |
1744 | */ | |
1745 | static | |
af3cbd45 MD |
1746 | int ja_node_clear_ptr(struct cds_ja *ja, |
1747 | struct cds_ja_inode_flag **node_flag_ptr, /* Pointer to location to nullify */ | |
1748 | struct cds_ja_inode_flag **parent_node_flag_ptr, /* Address of parent ptr in its parent */ | |
1749 | struct cds_ja_shadow_node *shadow_node, /* of parent */ | |
48cbe001 | 1750 | uint8_t n, int level) |
2e313670 MD |
1751 | { |
1752 | int ret; | |
1753 | unsigned int type_index; | |
1754 | const struct cds_ja_type *type; | |
1755 | struct cds_ja_inode *node; | |
1756 | ||
af3cbd45 MD |
1757 | dbg_printf("ja_node_clear_ptr for node %p, shadow %p, target ptr %p\n", |
1758 | ja_node_ptr(*parent_node_flag_ptr), shadow_node, node_flag_ptr); | |
2e313670 | 1759 | |
af3cbd45 MD |
1760 | node = ja_node_ptr(*parent_node_flag_ptr); |
1761 | type_index = ja_node_type(*parent_node_flag_ptr); | |
2e313670 | 1762 | type = &ja_types[type_index]; |
19ddcd04 | 1763 | ret = _ja_node_clear_ptr(type, node, *parent_node_flag_ptr, shadow_node, node_flag_ptr, n); |
2e313670 | 1764 | if (ret == -EFBIG) { |
19ddcd04 | 1765 | /* Should try recompaction. */ |
2e313670 | 1766 | ret = ja_node_recompact(JA_RECOMPACT_DEL, ja, type_index, type, node, |
af3cbd45 | 1767 | shadow_node, parent_node_flag_ptr, n, NULL, |
48cbe001 | 1768 | node_flag_ptr, level); |
7a0b2331 MD |
1769 | } |
1770 | return ret; | |
1771 | } | |
be9a7474 | 1772 | |
03ec1aeb | 1773 | struct cds_ja_node *cds_ja_lookup(struct cds_ja *ja, uint64_t key) |
b4540e8a | 1774 | { |
41975c12 MD |
1775 | unsigned int tree_depth, i; |
1776 | struct cds_ja_inode_flag *node_flag; | |
1777 | ||
1778 | if (caa_unlikely(key > ja->key_max)) | |
03ec1aeb | 1779 | return NULL; |
41975c12 | 1780 | tree_depth = ja->tree_depth; |
5a9a87dd | 1781 | node_flag = rcu_dereference(ja->root); |
41975c12 | 1782 | |
5a9a87dd MD |
1783 | /* level 0: root node */ |
1784 | if (!ja_node_ptr(node_flag)) | |
03ec1aeb | 1785 | return NULL; |
5a9a87dd MD |
1786 | |
1787 | for (i = 1; i < tree_depth; i++) { | |
79b41067 MD |
1788 | uint8_t iter_key; |
1789 | ||
1790 | iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (tree_depth - i - 1))); | |
48cbe001 | 1791 | node_flag = ja_node_get_nth(node_flag, NULL, iter_key); |
582a6ade MD |
1792 | dbg_printf("cds_ja_lookup iter key lookup %u finds node_flag %p\n", |
1793 | (unsigned int) iter_key, node_flag); | |
41975c12 | 1794 | if (!ja_node_ptr(node_flag)) |
03ec1aeb | 1795 | return NULL; |
41975c12 MD |
1796 | } |
1797 | ||
5a9a87dd | 1798 | /* Last level lookup succeded. We got an actual match. */ |
03ec1aeb | 1799 | return (struct cds_ja_node *) node_flag; |
5a9a87dd MD |
1800 | } |
1801 | ||
b023ba9f MD |
1802 | static |
1803 | struct cds_ja_node *cds_ja_lookup_inequality(struct cds_ja *ja, uint64_t key, | |
1804 | enum ja_lookup_inequality mode) | |
291b2543 MD |
1805 | { |
1806 | int tree_depth, level; | |
1807 | struct cds_ja_inode_flag *node_flag, *cur_node_depth[JA_MAX_DEPTH]; | |
b023ba9f | 1808 | enum ja_direction dir; |
291b2543 | 1809 | |
b023ba9f MD |
1810 | switch (mode) { |
1811 | case JA_LOOKUP_BE: | |
1812 | if (caa_unlikely(key > ja->key_max || key == 0)) | |
1813 | return NULL; | |
1814 | break; | |
1815 | case JA_LOOKUP_AE: | |
1816 | if (caa_unlikely(key >= ja->key_max)) | |
1817 | return NULL; | |
1818 | break; | |
1819 | default: | |
03ec1aeb | 1820 | return NULL; |
b023ba9f | 1821 | } |
291b2543 MD |
1822 | |
1823 | memset(cur_node_depth, 0, sizeof(cur_node_depth)); | |
1824 | tree_depth = ja->tree_depth; | |
1825 | node_flag = rcu_dereference(ja->root); | |
1826 | cur_node_depth[0] = node_flag; | |
1827 | ||
1828 | /* level 0: root node */ | |
1829 | if (!ja_node_ptr(node_flag)) | |
03ec1aeb | 1830 | return NULL; |
291b2543 MD |
1831 | |
1832 | for (level = 1; level < tree_depth; level++) { | |
1833 | uint8_t iter_key; | |
1834 | ||
1835 | iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (tree_depth - level - 1))); | |
1836 | node_flag = ja_node_get_nth(node_flag, NULL, iter_key); | |
1837 | if (!ja_node_ptr(node_flag)) | |
1838 | break; | |
1839 | cur_node_depth[level] = node_flag; | |
b023ba9f | 1840 | dbg_printf("cds_ja_lookup_inequality iter key lookup %u finds node_flag %p\n", |
291b2543 MD |
1841 | (unsigned int) iter_key, node_flag); |
1842 | } | |
1843 | ||
1844 | if (level == tree_depth) { | |
1845 | /* Last level lookup succeded. We got an equal match. */ | |
03ec1aeb | 1846 | return (struct cds_ja_node *) node_flag; |
291b2543 MD |
1847 | } |
1848 | ||
1849 | /* | |
b023ba9f | 1850 | * Find highest value left/right of current node. |
291b2543 | 1851 | * Current node is cur_node_depth[level]. |
b023ba9f MD |
1852 | * Start at current level. If we cannot find any key left/right |
1853 | * of ours, go one level up, seek highest value left/right of | |
1854 | * current (recursively), and when we find one, get the | |
1855 | * rightmost/leftmost child of its rightmost/leftmost child | |
1856 | * (recursively). | |
291b2543 | 1857 | */ |
b023ba9f MD |
1858 | switch (mode) { |
1859 | case JA_LOOKUP_BE: | |
1860 | dir = JA_LEFT; | |
1861 | break; | |
1862 | case JA_LOOKUP_AE: | |
1863 | dir = JA_RIGHT; | |
1864 | break; | |
1865 | default: | |
1866 | assert(0); | |
1867 | } | |
291b2543 MD |
