Commit | Line | Data |
---|---|---|
61009379 MD |
1 | /* |
2 | * rcuja/rcuja.c | |
3 | * | |
4 | * Userspace RCU library - RCU Judy Array | |
5 | * | |
6 | * Copyright 2012 - Mathieu Desnoyers <mathieu.desnoyers@efficios.com> | |
7 | * | |
8 | * This library is free software; you can redistribute it and/or | |
9 | * modify it under the terms of the GNU Lesser General Public | |
10 | * License as published by the Free Software Foundation; either | |
11 | * version 2.1 of the License, or (at your option) any later version. | |
12 | * | |
13 | * This library is distributed in the hope that it will be useful, | |
14 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
15 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | |
16 | * Lesser General Public License for more details. | |
17 | * | |
18 | * You should have received a copy of the GNU Lesser General Public | |
19 | * License along with this library; if not, write to the Free Software | |
20 | * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA | |
21 | */ | |
22 | ||
195e72d3 | 23 | #define _LGPL_SOURCE |
e5227865 | 24 | #include <stdint.h> |
8e519e3c | 25 | #include <errno.h> |
d68c6810 | 26 | #include <limits.h> |
61009379 | 27 | #include <urcu/rcuja.h> |
d68c6810 MD |
28 | #include <urcu/compiler.h> |
29 | #include <urcu/arch.h> | |
30 | #include <assert.h> | |
8e519e3c | 31 | #include <urcu-pointer.h> |
f07b240f | 32 | #include <urcu/uatomic.h> |
b4540e8a | 33 | #include <stdint.h> |
8e519e3c | 34 | |
61009379 | 35 | #include "rcuja-internal.h" |
d68c6810 | 36 | #include "bitfield.h" |
61009379 | 37 | |
d96bfb0d | 38 | enum cds_ja_type_class { |
e5227865 | 39 | RCU_JA_LINEAR = 0, /* Type A */ |
fd800776 MD |
40 | /* 32-bit: 1 to 25 children, 8 to 128 bytes */ |
41 | /* 64-bit: 1 to 28 children, 16 to 256 bytes */ | |
42 | RCU_JA_POOL = 1, /* Type B */ | |
43 | /* 32-bit: 26 to 100 children, 256 to 512 bytes */ | |
44 | /* 64-bit: 29 to 112 children, 512 to 1024 bytes */ | |
e5227865 | 45 | RCU_JA_PIGEON = 2, /* Type C */ |
fd800776 MD |
46 | /* 32-bit: 101 to 256 children, 1024 bytes */ |
47 | /* 64-bit: 113 to 256 children, 2048 bytes */ | |
e5227865 | 48 | /* Leaf nodes are implicit from their height in the tree */ |
1db4943c | 49 | RCU_JA_NR_TYPES, |
e1db2db5 MD |
50 | |
51 | RCU_JA_NULL, /* not an encoded type, but keeps code regular */ | |
e5227865 MD |
52 | }; |
53 | ||
d96bfb0d MD |
54 | struct cds_ja_type { |
55 | enum cds_ja_type_class type_class; | |
8e519e3c MD |
56 | uint16_t min_child; /* minimum number of children: 1 to 256 */ |
57 | uint16_t max_child; /* maximum number of children: 1 to 256 */ | |
58 | uint16_t max_linear_child; /* per-pool max nr. children: 1 to 256 */ | |
59 | uint16_t order; /* node size is (1 << order), in bytes */ | |
fd800776 MD |
60 | uint16_t nr_pool_order; /* number of pools */ |
61 | uint16_t pool_size_order; /* pool size */ | |
e5227865 MD |
62 | }; |
63 | ||
64 | /* | |
65 | * Iteration on the array to find the right node size for the number of | |
d68c6810 | 66 | * children stops when it reaches .max_child == 256 (this is the largest |
e5227865 | 67 | * possible node size, which contains 256 children). |
d68c6810 MD |
68 | * The min_child overlaps with the previous max_child to provide an |
69 | * hysteresis loop to reallocation for patterns of cyclic add/removal | |
70 | * within the same node. | |
71 | * The node the index within the following arrays is represented on 3 | |
