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 | 36 | #include "rcuja-internal.h" |
d68c6810 | 37 | #include "bitfield.h" |
61009379 | 38 | |
b1a90ce3 MD |
39 | #ifndef abs |
40 | #define abs_int(a) ((int) (a) > 0 ? (int) (a) : -((int) (a))) | |
41 | #endif | |
42 | ||
d96bfb0d | 43 | enum cds_ja_type_class { |
e5227865 | 44 | RCU_JA_LINEAR = 0, /* Type A */ |
fd800776 MD |
45 | /* 32-bit: 1 to 25 children, 8 to 128 bytes */ |
46 | /* 64-bit: 1 to 28 children, 16 to 256 bytes */ | |
47 | RCU_JA_POOL = 1, /* Type B */ | |
48 | /* 32-bit: 26 to 100 children, 256 to 512 bytes */ | |
49 | /* 64-bit: 29 to 112 children, 512 to 1024 bytes */ | |
e5227865 | 50 | RCU_JA_PIGEON = 2, /* Type C */ |
fd800776 MD |
51 | /* 32-bit: 101 to 256 children, 1024 bytes */ |
52 | /* 64-bit: 113 to 256 children, 2048 bytes */ | |
e5227865 | 53 | /* Leaf nodes are implicit from their height in the tree */ |
1db4943c | 54 | RCU_JA_NR_TYPES, |
e1db2db5 MD |
55 | |
56 | RCU_JA_NULL, /* not an encoded type, but keeps code regular */ | |
e5227865 MD |
57 | }; |
58 | ||
d96bfb0d MD |
59 | struct cds_ja_type { |
60 | enum cds_ja_type_class type_class; | |
8e519e3c MD |
61 | uint16_t min_child; /* minimum number of children: 1 to 256 */ |
62 | uint16_t max_child; /* maximum number of children: 1 to 256 */ | |
63 | uint16_t max_linear_child; /* per-pool max nr. children: 1 to 256 */ | |
64 | uint16_t order; /* node size is (1 << order), in bytes */ | |
fd800776 MD |
65 | uint16_t nr_pool_order; /* number of pools */ |
66 | uint16_t pool_size_order; /* pool size */ | |
e5227865 MD |
67 | }; |
68 | ||
69 | /* | |
70 | * Iteration on the array to find the right node size for the number of | |
d68c6810 | 71 | * children stops when it reaches .max_child == 256 (this is the largest |
e5227865 | 72 | * possible node size, which contains 256 children). |
d68c6810 MD |
73 | * The min_child overlaps with the previous max_child to provide an |
74 | * hysteresis loop to reallocation for patterns of cyclic add/removal | |
75 | * within the same node. | |
76 | * The node the index within the following arrays is represented on 3 | |
77 | * bits. It identifies the node type, min/max number of children, and | |
78 | * the size order. | |
3d45251f MD |
79 | * The max_child values for the RCU_JA_POOL below result from |
80 | * statistical approximation: over million populations, the max_child | |
81 | * covers between 97% and 99% of the populations generated. Therefore, a | |
82 | * fallback should exist to cover the rare extreme population unbalance | |
83 | * cases, but it will not have a major impact on speed nor space | |
84 | * consumption, since those are rare cases. | |
e5227865 | 85 | */ |
e5227865 | 86 | |
d68c6810 MD |
87 | #if (CAA_BITS_PER_LONG < 64) |
88 | /* 32-bit pointers */ | |
1db4943c MD |
89 | enum { |
90 | ja_type_0_max_child = 1, | |
91 | ja_type_1_max_child = 3, | |
92 | ja_type_2_max_child = 6, | |
93 | ja_type_3_max_child = 12, | |
94 | ja_type_4_max_child = 25, | |
95 | ja_type_5_max_child = 48, | |
96 | ja_type_6_max_child = 92, | |
97 | ja_type_7_max_child = 256, | |
e1db2db5 | 98 | ja_type_8_max_child = 0, /* NULL */ |
1db4943c MD |
99 | }; |
100 | ||
8e519e3c MD |
101 | enum { |
102 | ja_type_0_max_linear_child = 1, | |
103 | ja_type_1_max_linear_child = 3, | |
104 | ja_type_2_max_linear_child = 6, | |
105 | ja_type_3_max_linear_child = 12, | |
106 | ja_type_4_max_linear_child = 25, | |
107 | ja_type_5_max_linear_child = 24, | |
108 | ja_type_6_max_linear_child = 23, | |
109 | }; | |
110 | ||
1db4943c MD |
111 | enum { |
112 | ja_type_5_nr_pool_order = 1, | |
113 | ja_type_6_nr_pool_order = 2, | |
114 | }; | |
115 | ||
d96bfb0d | 116 | const struct cds_ja_type ja_types[] = { |
8e519e3c MD |
117 | { .type_class = RCU_JA_LINEAR, .min_child = 1, .max_child = ja_type_0_max_child, .max_linear_child = ja_type_0_max_linear_child, .order = 3, }, |
118 | { .type_class = RCU_JA_LINEAR, .min_child = 1, .max_child = ja_type_1_max_child, .max_linear_child = ja_type_1_max_linear_child, .order = 4, }, | |
119 | { .type_class = RCU_JA_LINEAR, .min_child = 3, .max_child = ja_type_2_max_child, .max_linear_child = ja_type_2_max_linear_child, .order = 5, }, | |
120 | { .type_class = RCU_JA_LINEAR, .min_child = 4, .max_child = ja_type_3_max_child, .max_linear_child = ja_type_3_max_linear_child, .order = 6, }, | |
121 | { .type_class = RCU_JA_LINEAR, .min_child = 10, .max_child = ja_type_4_max_child, .max_linear_child = ja_type_4_max_linear_child, .order = 7, }, | |
e5227865 | 122 | |
fd800776 | 123 | /* Pools may fill sooner than max_child */ |
8e519e3c MD |
124 | { .type_class = RCU_JA_POOL, .min_child = 20, .max_child = ja_type_5_max_child, .max_linear_child = ja_type_5_max_linear_child, .order = 8, .nr_pool_order = ja_type_5_nr_pool_order, .pool_size_order = 7, }, |
125 | { .type_class = RCU_JA_POOL, .min_child = 45, .max_child = ja_type_6_max_child, .max_linear_child = ja_type_6_max_linear_child, .order = 9, .nr_pool_order = ja_type_6_nr_pool_order, .pool_size_order = 7, }, | |
3d45251f MD |
126 | |
127 | /* | |
b1a90ce3 MD |
128 | * Upon node removal below min_child, if child pool is filled |
129 | * beyond capacity, we roll back to pigeon. | |
3d45251f | 130 | */ |
1db4943c | 131 | { .type_class = RCU_JA_PIGEON, .min_child = 89, .max_child = ja_type_7_max_child, .order = 10, }, |
e1db2db5 MD |
132 | |
133 | { .type_class = RCU_JA_NULL, .min_child = 0, .max_child = ja_type_8_max_child, }, | |
d68c6810 | 134 | }; |
d68c6810 MD |
135 | #else /* !(CAA_BITS_PER_LONG < 64) */ |
136 | /* 64-bit pointers */ | |
1db4943c MD |
137 | enum { |
138 | ja_type_0_max_child = 1, | |
139 | ja_type_1_max_child = 3, | |
140 | ja_type_2_max_child = 7, | |
141 | ja_type_3_max_child = 14, | |
142 | ja_type_4_max_child = 28, | |
143 | ja_type_5_max_child = 54, | |
144 | ja_type_6_max_child = 104, | |
145 | ja_type_7_max_child = 256, | |
e1db2db5 | 146 | ja_type_8_max_child = 256, |
1db4943c MD |
147 | }; |
148 | ||
8e519e3c MD |
149 | enum { |
150 | ja_type_0_max_linear_child = 1, | |
151 | ja_type_1_max_linear_child = 3, | |
152 | ja_type_2_max_linear_child = 7, | |
153 | ja_type_3_max_linear_child = 14, | |
154 | ja_type_4_max_linear_child = 28, | |
155 | ja_type_5_max_linear_child = 27, | |
156 | ja_type_6_max_linear_child = 26, | |
157 | }; | |
158 | ||
1db4943c MD |
159 | enum { |
160 | ja_type_5_nr_pool_order = 1, | |
161 | ja_type_6_nr_pool_order = 2, | |
162 | }; | |
163 | ||
d96bfb0d | 164 | const struct cds_ja_type ja_types[] = { |
8e519e3c MD |
165 | { .type_class = RCU_JA_LINEAR, .min_child = 1, .max_child = ja_type_0_max_child, .max_linear_child = ja_type_0_max_linear_child, .order = 4, }, |
166 | { .type_class = RCU_JA_LINEAR, .min_child = 1, .max_child = ja_type_1_max_child, .max_linear_child = ja_type_1_max_linear_child, .order = 5, }, | |
167 | { .type_class = RCU_JA_LINEAR, .min_child = 3, .max_child = ja_type_2_max_child, .max_linear_child = ja_type_2_max_linear_child, .order = 6, }, | |
168 | { .type_class = RCU_JA_LINEAR, .min_child = 5, .max_child = ja_type_3_max_child, .max_linear_child = ja_type_3_max_linear_child, .order = 7, }, | |
169 | { .type_class = RCU_JA_LINEAR, .min_child = 10, .max_child = ja_type_4_max_child, .max_linear_child = ja_type_4_max_linear_child, .order = 8, }, | |
e5227865 | 170 | |
3d45251f | 171 | /* Pools may fill sooner than max_child. */ |
8e519e3c MD |
172 | { .type_class = RCU_JA_POOL, .min_child = 22, .max_child = ja_type_5_max_child, .max_linear_child = ja_type_5_max_linear_child, .order = 9, .nr_pool_order = ja_type_5_nr_pool_order, .pool_size_order = 8, }, |
173 | { .type_class = RCU_JA_POOL, .min_child = 51, .max_child = ja_type_6_max_child, .max_linear_child = ja_type_6_max_linear_child, .order = 10, .nr_pool_order = ja_type_6_nr_pool_order, .pool_size_order = 8, }, | |
e5227865 | 174 | |
3d45251f | 175 | /* |
b1a90ce3 MD |
176 | * Upon node removal below min_child, if child pool is filled |
177 | * beyond capacity, we roll back to pigeon. | |
3d45251f | 178 | */ |
1db4943c | 179 | { .type_class = RCU_JA_PIGEON, .min_child = 101, .max_child = ja_type_7_max_child, .order = 11, }, |
e1db2db5 MD |
180 | |
181 | { .type_class = RCU_JA_NULL, .min_child = 0, .max_child = ja_type_8_max_child, }, | |
e5227865 | 182 | }; |
d68c6810 | 183 | #endif /* !(BITS_PER_LONG < 64) */ |
e5227865 | 184 | |
1db4943c MD |
185 | static inline __attribute__((unused)) |
186 | void static_array_size_check(void) | |
187 | { | |
e1db2db5 | 188 | CAA_BUILD_BUG_ON(CAA_ARRAY_SIZE(ja_types) < JA_TYPE_MAX_NR); |
1db4943c MD |
189 | } |
190 | ||
e5227865 | 191 | /* |
d96bfb0d | 192 | * The cds_ja_node contains the compressed node data needed for |
1db4943c MD |
193 | * read-side. For linear and pool node configurations, it starts with a |
194 | * byte counting the number of children in the node. Then, the | |
195 | * node-specific data is placed. | |
196 | * The node mutex, if any is needed, protecting concurrent updated of | |
197 | * each node is placed in a separate hash table indexed by node address. | |
198 | * For the pigeon configuration, the number of children is also kept in | |
199 | * a separate hash table, indexed by node address, because it is only | |
200 | * required for updates. | |
e5227865 | 201 | */ |
1db4943c | 202 | |
ff38c745 MD |
203 | #define DECLARE_LINEAR_NODE(index) \ |
204 | struct { \ | |
205 | uint8_t nr_child; \ | |
206 | uint8_t child_value[ja_type_## index ##_max_linear_child]; \ | |
b4540e8a | 207 | struct cds_ja_inode_flag *child_ptr[ja_type_## index ##_max_linear_child]; \ |
ff38c745 MD |
208 | } |
209 | ||
210 | #define DECLARE_POOL_NODE(index) \ | |
211 | struct { \ | |
212 | struct { \ | |
213 | uint8_t nr_child; \ | |
214 | uint8_t child_value[ja_type_## index ##_max_linear_child]; \ | |
b4540e8a | 215 | struct cds_ja_inode_flag *child_ptr[ja_type_## index ##_max_linear_child]; \ |
ff38c745 MD |
216 | } linear[1U << ja_type_## index ##_nr_pool_order]; \ |
217 | } | |
1db4943c | 218 | |
b4540e8a | 219 | struct cds_ja_inode { |
1db4943c MD |
220 | union { |