1868 | for (; level > 0; level--) { |
1869 | uint8_t iter_key; | |
1870 | ||
1871 | iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (tree_depth - level - 1))); | |
b023ba9f MD |
1872 | node_flag = ja_node_get_leftright(cur_node_depth[level - 1], |
1873 | iter_key, dir); | |
1874 | /* If found left sibling, find rightmost/leftmost child. */ | |
291b2543 MD |
1875 | if (ja_node_ptr(node_flag)) |
1876 | break; | |
1877 | } | |
1878 | ||
1879 | if (!level) { | |
b023ba9f | 1880 | /* Reached the root and could not find a left/right sibling. */ |
03ec1aeb | 1881 | return NULL; |
291b2543 MD |
1882 | } |
1883 | ||
1884 | level++; | |
3c52f0f9 MD |
1885 | |
1886 | /* | |
4cef6f97 | 1887 | * From this point, we are guaranteed to be able to find a |
b023ba9f MD |
1888 | * "below than"/"above than" match. ja_attach_node() and |
1889 | * ja_detach_node() both guarantee that it is not possible for a | |
1890 | * lookup to reach a dead-end. | |
3c52f0f9 MD |
1891 | */ |
1892 | ||
b023ba9f MD |
1893 | /* |
1894 | * Find rightmost/leftmost child of rightmost/leftmost child | |
1895 | * (recursively). | |
1896 | */ | |
1897 | switch (mode) { | |
1898 | case JA_LOOKUP_BE: | |
1899 | dir = JA_RIGHTMOST; | |
1900 | break; | |
1901 | case JA_LOOKUP_AE: | |
1902 | dir = JA_LEFTMOST; | |
1903 | break; | |
1904 | default: | |
1905 | assert(0); | |
1906 | } | |
291b2543 | 1907 | for (; level < tree_depth; level++) { |
b023ba9f | 1908 | node_flag = ja_node_get_minmax(node_flag, dir); |
291b2543 MD |
1909 | if (!ja_node_ptr(node_flag)) |
1910 | break; | |
1911 | } | |
1912 | ||
4cef6f97 | 1913 | assert(level == tree_depth); |
291b2543 | 1914 | |
03ec1aeb | 1915 | return (struct cds_ja_node *) node_flag; |
291b2543 MD |
1916 | } |
1917 | ||
b023ba9f MD |
1918 | struct cds_ja_node *cds_ja_lookup_below_equal(struct cds_ja *ja, uint64_t key) |
1919 | { | |
1920 | return cds_ja_lookup_inequality(ja, key, JA_LOOKUP_BE); | |
1921 | } | |
1922 | ||
1923 | struct cds_ja_node *cds_ja_lookup_above_equal(struct cds_ja *ja, uint64_t key) | |
1924 | { | |
1925 | return cds_ja_lookup_inequality(ja, key, JA_LOOKUP_AE); | |
1926 | } | |
1927 | ||
5a9a87dd MD |
1928 | /* |
1929 | * We reached an unpopulated node. Create it and the children we need, | |
1930 | * and then attach the entire branch to the current node. This may | |
1931 | * trigger recompaction of the current node. Locks needed: node lock | |
1932 | * (for add), and, possibly, parent node lock (to update pointer due to | |
1933 | * node recompaction). | |
1934 | * | |
1935 | * First take node lock, check if recompaction is needed, then take | |
1936 | * parent lock (if needed). Then we can proceed to create the new | |
1937 | * branch. Publish the new branch, and release locks. | |
1938 | * TODO: we currently always take the parent lock even when not needed. | |
47d2eab3 MD |
1939 | * |
1940 | * ja_attach_node() ensures that a lookup will _never_ see a branch that | |
1941 | * leads to a dead-end: before attaching a branch, the entire content of | |
1942 | * the new branch is populated, thus creating a cluster, before | |
1943 | * attaching the cluster to the rest of the tree, thus making it visible | |
1944 | * to lookups. | |
5a9a87dd MD |
1945 | */ |
1946 | static | |
1947 | int ja_attach_node(struct cds_ja *ja, | |
b0ca2d21 | 1948 | struct cds_ja_inode_flag **attach_node_flag_ptr, |
b62a8d0c | 1949 | struct cds_ja_inode_flag *attach_node_flag, |
48cbe001 MD |
1950 | struct cds_ja_inode_flag *parent_attach_node_flag, |
1951 | struct cds_ja_inode_flag **old_node_flag_ptr, | |
1952 | struct cds_ja_inode_flag *old_node_flag, | |
5a9a87dd | 1953 | uint64_t key, |
79b41067 | 1954 | unsigned int level, |
5a9a87dd MD |
1955 | struct cds_ja_node *child_node) |
1956 | { | |
1957 | struct cds_ja_shadow_node *shadow_node = NULL, | |
af3cbd45 | 1958 | *parent_shadow_node = NULL; |
5a9a87dd MD |
1959 | struct cds_ja_inode_flag *iter_node_flag, *iter_dest_node_flag; |
1960 | int ret, i; | |
a2a7ff59 | 1961 | struct cds_ja_inode_flag *created_nodes[JA_MAX_DEPTH]; |
5a9a87dd MD |
1962 | int nr_created_nodes = 0; |
1963 | ||
48cbe001 MD |
1964 | 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", |
1965 | level, old_node_flag, attach_node_flag_ptr, attach_node_flag, parent_attach_node_flag); | |
a2a7ff59 | 1966 | |
48cbe001 MD |
1967 | assert(!old_node_flag); |
1968 | if (attach_node_flag) { | |
1969 | shadow_node = rcuja_shadow_lookup_lock(ja->ht, attach_node_flag); | |
1970 | if (!shadow_node) { | |
1971 | ret = -EAGAIN; | |
1972 | goto end; | |
1973 | } | |
5a9a87dd | 1974 | } |
48cbe001 | 1975 | if (parent_attach_node_flag) { |
5a9a87dd | 1976 | parent_shadow_node = rcuja_shadow_lookup_lock(ja->ht, |
48cbe001 | 1977 | parent_attach_node_flag); |
5a9a87dd | 1978 | if (!parent_shadow_node) { |
2e313670 | 1979 | ret = -EAGAIN; |
5a9a87dd MD |
1980 | goto unlock_shadow; |
1981 | } | |
1982 | } | |
1983 | ||
48cbe001 | 1984 | if (old_node_flag_ptr && ja_node_ptr(*old_node_flag_ptr)) { |
b306a0fe | 1985 | /* |
c112acaa MD |
1986 | * Target node has been updated between RCU lookup and |
1987 | * lock acquisition. We need to re-try lookup and | |
1988 | * attach. | |
1989 | */ | |
1990 | ret = -EAGAIN; | |
1991 | goto unlock_parent; | |
1992 | } | |
1993 | ||