72 | * bits. It identifies the node type, min/max number of children, and | |
73 | * the size order. | |
3d45251f MD |
74 | * The max_child values for the RCU_JA_POOL below result from |
75 | * statistical approximation: over million populations, the max_child | |
76 | * covers between 97% and 99% of the populations generated. Therefore, a | |
77 | * fallback should exist to cover the rare extreme population unbalance | |
78 | * cases, but it will not have a major impact on speed nor space | |
79 | * consumption, since those are rare cases. | |
e5227865 | 80 | */ |
e5227865 | 81 | |
d68c6810 MD |
82 | #if (CAA_BITS_PER_LONG < 64) |
83 | /* 32-bit pointers */ | |
1db4943c MD |
84 | enum { |
85 | ja_type_0_max_child = 1, | |
86 | ja_type_1_max_child = 3, | |
87 | ja_type_2_max_child = 6, | |
88 | ja_type_3_max_child = 12, | |
89 | ja_type_4_max_child = 25, | |
90 | ja_type_5_max_child = 48, | |
91 | ja_type_6_max_child = 92, | |
92 | ja_type_7_max_child = 256, | |
e1db2db5 | 93 | ja_type_8_max_child = 0, /* NULL */ |
1db4943c MD |
94 | }; |
95 | ||
8e519e3c MD |
96 | enum { |
97 | ja_type_0_max_linear_child = 1, | |
98 | ja_type_1_max_linear_child = 3, | |
99 | ja_type_2_max_linear_child = 6, | |
100 | ja_type_3_max_linear_child = 12, | |
101 | ja_type_4_max_linear_child = 25, | |
102 | ja_type_5_max_linear_child = 24, | |
103 | ja_type_6_max_linear_child = 23, | |
104 | }; | |
105 | ||
1db4943c MD |
106 | enum { |
107 | ja_type_5_nr_pool_order = 1, | |
108 | ja_type_6_nr_pool_order = 2, | |
109 | }; | |
110 | ||
d96bfb0d | 111 | const struct cds_ja_type ja_types[] = { |
8e519e3c MD |
112 | { .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, }, |
113 | { .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, }, | |
114 | { .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, }, | |
115 | { .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, }, | |
116 | { .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 | 117 | |
fd800776 | 118 | /* Pools may fill sooner than max_child */ |
8e519e3c MD |
119 | { .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, }, |
120 | { .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 |
121 | |
122 | /* | |
123 | * TODO: Upon node removal below min_child, if child pool is | |
124 | * filled beyond capacity, we need to roll back to pigeon. | |
125 | */ | |
1db4943c | 126 | { .type_class = RCU_JA_PIGEON, .min_child = 89, .max_child = ja_type_7_max_child, .order = 10, }, |
e1db2db5 MD |
127 | |
128 | { .type_class = RCU_JA_NULL, .min_child = 0, .max_child = ja_type_8_max_child, }, | |
d68c6810 | 129 | }; |
d68c6810 MD |
130 | #else /* !(CAA_BITS_PER_LONG < 64) */ |
131 | /* 64-bit pointers */ | |
1db4943c MD |
132 | enum { |
133 | ja_type_0_max_child = 1, | |
134 | ja_type_1_max_child = 3, | |
135 | ja_type_2_max_child = 7, | |
136 | ja_type_3_max_child = 14, | |
137 | ja_type_4_max_child = 28, | |
138 | ja_type_5_max_child = 54, | |
139 | ja_type_6_max_child = 104, | |
140 | ja_type_7_max_child = 256, | |
e1db2db5 | 141 | ja_type_8_max_child = 256, |
1db4943c MD |
142 | }; |
143 | ||
8e519e3c MD |
144 | enum { |
145 | ja_type_0_max_linear_child = 1, | |
146 | ja_type_1_max_linear_child = 3, | |
147 | ja_type_2_max_linear_child = 7, | |
148 | ja_type_3_max_linear_child = 14, | |
149 | ja_type_4_max_linear_child = 28, | |
150 | ja_type_5_max_linear_child = 27, | |