221 | /* Linear configuration */ | |
222 | DECLARE_LINEAR_NODE(0) conf_0; | |
223 | DECLARE_LINEAR_NODE(1) conf_1; | |
224 | DECLARE_LINEAR_NODE(2) conf_2; | |
225 | DECLARE_LINEAR_NODE(3) conf_3; | |
226 | DECLARE_LINEAR_NODE(4) conf_4; | |
227 | ||
228 | /* Pool configuration */ | |
229 | DECLARE_POOL_NODE(5) conf_5; | |
230 | DECLARE_POOL_NODE(6) conf_6; | |
231 | ||
232 | /* Pigeon configuration */ | |
233 | struct { | |
b4540e8a | 234 | struct cds_ja_inode_flag *child[ja_type_7_max_child]; |
1db4943c MD |
235 | } conf_7; |
236 | /* data aliasing nodes for computed accesses */ | |
b4540e8a | 237 | uint8_t data[sizeof(struct cds_ja_inode_flag *) * ja_type_7_max_child]; |
1db4943c | 238 | } u; |
e5227865 MD |
239 | }; |
240 | ||
2e313670 MD |
241 | enum ja_recompact { |
242 | JA_RECOMPACT, | |
243 | JA_RECOMPACT_ADD, | |
244 | JA_RECOMPACT_DEL, | |
245 | }; | |
246 | ||
b1a90ce3 MD |
247 | static |
248 | struct cds_ja_inode *_ja_node_mask_ptr(struct cds_ja_inode_flag *node) | |
249 | { | |
250 | return (struct cds_ja_inode *) (((unsigned long) node) & JA_PTR_MASK); | |
251 | } | |
252 | ||
253 | unsigned long ja_node_type(struct cds_ja_inode_flag *node) | |
254 | { | |
255 | unsigned long type; | |
256 | ||
257 | if (_ja_node_mask_ptr(node) == NULL) { | |
258 | return NODE_INDEX_NULL; | |
259 | } | |
260 | type = (unsigned int) ((unsigned long) node & JA_TYPE_MASK); | |
261 | assert(type < (1UL << JA_TYPE_BITS)); | |
262 | return type; | |
263 | } | |
264 | ||
265 | struct cds_ja_inode *ja_node_ptr(struct cds_ja_inode_flag *node) | |
266 | { | |
267 | unsigned long type_index = ja_node_type(node); | |
268 | const struct cds_ja_type *type; | |
269 | ||
270 | type = &ja_types[type_index]; | |
271 | switch (type->type_class) { | |
272 | case RCU_JA_LINEAR: | |
273 | case RCU_JA_PIGEON: /* fall-through */ | |
274 | case RCU_JA_NULL: /* fall-through */ | |
275 | default: /* fall-through */ | |
276 | return _ja_node_mask_ptr(node); | |
277 | case RCU_JA_POOL: | |
278 | switch (type->nr_pool_order) { | |
279 | case 1: | |
280 | return (struct cds_ja_inode *) (((unsigned long) node) & ~(JA_POOL_1D_MASK | JA_TYPE_MASK)); | |
281 | case 2: | |
282 | return (struct cds_ja_inode *) (((unsigned long) node) & ~(JA_POOL_2D_MASK | JA_POOL_1D_MASK | JA_TYPE_MASK)); | |
283 | default: | |
284 | assert(0); | |
285 | } | |
286 | } | |
287 | } | |
288 | ||
b4540e8a | 289 | struct cds_ja_inode *alloc_cds_ja_node(const struct cds_ja_type *ja_type) |
e5227865 | 290 | { |
b1a90ce3 MD |
291 | size_t len = 1U << ja_type->order; |
292 | void *p; | |
293 | int ret; | |
294 | ||
295 | ret = posix_memalign(&p, len, len); | |
296 | if (ret || !p) { | |
297 | return NULL; | |
298 | } | |
299 | memset(p, 0, len); | |
300 | return p; | |
e5227865 MD |
301 | } |
302 | ||
b4540e8a | 303 | void free_cds_ja_node(struct cds_ja_inode *node) |
e5227865 MD |
304 | { |
305 | free(node); | |
306 | } | |
307 | ||
d68c6810 MD |
308 | #define __JA_ALIGN_MASK(v, mask) (((v) + (mask)) & ~(mask)) |
309 | #define JA_ALIGN(v, align) __JA_ALIGN_MASK(v, (typeof(v)) (align) - 1) | |
310 | #define __JA_FLOOR_MASK(v, mask) ((v) & ~(mask)) | |
311 | #define JA_FLOOR(v, align) __JA_FLOOR_MASK(v, (typeof(v)) (align) - 1) | |
312 | ||
313 | static | |
1db4943c | 314 | uint8_t *align_ptr_size(uint8_t *ptr) |
d68c6810 | 315 | { |
1db4943c | 316 | return (uint8_t *) JA_ALIGN((unsigned long) ptr, sizeof(void *)); |
d68c6810 MD |
317 | } |
318 | ||
11c5e016 | 319 | static |
d96bfb0d | 320 | uint8_t ja_linear_node_get_nr_child(const struct cds_ja_type *type, |
b4540e8a | 321 | struct cds_ja_inode *node) |
11c5e016 MD |
322 | { |
323 | assert(type->type_class == RCU_JA_LINEAR || type->type_class == RCU_JA_POOL); | |
2e313670 | 324 | return rcu_dereference(node->u.data[0]); |
11c5e016 MD |
325 | } |
326 | ||
13a7f5a6 MD |
327 | /* |
328 | * The order in which values and pointers are does does not matter: if | |
329 | * a value is missing, we return NULL. If a value is there, but its | |
330 | * associated pointers is still NULL, we return NULL too. | |
331 | */ | |
d68c6810 | 332 | static |
b4540e8a MD |
333 | struct cds_ja_inode_flag *ja_linear_node_get_nth(const struct cds_ja_type *type, |
334 | struct cds_ja_inode *node, | |
5a9a87dd | 335 | struct cds_ja_inode_flag ***child_node_flag_ptr, |
b62a8d0c | 336 | struct cds_ja_inode_flag **child_node_flag_v, |
b0ca2d21 | 337 | struct cds_ja_inode_flag ***node_flag_ptr, |
8e519e3c | 338 | uint8_t n) |
d68c6810 MD |
339 | { |
340 | uint8_t nr_child; | |
341 | uint8_t *values; | |
b4540e8a MD |
342 | struct cds_ja_inode_flag **pointers; |
343 | struct cds_ja_inode_flag *ptr; | |
d68c6810 MD |
344 | unsigned int i; |
345 | ||
8e519e3c | 346 | assert(type->type_class == RCU_JA_LINEAR || type->type_class == RCU_JA_POOL); |
d68c6810 | 347 | |
11c5e016 | 348 | nr_child = ja_linear_node_get_nr_child(type, node); |
13a7f5a6 | 349 | cmm_smp_rmb(); /* read nr_child before values and pointers */ |
8e519e3c MD |
350 | assert(nr_child <= type->max_linear_child); |
351 | assert(type->type_class != RCU_JA_LINEAR || nr_child >= type->min_child); | |
d68c6810 | 352 | |
1db4943c | 353 | values = &node->u.data[1]; |
d68c6810 | 354 | for (i = 0; i < nr_child; i++) { |
13a7f5a6 | 355 | if (CMM_LOAD_SHARED(values[i]) == n) |
d68c6810 MD |
356 | break; |
357 | } | |
b0ca2d21 MD |
358 | if (i >= nr_child) { |
359 | if (caa_unlikely(node_flag_ptr)) | |
360 | *node_flag_ptr = NULL; | |
d68c6810 | 361 | return NULL; |
b0ca2d21 | 362 | } |
b4540e8a | 363 | pointers = (struct cds_ja_inode_flag **) align_ptr_size(&values[type->max_linear_child]); |
13a7f5a6 | 364 | ptr = rcu_dereference(pointers[i]); |
2e313670 MD |
365 | if (caa_unlikely(child_node_flag_ptr) && ptr) |
366 | *child_node_flag_ptr = &pointers[i]; | |
b62a8d0c MD |
367 | if (caa_unlikely(child_node_flag_v) && ptr) |
368 | *child_node_flag_v = ptr; | |
b0ca2d21 MD |
369 | if (caa_unlikely(node_flag_ptr)) |
370 | *node_flag_ptr = &pointers[i]; | |
d68c6810 MD |
371 | return ptr; |
372 | } | |
373 | ||
11c5e016 | 374 | static |
5a9a87dd | 375 | void ja_linear_node_get_ith_pos(const struct cds_ja_type *type, |
b4540e8a | 376 | struct cds_ja_inode *node, |
11c5e016 MD |
377 | uint8_t i, |
378 | uint8_t *v, | |
b4540e8a | 379 | struct cds_ja_inode_flag **iter) |
11c5e016 MD |
380 | { |
381 | uint8_t *values; | |
b4540e8a | 382 | struct cds_ja_inode_flag **pointers; |
11c5e016 MD |
383 | |
384 | assert(type->type_class == RCU_JA_LINEAR || type->type_class == RCU_JA_POOL); | |
385 | assert(i < ja_linear_node_get_nr_child(type, node)); | |
386 | ||
387 | values = &node->u.data[1]; | |
388 | *v = values[i]; | |
b4540e8a | 389 | pointers = (struct cds_ja_inode_flag **) align_ptr_size(&values[type->max_linear_child]); |
11c5e016 MD |
390 | *iter = pointers[i]; |
391 | } | |
392 | ||
d68c6810 | 393 | static |
b4540e8a MD |
394 | struct cds_ja_inode_flag *ja_pool_node_get_nth(const struct cds_ja_type *type, |
395 | struct cds_ja_inode *node, | |
b1a90ce3 | 396 | struct cds_ja_inode_flag *node_flag, |
5a9a87dd | 397 | struct cds_ja_inode_flag ***child_node_flag_ptr, |
b62a8d0c | 398 | struct cds_ja_inode_flag **child_node_flag_v, |
b0ca2d21 | 399 | struct cds_ja_inode_flag ***node_flag_ptr, |
8e519e3c | 400 | uint8_t n) |
d68c6810 | 401 | { |
b4540e8a | 402 | struct cds_ja_inode *linear; |
d68c6810 | 403 | |
fd800776 | 404 | assert(type->type_class == RCU_JA_POOL); |
b1a90ce3 MD |
405 | |
406 | switch (type->nr_pool_order) { | |
407 | case 1: | |
408 | { | |
409 | unsigned long bitsel, index; | |
410 | ||
411 | bitsel = ja_node_pool_1d_bitsel(node_flag); | |
412 | assert(bitsel < CHAR_BIT); | |
413 | index = ((unsigned long) n >> bitsel) & type->nr_pool_order; | |
414 | linear = (struct cds_ja_inode *) &node->u.data[index << type->pool_size_order]; | |
415 | break; | |
416 | } | |
417 | case 2: | |
418 | { | |
419 | /* | |
420 | * TODO: currently, we select the pool by highest bits. We | |
421 | * should support various encodings. | |
422 | */ | |
423 | linear = (struct cds_ja_inode *) | |
424 | &node->u.data[((unsigned long) n >> (CHAR_BIT - type->nr_pool_order)) << type->pool_size_order]; | |
425 | break; | |
426 | } | |
427 | default: | |
428 | linear = NULL; | |
429 | assert(0); | |
430 | } | |
b0ca2d21 | 431 | return ja_linear_node_get_nth(type, linear, child_node_flag_ptr, |
b62a8d0c | 432 | child_node_flag_v, node_flag_ptr, n); |
d68c6810 MD |
433 | } |
434 | ||
11c5e016 | 435 | static |
b4540e8a MD |
436 | struct cds_ja_inode *ja_pool_node_get_ith_pool(const struct cds_ja_type *type, |
437 | struct cds_ja_inode *node, | |
11c5e016 MD |
438 | uint8_t i) |
439 | { | |
440 | assert(type->type_class == RCU_JA_POOL); | |
b4540e8a | 441 | return (struct cds_ja_inode *) |
11c5e016 MD |
442 | &node->u.data[(unsigned int) i << type->pool_size_order]; |
443 | } | |
444 | ||
d68c6810 | 445 | static |
b4540e8a MD |
446 | struct cds_ja_inode_flag *ja_pigeon_node_get_nth(const struct cds_ja_type *type, |
447 | struct cds_ja_inode *node, | |
5a9a87dd | 448 | struct cds_ja_inode_flag ***child_node_flag_ptr, |
b62a8d0c | 449 | struct cds_ja_inode_flag **child_node_flag_v, |
b0ca2d21 | 450 | struct cds_ja_inode_flag ***node_flag_ptr, |
8e519e3c | 451 | uint8_t n) |
d68c6810 | 452 | { |
5a9a87dd | 453 | struct cds_ja_inode_flag **child_node_flag; |
b62a8d0c | 454 | struct cds_ja_inode_flag *child_node_flag_read; |
5a9a87dd | 455 | |
d68c6810 | 456 | assert(type->type_class == RCU_JA_PIGEON); |
5a9a87dd | 457 | child_node_flag = &((struct cds_ja_inode_flag **) node->u.data)[n]; |
b62a8d0c | 458 | child_node_flag_read = rcu_dereference(*child_node_flag); |
582a6ade MD |
459 | dbg_printf("ja_pigeon_node_get_nth child_node_flag_ptr %p\n", |
460 | child_node_flag); | |
b62a8d0c | 461 | if (caa_unlikely(child_node_flag_ptr) && child_node_flag_read) |
5a9a87dd | 462 | *child_node_flag_ptr = child_node_flag; |
b62a8d0c MD |
463 | if (caa_unlikely(child_node_flag_v) && child_node_flag_read) |
464 | *child_node_flag_v = child_node_flag_read; | |
b0ca2d21 MD |
465 | if (caa_unlikely(node_flag_ptr)) |