9be99d4a MD |
1994 | /* |
1995 | * Perform a lookup query to handle the case where | |
1996 | * old_node_flag_ptr is NULL. We cannot use it to check if the | |
1997 | * node has been populated between RCU lookup and mutex | |
1998 | * acquisition. | |
1999 | */ | |
2000 | if (!old_node_flag_ptr) { | |
2001 | uint8_t iter_key; | |
2002 | struct cds_ja_inode_flag *lookup_node_flag; | |
2003 | struct cds_ja_inode_flag **lookup_node_flag_ptr; | |
2004 | ||
2005 | iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (ja->tree_depth - level))); | |
2006 | lookup_node_flag = ja_node_get_nth(attach_node_flag, | |
2007 | &lookup_node_flag_ptr, | |
2008 | iter_key); | |
2009 | if (lookup_node_flag) { | |
2010 | ret = -EEXIST; | |
2011 | goto unlock_parent; | |
2012 | } | |
2013 | } | |
2014 | ||
c112acaa | 2015 | if (attach_node_flag_ptr && ja_node_ptr(*attach_node_flag_ptr) != |
b62a8d0c | 2016 | ja_node_ptr(attach_node_flag)) { |
c112acaa MD |
2017 | /* |
2018 | * Target node has been updated between RCU lookup and | |
2019 | * lock acquisition. We need to re-try lookup and | |
2020 | * attach. | |
b306a0fe MD |
2021 | */ |
2022 | ret = -EAGAIN; | |
2023 | goto unlock_parent; | |
2024 | } | |
2025 | ||
a2a7ff59 | 2026 | /* Create new branch, starting from bottom */ |
03ec1aeb | 2027 | iter_node_flag = (struct cds_ja_inode_flag *) child_node; |
5a9a87dd | 2028 | |
48cbe001 | 2029 | for (i = ja->tree_depth - 1; i >= (int) level; i--) { |
79b41067 MD |
2030 | uint8_t iter_key; |
2031 | ||
48cbe001 | 2032 | iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (ja->tree_depth - i - 1))); |
79b41067 | 2033 | dbg_printf("branch creation level %d, key %u\n", |
48cbe001 | 2034 | i, (unsigned int) iter_key); |
5a9a87dd MD |
2035 | iter_dest_node_flag = NULL; |
2036 | ret = ja_node_set_nth(ja, &iter_dest_node_flag, | |
79b41067 | 2037 | iter_key, |
5a9a87dd | 2038 | iter_node_flag, |
48cbe001 | 2039 | NULL, i); |
9be99d4a MD |
2040 | if (ret) { |
2041 | dbg_printf("branch creation error %d\n", ret); | |
5a9a87dd | 2042 | goto check_error; |
9be99d4a | 2043 | } |
5a9a87dd MD |
2044 | created_nodes[nr_created_nodes++] = iter_dest_node_flag; |
2045 | iter_node_flag = iter_dest_node_flag; | |
2046 | } | |
48cbe001 | 2047 | assert(level > 0); |
5a9a87dd | 2048 | |
48cbe001 MD |
2049 | /* Publish branch */ |
2050 | if (level == 1) { | |
2051 | /* | |
2052 | * Attaching to root node. | |
2053 | */ | |
2054 | rcu_assign_pointer(ja->root, iter_node_flag); | |
2055 | } else { | |
79b41067 MD |
2056 | uint8_t iter_key; |
2057 | ||
2058 | iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (ja->tree_depth - level))); | |
48cbe001 MD |
2059 | dbg_printf("publish branch at level %d, key %u\n", |
2060 | level - 1, (unsigned int) iter_key); | |
a2a7ff59 | 2061 | /* We need to use set_nth on the previous level. */ |
48cbe001 | 2062 | iter_dest_node_flag = attach_node_flag; |
a2a7ff59 | 2063 | ret = ja_node_set_nth(ja, &iter_dest_node_flag, |
79b41067 | 2064 | iter_key, |
a2a7ff59 | 2065 | iter_node_flag, |
48cbe001 | 2066 | shadow_node, level - 1); |
9be99d4a MD |
2067 | if (ret) { |
2068 | dbg_printf("branch publish error %d\n", ret); | |
a2a7ff59 | 2069 | goto check_error; |
9be99d4a | 2070 | } |
48cbe001 MD |
2071 | /* |
2072 | * Attach branch | |
2073 | */ | |
2074 | rcu_assign_pointer(*attach_node_flag_ptr, iter_dest_node_flag); | |
a2a7ff59 MD |
2075 | } |
2076 | ||
5a9a87dd MD |
2077 | /* Success */ |
2078 | ret = 0; | |
2079 | ||
2080 | check_error: | |
2081 | if (ret) { | |
2082 | for (i = 0; i < nr_created_nodes; i++) { | |
2083 | int tmpret; | |
a2a7ff59 MD |
2084 | int flags; |
2085 | ||
2086 | flags = RCUJA_SHADOW_CLEAR_FREE_LOCK; | |
2087 | if (i) | |
2088 | flags |= RCUJA_SHADOW_CLEAR_FREE_NODE; | |
5a9a87dd | 2089 | tmpret = rcuja_shadow_clear(ja->ht, |
3d8fe307 | 2090 | created_nodes[i], |
a2a7ff59 MD |
2091 | NULL, |
2092 | flags); | |
5a9a87dd MD |
2093 | assert(!tmpret); |
2094 | } | |
2095 | } | |
b306a0fe | 2096 | unlock_parent: |
5a9a87dd MD |
2097 | if (parent_shadow_node) |
2098 | rcuja_shadow_unlock(parent_shadow_node); | |
2099 | unlock_shadow: | |
2100 | if (shadow_node) | |
2101 | rcuja_shadow_unlock(shadow_node); | |
2102 | end: | |
2103 | return ret; | |
2104 | } | |
2105 | ||
2106 | /* | |
03ec1aeb MD |
2107 | * Lock the parent containing the pointer to list of duplicates, and add |
2108 | * node to this list. Failure can happen if concurrent update changes | |
2109 | * the parent before we get the lock. We return -EAGAIN in that case. | |
5a9a87dd MD |
2110 | * Return 0 on success, negative error value on failure. |
2111 | */ | |
2112 | static | |
2113 | int ja_chain_node(struct cds_ja *ja, | |
af3cbd45 | 2114 | struct cds_ja_inode_flag *parent_node_flag, |
fa112799 | 2115 | struct cds_ja_inode_flag **node_flag_ptr, |
c112acaa | 2116 | struct cds_ja_inode_flag *node_flag, |
5a9a87dd MD |
2117 | struct cds_ja_node *node) |
2118 | { | |
2119 | struct cds_ja_shadow_node *shadow_node; | |
fa112799 | 2120 | int ret = 0; |
5a9a87dd | 2121 | |
3d8fe307 | 2122 | shadow_node = rcuja_shadow_lookup_lock(ja->ht, parent_node_flag); |
b306a0fe | 2123 | if (!shadow_node) { |
2e313670 | 2124 | return -EAGAIN; |
b306a0fe | 2125 | } |
c112acaa | 2126 | if (ja_node_ptr(*node_flag_ptr) != ja_node_ptr(node_flag)) { |