151 | ja_type_6_max_linear_child = 26, | |
152 | }; | |
153 | ||
1db4943c MD |
154 | enum { |
155 | ja_type_5_nr_pool_order = 1, | |
156 | ja_type_6_nr_pool_order = 2, | |
157 | }; | |
158 | ||
d96bfb0d | 159 | const struct cds_ja_type ja_types[] = { |
8e519e3c MD |
160 | { .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, }, |
161 | { .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, }, | |
162 | { .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, }, | |
163 | { .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, }, | |
164 | { .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 | 165 | |
3d45251f | 166 | /* Pools may fill sooner than max_child. */ |
8e519e3c MD |
167 | { .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, }, |
168 | { .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 | 169 | |
3d45251f MD |
170 | /* |
171 | * TODO: Upon node removal below min_child, if child pool is | |
172 | * filled beyond capacity, we need to roll back to pigeon. | |
173 | */ | |
1db4943c | 174 | { .type_class = RCU_JA_PIGEON, .min_child = 101, .max_child = ja_type_7_max_child, .order = 11, }, |
e1db2db5 MD |
175 | |
176 | { .type_class = RCU_JA_NULL, .min_child = 0, .max_child = ja_type_8_max_child, }, | |
e5227865 | 177 | }; |
d68c6810 | 178 | #endif /* !(BITS_PER_LONG < 64) */ |
e5227865 | 179 | |
1db4943c MD |
180 | static inline __attribute__((unused)) |
181 | void static_array_size_check(void) | |
182 | { | |
e1db2db5 | 183 | CAA_BUILD_BUG_ON(CAA_ARRAY_SIZE(ja_types) < JA_TYPE_MAX_NR); |
1db4943c MD |
184 | } |
185 | ||
e5227865 | 186 | /* |
d96bfb0d | 187 | * The cds_ja_node contains the compressed node data needed for |
1db4943c MD |
188 | * read-side. For linear and pool node configurations, it starts with a |
189 | * byte counting the number of children in the node. Then, the | |
190 | * node-specific data is placed. | |
191 | * The node mutex, if any is needed, protecting concurrent updated of | |
192 | * each node is placed in a separate hash table indexed by node address. | |
193 | * For the pigeon configuration, the number of children is also kept in | |
194 | * a separate hash table, indexed by node address, because it is only | |
195 | * required for updates. | |
e5227865 | 196 | */ |
1db4943c | 197 | |
ff38c745 MD |
198 | #define DECLARE_LINEAR_NODE(index) \ |
199 | struct { \ | |
200 | uint8_t nr_child; \ | |
201 | uint8_t child_value[ja_type_## index ##_max_linear_child]; \ | |
b4540e8a | 202 | struct cds_ja_inode_flag *child_ptr[ja_type_## index ##_max_linear_child]; \ |
ff38c745 MD |
203 | } |
204 | ||
205 | #define DECLARE_POOL_NODE(index) \ | |
206 | struct { \ | |
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 | } linear[1U << ja_type_## index ##_nr_pool_order]; \ |
212 | } | |
1db4943c | 213 | |
b4540e8a | 214 | struct cds_ja_inode { |
1db4943c MD |
215 | union { |
216 | /* Linear configuration */ | |
217 | DECLARE_LINEAR_NODE(0) conf_0; | |
218 | DECLARE_LINEAR_NODE(1) conf_1; | |
219 | DECLARE_LINEAR_NODE(2) conf_2; | |
220 | DECLARE_LINEAR_NODE(3) conf_3; | |
221 | DECLARE_LINEAR_NODE(4) conf_4; | |
222 | ||
223 | /* Pool configuration */ | |
224 | DECLARE_POOL_NODE(5) conf_5; | |
225 | DECLARE_POOL_NODE(6) conf_6; | |
226 | ||
227 | /* Pigeon configuration */ | |
228 | struct { | |
b4540e8a | 229 | struct cds_ja_inode_flag *child[ja_type_7_max_child]; |
1db4943c MD |
230 | } conf_7; |