466 | *node_flag_ptr = child_node_flag; | |
b62a8d0c | 467 | return child_node_flag_read; |
d68c6810 MD |
468 | } |
469 | ||
2e313670 MD |
470 | static |
471 | struct cds_ja_inode_flag *ja_pigeon_node_get_ith_pos(const struct cds_ja_type *type, | |
472 | struct cds_ja_inode *node, | |
473 | uint8_t i) | |
474 | { | |
b62a8d0c | 475 | return ja_pigeon_node_get_nth(type, node, NULL, NULL, NULL, i); |
2e313670 MD |
476 | } |
477 | ||
13a7f5a6 MD |
478 | /* |
479 | * ja_node_get_nth: get nth item from a node. | |
480 | * node_flag is already rcu_dereference'd. | |
481 | */ | |
d68c6810 | 482 | static |
b62a8d0c | 483 | struct cds_ja_inode_flag *ja_node_get_nth(struct cds_ja_inode_flag *node_flag, |
5a9a87dd | 484 | struct cds_ja_inode_flag ***child_node_flag_ptr, |
b62a8d0c | 485 | struct cds_ja_inode_flag **child_node_flag, |
b0ca2d21 | 486 | struct cds_ja_inode_flag ***node_flag_ptr, |
8e519e3c | 487 | uint8_t n) |
d68c6810 MD |
488 | { |
489 | unsigned int type_index; | |
b4540e8a | 490 | struct cds_ja_inode *node; |
d96bfb0d | 491 | const struct cds_ja_type *type; |
d68c6810 | 492 | |
d68c6810 | 493 | node = ja_node_ptr(node_flag); |
5a9a87dd | 494 | assert(node != NULL); |
d68c6810 MD |
495 | type_index = ja_node_type(node_flag); |
496 | type = &ja_types[type_index]; | |
497 | ||
498 | switch (type->type_class) { | |
499 | case RCU_JA_LINEAR: | |
5a9a87dd | 500 | return ja_linear_node_get_nth(type, node, |
b62a8d0c MD |
501 | child_node_flag_ptr, child_node_flag, |
502 | node_flag_ptr, n); | |
fd800776 | 503 | case RCU_JA_POOL: |
b1a90ce3 | 504 | return ja_pool_node_get_nth(type, node, node_flag, |
b62a8d0c MD |
505 | child_node_flag_ptr, child_node_flag, |
506 | node_flag_ptr, n); | |
d68c6810 | 507 | case RCU_JA_PIGEON: |
5a9a87dd | 508 | return ja_pigeon_node_get_nth(type, node, |
b62a8d0c MD |
509 | child_node_flag_ptr, child_node_flag, |
510 | node_flag_ptr, n); | |
d68c6810 MD |
511 | default: |
512 | assert(0); | |
513 | return (void *) -1UL; | |
514 | } | |
515 | } | |
516 | ||
8e519e3c | 517 | static |
d96bfb0d | 518 | int ja_linear_node_set_nth(const struct cds_ja_type *type, |
b4540e8a | 519 | struct cds_ja_inode *node, |
d96bfb0d | 520 | struct cds_ja_shadow_node *shadow_node, |
8e519e3c | 521 | uint8_t n, |
b4540e8a | 522 | struct cds_ja_inode_flag *child_node_flag) |
8e519e3c MD |
523 | { |
524 | uint8_t nr_child; | |
525 | uint8_t *values, *nr_child_ptr; | |
b4540e8a | 526 | struct cds_ja_inode_flag **pointers; |
2e313670 | 527 | unsigned int i, unused = 0; |
8e519e3c MD |
528 | |
529 | assert(type->type_class == RCU_JA_LINEAR || type->type_class == RCU_JA_POOL); | |
530 | ||
531 | nr_child_ptr = &node->u.data[0]; | |
a2a7ff59 | 532 | dbg_printf("linear set nth: nr_child_ptr %p\n", nr_child_ptr); |
8e519e3c MD |
533 | nr_child = *nr_child_ptr; |
534 | assert(nr_child <= type->max_linear_child); | |
8e519e3c MD |
535 | |
536 | values = &node->u.data[1]; | |
2e313670 MD |
537 | pointers = (struct cds_ja_inode_flag **) align_ptr_size(&values[type->max_linear_child]); |
538 | /* Check if node value is already populated */ | |
8e519e3c | 539 | for (i = 0; i < nr_child; i++) { |
2e313670 MD |
540 | if (values[i] == n) { |
541 | if (pointers[i]) | |
542 | return -EEXIST; | |
543 | else | |
544 | break; | |
545 | } else { | |
546 | if (!pointers[i]) | |
547 | unused++; | |
548 | } | |
8e519e3c | 549 | } |
2e313670 MD |
550 | if (i == nr_child && nr_child >= type->max_linear_child) { |
551 | if (unused) | |
552 | return -ERANGE; /* recompact node */ | |
553 | else | |
554 | return -ENOSPC; /* No space left in this node type */ | |
555 | } | |
556 | ||
557 | assert(pointers[i] == NULL); | |
558 | rcu_assign_pointer(pointers[i], child_node_flag); | |
559 | /* If we expanded the nr_child, increment it */ | |
560 | if (i == nr_child) { | |
561 | CMM_STORE_SHARED(values[nr_child], n); | |
562 | /* write pointer and value before nr_child */ | |
563 | cmm_smp_wmb(); | |
564 | CMM_STORE_SHARED(*nr_child_ptr, nr_child + 1); | |
8e519e3c | 565 | } |
e1db2db5 | 566 | shadow_node->nr_child++; |
a2a7ff59 MD |
567 | dbg_printf("linear set nth: %u child, shadow: %u child, for node %p shadow %p\n", |
568 | (unsigned int) CMM_LOAD_SHARED(*nr_child_ptr), | |
569 | (unsigned int) shadow_node->nr_child, | |
570 | node, shadow_node); | |
571 | ||
8e519e3c MD |
572 | return 0; |
573 | } | |
574 | ||
575 | static | |
d96bfb0d | 576 | int ja_pool_node_set_nth(const struct cds_ja_type *type, |
b4540e8a | 577 | struct cds_ja_inode *node, |
b1a90ce3 | 578 | struct cds_ja_inode_flag *node_flag, |
d96bfb0d | 579 | struct cds_ja_shadow_node *shadow_node, |
8e519e3c | 580 | uint8_t n, |
b4540e8a | 581 | struct cds_ja_inode_flag *child_node_flag) |
8e519e3c | 582 | { |
b4540e8a | 583 | struct cds_ja_inode *linear; |
8e519e3c MD |
584 | |
585 | assert(type->type_class == RCU_JA_POOL); | |
b1a90ce3 MD |
586 | |
587 | switch (type->nr_pool_order) { | |
588 | case 1: | |
589 | { | |
590 | unsigned long bitsel, index; | |
591 | ||
592 | bitsel = ja_node_pool_1d_bitsel(node_flag); | |
593 | assert(bitsel < CHAR_BIT); | |
594 | index = ((unsigned long) n >> bitsel) & type->nr_pool_order; | |
595 | linear = (struct cds_ja_inode *) &node->u.data[index << type->pool_size_order]; | |
596 | break; | |
597 | } | |
598 | case 2: | |
599 | { | |
600 | /* | |
601 | * TODO: currently, we select the pool by highest bits. We | |
602 | * should support various encodings. | |
603 | */ | |
604 | linear = (struct cds_ja_inode *) | |
605 | &node->u.data[((unsigned long) n >> (CHAR_BIT - type->nr_pool_order)) << type->pool_size_order]; | |
606 | break; | |
607 | } | |
608 | default: | |
609 | linear = NULL; | |
610 | assert(0); | |
611 | } | |
612 | ||
e1db2db5 MD |
613 | return ja_linear_node_set_nth(type, linear, shadow_node, |
614 | n, child_node_flag); | |
8e519e3c MD |
615 | } |
616 | ||
617 | static | |
d96bfb0d | 618 | int ja_pigeon_node_set_nth(const struct cds_ja_type *type, |
b4540e8a | 619 | struct cds_ja_inode *node, |
d96bfb0d | 620 | struct cds_ja_shadow_node *shadow_node, |
8e519e3c | 621 | uint8_t n, |
b4540e8a | 622 | struct cds_ja_inode_flag *child_node_flag) |
8e519e3c | 623 | { |
b4540e8a | 624 | struct cds_ja_inode_flag **ptr; |
8e519e3c MD |
625 | |
626 | assert(type->type_class == RCU_JA_PIGEON); | |
b4540e8a | 627 | ptr = &((struct cds_ja_inode_flag **) node->u.data)[n]; |
5a9a87dd | 628 | if (*ptr) |
8e519e3c MD |
629 | return -EEXIST; |
630 | rcu_assign_pointer(*ptr, child_node_flag); | |
e1db2db5 | 631 | shadow_node->nr_child++; |
8e519e3c MD |
632 | return 0; |
633 | } | |
634 | ||
d68c6810 | 635 | /* |
7a0b2331 | 636 | * _ja_node_set_nth: set nth item within a node. Return an error |
8e519e3c | 637 | * (negative error value) if it is already there. |
d68c6810 | 638 | */ |
8e519e3c | 639 | static |
d96bfb0d | 640 | int _ja_node_set_nth(const struct cds_ja_type *type, |
b4540e8a | 641 | struct cds_ja_inode *node, |
b1a90ce3 | 642 | struct cds_ja_inode_flag *node_flag, |
d96bfb0d | 643 | struct cds_ja_shadow_node *shadow_node, |
e1db2db5 | 644 | uint8_t n, |
b4540e8a | 645 | struct cds_ja_inode_flag *child_node_flag) |
8e519e3c | 646 | { |
8e519e3c MD |
647 | switch (type->type_class) { |
648 | case RCU_JA_LINEAR: | |
e1db2db5 | 649 | return ja_linear_node_set_nth(type, node, shadow_node, n, |
8e519e3c MD |
650 | child_node_flag); |
651 | case RCU_JA_POOL: | |
b1a90ce3 | 652 | return ja_pool_node_set_nth(type, node, node_flag, shadow_node, n, |
8e519e3c MD |
653 | child_node_flag); |
654 | case RCU_JA_PIGEON: | |
e1db2db5 | 655 | return ja_pigeon_node_set_nth(type, node, shadow_node, n, |
8e519e3c | 656 | child_node_flag); |
e1db2db5 MD |
657 | case RCU_JA_NULL: |
658 | return -ENOSPC; | |
8e519e3c MD |
659 | default: |
660 | assert(0); | |
661 | return -EINVAL; | |
662 | } | |
663 | ||
664 | return 0; | |
665 | } | |
7a0b2331 | 666 | |
2e313670 | 667 | static |
af3cbd45 | 668 | int ja_linear_node_clear_ptr(const struct cds_ja_type *type, |
2e313670 MD |
669 | struct cds_ja_inode *node, |
670 | struct cds_ja_shadow_node *shadow_node, | |
af3cbd45 | 671 | struct cds_ja_inode_flag **node_flag_ptr) |
2e313670 MD |
672 | { |
673 | uint8_t nr_child; | |
af3cbd45 | 674 | uint8_t *nr_child_ptr; |
2e313670 MD |
675 | |
676 | assert(type->type_class == RCU_JA_LINEAR || type->type_class == RCU_JA_POOL); | |
677 | ||
678 | nr_child_ptr = &node->u.data[0]; | |
af3cbd45 | 679 | dbg_printf("linear clear ptr: nr_child_ptr %p\n", nr_child_ptr); |
2e313670 MD |
680 | nr_child = *nr_child_ptr; |
681 | assert(nr_child <= type->max_linear_child); | |
682 | ||
2e313670 MD |
683 | if (shadow_node->fallback_removal_count) { |
684 | shadow_node->fallback_removal_count--; | |
685 | } else { | |
686 | if (shadow_node->nr_child <= type->min_child) { | |
687 | /* We need to try recompacting the node */ | |
688 | return -EFBIG; | |
689 | } | |
690 | } | |
af3cbd45 MD |
691 | assert(*node_flag_ptr != NULL); |
692 | rcu_assign_pointer(*node_flag_ptr, NULL); | |
2e313670 MD |
693 | /* |
694 | * Value and nr_child are never changed (would cause ABA issue). | |
695 | * Instead, we leave the pointer to NULL and recompact the node | |
696 | * once in a while. It is allowed to set a NULL pointer to a new | |
697 | * value without recompaction though. | |
698 | * Only update the shadow node accounting. | |
699 | */ | |
700 | shadow_node->nr_child--; | |
af3cbd45 | 701 | dbg_printf("linear clear ptr: %u child, shadow: %u child, for node %p shadow %p\n", |
2e313670 MD |
702 | (unsigned int) CMM_LOAD_SHARED(*nr_child_ptr), |
703 | (unsigned int) shadow_node->nr_child, | |
704 | node, shadow_node); | |
705 | ||
706 | return 0; | |
707 | } | |
708 | ||
709 | static | |
af3cbd45 | 710 | int ja_pool_node_clear_ptr(const struct cds_ja_type *type, |
2e313670 MD |
711 | struct cds_ja_inode *node, |
712 | struct cds_ja_shadow_node *shadow_node, | |
af3cbd45 | 713 | struct cds_ja_inode_flag **node_flag_ptr, |
2e313670 MD |
714 | uint8_t n) |
715 | { | |
716 | struct cds_ja_inode *linear; | |
717 | ||