fa112799 MD |
2127 | ret = -EAGAIN; |
2128 | goto end; | |
2129 | } | |
03ec1aeb MD |
2130 | /* |
2131 | * Add node to head of list. Safe against concurrent RCU read | |
2132 | * traversals. | |
2133 | */ | |
2134 | node->next = (struct cds_ja_node *) node_flag; | |
2135 | rcu_assign_pointer(*node_flag_ptr, (struct cds_ja_inode_flag *) node); | |
fa112799 | 2136 | end: |
5a9a87dd | 2137 | rcuja_shadow_unlock(shadow_node); |
fa112799 | 2138 | return ret; |
5a9a87dd MD |
2139 | } |
2140 | ||
75d573aa MD |
2141 | static |
2142 | int _cds_ja_add(struct cds_ja *ja, uint64_t key, | |
6475613c | 2143 | struct cds_ja_node *node, |
75d573aa | 2144 | struct cds_ja_node **unique_node_ret) |
5a9a87dd MD |
2145 | { |
2146 | unsigned int tree_depth, i; | |
48cbe001 | 2147 | struct cds_ja_inode_flag *attach_node_flag, |
5a9a87dd | 2148 | *parent_node_flag, |
b62a8d0c | 2149 | *parent2_node_flag, |
48cbe001 MD |
2150 | *node_flag, |
2151 | *parent_attach_node_flag; | |
2152 | struct cds_ja_inode_flag **attach_node_flag_ptr, | |
2153 | **parent_node_flag_ptr, | |
2154 | **node_flag_ptr; | |
5a9a87dd MD |
2155 | int ret; |
2156 | ||
b306a0fe | 2157 | if (caa_unlikely(key > ja->key_max)) { |
5a9a87dd | 2158 | return -EINVAL; |
b306a0fe | 2159 | } |
5a9a87dd MD |
2160 | tree_depth = ja->tree_depth; |
2161 | ||
2162 | retry: | |
a2a7ff59 | 2163 | dbg_printf("cds_ja_add attempt: key %" PRIu64 ", node %p\n", |
6475613c | 2164 | key, node); |
5a9a87dd | 2165 | parent2_node_flag = NULL; |
b0f74e47 MD |
2166 | parent_node_flag = |
2167 | (struct cds_ja_inode_flag *) &ja->root; /* Use root ptr address as key for mutex */ | |
48cbe001 | 2168 | parent_node_flag_ptr = NULL; |
35170a44 | 2169 | node_flag = rcu_dereference(ja->root); |
48cbe001 | 2170 | node_flag_ptr = &ja->root; |
5a9a87dd MD |
2171 | |
2172 | /* Iterate on all internal levels */ | |
a2a7ff59 | 2173 | for (i = 1; i < tree_depth; i++) { |
79b41067 MD |
2174 | uint8_t iter_key; |
2175 | ||
48cbe001 MD |
2176 | if (!ja_node_ptr(node_flag)) |
2177 | break; | |
2178 | dbg_printf("cds_ja_add iter parent2_node_flag %p parent_node_flag %p node_flag_ptr %p node_flag %p\n", | |
2179 | parent2_node_flag, parent_node_flag, node_flag_ptr, node_flag); | |
79b41067 | 2180 | iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (tree_depth - i - 1))); |
5a9a87dd MD |
2181 | parent2_node_flag = parent_node_flag; |
2182 | parent_node_flag = node_flag; | |
48cbe001 | 2183 | parent_node_flag_ptr = node_flag_ptr; |
5a9a87dd MD |
2184 | node_flag = ja_node_get_nth(node_flag, |
2185 | &node_flag_ptr, | |
79b41067 | 2186 | iter_key); |
5a9a87dd MD |
2187 | } |
2188 | ||
2189 | /* | |
48cbe001 MD |
2190 | * We reached either bottom of tree or internal NULL node, |
2191 | * simply add node to last internal level, or chain it if key is | |
2192 | * already present. | |
5a9a87dd MD |
2193 | */ |
2194 | if (!ja_node_ptr(node_flag)) { | |
48cbe001 MD |
2195 | dbg_printf("cds_ja_add NULL parent2_node_flag %p parent_node_flag %p node_flag_ptr %p node_flag %p\n", |
2196 | parent2_node_flag, parent_node_flag, node_flag_ptr, node_flag); | |
75d573aa | 2197 | |
48cbe001 MD |
2198 | attach_node_flag = parent_node_flag; |
2199 | attach_node_flag_ptr = parent_node_flag_ptr; | |
2200 | parent_attach_node_flag = parent2_node_flag; | |
2201 | ||
b0ca2d21 | 2202 | ret = ja_attach_node(ja, attach_node_flag_ptr, |
b62a8d0c | 2203 | attach_node_flag, |
48cbe001 MD |
2204 | parent_attach_node_flag, |
2205 | node_flag_ptr, | |
2206 | node_flag, | |
6475613c | 2207 | key, i, node); |
5a9a87dd | 2208 | } else { |
75d573aa MD |
2209 | if (unique_node_ret) { |
2210 | *unique_node_ret = (struct cds_ja_node *) ja_node_ptr(node_flag); | |
2211 | return -EEXIST; | |
2212 | } | |
2213 | ||
48cbe001 MD |
2214 | dbg_printf("cds_ja_add duplicate parent2_node_flag %p parent_node_flag %p node_flag_ptr %p node_flag %p\n", |
2215 | parent2_node_flag, parent_node_flag, node_flag_ptr, node_flag); | |
75d573aa | 2216 | |
48cbe001 MD |
2217 | attach_node_flag = node_flag; |
2218 | attach_node_flag_ptr = node_flag_ptr; | |
2219 | parent_attach_node_flag = parent_node_flag; | |
2220 | ||
5a9a87dd | 2221 | ret = ja_chain_node(ja, |
48cbe001 MD |
2222 | parent_attach_node_flag, |
2223 | attach_node_flag_ptr, | |
2224 | attach_node_flag, | |
6475613c | 2225 | node); |
5a9a87dd | 2226 | } |
b306a0fe | 2227 | if (ret == -EAGAIN || ret == -EEXIST) |
5a9a87dd | 2228 | goto retry; |
48cbe001 | 2229 | |
5a9a87dd | 2230 | return ret; |
b4540e8a MD |
2231 | } |
2232 | ||
75d573aa | 2233 | int cds_ja_add(struct cds_ja *ja, uint64_t key, |
6475613c | 2234 | struct cds_ja_node *node) |
75d573aa | 2235 | { |
6475613c | 2236 | return _cds_ja_add(ja, key, node, NULL); |
75d573aa MD |
2237 | } |
2238 | ||
2239 | struct cds_ja_node *cds_ja_add_unique(struct cds_ja *ja, uint64_t key, | |
6475613c | 2240 | struct cds_ja_node *node) |
75d573aa MD |
2241 | { |
2242 | int ret; | |
2243 | struct cds_ja_node *ret_node; | |
2244 | ||
6475613c | 2245 | ret = _cds_ja_add(ja, key, node, &ret_node); |
75d573aa MD |
2246 | if (ret == -EEXIST) |
2247 | return ret_node; | |
2248 | else | |
6475613c | 2249 | return node; |
75d573aa MD |
2250 | } |
2251 | ||
af3cbd45 MD |
2252 | /* |