231 | /* data aliasing nodes for computed accesses */ | |
b4540e8a | 232 | uint8_t data[sizeof(struct cds_ja_inode_flag *) * ja_type_7_max_child]; |
1db4943c | 233 | } u; |
e5227865 MD |
234 | }; |
235 | ||
2e313670 MD |
236 | enum ja_recompact { |
237 | JA_RECOMPACT, | |
238 | JA_RECOMPACT_ADD, | |
239 | JA_RECOMPACT_DEL, | |
240 | }; | |
241 | ||
b4540e8a | 242 | struct cds_ja_inode *alloc_cds_ja_node(const struct cds_ja_type *ja_type) |
e5227865 | 243 | { |
1db4943c | 244 | return calloc(1U << ja_type->order, sizeof(char)); |
e5227865 MD |
245 | } |
246 | ||
b4540e8a | 247 | void free_cds_ja_node(struct cds_ja_inode *node) |
e5227865 MD |
248 | { |
249 | free(node); | |
250 | } | |
251 | ||
d68c6810 MD |
252 | #define __JA_ALIGN_MASK(v, mask) (((v) + (mask)) & ~(mask)) |
253 | #define JA_ALIGN(v, align) __JA_ALIGN_MASK(v, (typeof(v)) (align) - 1) | |
254 | #define __JA_FLOOR_MASK(v, mask) ((v) & ~(mask)) | |
255 | #define JA_FLOOR(v, align) __JA_FLOOR_MASK(v, (typeof(v)) (align) - 1) | |
256 | ||
257 | static | |
1db4943c | 258 | uint8_t *align_ptr_size(uint8_t *ptr) |
d68c6810 | 259 | { |
1db4943c | 260 | return (uint8_t *) JA_ALIGN((unsigned long) ptr, sizeof(void *)); |
d68c6810 MD |
261 | } |
262 | ||
11c5e016 | 263 | static |
d96bfb0d | 264 | uint8_t ja_linear_node_get_nr_child(const struct cds_ja_type *type, |
b4540e8a | 265 | struct cds_ja_inode *node) |
11c5e016 MD |
266 | { |
267 | assert(type->type_class == RCU_JA_LINEAR || type->type_class == RCU_JA_POOL); | |
2e313670 | 268 | return rcu_dereference(node->u.data[0]); |
11c5e016 MD |
269 | } |
270 | ||
13a7f5a6 MD |
271 | /* |
272 | * The order in which values and pointers are does does not matter: if | |
273 | * a value is missing, we return NULL. If a value is there, but its | |
274 | * associated pointers is still NULL, we return NULL too. | |
275 | */ | |
d68c6810 | 276 | static |
b4540e8a MD |
277 | struct cds_ja_inode_flag *ja_linear_node_get_nth(const struct cds_ja_type *type, |
278 | struct cds_ja_inode *node, | |
5a9a87dd | 279 | struct cds_ja_inode_flag ***child_node_flag_ptr, |
b0ca2d21 | 280 | struct cds_ja_inode_flag ***node_flag_ptr, |
8e519e3c | 281 | uint8_t n) |
d68c6810 MD |
282 | { |
283 | uint8_t nr_child; | |
284 | uint8_t *values; | |
b4540e8a MD |
285 | struct cds_ja_inode_flag **pointers; |
286 | struct cds_ja_inode_flag *ptr; | |
d68c6810 MD |
287 | unsigned int i; |
288 | ||
8e519e3c | 289 | assert(type->type_class == RCU_JA_LINEAR || type->type_class == RCU_JA_POOL); |
d68c6810 | 290 | |
11c5e016 | 291 | nr_child = ja_linear_node_get_nr_child(type, node); |
13a7f5a6 | 292 | cmm_smp_rmb(); /* read nr_child before values and pointers */ |
8e519e3c MD |
293 | assert(nr_child <= type->max_linear_child); |
294 | assert(type->type_class != RCU_JA_LINEAR || nr_child >= type->min_child); | |
d68c6810 | 295 | |
1db4943c | 296 | values = &node->u.data[1]; |
d68c6810 | 297 | for (i = 0; i < nr_child; i++) { |
13a7f5a6 | 298 | if (CMM_LOAD_SHARED(values[i]) == n) |
d68c6810 MD |
299 | break; |
300 | } | |
b0ca2d21 MD |
301 | if (i >= nr_child) { |
302 | if (caa_unlikely(node_flag_ptr)) | |
303 | *node_flag_ptr = NULL; | |
d68c6810 | 304 | return NULL; |
b0ca2d21 | 305 | } |
b4540e8a | 306 | pointers = (struct cds_ja_inode_flag **) align_ptr_size(&values[type->max_linear_child]); |
13a7f5a6 | 307 | ptr = rcu_dereference(pointers[i]); |
2e313670 MD |
308 | if (caa_unlikely(child_node_flag_ptr) && ptr) |