718 | assert(type->type_class == RCU_JA_POOL); | |
719 | linear = (struct cds_ja_inode *) | |
720 | &node->u.data[((unsigned long) n >> (CHAR_BIT - type->nr_pool_order)) << type->pool_size_order]; | |
af3cbd45 | 721 | return ja_linear_node_clear_ptr(type, linear, shadow_node, node_flag_ptr); |
2e313670 MD |
722 | } |
723 | ||
724 | static | |
af3cbd45 | 725 | int ja_pigeon_node_clear_ptr(const struct cds_ja_type *type, |
2e313670 MD |
726 | struct cds_ja_inode *node, |
727 | struct cds_ja_shadow_node *shadow_node, | |
af3cbd45 | 728 | struct cds_ja_inode_flag **node_flag_ptr) |
2e313670 | 729 | { |
2e313670 | 730 | assert(type->type_class == RCU_JA_PIGEON); |
4d6ef45e | 731 | dbg_printf("ja_pigeon_node_clear_ptr: clearing ptr: %p\n", *node_flag_ptr); |
af3cbd45 | 732 | rcu_assign_pointer(*node_flag_ptr, NULL); |
2e313670 MD |
733 | shadow_node->nr_child--; |
734 | return 0; | |
735 | } | |
736 | ||
737 | /* | |
af3cbd45 | 738 | * _ja_node_clear_ptr: clear ptr item within a node. Return an error |
2e313670 MD |
739 | * (negative error value) if it is not found (-ENOENT). |
740 | */ | |
741 | static | |
af3cbd45 | 742 | int _ja_node_clear_ptr(const struct cds_ja_type *type, |
2e313670 MD |
743 | struct cds_ja_inode *node, |
744 | struct cds_ja_shadow_node *shadow_node, | |
af3cbd45 | 745 | struct cds_ja_inode_flag **node_flag_ptr, |
2e313670 MD |
746 | uint8_t n) |
747 | { | |
748 | switch (type->type_class) { | |
749 | case RCU_JA_LINEAR: | |
af3cbd45 | 750 | return ja_linear_node_clear_ptr(type, node, shadow_node, node_flag_ptr); |
2e313670 | 751 | case RCU_JA_POOL: |
af3cbd45 | 752 | return ja_pool_node_clear_ptr(type, node, shadow_node, node_flag_ptr, n); |
2e313670 | 753 | case RCU_JA_PIGEON: |
af3cbd45 | 754 | return ja_pigeon_node_clear_ptr(type, node, shadow_node, node_flag_ptr); |
2e313670 MD |
755 | case RCU_JA_NULL: |
756 | return -ENOENT; | |
757 | default: | |
758 | assert(0); | |
759 | return -EINVAL; | |
760 | } | |
761 | ||
762 | return 0; | |
763 | } | |
764 | ||
b1a90ce3 MD |
765 | /* |
766 | * Calculate bit distribution. Returns the bit (0 to 7) that splits the | |
767 | * distribution in two sub-distributions containing as much elements one | |
768 | * compared to the other. | |
769 | */ | |
770 | static | |
771 | unsigned int ja_node_sum_distribution_1d(enum ja_recompact mode, | |
772 | struct cds_ja *ja, | |
773 | unsigned int type_index, | |
774 | const struct cds_ja_type *type, | |
775 | struct cds_ja_inode *node, | |
776 | struct cds_ja_shadow_node *shadow_node, | |
777 | uint8_t n, | |
778 | struct cds_ja_inode_flag *child_node_flag, | |
779 | struct cds_ja_inode_flag **nullify_node_flag_ptr) | |
780 | { | |
781 | uint8_t nr_one[JA_BITS_PER_BYTE]; | |
782 | unsigned int bitsel = 0, bit_i, overall_best_distance = UINT_MAX; | |
783 | unsigned int distrib_nr_child = 0; | |
784 | ||
785 | memset(nr_one, 0, sizeof(nr_one)); | |
786 | ||
787 | switch (type->type_class) { | |
788 | case RCU_JA_LINEAR: | |
789 | { | |
790 | uint8_t nr_child = | |
791 | ja_linear_node_get_nr_child(type, node); | |
792 | unsigned int i; | |
793 | ||
794 | for (i = 0; i < nr_child; i++) { | |
795 | struct cds_ja_inode_flag *iter; | |
796 | unsigned int bit; | |
797 | uint8_t v; | |
798 | ||
799 | ja_linear_node_get_ith_pos(type, node, i, &v, &iter); | |
800 | if (!iter) | |
801 | continue; | |
802 | if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) | |
803 | continue; | |
804 | for (bit = 0; bit < JA_BITS_PER_BYTE; bit++) { | |
805 | if (v & (1U << bit)) | |
806 | nr_one[bit]++; | |
807 | } | |
808 | distrib_nr_child++; | |
809 | } | |
810 | break; | |
811 | } | |
812 | case RCU_JA_POOL: | |
813 | { | |
814 | unsigned int pool_nr; | |
815 | ||
816 | for (pool_nr = 0; pool_nr < (1U << type->nr_pool_order); pool_nr++) { | |
817 | struct cds_ja_inode *pool = | |
818 | ja_pool_node_get_ith_pool(type, | |
819 | node, pool_nr); | |
820 | uint8_t nr_child = | |
821 | ja_linear_node_get_nr_child(type, pool); | |
822 | unsigned int j; | |
823 | ||
824 | for (j = 0; j < nr_child; j++) { | |
825 | struct cds_ja_inode_flag *iter; | |
826 | unsigned int bit; | |
827 | uint8_t v; | |
828 | ||
829 | ja_linear_node_get_ith_pos(type, pool, | |
830 | j, &v, &iter); | |
831 | if (!iter) | |
832 | continue; | |
833 | if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) | |
834 | continue; | |
835 | for (bit = 0; bit < JA_BITS_PER_BYTE; bit++) { | |
836 | if (v & (1U << bit)) | |
837 | nr_one[bit]++; | |
838 | } | |
839 | distrib_nr_child++; | |
840 | } | |
841 | } | |
842 | break; | |
843 | } | |
844 | case RCU_JA_PIGEON: | |
845 | { | |
846 | uint8_t nr_child; | |
847 | unsigned int i; | |
848 | ||
849 | assert(mode == JA_RECOMPACT_DEL); | |
850 | nr_child = shadow_node->nr_child; | |
851 | for (i = 0; i < nr_child; i++) { | |
852 | struct cds_ja_inode_flag *iter; | |
853 | unsigned int bit; | |
854 | ||
855 | iter = ja_pigeon_node_get_ith_pos(type, node, i); | |
856 | if (!iter) | |
857 | continue; | |
858 | if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) | |
859 | continue; | |
860 | for (bit = 0; bit < JA_BITS_PER_BYTE; bit++) { | |
861 | if (i & (1U << bit)) | |
862 | nr_one[bit]++; | |
863 | } | |
864 | distrib_nr_child++; | |
865 | } | |
866 | break; | |
867 | } | |
868 | case RCU_JA_NULL: | |
869 | assert(mode == JA_RECOMPACT_ADD); | |
870 | break; | |
871 | default: | |
872 | assert(0); | |
873 | break; | |
874 | } | |
875 | ||
876 | if (mode == JA_RECOMPACT_ADD) { | |
877 | unsigned int bit; | |
878 | ||
879 | for (bit = 0; bit < JA_BITS_PER_BYTE; bit++) { | |
880 | if (n & (1U << bit)) | |
881 | nr_one[bit]++; | |
882 | } | |
883 | distrib_nr_child++; | |
884 | } | |
885 | ||
886 | /* | |
887 | * The best bit selector is that for which the number of ones is | |
888 | * closest to half of the number of children in the | |
889 | * distribution. | |
890 | */ | |
891 | for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) { | |
892 | unsigned int distance_to_best; | |
893 | ||
894 | distance_to_best = abs_int(nr_one[bit_i] - (distrib_nr_child >> 1U)); | |
895 | if (distance_to_best < overall_best_distance) { | |
896 | overall_best_distance = distance_to_best; | |
897 | bitsel = bit_i; | |
898 | } | |
899 | } | |
900 | dbg_printf("1 dimension pool bit selection: (%u)\n", bitsel); | |
901 | return bitsel; | |
902 | } | |
903 | ||
7a0b2331 MD |
904 | /* |
905 | * ja_node_recompact_add: recompact a node, adding a new child. | |
2e313670 | 906 | * Return 0 on success, -EAGAIN if need to retry, or other negative |
5a9a87dd | 907 | * error value otherwise. |
7a0b2331 MD |
908 | */ |
909 | static | |
2e313670 MD |
910 | int ja_node_recompact(enum ja_recompact mode, |
911 | struct cds_ja *ja, | |
e1db2db5 | 912 | unsigned int old_type_index, |
d96bfb0d | 913 | const struct cds_ja_type *old_type, |
b4540e8a | 914 | struct cds_ja_inode *old_node, |
5a9a87dd | 915 | struct cds_ja_shadow_node *shadow_node, |
3d8fe307 | 916 | struct cds_ja_inode_flag **old_node_flag_ptr, uint8_t n, |
af3cbd45 MD |
917 | struct cds_ja_inode_flag *child_node_flag, |
918 | struct cds_ja_inode_flag **nullify_node_flag_ptr) | |
7a0b2331 | 919 | { |
e1db2db5 | 920 | unsigned int new_type_index; |
b4540e8a | 921 | struct cds_ja_inode *new_node; |
af3cbd45 | 922 | struct cds_ja_shadow_node *new_shadow_node = NULL; |
d96bfb0d | 923 | const struct cds_ja_type *new_type; |
3d8fe307 | 924 | struct cds_ja_inode_flag *new_node_flag, *old_node_flag; |
7a0b2331 | 925 | int ret; |
f07b240f | 926 | int fallback = 0; |
7a0b2331 | 927 | |
3d8fe307 MD |
928 | old_node_flag = *old_node_flag_ptr; |
929 | ||
2e313670 MD |
930 | switch (mode) { |
931 | case JA_RECOMPACT: | |
932 | new_type_index = old_type_index; | |
933 | break; | |
934 | case JA_RECOMPACT_ADD: | |
935 | if (!shadow_node || old_type_index == NODE_INDEX_NULL) { | |
936 | new_type_index = 0; | |
937 | } else { | |
938 | new_type_index = old_type_index + 1; | |
939 | } | |
940 | break; | |
941 | case JA_RECOMPACT_DEL: | |
942 | if (old_type_index == 0) { | |
943 | new_type_index = NODE_INDEX_NULL; | |
944 | } else { | |
945 | new_type_index = old_type_index - 1; | |
946 | } | |
947 | break; | |
948 | default: | |
949 | assert(0); | |
7a0b2331 | 950 | } |
a2a7ff59 | 951 | |
f07b240f | 952 | retry: /* for fallback */ |
582a6ade MD |
953 | dbg_printf("Recompact from type %d to type %d\n", |
954 | old_type_index, new_type_index); | |
7a0b2331 | 955 | new_type = &ja_types[new_type_index]; |
2e313670 MD |
956 | if (new_type_index != NODE_INDEX_NULL) { |
957 | new_node = alloc_cds_ja_node(new_type); | |
958 | if (!new_node) | |
959 | return -ENOMEM; | |
b1a90ce3 MD |
960 | |
961 | if (new_type->type_class == RCU_JA_POOL) { | |
962 | switch (new_type->nr_pool_order) { | |
963 | case 1: | |
964 | { | |
965 | unsigned int node_distrib_bitsel = 0; | |
966 | node_distrib_bitsel = | |
967 | ja_node_sum_distribution_1d(mode, ja, | |
968 | old_type_index, old_type, | |
969 | old_node, shadow_node, | |
970 | n, child_node_flag, | |
971 | nullify_node_flag_ptr); | |
972 | assert(!((unsigned long) new_node & JA_POOL_1D_MASK)); | |
973 | new_node_flag = ja_node_flag_pool_1d(new_node, | |
974 | new_type_index, node_distrib_bitsel); | |
975 | break; | |
976 | } | |
977 | case 2: | |
978 | { | |
979 | /* TODO: pool order 2 in 2d */ | |
980 | assert(!((unsigned long) new_node & JA_POOL_1D_MASK)); | |
981 | assert(!((unsigned long) new_node & JA_POOL_2D_MASK)); | |
982 | new_node_flag = ja_node_flag(new_node, new_type_index); | |
983 | break; | |
984 | } | |
985 | default: | |
986 | assert(0); | |
987 | } | |
988 | } else { | |
989 | new_node_flag = ja_node_flag(new_node, new_type_index); | |
990 | } | |
991 | ||
2e313670 | 992 | dbg_printf("Recompact inherit lock from %p\n", shadow_node); |
3d8fe307 | 993 | new_shadow_node = rcuja_shadow_set(ja->ht, new_node_flag, shadow_node, ja); |
2e313670 MD |
994 | if (!new_shadow_node) { |
995 | free(new_node); | |
996 | return -ENOMEM; | |
997 | } | |
998 | if (fallback) | |
999 | new_shadow_node->fallback_removal_count = | |
1000 | JA_FALLBACK_REMOVAL_COUNT; | |