2253 | * Note: there is no need to lookup the pointer address associated with | |
2254 | * each node's nth item after taking the lock: it's already been done by | |
2255 | * cds_ja_del while holding the rcu read-side lock, and our node rules | |
2256 | * ensure that when a match value -> pointer is found in a node, it is | |
2257 | * _NEVER_ changed for that node without recompaction, and recompaction | |
2258 | * reallocates the node. | |
b306a0fe MD |
2259 | * However, when a child is removed from "linear" nodes, its pointer |
2260 | * is set to NULL. We therefore check, while holding the locks, if this | |
2261 | * pointer is NULL, and return -ENOENT to the caller if it is the case. | |
4cef6f97 MD |
2262 | * |
2263 | * ja_detach_node() ensures that a lookup will _never_ see a branch that | |
2264 | * leads to a dead-end: when removing branch, it makes sure to perform | |
2265 | * the "cut" at the highest node that has only one child, effectively | |
2266 | * replacing it with a NULL pointer. | |
af3cbd45 | 2267 | */ |
35170a44 MD |
2268 | static |
2269 | int ja_detach_node(struct cds_ja *ja, | |
2270 | struct cds_ja_inode_flag **snapshot, | |
af3cbd45 MD |
2271 | struct cds_ja_inode_flag ***snapshot_ptr, |
2272 | uint8_t *snapshot_n, | |
35170a44 MD |
2273 | int nr_snapshot, |
2274 | uint64_t key, | |
2275 | struct cds_ja_node *node) | |
2276 | { | |
af3cbd45 MD |
2277 | struct cds_ja_shadow_node *shadow_nodes[JA_MAX_DEPTH]; |
2278 | struct cds_ja_inode_flag **node_flag_ptr = NULL, | |
2279 | *parent_node_flag = NULL, | |
2280 | **parent_node_flag_ptr = NULL; | |
b62a8d0c | 2281 | struct cds_ja_inode_flag *iter_node_flag; |
4d6ef45e MD |
2282 | int ret, i, nr_shadow = 0, nr_clear = 0, nr_branch = 0; |
2283 | uint8_t n = 0; | |
35170a44 | 2284 | |
4d6ef45e | 2285 | assert(nr_snapshot == ja->tree_depth + 1); |
35170a44 | 2286 | |
af3cbd45 MD |
2287 | /* |
2288 | * From the last internal level node going up, get the node | |
2289 | * lock, check if the node has only one child left. If it is the | |
2290 | * case, we continue iterating upward. When we reach a node | |
2291 | * which has more that one child left, we lock the parent, and | |
2292 | * proceed to the node deletion (removing its children too). | |
2293 | */ | |
4d6ef45e | 2294 | for (i = nr_snapshot - 2; i >= 1; i--) { |
af3cbd45 MD |
2295 | struct cds_ja_shadow_node *shadow_node; |
2296 | ||
2297 | shadow_node = rcuja_shadow_lookup_lock(ja->ht, | |
3d8fe307 | 2298 | snapshot[i]); |
af3cbd45 MD |
2299 | if (!shadow_node) { |
2300 | ret = -EAGAIN; | |
2301 | goto end; | |
2302 | } | |
af3cbd45 | 2303 | shadow_nodes[nr_shadow++] = shadow_node; |
b62a8d0c MD |
2304 | |
2305 | /* | |
2306 | * Check if node has been removed between RCU | |
2307 | * lookup and lock acquisition. | |
2308 | */ | |
2309 | assert(snapshot_ptr[i + 1]); | |
2310 | if (ja_node_ptr(*snapshot_ptr[i + 1]) | |
2311 | != ja_node_ptr(snapshot[i + 1])) { | |
2312 | ret = -ENOENT; | |
2313 | goto end; | |
2314 | } | |
2315 | ||
2316 | assert(shadow_node->nr_child > 0); | |
d810c97f | 2317 | if (shadow_node->nr_child == 1 && i > 1) |
4d6ef45e MD |
2318 | nr_clear++; |
2319 | nr_branch++; | |
af3cbd45 MD |
2320 | if (shadow_node->nr_child > 1 || i == 1) { |
2321 | /* Lock parent and break */ | |
2322 | shadow_node = rcuja_shadow_lookup_lock(ja->ht, | |
3d8fe307 | 2323 | snapshot[i - 1]); |
af3cbd45 MD |
2324 | if (!shadow_node) { |
2325 | ret = -EAGAIN; | |
2326 | goto end; | |
2327 | } | |
2328 | shadow_nodes[nr_shadow++] = shadow_node; | |
b62a8d0c | 2329 | |
c112acaa MD |
2330 | /* |
2331 | * Check if node has been removed between RCU | |
2332 | * lookup and lock acquisition. | |
2333 | */ | |
b62a8d0c MD |
2334 | assert(snapshot_ptr[i]); |
2335 | if (ja_node_ptr(*snapshot_ptr[i]) | |
2336 | != ja_node_ptr(snapshot[i])) { | |
c112acaa MD |
2337 | ret = -ENOENT; |
2338 | goto end; | |
2339 | } | |
2340 | ||
b62a8d0c | 2341 | node_flag_ptr = snapshot_ptr[i + 1]; |
4d6ef45e MD |
2342 | n = snapshot_n[i + 1]; |
2343 | parent_node_flag_ptr = snapshot_ptr[i]; | |
2344 | parent_node_flag = snapshot[i]; | |
c112acaa | 2345 | |
af3cbd45 MD |
2346 | if (i > 1) { |
2347 | /* | |
2348 | * Lock parent's parent, in case we need | |
2349 | * to recompact parent. | |
2350 | */ | |
2351 | shadow_node = rcuja_shadow_lookup_lock(ja->ht, | |
3d8fe307 | 2352 | snapshot[i - 2]); |
af3cbd45 MD |
2353 | if (!shadow_node) { |
2354 | ret = -EAGAIN; | |
2355 | goto end; | |
2356 | } | |
2357 | shadow_nodes[nr_shadow++] = shadow_node; | |
b62a8d0c MD |
2358 | |
2359 | /* | |
2360 | * Check if node has been removed between RCU | |
2361 | * lookup and lock acquisition. | |
2362 | */ | |
2363 | assert(snapshot_ptr[i - 1]); | |
2364 | if (ja_node_ptr(*snapshot_ptr[i - 1]) | |
2365 | != ja_node_ptr(snapshot[i - 1])) { | |
2366 | ret = -ENOENT; | |
2367 | goto end; | |
2368 | } | |
af3cbd45 | 2369 | } |
b62a8d0c | 2370 | |
af3cbd45 MD |
2371 | break; |
2372 | } | |
2373 | } | |
2374 | ||
2375 | /* | |
4d6ef45e MD |
2376 | * At this point, we want to delete all nodes that are about to |
2377 | * be removed from shadow_nodes (except the last one, which is | |
2378 | * either the root or the parent of the upmost node with 1 | |
b62a8d0c MD |
2379 | * child). OK to free lock here, because RCU read lock is held, |