309 | *child_node_flag_ptr = &pointers[i]; | |
b0ca2d21 MD |
310 | if (caa_unlikely(node_flag_ptr)) |
311 | *node_flag_ptr = &pointers[i]; | |
d68c6810 MD |
312 | return ptr; |
313 | } | |
314 | ||
11c5e016 | 315 | static |
5a9a87dd | 316 | void ja_linear_node_get_ith_pos(const struct cds_ja_type *type, |
b4540e8a | 317 | struct cds_ja_inode *node, |
11c5e016 MD |
318 | uint8_t i, |
319 | uint8_t *v, | |
b4540e8a | 320 | struct cds_ja_inode_flag **iter) |
11c5e016 MD |
321 | { |
322 | uint8_t *values; | |
b4540e8a | 323 | struct cds_ja_inode_flag **pointers; |
11c5e016 MD |
324 | |
325 | assert(type->type_class == RCU_JA_LINEAR || type->type_class == RCU_JA_POOL); | |
326 | assert(i < ja_linear_node_get_nr_child(type, node)); | |
327 | ||
328 | values = &node->u.data[1]; | |
329 | *v = values[i]; | |
b4540e8a | 330 | pointers = (struct cds_ja_inode_flag **) align_ptr_size(&values[type->max_linear_child]); |
11c5e016 MD |
331 | *iter = pointers[i]; |
332 | } | |
333 | ||
d68c6810 | 334 | static |
b4540e8a MD |
335 | struct cds_ja_inode_flag *ja_pool_node_get_nth(const struct cds_ja_type *type, |
336 | struct cds_ja_inode *node, | |
5a9a87dd | 337 | struct cds_ja_inode_flag ***child_node_flag_ptr, |
b0ca2d21 | 338 | struct cds_ja_inode_flag ***node_flag_ptr, |
8e519e3c | 339 | uint8_t n) |
d68c6810 | 340 | { |
b4540e8a | 341 | struct cds_ja_inode *linear; |
d68c6810 | 342 | |
fd800776 | 343 | assert(type->type_class == RCU_JA_POOL); |
e1db2db5 MD |
344 | /* |
345 | * TODO: currently, we select the pool by highest bits. We | |
346 | * should support various encodings. | |
347 | */ | |
b4540e8a | 348 | linear = (struct cds_ja_inode *) |
1db4943c | 349 | &node->u.data[((unsigned long) n >> (CHAR_BIT - type->nr_pool_order)) << type->pool_size_order]; |
b0ca2d21 MD |
350 | return ja_linear_node_get_nth(type, linear, child_node_flag_ptr, |
351 | node_flag_ptr, n); | |
d68c6810 MD |
352 | } |
353 | ||
11c5e016 | 354 | static |
b4540e8a MD |
355 | struct cds_ja_inode *ja_pool_node_get_ith_pool(const struct cds_ja_type *type, |
356 | struct cds_ja_inode *node, | |
11c5e016 MD |
357 | uint8_t i) |
358 | { | |
359 | assert(type->type_class == RCU_JA_POOL); | |
b4540e8a | 360 | return (struct cds_ja_inode *) |
11c5e016 MD |
361 | &node->u.data[(unsigned int) i << type->pool_size_order]; |
362 | } | |
363 | ||
d68c6810 | 364 | static |
b4540e8a MD |
365 | struct cds_ja_inode_flag *ja_pigeon_node_get_nth(const struct cds_ja_type *type, |
366 | struct cds_ja_inode *node, | |
5a9a87dd | 367 | struct cds_ja_inode_flag ***child_node_flag_ptr, |
b0ca2d21 | 368 | struct cds_ja_inode_flag ***node_flag_ptr, |
8e519e3c | 369 | uint8_t n) |
d68c6810 | 370 | { |
5a9a87dd MD |
371 | struct cds_ja_inode_flag **child_node_flag; |
372 | ||
d68c6810 | 373 | assert(type->type_class == RCU_JA_PIGEON); |
5a9a87dd | 374 | child_node_flag = &((struct cds_ja_inode_flag **) node->u.data)[n]; |
582a6ade MD |
375 | dbg_printf("ja_pigeon_node_get_nth child_node_flag_ptr %p\n", |
376 | child_node_flag); | |
377 | if (caa_unlikely(child_node_flag_ptr) && *child_node_flag) | |
5a9a87dd | 378 | *child_node_flag_ptr = child_node_flag; |
b0ca2d21 MD |
379 | if (caa_unlikely(node_flag_ptr)) |
380 | *node_flag_ptr = child_node_flag; | |
5a9a87dd | 381 | return rcu_dereference(*child_node_flag); |
d68c6810 MD |
382 | } |
383 | ||
2e313670 MD |
384 | static |