1001 | } else { | |
1002 | new_node = NULL; | |
1003 | new_node_flag = NULL; | |
e1db2db5 | 1004 | } |
11c5e016 | 1005 | |
2e313670 MD |
1006 | assert(mode != JA_RECOMPACT_ADD || old_type->type_class != RCU_JA_PIGEON); |
1007 | ||
1008 | if (new_type_index == NODE_INDEX_NULL) | |
1009 | goto skip_copy; | |
1010 | ||
11c5e016 MD |
1011 | switch (old_type->type_class) { |
1012 | case RCU_JA_LINEAR: | |
1013 | { | |
1014 | uint8_t nr_child = | |
1015 | ja_linear_node_get_nr_child(old_type, old_node); | |
1016 | unsigned int i; | |
1017 | ||
1018 | for (i = 0; i < nr_child; i++) { | |
b4540e8a | 1019 | struct cds_ja_inode_flag *iter; |
11c5e016 MD |
1020 | uint8_t v; |
1021 | ||
1022 | ja_linear_node_get_ith_pos(old_type, old_node, i, &v, &iter); | |
1023 | if (!iter) | |
1024 | continue; | |
af3cbd45 | 1025 | if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) |
2e313670 | 1026 | continue; |
b1a90ce3 | 1027 | ret = _ja_node_set_nth(new_type, new_node, new_node_flag, |
f07b240f | 1028 | new_shadow_node, |
11c5e016 | 1029 | v, iter); |
f07b240f MD |
1030 | if (new_type->type_class == RCU_JA_POOL && ret) { |
1031 | goto fallback_toosmall; | |
1032 | } | |
11c5e016 MD |
1033 | assert(!ret); |
1034 | } | |
1035 | break; | |
1036 | } | |
1037 | case RCU_JA_POOL: | |
1038 | { | |
1039 | unsigned int pool_nr; | |
1040 | ||
1041 | for (pool_nr = 0; pool_nr < (1U << old_type->nr_pool_order); pool_nr++) { | |
b4540e8a | 1042 | struct cds_ja_inode *pool = |
11c5e016 MD |
1043 | ja_pool_node_get_ith_pool(old_type, |
1044 | old_node, pool_nr); | |
1045 | uint8_t nr_child = | |
1046 | ja_linear_node_get_nr_child(old_type, pool); | |
1047 | unsigned int j; | |
1048 | ||
1049 | for (j = 0; j < nr_child; j++) { | |
b4540e8a | 1050 | struct cds_ja_inode_flag *iter; |
11c5e016 MD |
1051 | uint8_t v; |
1052 | ||
1053 | ja_linear_node_get_ith_pos(old_type, pool, | |
1054 | j, &v, &iter); | |
1055 | if (!iter) | |
1056 | continue; | |
af3cbd45 | 1057 | if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) |
2e313670 | 1058 | continue; |
b1a90ce3 | 1059 | ret = _ja_node_set_nth(new_type, new_node, new_node_flag, |
f07b240f | 1060 | new_shadow_node, |
11c5e016 | 1061 | v, iter); |
f07b240f MD |
1062 | if (new_type->type_class == RCU_JA_POOL |
1063 | && ret) { | |
1064 | goto fallback_toosmall; | |
1065 | } | |
11c5e016 MD |
1066 | assert(!ret); |
1067 | } | |
1068 | } | |
1069 | break; | |
7a0b2331 | 1070 | } |
a2a7ff59 | 1071 | case RCU_JA_NULL: |
2e313670 | 1072 | assert(mode == JA_RECOMPACT_ADD); |
a2a7ff59 | 1073 | break; |
11c5e016 | 1074 | case RCU_JA_PIGEON: |
2e313670 MD |
1075 | { |
1076 | uint8_t nr_child; | |
1077 | unsigned int i; | |
1078 | ||
1079 | assert(mode == JA_RECOMPACT_DEL); | |
1080 | nr_child = shadow_node->nr_child; | |
1081 | for (i = 0; i < nr_child; i++) { | |
1082 | struct cds_ja_inode_flag *iter; | |
1083 | ||
1084 | iter = ja_pigeon_node_get_ith_pos(old_type, old_node, i); | |
1085 | if (!iter) | |
1086 | continue; | |
af3cbd45 | 1087 | if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) |
2e313670 | 1088 | continue; |
b1a90ce3 | 1089 | ret = _ja_node_set_nth(new_type, new_node, new_node_flag, |
2e313670 MD |
1090 | new_shadow_node, |
1091 | i, iter); | |
1092 | if (new_type->type_class == RCU_JA_POOL && ret) { | |
1093 | goto fallback_toosmall; | |
1094 | } | |
1095 | assert(!ret); | |
1096 | } | |
1097 | break; | |
1098 | } | |
11c5e016 MD |
1099 | default: |
1100 | assert(0); | |
5a9a87dd | 1101 | ret = -EINVAL; |
f07b240f | 1102 | goto end; |
11c5e016 | 1103 | } |
2e313670 | 1104 | skip_copy: |
11c5e016 | 1105 | |
9a1c1915 | 1106 | if (mode == JA_RECOMPACT_ADD) { |
2e313670 | 1107 | /* add node */ |
b1a90ce3 | 1108 | ret = _ja_node_set_nth(new_type, new_node, new_node_flag, |
2e313670 MD |
1109 | new_shadow_node, |
1110 | n, child_node_flag); | |
7b413155 MD |
1111 | if (new_type->type_class == RCU_JA_POOL && ret) { |
1112 | goto fallback_toosmall; | |
1113 | } | |
2e313670 MD |
1114 | assert(!ret); |
1115 | } | |
3d8fe307 MD |
1116 | /* Return pointer to new recompacted node through old_node_flag_ptr */ |
1117 | *old_node_flag_ptr = new_node_flag; | |
a2a7ff59 | 1118 | if (old_node) { |
2e313670 MD |
1119 | int flags; |
1120 | ||
1121 | flags = RCUJA_SHADOW_CLEAR_FREE_NODE; | |
1122 | /* | |
1123 | * It is OK to free the lock associated with a node | |
1124 | * going to NULL, since we are holding the parent lock. | |
1125 | * This synchronizes removal with re-add of that node. | |
1126 | */ | |
1127 | if (new_type_index == NODE_INDEX_NULL) | |
1128 | flags = RCUJA_SHADOW_CLEAR_FREE_LOCK; | |
3d8fe307 | 1129 | ret = rcuja_shadow_clear(ja->ht, old_node_flag, shadow_node, |
2e313670 | 1130 | flags); |
a2a7ff59 MD |
1131 | assert(!ret); |
1132 | } | |
5a9a87dd MD |
1133 | |
1134 | ret = 0; | |
f07b240f | 1135 | end: |
5a9a87dd | 1136 | return ret; |
f07b240f MD |
1137 | |
1138 | fallback_toosmall: | |
1139 | /* fallback if next pool is too small */ | |
af3cbd45 | 1140 | assert(new_shadow_node); |
3d8fe307 | 1141 | ret = rcuja_shadow_clear(ja->ht, new_node_flag, new_shadow_node, |
f07b240f MD |
1142 | RCUJA_SHADOW_CLEAR_FREE_NODE); |
1143 | assert(!ret); | |
1144 | ||
2e313670 | 1145 | /* Choose fallback type: pigeon */ |
f07b240f MD |
1146 | new_type_index = (1UL << JA_TYPE_BITS) - 1; |
1147 | dbg_printf("Fallback to type %d\n", new_type_index); | |
1148 | uatomic_inc(&ja->nr_fallback); | |
1149 | fallback = 1; | |
1150 | goto retry; | |
7a0b2331 MD |
1151 | } |
1152 | ||
5a9a87dd | 1153 | /* |
2e313670 | 1154 | * Return 0 on success, -EAGAIN if need to retry, or other negative |
5a9a87dd MD |
1155 | * error value otherwise. |
1156 | */ | |
7a0b2331 | 1157 | static |
d96bfb0d | 1158 | int ja_node_set_nth(struct cds_ja *ja, |
b4540e8a | 1159 | struct cds_ja_inode_flag **node_flag, uint8_t n, |
5a9a87dd MD |
1160 | struct cds_ja_inode_flag *child_node_flag, |
1161 | struct cds_ja_shadow_node *shadow_node) | |
7a0b2331 MD |
1162 | { |
1163 | int ret; | |
e1db2db5 | 1164 | unsigned int type_index; |
d96bfb0d | 1165 | const struct cds_ja_type *type; |
b4540e8a | 1166 | struct cds_ja_inode *node; |
7a0b2331 | 1167 | |
a2a7ff59 MD |
1168 | dbg_printf("ja_node_set_nth for n=%u, node %p, shadow %p\n", |
1169 | (unsigned int) n, ja_node_ptr(*node_flag), shadow_node); | |
1170 | ||
e1db2db5 MD |
1171 | node = ja_node_ptr(*node_flag); |
1172 | type_index = ja_node_type(*node_flag); | |
1173 | type = &ja_types[type_index]; | |
b1a90ce3 | 1174 | ret = _ja_node_set_nth(type, node, *node_flag, shadow_node, |
e1db2db5 | 1175 | n, child_node_flag); |
2e313670 MD |
1176 | switch (ret) { |
1177 | case -ENOSPC: | |
e1db2db5 | 1178 | /* Not enough space in node, need to recompact. */ |
2e313670 | 1179 | ret = ja_node_recompact(JA_RECOMPACT_ADD, ja, type_index, type, node, |
af3cbd45 | 1180 | shadow_node, node_flag, n, child_node_flag, NULL); |
2e313670 MD |
1181 | break; |
1182 | case -ERANGE: | |
1183 | /* Node needs to be recompacted. */ | |
1184 | ret = ja_node_recompact(JA_RECOMPACT, ja, type_index, type, node, | |
af3cbd45 | 1185 | shadow_node, node_flag, n, child_node_flag, NULL); |
2e313670 MD |
1186 | break; |
1187 | } | |
1188 | return ret; | |
1189 | } | |
1190 | ||
1191 | /* | |
1192 | * Return 0 on success, -EAGAIN if need to retry, or other negative | |
1193 | * error value otherwise. | |
1194 | */ | |
1195 | static | |
af3cbd45 MD |
1196 | int ja_node_clear_ptr(struct cds_ja *ja, |
1197 | struct cds_ja_inode_flag **node_flag_ptr, /* Pointer to location to nullify */ | |
1198 | struct cds_ja_inode_flag **parent_node_flag_ptr, /* Address of parent ptr in its parent */ | |
1199 | struct cds_ja_shadow_node *shadow_node, /* of parent */ | |
1200 | uint8_t n) | |
2e313670 MD |
1201 | { |
1202 | int ret; | |
1203 | unsigned int type_index; | |
1204 | const struct cds_ja_type *type; | |
1205 | struct cds_ja_inode *node; | |
1206 | ||
af3cbd45 MD |
1207 | dbg_printf("ja_node_clear_ptr for node %p, shadow %p, target ptr %p\n", |
1208 | ja_node_ptr(*parent_node_flag_ptr), shadow_node, node_flag_ptr); | |
2e313670 | 1209 | |
af3cbd45 MD |
1210 | node = ja_node_ptr(*parent_node_flag_ptr); |
1211 | type_index = ja_node_type(*parent_node_flag_ptr); | |
2e313670 | 1212 | type = &ja_types[type_index]; |
af3cbd45 | 1213 | ret = _ja_node_clear_ptr(type, node, shadow_node, node_flag_ptr, n); |
2e313670 MD |
1214 | if (ret == -EFBIG) { |
1215 | /* Should to try recompaction. */ | |
1216 | ret = ja_node_recompact(JA_RECOMPACT_DEL, ja, type_index, type, node, | |
af3cbd45 MD |
1217 | shadow_node, parent_node_flag_ptr, n, NULL, |
1218 | node_flag_ptr); | |
7a0b2331 MD |
1219 | } |
1220 | return ret; | |
1221 | } | |
be9a7474 | 1222 | |
af3cbd45 | 1223 | struct cds_hlist_head cds_ja_lookup(struct cds_ja *ja, uint64_t key) |
b4540e8a | 1224 | { |
41975c12 MD |
1225 | unsigned int tree_depth, i; |
1226 | struct cds_ja_inode_flag *node_flag; | |
af3cbd45 | 1227 | struct cds_hlist_head head = { NULL }; |
41975c12 MD |
1228 | |
1229 | if (caa_unlikely(key > ja->key_max)) | |
af3cbd45 | 1230 | return head; |
41975c12 | 1231 | tree_depth = ja->tree_depth; |
5a9a87dd | 1232 | node_flag = rcu_dereference(ja->root); |
41975c12 | 1233 | |
5a9a87dd MD |
1234 | /* level 0: root node */ |
1235 | if (!ja_node_ptr(node_flag)) | |
af3cbd45 | 1236 | return head; |
5a9a87dd MD |
1237 | |
1238 | for (i = 1; i < tree_depth; i++) { | |
79b41067 MD |
1239 | uint8_t iter_key; |
1240 | ||
1241 | iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (tree_depth - i - 1))); | |
b62a8d0c | 1242 | node_flag = ja_node_get_nth(node_flag, NULL, NULL, NULL, |
79b41067 | 1243 | iter_key); |
582a6ade MD |
1244 | dbg_printf("cds_ja_lookup iter key lookup %u finds node_flag %p\n", |
1245 | (unsigned int) iter_key, node_flag); | |
41975c12 | 1246 | if (!ja_node_ptr(node_flag)) |
af3cbd45 | 1247 | return head; |
41975c12 MD |
1248 | } |
1249 | ||
5a9a87dd | 1250 | /* Last level lookup succeded. We got an actual match. */ |
af3cbd45 MD |
1251 | head.next = (struct cds_hlist_node *) node_flag; |
1252 | return head; | |