2380 | * and free only performed in call_rcu. | |
af3cbd45 MD |
2381 | */ |
2382 | ||
2383 | for (i = 0; i < nr_clear; i++) { | |
2384 | ret = rcuja_shadow_clear(ja->ht, | |
3d8fe307 | 2385 | shadow_nodes[i]->node_flag, |
af3cbd45 MD |
2386 | shadow_nodes[i], |
2387 | RCUJA_SHADOW_CLEAR_FREE_NODE | |
2388 | | RCUJA_SHADOW_CLEAR_FREE_LOCK); | |
2389 | assert(!ret); | |
2390 | } | |
2391 | ||
2392 | iter_node_flag = parent_node_flag; | |
2393 | /* Remove from parent */ | |
2394 | ret = ja_node_clear_ptr(ja, | |
2395 | node_flag_ptr, /* Pointer to location to nullify */ | |
2396 | &iter_node_flag, /* Old new parent ptr in its parent */ | |
4d6ef45e | 2397 | shadow_nodes[nr_branch - 1], /* of parent */ |
48cbe001 | 2398 | n, nr_branch - 1); |
b306a0fe MD |
2399 | if (ret) |
2400 | goto end; | |
af3cbd45 | 2401 | |
4d6ef45e MD |
2402 | dbg_printf("ja_detach_node: publish %p instead of %p\n", |
2403 | iter_node_flag, *parent_node_flag_ptr); | |
af3cbd45 MD |
2404 | /* Update address of parent ptr in its parent */ |
2405 | rcu_assign_pointer(*parent_node_flag_ptr, iter_node_flag); | |
2406 | ||
2407 | end: | |
2408 | for (i = 0; i < nr_shadow; i++) | |
2409 | rcuja_shadow_unlock(shadow_nodes[i]); | |
35170a44 MD |
2410 | return ret; |
2411 | } | |
2412 | ||
af3cbd45 MD |
2413 | static |
2414 | int ja_unchain_node(struct cds_ja *ja, | |
2415 | struct cds_ja_inode_flag *parent_node_flag, | |
fa112799 | 2416 | struct cds_ja_inode_flag **node_flag_ptr, |
013a6083 | 2417 | struct cds_ja_inode_flag *node_flag, |
af3cbd45 MD |
2418 | struct cds_ja_node *node) |
2419 | { | |
2420 | struct cds_ja_shadow_node *shadow_node; | |
03ec1aeb | 2421 | struct cds_ja_node *iter_node, **iter_node_ptr, **prev_node_ptr = NULL; |
013a6083 | 2422 | int ret = 0, count = 0, found = 0; |
af3cbd45 | 2423 | |
3d8fe307 | 2424 | shadow_node = rcuja_shadow_lookup_lock(ja->ht, parent_node_flag); |
af3cbd45 MD |
2425 | if (!shadow_node) |
2426 | return -EAGAIN; | |
013a6083 | 2427 | if (ja_node_ptr(*node_flag_ptr) != ja_node_ptr(node_flag)) { |
fa112799 MD |
2428 | ret = -EAGAIN; |
2429 | goto end; | |
2430 | } | |
af3cbd45 | 2431 | /* |
03ec1aeb MD |
2432 | * Find the previous node's next pointer pointing to our node, |
2433 | * so we can update it. Retry if another thread removed all but | |
2434 | * one of duplicates since check (this check was performed | |
2435 | * without lock). Ensure that the node we are about to remove is | |
2436 | * still in the list (while holding lock). No need for RCU | |
2437 | * traversal here since we hold the lock on the parent. | |
af3cbd45 | 2438 | */ |
03ec1aeb MD |
2439 | iter_node_ptr = (struct cds_ja_node **) node_flag_ptr; |
2440 | iter_node = (struct cds_ja_node *) ja_node_ptr(node_flag); | |
2441 | cds_ja_for_each_duplicate(iter_node) { | |
f2758d14 | 2442 | count++; |
03ec1aeb MD |
2443 | if (iter_node == node) { |
2444 | prev_node_ptr = iter_node_ptr; | |
013a6083 | 2445 | found++; |
03ec1aeb MD |
2446 | } |
2447 | iter_node_ptr = &iter_node->next; | |
f2758d14 | 2448 | } |
013a6083 MD |
2449 | assert(found <= 1); |
2450 | if (!found || count == 1) { | |
af3cbd45 MD |
2451 | ret = -EAGAIN; |
2452 | goto end; | |
2453 | } | |
03ec1aeb | 2454 | CMM_STORE_SHARED(*prev_node_ptr, node->next); |
ade342cb MD |
2455 | /* |
2456 | * Validate that we indeed removed the node from linked list. | |
2457 | */ | |
2458 | assert(ja_node_ptr(*node_flag_ptr) != (struct cds_ja_inode *) node); | |
af3cbd45 MD |
2459 | end: |
2460 | rcuja_shadow_unlock(shadow_node); | |
2461 | return ret; | |
2462 | } | |
2463 | ||
2464 | /* | |
2465 | * Called with RCU read lock held. | |
2466 | */ | |
35170a44 MD |
2467 | int cds_ja_del(struct cds_ja *ja, uint64_t key, |
2468 | struct cds_ja_node *node) | |
2469 | { | |
2470 | unsigned int tree_depth, i; | |
2471 | struct cds_ja_inode_flag *snapshot[JA_MAX_DEPTH]; | |
af3cbd45 MD |
2472 | struct cds_ja_inode_flag **snapshot_ptr[JA_MAX_DEPTH]; |
2473 | uint8_t snapshot_n[JA_MAX_DEPTH]; | |
35170a44 | 2474 | struct cds_ja_inode_flag *node_flag; |
fa112799 MD |
2475 | struct cds_ja_inode_flag **prev_node_flag_ptr, |
2476 | **node_flag_ptr; | |
4d6ef45e | 2477 | int nr_snapshot; |
35170a44 MD |
2478 | int ret; |
2479 | ||
2480 | if (caa_unlikely(key > ja->key_max)) | |
2481 | return -EINVAL; | |
2482 | tree_depth = ja->tree_depth; | |
2483 | ||
2484 | retry: | |
4d6ef45e | 2485 | nr_snapshot = 0; |
35170a44 MD |
2486 | dbg_printf("cds_ja_del attempt: key %" PRIu64 ", node %p\n", |
2487 | key, node); | |
2488 | ||
2489 | /* snapshot for level 0 is only for shadow node lookup */ | |
4d6ef45e MD |
2490 | snapshot_n[0] = 0; |
2491 | snapshot_n[1] = 0; | |
af3cbd45 | 2492 | snapshot_ptr[nr_snapshot] = NULL; |
35170a44 MD |
2493 | snapshot[nr_snapshot++] = (struct cds_ja_inode_flag *) &ja->root; |
2494 | node_flag = rcu_dereference(ja->root); | |
af3cbd45 | 2495 | prev_node_flag_ptr = &ja->root; |
fa112799 | 2496 | node_flag_ptr = &ja->root; |
35170a44 MD |
2497 | |
2498 | /* Iterate on all internal levels */ | |
2499 | for (i = 1; i < tree_depth; i++) { | |
2500 | uint8_t iter_key; | |
2501 | ||
2502 | dbg_printf("cds_ja_del iter node_flag %p\n", | |
2503 | node_flag); | |
2504 | if (!ja_node_ptr(node_flag)) { | |
2505 | return -ENOENT; | |
2506 | } | |