385 | struct cds_ja_inode_flag *ja_pigeon_node_get_ith_pos(const struct cds_ja_type *type, | |
386 | struct cds_ja_inode *node, | |
387 | uint8_t i) | |
388 | { | |
b0ca2d21 | 389 | return ja_pigeon_node_get_nth(type, node, NULL, NULL, i); |
2e313670 MD |
390 | } |
391 | ||
13a7f5a6 MD |
392 | /* |
393 | * ja_node_get_nth: get nth item from a node. | |
394 | * node_flag is already rcu_dereference'd. | |
395 | */ | |
d68c6810 | 396 | static |
41975c12 | 397 | struct cds_ja_inode_flag * ja_node_get_nth(struct cds_ja_inode_flag *node_flag, |
5a9a87dd | 398 | struct cds_ja_inode_flag ***child_node_flag_ptr, |
b0ca2d21 | 399 | struct cds_ja_inode_flag ***node_flag_ptr, |
8e519e3c | 400 | uint8_t n) |
d68c6810 MD |
401 | { |
402 | unsigned int type_index; | |
b4540e8a | 403 | struct cds_ja_inode *node; |
d96bfb0d | 404 | const struct cds_ja_type *type; |
d68c6810 | 405 | |
d68c6810 | 406 | node = ja_node_ptr(node_flag); |
5a9a87dd | 407 | assert(node != NULL); |
d68c6810 MD |
408 | type_index = ja_node_type(node_flag); |
409 | type = &ja_types[type_index]; | |
410 | ||
411 | switch (type->type_class) { | |
412 | case RCU_JA_LINEAR: | |
5a9a87dd | 413 | return ja_linear_node_get_nth(type, node, |
b0ca2d21 | 414 | child_node_flag_ptr, node_flag_ptr, n); |
fd800776 | 415 | case RCU_JA_POOL: |
5a9a87dd | 416 | return ja_pool_node_get_nth(type, node, |
b0ca2d21 | 417 | child_node_flag_ptr, node_flag_ptr, n); |
d68c6810 | 418 | case RCU_JA_PIGEON: |
5a9a87dd | 419 | return ja_pigeon_node_get_nth(type, node, |
b0ca2d21 | 420 | child_node_flag_ptr, node_flag_ptr, n); |
d68c6810 MD |
421 | default: |
422 | assert(0); | |
423 | return (void *) -1UL; | |
424 | } | |
425 | } | |
426 | ||
8e519e3c | 427 | static |
d96bfb0d | 428 | int ja_linear_node_set_nth(const struct cds_ja_type *type, |
b4540e8a | 429 | struct cds_ja_inode *node, |
d96bfb0d | 430 | struct cds_ja_shadow_node *shadow_node, |
8e519e3c | 431 | uint8_t n, |
b4540e8a | 432 | struct cds_ja_inode_flag *child_node_flag) |
8e519e3c MD |
433 | { |
434 | uint8_t nr_child; | |
435 | uint8_t *values, *nr_child_ptr; | |
b4540e8a | 436 | struct cds_ja_inode_flag **pointers; |
2e313670 | 437 | unsigned int i, unused = 0; |
8e519e3c MD |
438 | |
439 | assert(type->type_class == RCU_JA_LINEAR || type->type_class == RCU_JA_POOL); | |
440 | ||
441 | nr_child_ptr = &node->u.data[0]; | |
a2a7ff59 | 442 | dbg_printf("linear set nth: nr_child_ptr %p\n", nr_child_ptr); |
8e519e3c MD |
443 | nr_child = *nr_child_ptr; |
444 | assert(nr_child <= type->max_linear_child); | |
8e519e3c MD |
445 | |
446 | values = &node->u.data[1]; | |
2e313670 MD |
447 | pointers = (struct cds_ja_inode_flag **) align_ptr_size(&values[type->max_linear_child]); |
448 | /* Check if node value is already populated */ | |
8e519e3c | 449 | for (i = 0; i < nr_child; i++) { |
2e313670 MD |
450 | if (values[i] == n) { |
451 | if (pointers[i]) | |
452 | return -EEXIST; | |
453 | else | |
454 | break; | |
455 | } else { | |
456 | if (!pointers[i]) | |
457 | unused++; | |
458 | } | |
8e519e3c | 459 | } |
2e313670 MD |
460 | if (i == nr_child && nr_child >= type->max_linear_child) { |
461 | if (unused) | |
462 | return -ERANGE; /* recompact node */ | |
463 | else | |
464 | return -ENOSPC; /* No space left in this node type */ | |
465 | } | |
466 | ||
467 | assert(pointers[i] == NULL); | |
468 | rcu_assign_pointer(pointers[i], child_node_flag); | |
469 | /* If we expanded the nr_child, increment it */ | |