5a9a87dd MD |
1253 | } |
1254 | ||
1255 | /* | |
1256 | * We reached an unpopulated node. Create it and the children we need, | |
1257 | * and then attach the entire branch to the current node. This may | |
1258 | * trigger recompaction of the current node. Locks needed: node lock | |
1259 | * (for add), and, possibly, parent node lock (to update pointer due to | |
1260 | * node recompaction). | |
1261 | * | |
1262 | * First take node lock, check if recompaction is needed, then take | |
1263 | * parent lock (if needed). Then we can proceed to create the new | |
1264 | * branch. Publish the new branch, and release locks. | |
1265 | * TODO: we currently always take the parent lock even when not needed. | |
1266 | */ | |
1267 | static | |
1268 | int ja_attach_node(struct cds_ja *ja, | |
b0ca2d21 | 1269 | struct cds_ja_inode_flag **attach_node_flag_ptr, |
b62a8d0c | 1270 | struct cds_ja_inode_flag *attach_node_flag, |
5a9a87dd MD |
1271 | struct cds_ja_inode_flag **node_flag_ptr, |
1272 | struct cds_ja_inode_flag *node_flag, | |
1273 | struct cds_ja_inode_flag *parent_node_flag, | |
1274 | uint64_t key, | |
79b41067 | 1275 | unsigned int level, |
5a9a87dd MD |
1276 | struct cds_ja_node *child_node) |
1277 | { | |
1278 | struct cds_ja_shadow_node *shadow_node = NULL, | |
af3cbd45 | 1279 | *parent_shadow_node = NULL; |
5a9a87dd MD |
1280 | struct cds_ja_inode *node = ja_node_ptr(node_flag); |
1281 | struct cds_ja_inode *parent_node = ja_node_ptr(parent_node_flag); | |
1282 | struct cds_hlist_head head; | |
1283 | struct cds_ja_inode_flag *iter_node_flag, *iter_dest_node_flag; | |
1284 | int ret, i; | |
a2a7ff59 | 1285 | struct cds_ja_inode_flag *created_nodes[JA_MAX_DEPTH]; |
5a9a87dd MD |
1286 | int nr_created_nodes = 0; |
1287 | ||
582a6ade MD |
1288 | dbg_printf("Attach node at level %u (node %p, node_flag %p)\n", |
1289 | level, node, node_flag); | |
a2a7ff59 | 1290 | |
5a9a87dd | 1291 | assert(node); |
3d8fe307 | 1292 | shadow_node = rcuja_shadow_lookup_lock(ja->ht, node_flag); |
5a9a87dd | 1293 | if (!shadow_node) { |
2e313670 | 1294 | ret = -EAGAIN; |
5a9a87dd MD |
1295 | goto end; |
1296 | } | |
1297 | if (parent_node) { | |
1298 | parent_shadow_node = rcuja_shadow_lookup_lock(ja->ht, | |
3d8fe307 | 1299 | parent_node_flag); |
5a9a87dd | 1300 | if (!parent_shadow_node) { |
2e313670 | 1301 | ret = -EAGAIN; |
5a9a87dd MD |
1302 | goto unlock_shadow; |
1303 | } | |
1304 | } | |
1305 | ||
b62a8d0c | 1306 | if (node_flag_ptr && ja_node_ptr(*node_flag_ptr)) { |
b306a0fe | 1307 | /* |
c112acaa MD |
1308 | * Target node has been updated between RCU lookup and |
1309 | * lock acquisition. We need to re-try lookup and | |
1310 | * attach. | |
1311 | */ | |
1312 | ret = -EAGAIN; | |
1313 | goto unlock_parent; | |
1314 | } | |
1315 | ||
1316 | if (attach_node_flag_ptr && ja_node_ptr(*attach_node_flag_ptr) != | |
b62a8d0c | 1317 | ja_node_ptr(attach_node_flag)) { |
c112acaa MD |
1318 | /* |
1319 | * Target node has been updated between RCU lookup and | |
1320 | * lock acquisition. We need to re-try lookup and | |
1321 | * attach. | |
b306a0fe MD |
1322 | */ |
1323 | ret = -EAGAIN; | |
1324 | goto unlock_parent; | |
1325 | } | |
1326 | ||
a2a7ff59 | 1327 | /* Create new branch, starting from bottom */ |
5a9a87dd MD |
1328 | CDS_INIT_HLIST_HEAD(&head); |
1329 | cds_hlist_add_head_rcu(&child_node->list, &head); | |
a2a7ff59 | 1330 | iter_node_flag = (struct cds_ja_inode_flag *) head.next; |
5a9a87dd | 1331 | |
79b41067 MD |
1332 | for (i = ja->tree_depth; i > (int) level; i--) { |
1333 | uint8_t iter_key; | |
1334 | ||
1335 | iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (ja->tree_depth - i))); | |
1336 | dbg_printf("branch creation level %d, key %u\n", | |
1337 | i - 1, (unsigned int) iter_key); | |
5a9a87dd MD |
1338 | iter_dest_node_flag = NULL; |
1339 | ret = ja_node_set_nth(ja, &iter_dest_node_flag, | |
79b41067 | 1340 | iter_key, |
5a9a87dd MD |
1341 | iter_node_flag, |
1342 | NULL); | |
1343 | if (ret) | |
1344 | goto check_error; | |
1345 | created_nodes[nr_created_nodes++] = iter_dest_node_flag; | |
1346 | iter_node_flag = iter_dest_node_flag; | |
1347 | } | |
1348 | ||
79b41067 MD |
1349 | if (level > 1) { |
1350 | uint8_t iter_key; | |
1351 | ||
1352 | iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (ja->tree_depth - level))); | |
a2a7ff59 MD |
1353 | /* We need to use set_nth on the previous level. */ |
1354 | iter_dest_node_flag = node_flag; | |
1355 | ret = ja_node_set_nth(ja, &iter_dest_node_flag, | |
79b41067 | 1356 | iter_key, |
a2a7ff59 MD |
1357 | iter_node_flag, |
1358 | shadow_node); | |
1359 | if (ret) | |
1360 | goto check_error; | |
1361 | created_nodes[nr_created_nodes++] = iter_dest_node_flag; | |
1362 | iter_node_flag = iter_dest_node_flag; | |
1363 | } | |
1364 | ||
5a9a87dd | 1365 | /* Publish new branch */ |
a2a7ff59 | 1366 | dbg_printf("Publish branch %p, replacing %p\n", |
b0ca2d21 MD |
1367 | iter_node_flag, *attach_node_flag_ptr); |
1368 | rcu_assign_pointer(*attach_node_flag_ptr, iter_node_flag); | |
5a9a87dd MD |
1369 | |
1370 | /* Success */ | |
1371 | ret = 0; | |
1372 | ||
1373 | check_error: | |
1374 | if (ret) { | |
1375 | for (i = 0; i < nr_created_nodes; i++) { | |
1376 | int tmpret; | |
a2a7ff59 MD |
1377 | int flags; |
1378 | ||
1379 | flags = RCUJA_SHADOW_CLEAR_FREE_LOCK; | |
1380 | if (i) | |
1381 | flags |= RCUJA_SHADOW_CLEAR_FREE_NODE; | |
5a9a87dd | 1382 | tmpret = rcuja_shadow_clear(ja->ht, |
3d8fe307 | 1383 | created_nodes[i], |
a2a7ff59 MD |
1384 | NULL, |
1385 | flags); | |
5a9a87dd MD |
1386 | assert(!tmpret); |
1387 | } | |
1388 | } | |
b306a0fe | 1389 | unlock_parent: |
5a9a87dd MD |
1390 | if (parent_shadow_node) |
1391 | rcuja_shadow_unlock(parent_shadow_node); | |
1392 | unlock_shadow: | |
1393 | if (shadow_node) | |
1394 | rcuja_shadow_unlock(shadow_node); | |
1395 | end: | |
1396 | return ret; | |
1397 | } | |
1398 | ||
1399 | /* | |
af3cbd45 MD |
1400 | * Lock the parent containing the hlist head pointer, and add node to list of |
1401 | * duplicates. Failure can happen if concurrent update changes the | |
1402 | * parent before we get the lock. We return -EAGAIN in that case. | |
5a9a87dd MD |
1403 | * Return 0 on success, negative error value on failure. |
1404 | */ | |
1405 | static | |
1406 | int ja_chain_node(struct cds_ja *ja, | |
af3cbd45 | 1407 | struct cds_ja_inode_flag *parent_node_flag, |
fa112799 | 1408 | struct cds_ja_inode_flag **node_flag_ptr, |
c112acaa | 1409 | struct cds_ja_inode_flag *node_flag, |
5a9a87dd MD |
1410 | struct cds_hlist_head *head, |
1411 | struct cds_ja_node *node) | |
1412 | { | |
1413 | struct cds_ja_shadow_node *shadow_node; | |
fa112799 | 1414 | int ret = 0; |
5a9a87dd | 1415 | |
3d8fe307 | 1416 | shadow_node = rcuja_shadow_lookup_lock(ja->ht, parent_node_flag); |
b306a0fe | 1417 | if (!shadow_node) { |
2e313670 | 1418 | return -EAGAIN; |
b306a0fe | 1419 | } |
c112acaa | 1420 | if (ja_node_ptr(*node_flag_ptr) != ja_node_ptr(node_flag)) { |
fa112799 MD |
1421 | ret = -EAGAIN; |
1422 | goto end; | |
1423 | } | |
5a9a87dd | 1424 | cds_hlist_add_head_rcu(&node->list, head); |
fa112799 | 1425 | end: |
5a9a87dd | 1426 | rcuja_shadow_unlock(shadow_node); |
fa112799 | 1427 | return ret; |
5a9a87dd MD |
1428 | } |
1429 | ||
1430 | int cds_ja_add(struct cds_ja *ja, uint64_t key, | |
1431 | struct cds_ja_node *new_node) | |
1432 | { | |
1433 | unsigned int tree_depth, i; | |
b0ca2d21 MD |
1434 | struct cds_ja_inode_flag **attach_node_flag_ptr, |
1435 | **node_flag_ptr; | |
5a9a87dd MD |
1436 | struct cds_ja_inode_flag *node_flag, |
1437 | *parent_node_flag, | |
b62a8d0c MD |
1438 | *parent2_node_flag, |
1439 | *attach_node_flag; | |
5a9a87dd MD |
1440 | int ret; |
1441 | ||
b306a0fe | 1442 | if (caa_unlikely(key > ja->key_max)) { |
5a9a87dd | 1443 | return -EINVAL; |
b306a0fe | 1444 | } |
5a9a87dd MD |
1445 | tree_depth = ja->tree_depth; |
1446 | ||
1447 | retry: | |
a2a7ff59 MD |
1448 | dbg_printf("cds_ja_add attempt: key %" PRIu64 ", node %p\n", |
1449 | key, new_node); | |
5a9a87dd | 1450 | parent2_node_flag = NULL; |
b0f74e47 MD |
1451 | parent_node_flag = |
1452 | (struct cds_ja_inode_flag *) &ja->root; /* Use root ptr address as key for mutex */ | |
b0ca2d21 | 1453 | attach_node_flag_ptr = &ja->root; |
b62a8d0c | 1454 | attach_node_flag = rcu_dereference(ja->root); |
5a9a87dd | 1455 | node_flag_ptr = &ja->root; |
35170a44 | 1456 | node_flag = rcu_dereference(ja->root); |
5a9a87dd MD |
1457 | |
1458 | /* Iterate on all internal levels */ | |
a2a7ff59 | 1459 | for (i = 1; i < tree_depth; i++) { |
79b41067 MD |
1460 | uint8_t iter_key; |
1461 | ||
b0ca2d21 | 1462 | dbg_printf("cds_ja_add iter attach_node_flag_ptr %p node_flag_ptr %p node_flag %p\n", |
c112acaa | 1463 | attach_node_flag_ptr, node_flag_ptr, node_flag); |
5a9a87dd | 1464 | if (!ja_node_ptr(node_flag)) { |
b0ca2d21 | 1465 | ret = ja_attach_node(ja, attach_node_flag_ptr, |
b62a8d0c | 1466 | attach_node_flag, |
b0ca2d21 | 1467 | node_flag_ptr, |
c112acaa MD |
1468 | parent_node_flag, |
1469 | parent2_node_flag, | |
5a9a87dd | 1470 | key, i, new_node); |
2e313670 | 1471 | if (ret == -EAGAIN || ret == -EEXIST) |
5a9a87dd MD |
1472 | goto retry; |
1473 | else | |
1474 | goto end; | |
1475 | } | |
79b41067 | 1476 | iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (tree_depth - i - 1))); |
5a9a87dd MD |
1477 | parent2_node_flag = parent_node_flag; |
1478 | parent_node_flag = node_flag; | |
1479 | node_flag = ja_node_get_nth(node_flag, | |
b0ca2d21 | 1480 | &attach_node_flag_ptr, |
b62a8d0c | 1481 | &attach_node_flag, |
5a9a87dd | 1482 | &node_flag_ptr, |
79b41067 | 1483 | iter_key); |
b0ca2d21 MD |
1484 | dbg_printf("cds_ja_add iter key lookup %u finds node_flag %p attach_node_flag_ptr %p node_flag_ptr %p\n", |
1485 | (unsigned int) iter_key, node_flag, | |
c112acaa MD |
1486 | attach_node_flag_ptr, |
1487 | node_flag_ptr); | |
5a9a87dd MD |
1488 | } |
1489 | ||
1490 | /* | |
1491 | * We reached bottom of tree, simply add node to last internal | |
1492 | * level, or chain it if key is already present. | |