35170a44 | 2507 | iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (tree_depth - i - 1))); |
4d6ef45e | 2508 | snapshot_n[nr_snapshot + 1] = iter_key; |
af3cbd45 MD |
2509 | snapshot_ptr[nr_snapshot] = prev_node_flag_ptr; |
2510 | snapshot[nr_snapshot++] = node_flag; | |
35170a44 | 2511 | node_flag = ja_node_get_nth(node_flag, |
fa112799 | 2512 | &node_flag_ptr, |
35170a44 | 2513 | iter_key); |
48cbe001 MD |
2514 | if (node_flag) |
2515 | prev_node_flag_ptr = node_flag_ptr; | |
af3cbd45 MD |
2516 | dbg_printf("cds_ja_del iter key lookup %u finds node_flag %p, prev_node_flag_ptr %p\n", |
2517 | (unsigned int) iter_key, node_flag, | |
2518 | prev_node_flag_ptr); | |
35170a44 | 2519 | } |
35170a44 MD |
2520 | /* |
2521 | * We reached bottom of tree, try to find the node we are trying | |
2522 | * to remove. Fail if we cannot find it. | |
2523 | */ | |
2524 | if (!ja_node_ptr(node_flag)) { | |
4d6ef45e MD |
2525 | dbg_printf("cds_ja_del: no node found for key %" PRIu64 "\n", |
2526 | key); | |
35170a44 MD |
2527 | return -ENOENT; |
2528 | } else { | |
03ec1aeb | 2529 | struct cds_ja_node *iter_node, *match = NULL; |
af3cbd45 | 2530 | int count = 0; |
35170a44 | 2531 | |
03ec1aeb MD |
2532 | iter_node = (struct cds_ja_node *) ja_node_ptr(node_flag); |
2533 | cds_ja_for_each_duplicate_rcu(iter_node) { | |
2534 | dbg_printf("cds_ja_del: compare %p with iter_node %p\n", node, iter_node); | |
2535 | if (iter_node == node) | |
2536 | match = iter_node; | |
af3cbd45 | 2537 | count++; |
35170a44 | 2538 | } |
03ec1aeb | 2539 | |
4d6ef45e MD |
2540 | if (!match) { |
2541 | dbg_printf("cds_ja_del: no node match for node %p key %" PRIu64 "\n", node, key); | |
35170a44 | 2542 | return -ENOENT; |
4d6ef45e | 2543 | } |
af3cbd45 MD |
2544 | assert(count > 0); |
2545 | if (count == 1) { | |
2546 | /* | |
4d6ef45e MD |
2547 | * Removing last of duplicates. Last snapshot |
2548 | * does not have a shadow node (external leafs). | |
af3cbd45 MD |
2549 | */ |
2550 | snapshot_ptr[nr_snapshot] = prev_node_flag_ptr; | |
2551 | snapshot[nr_snapshot++] = node_flag; | |
2552 | ret = ja_detach_node(ja, snapshot, snapshot_ptr, | |
2553 | snapshot_n, nr_snapshot, key, node); | |
2554 | } else { | |
f2758d14 | 2555 | ret = ja_unchain_node(ja, snapshot[nr_snapshot - 1], |
013a6083 | 2556 | node_flag_ptr, node_flag, match); |
af3cbd45 | 2557 | } |
35170a44 | 2558 | } |
b306a0fe MD |
2559 | /* |
2560 | * Explanation of -ENOENT handling: caused by concurrent delete | |
2561 | * between RCU lookup and actual removal. Need to re-do the | |
2562 | * lookup and removal attempt. | |
2563 | */ | |
2564 | if (ret == -EAGAIN || ret == -ENOENT) | |
35170a44 MD |
2565 | goto retry; |
2566 | return ret; | |
2567 | } | |
2568 | ||
b4540e8a MD |
2569 | struct cds_ja *_cds_ja_new(unsigned int key_bits, |
2570 | const struct rcu_flavor_struct *flavor) | |
be9a7474 MD |
2571 | { |
2572 | struct cds_ja *ja; | |
b0f74e47 | 2573 | int ret; |
f07b240f | 2574 | struct cds_ja_shadow_node *root_shadow_node; |
be9a7474 MD |
2575 | |
2576 | ja = calloc(sizeof(*ja), 1); | |
2577 | if (!ja) | |
2578 | goto ja_error; | |
b4540e8a MD |
2579 | |
2580 | switch (key_bits) { | |
2581 | case 8: | |
b4540e8a | 2582 | case 16: |
1216b3d2 | 2583 | case 24: |
b4540e8a | 2584 | case 32: |
1216b3d2 MD |
2585 | case 40: |
2586 | case 48: | |
2587 | case 56: | |
2588 | ja->key_max = (1ULL << key_bits) - 1; | |
b4540e8a MD |
2589 | break; |
2590 | case 64: | |
2591 | ja->key_max = UINT64_MAX; | |
2592 | break; | |
2593 | default: | |
2594 | goto check_error; | |
2595 | } | |
2596 | ||
be9a7474 | 2597 | /* ja->root is NULL */ |
5a9a87dd | 2598 | /* tree_depth 0 is for pointer to root node */ |
582a6ade | 2599 | ja->tree_depth = (key_bits >> JA_LOG2_BITS_PER_BYTE) + 1; |
a2a7ff59 | 2600 | assert(ja->tree_depth <= JA_MAX_DEPTH); |
be9a7474 MD |
2601 | ja->ht = rcuja_create_ht(flavor); |
2602 | if (!ja->ht) | |
2603 | goto ht_error; | |
b0f74e47 MD |
2604 | |
2605 | /* | |
2606 | * Note: we should not free this node until judy array destroy. | |
2607 | */ | |
f07b240f | 2608 | root_shadow_node = rcuja_shadow_set(ja->ht, |
3d8fe307 | 2609 | (struct cds_ja_inode_flag *) &ja->root, |
48cbe001 | 2610 | NULL, ja, 0); |
f07b240f MD |
2611 | if (!root_shadow_node) { |
2612 | ret = -ENOMEM; | |
b0f74e47 | 2613 | goto ht_node_error; |
f07b240f | 2614 | } |
b0f74e47 | 2615 | |
be9a7474 MD |
2616 | return ja; |
2617 | ||
b0f74e47 MD |
2618 | ht_node_error: |
2619 | ret = rcuja_delete_ht(ja->ht); | |
2620 | assert(!ret); | |
be9a7474 | 2621 | ht_error: |
b4540e8a | 2622 | check_error: |
be9a7474 MD |
2623 | free(ja); |
2624 | ja_error: | |
2625 | return NULL; | |
2626 | } | |
2627 | ||
3d8fe307 MD |
2628 | /* |
2629 | * Called from RCU read-side CS. | |
2630 | */ | |
2631 | __attribute__((visibility("protected"))) | |
2632 | void rcuja_free_all_children(struct cds_ja_shadow_node *shadow_node, | |
2633 | struct cds_ja_inode_flag *node_flag, | |
21ac4c56 | 2634 | void (*rcu_free_node)(struct cds_ja_node *node)) |
3d8fe307 | 2635 | { |
3d8fe307 MD |
2636 | unsigned int type_index; |
2637 | struct cds_ja_inode *node; | |
2638 | const struct cds_ja_type *type; | |
2639 | ||
3d8fe307 MD |
2640 | node = ja_node_ptr(node_flag); |
2641 | assert(node != NULL); | |