470 | if (i == nr_child) { | |
471 | CMM_STORE_SHARED(values[nr_child], n); | |
472 | /* write pointer and value before nr_child */ | |
473 | cmm_smp_wmb(); | |
474 | CMM_STORE_SHARED(*nr_child_ptr, nr_child + 1); | |
8e519e3c | 475 | } |
e1db2db5 | 476 | shadow_node->nr_child++; |
a2a7ff59 MD |
477 | dbg_printf("linear set nth: %u child, shadow: %u child, for node %p shadow %p\n", |
478 | (unsigned int) CMM_LOAD_SHARED(*nr_child_ptr), | |
479 | (unsigned int) shadow_node->nr_child, | |
480 | node, shadow_node); | |
481 | ||
8e519e3c MD |
482 | return 0; |
483 | } | |
484 | ||
485 | static | |
d96bfb0d | 486 | int ja_pool_node_set_nth(const struct cds_ja_type *type, |
b4540e8a | 487 | struct cds_ja_inode *node, |
d96bfb0d | 488 | struct cds_ja_shadow_node *shadow_node, |
8e519e3c | 489 | uint8_t n, |
b4540e8a | 490 | struct cds_ja_inode_flag *child_node_flag) |
8e519e3c | 491 | { |
b4540e8a | 492 | struct cds_ja_inode *linear; |
8e519e3c MD |
493 | |
494 | assert(type->type_class == RCU_JA_POOL); | |
b4540e8a | 495 | linear = (struct cds_ja_inode *) |
8e519e3c | 496 | &node->u.data[((unsigned long) n >> (CHAR_BIT - type->nr_pool_order)) << type->pool_size_order]; |
e1db2db5 MD |
497 | return ja_linear_node_set_nth(type, linear, shadow_node, |
498 | n, child_node_flag); | |
8e519e3c MD |
499 | } |
500 | ||
501 | static | |
d96bfb0d | 502 | int ja_pigeon_node_set_nth(const struct cds_ja_type *type, |
b4540e8a | 503 | struct cds_ja_inode *node, |
d96bfb0d | 504 | struct cds_ja_shadow_node *shadow_node, |
8e519e3c | 505 | uint8_t n, |
b4540e8a | 506 | struct cds_ja_inode_flag *child_node_flag) |
8e519e3c | 507 | { |
b4540e8a | 508 | struct cds_ja_inode_flag **ptr; |
8e519e3c MD |
509 | |
510 | assert(type->type_class == RCU_JA_PIGEON); | |
b4540e8a | 511 | ptr = &((struct cds_ja_inode_flag **) node->u.data)[n]; |
5a9a87dd | 512 | if (*ptr) |
8e519e3c MD |
513 | return -EEXIST; |
514 | rcu_assign_pointer(*ptr, child_node_flag); | |
e1db2db5 | 515 | shadow_node->nr_child++; |
8e519e3c MD |
516 | return 0; |
517 | } | |
518 | ||
d68c6810 | 519 | /* |
7a0b2331 | 520 | * _ja_node_set_nth: set nth item within a node. Return an error |
8e519e3c | 521 | * (negative error value) if it is already there. |
d68c6810 | 522 | */ |
8e519e3c | 523 | static |
d96bfb0d | 524 | int _ja_node_set_nth(const struct cds_ja_type *type, |
b4540e8a | 525 | struct cds_ja_inode *node, |
d96bfb0d | 526 | struct cds_ja_shadow_node *shadow_node, |
e1db2db5 | 527 | uint8_t n, |
b4540e8a | 528 | struct cds_ja_inode_flag *child_node_flag) |
8e519e3c | 529 | { |
8e519e3c MD |
530 | switch (type->type_class) { |
531 | case RCU_JA_LINEAR: | |
e1db2db5 | 532 | return ja_linear_node_set_nth(type, node, shadow_node, n, |
8e519e3c MD |
533 | child_node_flag); |
534 | case RCU_JA_POOL: | |
e1db2db5 | 535 | return ja_pool_node_set_nth(type, node, shadow_node, n, |
8e519e3c MD |
536 | child_node_flag); |
537 | case RCU_JA_PIGEON: | |
e1db2db5 | 538 | return ja_pigeon_node_set_nth(type, node, shadow_node, n, |
8e519e3c | 539 | child_node_flag); |
e1db2db5 MD |
540 | case RCU_JA_NULL: |
541 | return -ENOSPC; | |
8e519e3c MD |
542 | default: |
543 | assert(0); | |
544 | return -EINVAL; | |
545 | } | |
546 | ||
547 | return 0; | |
548 | } | |
7a0b2331 | 549 | |
2e313670 | 550 | static |
af3cbd45 | 551 | int ja_linear_node_clear_ptr(const struct cds_ja_type *type, |
2e313670 MD |