1493 | */ | |
1494 | if (!ja_node_ptr(node_flag)) { | |
c112acaa MD |
1495 | dbg_printf("cds_ja_add attach_node_flag_ptr %p node_flag_ptr %p node_flag %p\n", |
1496 | attach_node_flag_ptr, node_flag_ptr, node_flag); | |
b0ca2d21 | 1497 | ret = ja_attach_node(ja, attach_node_flag_ptr, |
b62a8d0c | 1498 | attach_node_flag, |
b0ca2d21 | 1499 | node_flag_ptr, parent_node_flag, |
5a9a87dd MD |
1500 | parent2_node_flag, key, i, new_node); |
1501 | } else { | |
1502 | ret = ja_chain_node(ja, | |
af3cbd45 | 1503 | parent_node_flag, |
fa112799 | 1504 | node_flag_ptr, |
c112acaa | 1505 | node_flag, |
b0ca2d21 | 1506 | (struct cds_hlist_head *) attach_node_flag_ptr, |
5a9a87dd MD |
1507 | new_node); |
1508 | } | |
b306a0fe | 1509 | if (ret == -EAGAIN || ret == -EEXIST) |
5a9a87dd MD |
1510 | goto retry; |
1511 | end: | |
1512 | return ret; | |
b4540e8a MD |
1513 | } |
1514 | ||
af3cbd45 MD |
1515 | /* |
1516 | * Note: there is no need to lookup the pointer address associated with | |
1517 | * each node's nth item after taking the lock: it's already been done by | |
1518 | * cds_ja_del while holding the rcu read-side lock, and our node rules | |
1519 | * ensure that when a match value -> pointer is found in a node, it is | |
1520 | * _NEVER_ changed for that node without recompaction, and recompaction | |
1521 | * reallocates the node. | |
b306a0fe MD |
1522 | * However, when a child is removed from "linear" nodes, its pointer |
1523 | * is set to NULL. We therefore check, while holding the locks, if this | |
1524 | * pointer is NULL, and return -ENOENT to the caller if it is the case. | |
af3cbd45 | 1525 | */ |
35170a44 MD |
1526 | static |
1527 | int ja_detach_node(struct cds_ja *ja, | |
1528 | struct cds_ja_inode_flag **snapshot, | |
af3cbd45 MD |
1529 | struct cds_ja_inode_flag ***snapshot_ptr, |
1530 | uint8_t *snapshot_n, | |
35170a44 MD |
1531 | int nr_snapshot, |
1532 | uint64_t key, | |
1533 | struct cds_ja_node *node) | |
1534 | { | |
af3cbd45 MD |
1535 | struct cds_ja_shadow_node *shadow_nodes[JA_MAX_DEPTH]; |
1536 | struct cds_ja_inode_flag **node_flag_ptr = NULL, | |
1537 | *parent_node_flag = NULL, | |
1538 | **parent_node_flag_ptr = NULL; | |
b62a8d0c | 1539 | struct cds_ja_inode_flag *iter_node_flag; |
4d6ef45e MD |
1540 | int ret, i, nr_shadow = 0, nr_clear = 0, nr_branch = 0; |
1541 | uint8_t n = 0; | |
35170a44 | 1542 | |
4d6ef45e | 1543 | assert(nr_snapshot == ja->tree_depth + 1); |
35170a44 | 1544 | |
af3cbd45 MD |
1545 | /* |
1546 | * From the last internal level node going up, get the node | |
1547 | * lock, check if the node has only one child left. If it is the | |
1548 | * case, we continue iterating upward. When we reach a node | |
1549 | * which has more that one child left, we lock the parent, and | |
1550 | * proceed to the node deletion (removing its children too). | |
1551 | */ | |
4d6ef45e | 1552 | for (i = nr_snapshot - 2; i >= 1; i--) { |
af3cbd45 MD |
1553 | struct cds_ja_shadow_node *shadow_node; |
1554 | ||
1555 | shadow_node = rcuja_shadow_lookup_lock(ja->ht, | |
3d8fe307 | 1556 | snapshot[i]); |
af3cbd45 MD |
1557 | if (!shadow_node) { |
1558 | ret = -EAGAIN; | |
1559 | goto end; | |
1560 | } | |
af3cbd45 | 1561 | shadow_nodes[nr_shadow++] = shadow_node; |
b62a8d0c MD |
1562 | |
1563 | /* | |
1564 | * Check if node has been removed between RCU | |
1565 | * lookup and lock acquisition. | |
1566 | */ | |
1567 | assert(snapshot_ptr[i + 1]); | |
1568 | if (ja_node_ptr(*snapshot_ptr[i + 1]) | |
1569 | != ja_node_ptr(snapshot[i + 1])) { | |
1570 | ret = -ENOENT; | |
1571 | goto end; | |
1572 | } | |
1573 | ||
1574 | assert(shadow_node->nr_child > 0); | |
d810c97f | 1575 | if (shadow_node->nr_child == 1 && i > 1) |
4d6ef45e MD |
1576 | nr_clear++; |
1577 | nr_branch++; | |
af3cbd45 MD |
1578 | if (shadow_node->nr_child > 1 || i == 1) { |
1579 | /* Lock parent and break */ | |
1580 | shadow_node = rcuja_shadow_lookup_lock(ja->ht, | |
3d8fe307 | 1581 | snapshot[i - 1]); |
af3cbd45 MD |
1582 | if (!shadow_node) { |
1583 | ret = -EAGAIN; | |
1584 | goto end; | |
1585 | } | |
1586 | shadow_nodes[nr_shadow++] = shadow_node; | |
b62a8d0c | 1587 | |
c112acaa MD |
1588 | /* |
1589 | * Check if node has been removed between RCU | |
1590 | * lookup and lock acquisition. | |
1591 | */ | |
b62a8d0c MD |
1592 | assert(snapshot_ptr[i]); |
1593 | if (ja_node_ptr(*snapshot_ptr[i]) | |
1594 | != ja_node_ptr(snapshot[i])) { | |
c112acaa MD |
1595 | ret = -ENOENT; |
1596 | goto end; | |
1597 | } | |
1598 | ||
b62a8d0c | 1599 | node_flag_ptr = snapshot_ptr[i + 1]; |
4d6ef45e MD |
1600 | n = snapshot_n[i + 1]; |
1601 | parent_node_flag_ptr = snapshot_ptr[i]; | |
1602 | parent_node_flag = snapshot[i]; | |
c112acaa | 1603 | |
af3cbd45 MD |
1604 | if (i > 1) { |
1605 | /* | |
1606 | * Lock parent's parent, in case we need | |
1607 | * to recompact parent. | |
1608 | */ | |
1609 | shadow_node = rcuja_shadow_lookup_lock(ja->ht, | |
3d8fe307 | 1610 | snapshot[i - 2]); |
af3cbd45 MD |
1611 | if (!shadow_node) { |
1612 | ret = -EAGAIN; | |
1613 | goto end; | |
1614 | } | |
1615 | shadow_nodes[nr_shadow++] = shadow_node; | |
b62a8d0c MD |
1616 | |
1617 | /* | |
1618 | * Check if node has been removed between RCU | |
1619 | * lookup and lock acquisition. | |
1620 | */ | |
1621 | assert(snapshot_ptr[i - 1]); | |
1622 | if (ja_node_ptr(*snapshot_ptr[i - 1]) | |
1623 | != ja_node_ptr(snapshot[i - 1])) { | |
1624 | ret = -ENOENT; | |
1625 | goto end; | |
1626 | } | |
af3cbd45 | 1627 | } |
b62a8d0c | 1628 | |
af3cbd45 MD |
1629 | break; |
1630 | } | |
1631 | } | |
1632 | ||
1633 | /* | |
4d6ef45e MD |
1634 | * At this point, we want to delete all nodes that are about to |
1635 | * be removed from shadow_nodes (except the last one, which is | |
1636 | * either the root or the parent of the upmost node with 1 | |
b62a8d0c MD |
1637 | * child). OK to free lock here, because RCU read lock is held, |
1638 | * and free only performed in call_rcu. | |
af3cbd45 MD |
1639 | */ |
1640 | ||
1641 | for (i = 0; i < nr_clear; i++) { | |
1642 | ret = rcuja_shadow_clear(ja->ht, | |
3d8fe307 | 1643 | shadow_nodes[i]->node_flag, |
af3cbd45 MD |
1644 | shadow_nodes[i], |
1645 | RCUJA_SHADOW_CLEAR_FREE_NODE | |
1646 | | RCUJA_SHADOW_CLEAR_FREE_LOCK); | |
1647 | assert(!ret); | |
1648 | } | |
1649 | ||
1650 | iter_node_flag = parent_node_flag; | |
1651 | /* Remove from parent */ | |
1652 | ret = ja_node_clear_ptr(ja, | |
1653 | node_flag_ptr, /* Pointer to location to nullify */ | |
1654 | &iter_node_flag, /* Old new parent ptr in its parent */ | |
4d6ef45e | 1655 | shadow_nodes[nr_branch - 1], /* of parent */ |
af3cbd45 | 1656 | n); |
b306a0fe MD |
1657 | if (ret) |
1658 | goto end; | |
af3cbd45 | 1659 | |
4d6ef45e MD |
1660 | dbg_printf("ja_detach_node: publish %p instead of %p\n", |
1661 | iter_node_flag, *parent_node_flag_ptr); | |
af3cbd45 MD |
1662 | /* Update address of parent ptr in its parent */ |
1663 | rcu_assign_pointer(*parent_node_flag_ptr, iter_node_flag); | |
1664 | ||
1665 | end: | |
1666 | for (i = 0; i < nr_shadow; i++) | |
1667 | rcuja_shadow_unlock(shadow_nodes[i]); | |
35170a44 MD |
1668 | return ret; |
1669 | } | |
1670 | ||
af3cbd45 MD |
1671 | static |
1672 | int ja_unchain_node(struct cds_ja *ja, | |
1673 | struct cds_ja_inode_flag *parent_node_flag, | |
fa112799 | 1674 | struct cds_ja_inode_flag **node_flag_ptr, |
013a6083 | 1675 | struct cds_ja_inode_flag *node_flag, |
af3cbd45 MD |
1676 | struct cds_ja_node *node) |
1677 | { | |
1678 | struct cds_ja_shadow_node *shadow_node; | |
f2758d14 | 1679 | struct cds_hlist_node *hlist_node; |
013a6083 MD |
1680 | struct cds_hlist_head hlist_head; |
1681 | int ret = 0, count = 0, found = 0; | |
af3cbd45 | 1682 | |
3d8fe307 | 1683 | shadow_node = rcuja_shadow_lookup_lock(ja->ht, parent_node_flag); |
af3cbd45 MD |
1684 | if (!shadow_node) |
1685 | return -EAGAIN; | |
013a6083 | 1686 | if (ja_node_ptr(*node_flag_ptr) != ja_node_ptr(node_flag)) { |
fa112799 MD |
1687 | ret = -EAGAIN; |
1688 | goto end; | |
1689 | } | |
013a6083 | 1690 | hlist_head.next = (struct cds_hlist_node *) ja_node_ptr(node_flag); |
af3cbd45 MD |
1691 | /* |
1692 | * Retry if another thread removed all but one of duplicates | |
fa112799 | 1693 | * since check (this check was performed without lock). |
013a6083 MD |
1694 | * Ensure that the node we are about to remove is still in the |
1695 | * list (while holding lock). | |
af3cbd45 | 1696 | */ |
013a6083 | 1697 | cds_hlist_for_each_rcu(hlist_node, &hlist_head) { |
ade342cb MD |
1698 | if (count == 0) { |
1699 | /* FIXME: currently a work-around */ | |
1700 | hlist_node->prev = (struct cds_hlist_node *) node_flag_ptr; | |
1701 | } | |
f2758d14 | 1702 | count++; |
013a6083 MD |
1703 | if (hlist_node == &node->list) |
1704 | found++; | |
f2758d14 | 1705 | } |
013a6083 MD |
1706 | assert(found <= 1); |
1707 | if (!found || count == 1) { | |
af3cbd45 MD |
1708 | ret = -EAGAIN; |
1709 | goto end; | |
1710 | } | |
1711 | cds_hlist_del_rcu(&node->list); | |
ade342cb MD |
1712 | /* |
1713 | * Validate that we indeed removed the node from linked list. | |
1714 | */ | |
1715 | assert(ja_node_ptr(*node_flag_ptr) != (struct cds_ja_inode *) node); | |
af3cbd45 MD |
1716 | end: |
1717 | rcuja_shadow_unlock(shadow_node); | |
1718 | return ret; | |
1719 | } | |
1720 | ||
1721 | /* | |
1722 | * Called with RCU read lock held. | |
1723 | */ | |
35170a44 MD |
1724 | int cds_ja_del(struct cds_ja *ja, uint64_t key, |
1725 | struct cds_ja_node *node) | |
1726 | { | |
1727 | unsigned int tree_depth, i; | |
1728 | struct cds_ja_inode_flag *snapshot[JA_MAX_DEPTH]; | |
af3cbd45 MD |
1729 | struct cds_ja_inode_flag **snapshot_ptr[JA_MAX_DEPTH]; |
1730 | uint8_t snapshot_n[JA_MAX_DEPTH]; | |
35170a44 | 1731 | struct cds_ja_inode_flag *node_flag; |
fa112799 MD |
1732 | struct cds_ja_inode_flag **prev_node_flag_ptr, |
1733 | **node_flag_ptr; | |
4d6ef45e | 1734 | int nr_snapshot; |
35170a44 MD |
1735 | int ret; |
1736 | ||
1737 | if (caa_unlikely(key > ja->key_max)) | |
1738 | return -EINVAL; | |
1739 | tree_depth = ja->tree_depth; | |
1740 | ||