2642 | type_index = ja_node_type(node_flag); | |
2643 | type = &ja_types[type_index]; | |
2644 | ||
2645 | switch (type->type_class) { | |
2646 | case RCU_JA_LINEAR: | |
2647 | { | |
2648 | uint8_t nr_child = | |
2649 | ja_linear_node_get_nr_child(type, node); | |
2650 | unsigned int i; | |
2651 | ||
2652 | for (i = 0; i < nr_child; i++) { | |
2653 | struct cds_ja_inode_flag *iter; | |
03ec1aeb | 2654 | struct cds_ja_node *node_iter, *n; |
3d8fe307 MD |
2655 | uint8_t v; |
2656 | ||
2657 | ja_linear_node_get_ith_pos(type, node, i, &v, &iter); | |
03ec1aeb MD |
2658 | node_iter = (struct cds_ja_node *) iter; |
2659 | cds_ja_for_each_duplicate_safe(node_iter, n) { | |
2660 | rcu_free_node(node_iter); | |
3d8fe307 MD |
2661 | } |
2662 | } | |
2663 | break; | |
2664 | } | |
2665 | case RCU_JA_POOL: | |
2666 | { | |
2667 | unsigned int pool_nr; | |
2668 | ||
2669 | for (pool_nr = 0; pool_nr < (1U << type->nr_pool_order); pool_nr++) { | |
2670 | struct cds_ja_inode *pool = | |
2671 | ja_pool_node_get_ith_pool(type, node, pool_nr); | |
2672 | uint8_t nr_child = | |
2673 | ja_linear_node_get_nr_child(type, pool); | |
2674 | unsigned int j; | |
2675 | ||
2676 | for (j = 0; j < nr_child; j++) { | |
2677 | struct cds_ja_inode_flag *iter; | |
03ec1aeb | 2678 | struct cds_ja_node *node_iter, *n; |
3d8fe307 MD |
2679 | uint8_t v; |
2680 | ||
75d573aa | 2681 | ja_linear_node_get_ith_pos(type, pool, j, &v, &iter); |
03ec1aeb MD |
2682 | node_iter = (struct cds_ja_node *) iter; |
2683 | cds_ja_for_each_duplicate_safe(node_iter, n) { | |
2684 | rcu_free_node(node_iter); | |
3d8fe307 MD |
2685 | } |
2686 | } | |
2687 | } | |
2688 | break; | |
2689 | } | |
2690 | case RCU_JA_NULL: | |
2691 | break; | |
2692 | case RCU_JA_PIGEON: | |
2693 | { | |
3d8fe307 MD |
2694 | unsigned int i; |
2695 | ||
48cbe001 | 2696 | for (i = 0; i < JA_ENTRY_PER_NODE; i++) { |
3d8fe307 | 2697 | struct cds_ja_inode_flag *iter; |
03ec1aeb | 2698 | struct cds_ja_node *node_iter, *n; |
3d8fe307 MD |
2699 | |
2700 | iter = ja_pigeon_node_get_ith_pos(type, node, i); | |
03ec1aeb MD |
2701 | node_iter = (struct cds_ja_node *) iter; |
2702 | cds_ja_for_each_duplicate_safe(node_iter, n) { | |
2703 | rcu_free_node(node_iter); | |
3d8fe307 MD |
2704 | } |
2705 | } | |
2706 | break; | |
2707 | } | |
2708 | default: | |
2709 | assert(0); | |
2710 | } | |
2711 | } | |
2712 | ||
19ddcd04 | 2713 | static |
354981c2 | 2714 | void print_debug_fallback_distribution(struct cds_ja *ja) |
19ddcd04 MD |
2715 | { |
2716 | int i; | |
2717 | ||
2718 | fprintf(stderr, "Fallback node distribution:\n"); | |
2719 | for (i = 0; i < JA_ENTRY_PER_NODE; i++) { | |
354981c2 | 2720 | if (!ja->node_fallback_count_distribution[i]) |
19ddcd04 MD |
2721 | continue; |
2722 | fprintf(stderr, " %3u: %4lu\n", | |
354981c2 | 2723 | i, ja->node_fallback_count_distribution[i]); |
19ddcd04 MD |
2724 | } |
2725 | } | |
2726 | ||
021c72c0 | 2727 | static |
19a748d9 | 2728 | int ja_final_checks(struct cds_ja *ja) |
021c72c0 MD |
2729 | { |
2730 | double fallback_ratio; | |
2731 | unsigned long na, nf, nr_fallback; | |
19a748d9 | 2732 | int ret = 0; |
021c72c0 MD |
2733 | |
2734 | fallback_ratio = (double) uatomic_read(&ja->nr_fallback); | |
2735 | fallback_ratio /= (double) uatomic_read(&ja->nr_nodes_allocated); | |
2736 | nr_fallback = uatomic_read(&ja->nr_fallback); | |
2737 | if (nr_fallback) | |
2738 | fprintf(stderr, | |
2739 | "[warning] RCU Judy Array used %lu fallback node(s) (ratio: %g)\n", | |
2740 | uatomic_read(&ja->nr_fallback), | |
2741 | fallback_ratio); | |
2742 | ||
2743 | na = uatomic_read(&ja->nr_nodes_allocated); | |
2744 | nf = uatomic_read(&ja->nr_nodes_freed); | |
19a748d9 MD |
2745 | dbg_printf("Nodes allocated: %lu, Nodes freed: %lu.\n", na, nf); |
2746 | if (nr_fallback) | |
2747 | print_debug_fallback_distribution(ja); | |
2748 | ||
021c72c0 MD |
2749 | if (na != nf) { |
2750 | fprintf(stderr, "[error] Judy array leaked %ld nodes. Allocated: %lu, freed: %lu.\n", | |
2751 | (long) na - nf, na, nf); | |
19a748d9 | 2752 | ret = -1; |
021c72c0 | 2753 | } |
19a748d9 | 2754 | return ret; |
021c72c0 MD |
2755 | } |
2756 | ||
be9a7474 | 2757 | /* |
dc0e9798 MD |
2758 | * There should be no more concurrent add, delete, nor look-up performed |
2759 | * on the Judy array while it is being destroyed (ensured by the | |
2760 | * caller). | |
be9a7474 | 2761 | */ |
3d8fe307 | 2762 | int cds_ja_destroy(struct cds_ja *ja, |
dc0e9798 | 2763 | void (*free_node_cb)(struct cds_ja_node *node)) |
be9a7474 | 2764 | { |
48cbe001 | 2765 | const struct rcu_flavor_struct *flavor; |
b4540e8a MD |
2766 | int ret; |
2767 | ||
48cbe001 | 2768 | flavor = cds_lfht_rcu_flavor(ja->ht); |
be9a7474 | 2769 | rcuja_shadow_prune(ja->ht, |
3d8fe307 | 2770 | RCUJA_SHADOW_CLEAR_FREE_NODE | RCUJA_SHADOW_CLEAR_FREE_LOCK, |
dc0e9798 | 2771 | free_node_cb); |
48cbe001 | 2772 | flavor->thread_offline(); |
b4540e8a MD |
2773 | ret = rcuja_delete_ht(ja->ht); |
2774 | if (ret) | |
2775 | return ret; | |
f2ae7af7 MD |
2776 | |
2777 | /* Wait for in-flight call_rcu free to complete. */ | |
2778 | flavor->barrier(); | |
2779 | ||
48cbe001 | 2780 | flavor->thread_online(); |
19a748d9 | 2781 | ret = ja_final_checks(ja); |
b4540e8a | 2782 | free(ja); |
19a748d9 | 2783 | return ret; |
be9a7474 | 2784 | } |