1741 | retry: | |
4d6ef45e | 1742 | nr_snapshot = 0; |
35170a44 MD |
1743 | dbg_printf("cds_ja_del attempt: key %" PRIu64 ", node %p\n", |
1744 | key, node); | |
1745 | ||
1746 | /* snapshot for level 0 is only for shadow node lookup */ | |
4d6ef45e MD |
1747 | snapshot_n[0] = 0; |
1748 | snapshot_n[1] = 0; | |
af3cbd45 | 1749 | snapshot_ptr[nr_snapshot] = NULL; |
35170a44 MD |
1750 | snapshot[nr_snapshot++] = (struct cds_ja_inode_flag *) &ja->root; |
1751 | node_flag = rcu_dereference(ja->root); | |
af3cbd45 | 1752 | prev_node_flag_ptr = &ja->root; |
fa112799 | 1753 | node_flag_ptr = &ja->root; |
35170a44 MD |
1754 | |
1755 | /* Iterate on all internal levels */ | |
1756 | for (i = 1; i < tree_depth; i++) { | |
1757 | uint8_t iter_key; | |
1758 | ||
1759 | dbg_printf("cds_ja_del iter node_flag %p\n", | |
1760 | node_flag); | |
1761 | if (!ja_node_ptr(node_flag)) { | |
1762 | return -ENOENT; | |
1763 | } | |
35170a44 | 1764 | iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (tree_depth - i - 1))); |
4d6ef45e | 1765 | snapshot_n[nr_snapshot + 1] = iter_key; |
af3cbd45 MD |
1766 | snapshot_ptr[nr_snapshot] = prev_node_flag_ptr; |
1767 | snapshot[nr_snapshot++] = node_flag; | |
35170a44 | 1768 | node_flag = ja_node_get_nth(node_flag, |
af3cbd45 | 1769 | &prev_node_flag_ptr, |
b62a8d0c | 1770 | NULL, |
fa112799 | 1771 | &node_flag_ptr, |
35170a44 | 1772 | iter_key); |
af3cbd45 MD |
1773 | dbg_printf("cds_ja_del iter key lookup %u finds node_flag %p, prev_node_flag_ptr %p\n", |
1774 | (unsigned int) iter_key, node_flag, | |
1775 | prev_node_flag_ptr); | |
35170a44 | 1776 | } |
35170a44 MD |
1777 | /* |
1778 | * We reached bottom of tree, try to find the node we are trying | |
1779 | * to remove. Fail if we cannot find it. | |
1780 | */ | |
1781 | if (!ja_node_ptr(node_flag)) { | |
4d6ef45e MD |
1782 | dbg_printf("cds_ja_del: no node found for key %" PRIu64 "\n", |
1783 | key); | |
35170a44 MD |
1784 | return -ENOENT; |
1785 | } else { | |
4d6ef45e | 1786 | struct cds_hlist_head hlist_head; |
35170a44 | 1787 | struct cds_hlist_node *hlist_node; |
af3cbd45 MD |
1788 | struct cds_ja_node *entry, *match = NULL; |
1789 | int count = 0; | |
35170a44 | 1790 | |
4d6ef45e MD |
1791 | hlist_head.next = |
1792 | (struct cds_hlist_node *) ja_node_ptr(node_flag); | |
af3cbd45 | 1793 | cds_hlist_for_each_entry_rcu(entry, |
35170a44 | 1794 | hlist_node, |
4d6ef45e | 1795 | &hlist_head, |
35170a44 | 1796 | list) { |
4d6ef45e | 1797 | dbg_printf("cds_ja_del: compare %p with entry %p\n", node, entry); |
af3cbd45 MD |
1798 | if (entry == node) |
1799 | match = entry; | |
1800 | count++; | |
35170a44 | 1801 | } |
4d6ef45e MD |
1802 | if (!match) { |
1803 | dbg_printf("cds_ja_del: no node match for node %p key %" PRIu64 "\n", node, key); | |
35170a44 | 1804 | return -ENOENT; |
4d6ef45e | 1805 | } |
af3cbd45 MD |
1806 | assert(count > 0); |
1807 | if (count == 1) { | |
1808 | /* | |
4d6ef45e MD |
1809 | * Removing last of duplicates. Last snapshot |
1810 | * does not have a shadow node (external leafs). | |
af3cbd45 MD |
1811 | */ |
1812 | snapshot_ptr[nr_snapshot] = prev_node_flag_ptr; | |
1813 | snapshot[nr_snapshot++] = node_flag; | |
1814 | ret = ja_detach_node(ja, snapshot, snapshot_ptr, | |
1815 | snapshot_n, nr_snapshot, key, node); | |
1816 | } else { | |
f2758d14 | 1817 | ret = ja_unchain_node(ja, snapshot[nr_snapshot - 1], |
013a6083 | 1818 | node_flag_ptr, node_flag, match); |
af3cbd45 | 1819 | } |
35170a44 | 1820 | } |
b306a0fe MD |
1821 | /* |
1822 | * Explanation of -ENOENT handling: caused by concurrent delete | |
1823 | * between RCU lookup and actual removal. Need to re-do the | |
1824 | * lookup and removal attempt. | |
1825 | */ | |
1826 | if (ret == -EAGAIN || ret == -ENOENT) | |
35170a44 MD |
1827 | goto retry; |
1828 | return ret; | |
1829 | } | |
1830 | ||
b4540e8a MD |
1831 | struct cds_ja *_cds_ja_new(unsigned int key_bits, |
1832 | const struct rcu_flavor_struct *flavor) | |
be9a7474 MD |
1833 | { |
1834 | struct cds_ja *ja; | |
b0f74e47 | 1835 | int ret; |
f07b240f | 1836 | struct cds_ja_shadow_node *root_shadow_node; |
be9a7474 MD |
1837 | |
1838 | ja = calloc(sizeof(*ja), 1); | |
1839 | if (!ja) | |
1840 | goto ja_error; | |
b4540e8a MD |
1841 | |
1842 | switch (key_bits) { | |
1843 | case 8: | |
b4540e8a | 1844 | case 16: |
1216b3d2 | 1845 | case 24: |
b4540e8a | 1846 | case 32: |
1216b3d2 MD |
1847 | case 40: |
1848 | case 48: | |
1849 | case 56: | |
1850 | ja->key_max = (1ULL << key_bits) - 1; | |
b4540e8a MD |
1851 | break; |
1852 | case 64: | |
1853 | ja->key_max = UINT64_MAX; | |
1854 | break; | |
1855 | default: | |
1856 | goto check_error; | |
1857 | } | |
1858 | ||
be9a7474 | 1859 | /* ja->root is NULL */ |
5a9a87dd | 1860 | /* tree_depth 0 is for pointer to root node */ |
582a6ade | 1861 | ja->tree_depth = (key_bits >> JA_LOG2_BITS_PER_BYTE) + 1; |
a2a7ff59 | 1862 | assert(ja->tree_depth <= JA_MAX_DEPTH); |
be9a7474 MD |
1863 | ja->ht = rcuja_create_ht(flavor); |
1864 | if (!ja->ht) | |
1865 | goto ht_error; | |
b0f74e47 MD |
1866 | |
1867 | /* | |
1868 | * Note: we should not free this node until judy array destroy. | |
1869 | */ | |
f07b240f | 1870 | root_shadow_node = rcuja_shadow_set(ja->ht, |
3d8fe307 MD |
1871 | (struct cds_ja_inode_flag *) &ja->root, |
1872 | NULL, ja); | |
f07b240f MD |
1873 | if (!root_shadow_node) { |
1874 | ret = -ENOMEM; | |
b0f74e47 | 1875 | goto ht_node_error; |
f07b240f | 1876 | } |
3d8fe307 | 1877 | root_shadow_node->level = 0; |
b0f74e47 | 1878 | |
be9a7474 MD |
1879 | return ja; |
1880 | ||
b0f74e47 MD |
1881 | ht_node_error: |
1882 | ret = rcuja_delete_ht(ja->ht); | |
1883 | assert(!ret); | |
be9a7474 | 1884 | ht_error: |
b4540e8a | 1885 | check_error: |
be9a7474 MD |
1886 | free(ja); |
1887 | ja_error: | |
1888 | return NULL; | |
1889 | } | |
1890 | ||
3d8fe307 MD |
1891 | /* |
1892 | * Called from RCU read-side CS. | |
1893 | */ | |
1894 | __attribute__((visibility("protected"))) | |
1895 | void rcuja_free_all_children(struct cds_ja_shadow_node *shadow_node, | |
1896 | struct cds_ja_inode_flag *node_flag, | |
1897 | void (*free_node_cb)(struct rcu_head *head)) | |
1898 | { | |
1899 | const struct rcu_flavor_struct *flavor; | |
1900 | unsigned int type_index; | |
1901 | struct cds_ja_inode *node; | |
1902 | const struct cds_ja_type *type; | |
1903 | ||
1904 | flavor = cds_lfht_rcu_flavor(shadow_node->ja->ht); | |
1905 | node = ja_node_ptr(node_flag); | |
1906 | assert(node != NULL); | |
1907 | type_index = ja_node_type(node_flag); | |
1908 | type = &ja_types[type_index]; | |
1909 | ||
1910 | switch (type->type_class) { | |
1911 | case RCU_JA_LINEAR: | |
1912 | { | |
1913 | uint8_t nr_child = | |
1914 | ja_linear_node_get_nr_child(type, node); | |
1915 | unsigned int i; | |
1916 | ||
1917 | for (i = 0; i < nr_child; i++) { | |
1918 | struct cds_ja_inode_flag *iter; | |
1919 | struct cds_hlist_head head; | |
1920 | struct cds_ja_node *entry; | |
1921 | struct cds_hlist_node *pos; | |
1922 | uint8_t v; | |
1923 | ||
1924 | ja_linear_node_get_ith_pos(type, node, i, &v, &iter); | |
1925 | if (!iter) | |
1926 | continue; | |
1927 | head.next = (struct cds_hlist_node *) iter; | |
1928 | cds_hlist_for_each_entry_rcu(entry, pos, &head, list) { | |
1929 | flavor->update_call_rcu(&entry->head, free_node_cb); | |
1930 | } | |
1931 | } | |
1932 | break; | |
1933 | } | |
1934 | case RCU_JA_POOL: | |
1935 | { | |
1936 | unsigned int pool_nr; | |
1937 | ||
1938 | for (pool_nr = 0; pool_nr < (1U << type->nr_pool_order); pool_nr++) { | |
1939 | struct cds_ja_inode *pool = | |
1940 | ja_pool_node_get_ith_pool(type, node, pool_nr); | |
1941 | uint8_t nr_child = | |
1942 | ja_linear_node_get_nr_child(type, pool); | |
1943 | unsigned int j; | |
1944 | ||
1945 | for (j = 0; j < nr_child; j++) { | |
1946 | struct cds_ja_inode_flag *iter; | |
1947 | struct cds_hlist_head head; | |
1948 | struct cds_ja_node *entry; | |
1949 | struct cds_hlist_node *pos; | |
1950 | uint8_t v; | |
1951 | ||
1952 | ja_linear_node_get_ith_pos(type, node, j, &v, &iter); | |
1953 | if (!iter) | |
1954 | continue; | |
1955 | head.next = (struct cds_hlist_node *) iter; | |
1956 | cds_hlist_for_each_entry_rcu(entry, pos, &head, list) { | |
1957 | flavor->update_call_rcu(&entry->head, free_node_cb); | |
1958 | } | |
1959 | } | |
1960 | } | |
1961 | break; | |
1962 | } | |
1963 | case RCU_JA_NULL: | |
1964 | break; | |
1965 | case RCU_JA_PIGEON: | |
1966 | { | |
1967 | uint8_t nr_child; | |
1968 | unsigned int i; | |
1969 | ||
1970 | nr_child = shadow_node->nr_child; | |
1971 | for (i = 0; i < nr_child; i++) { | |
1972 | struct cds_ja_inode_flag *iter; | |
1973 | struct cds_hlist_head head; | |
1974 | struct cds_ja_node *entry; | |
1975 | struct cds_hlist_node *pos; | |
1976 | ||
1977 | iter = ja_pigeon_node_get_ith_pos(type, node, i); | |
1978 | if (!iter) | |
1979 | continue; | |
1980 | head.next = (struct cds_hlist_node *) iter; | |
1981 | cds_hlist_for_each_entry_rcu(entry, pos, &head, list) { | |
1982 | flavor->update_call_rcu(&entry->head, free_node_cb); | |
1983 | } | |
1984 | } | |
1985 | break; | |
1986 | } | |
1987 | default: | |
1988 | assert(0); | |
1989 | } | |
1990 | } | |
1991 | ||
be9a7474 MD |
1992 | /* |
1993 | * There should be no more concurrent add to the judy array while it is | |
1994 | * being destroyed (ensured by the caller). | |
1995 | */ | |
3d8fe307 MD |
1996 | int cds_ja_destroy(struct cds_ja *ja, |
1997 | void (*free_node_cb)(struct rcu_head *head)) | |
be9a7474 | 1998 | { |
b4540e8a MD |
1999 | int ret; |
2000 | ||
be9a7474 | 2001 | rcuja_shadow_prune(ja->ht, |
3d8fe307 MD |
2002 | RCUJA_SHADOW_CLEAR_FREE_NODE | RCUJA_SHADOW_CLEAR_FREE_LOCK, |
2003 | free_node_cb); | |
b4540e8a MD |
2004 | ret = rcuja_delete_ht(ja->ht); |
2005 | if (ret) | |
2006 | return ret; | |
f07b240f MD |
2007 | if (uatomic_read(&ja->nr_fallback)) |
2008 | fprintf(stderr, | |
2009 | "[warning] RCU Judy Array used %lu fallback node(s)\n", | |
2010 | uatomic_read(&ja->nr_fallback)); | |
b4540e8a | 2011 | free(ja); |
41975c12 | 2012 | return 0; |
be9a7474 | 2013 | } |