552 | struct cds_ja_inode *node, |
553 | struct cds_ja_shadow_node *shadow_node, | |
af3cbd45 | 554 | struct cds_ja_inode_flag **node_flag_ptr) |
2e313670 MD |
555 | { |
556 | uint8_t nr_child; | |
af3cbd45 | 557 | uint8_t *nr_child_ptr; |
2e313670 MD |
558 | |
559 | assert(type->type_class == RCU_JA_LINEAR || type->type_class == RCU_JA_POOL); | |
560 | ||
561 | nr_child_ptr = &node->u.data[0]; | |
af3cbd45 | 562 | dbg_printf("linear clear ptr: nr_child_ptr %p\n", nr_child_ptr); |
2e313670 MD |
563 | nr_child = *nr_child_ptr; |
564 | assert(nr_child <= type->max_linear_child); | |
565 | ||
2e313670 MD |
566 | if (shadow_node->fallback_removal_count) { |
567 | shadow_node->fallback_removal_count--; | |
568 | } else { | |
569 | if (shadow_node->nr_child <= type->min_child) { | |
570 | /* We need to try recompacting the node */ | |
571 | return -EFBIG; | |
572 | } | |
573 | } | |
af3cbd45 MD |
574 | assert(*node_flag_ptr != NULL); |
575 | rcu_assign_pointer(*node_flag_ptr, NULL); | |
2e313670 MD |
576 | /* |
577 | * Value and nr_child are never changed (would cause ABA issue). | |
578 | * Instead, we leave the pointer to NULL and recompact the node | |
579 | * once in a while. It is allowed to set a NULL pointer to a new | |
580 | * value without recompaction though. | |
581 | * Only update the shadow node accounting. | |
582 | */ | |
583 | shadow_node->nr_child--; | |
af3cbd45 | 584 | dbg_printf("linear clear ptr: %u child, shadow: %u child, for node %p shadow %p\n", |
2e313670 MD |
585 | (unsigned int) CMM_LOAD_SHARED(*nr_child_ptr), |
586 | (unsigned int) shadow_node->nr_child, | |
587 | node, shadow_node); | |
588 | ||
589 | return 0; | |
590 | } | |
591 | ||
592 | static | |
af3cbd45 | 593 | int ja_pool_node_clear_ptr(const struct cds_ja_type *type, |
2e313670 MD |
594 | struct cds_ja_inode *node, |
595 | struct cds_ja_shadow_node *shadow_node, | |
af3cbd45 | 596 | struct cds_ja_inode_flag **node_flag_ptr, |
2e313670 MD |
597 | uint8_t n) |
598 | { | |
599 | struct cds_ja_inode *linear; | |
600 | ||
601 | assert(type->type_class == RCU_JA_POOL); | |
602 | linear = (struct cds_ja_inode *) | |
603 | &node->u.data[((unsigned long) n >> (CHAR_BIT - type->nr_pool_order)) << type->pool_size_order]; | |
af3cbd45 | 604 | return ja_linear_node_clear_ptr(type, linear, shadow_node, node_flag_ptr); |
2e313670 MD |
605 | } |
606 | ||
607 | static | |
af3cbd45 | 608 | int ja_pigeon_node_clear_ptr(const struct cds_ja_type *type, |
2e313670 MD |
609 | struct cds_ja_inode *node, |
610 | struct cds_ja_shadow_node *shadow_node, | |
af3cbd45 | 611 | struct cds_ja_inode_flag **node_flag_ptr) |
2e313670 | 612 | { |
2e313670 | 613 | assert(type->type_class == RCU_JA_PIGEON); |
4d6ef45e | 614 | dbg_printf("ja_pigeon_node_clear_ptr: clearing ptr: %p\n", *node_flag_ptr); |
af3cbd45 | 615 | rcu_assign_pointer(*node_flag_ptr, NULL); |
2e313670 MD |
616 | shadow_node->nr_child--; |
617 | return 0; | |
618 | } | |
619 | ||
620 | /* | |
af3cbd45 | 621 | * _ja_node_clear_ptr: clear ptr item within a node. Return an error |
2e313670 MD |
622 | * (negative error value) if it is not found (-ENOENT). |
623 | */ | |
624 | static | |
af3cbd45 | 625 | int _ja_node_clear_ptr(const struct cds_ja_type *type, |
2e313670 MD |
626 | struct cds_ja_inode *node, |
627 | struct cds_ja_shadow_node *shadow_node, | |
af3cbd45 | 628 | struct cds_ja_inode_flag **node_flag_ptr, |
2e313670 MD |
629 | uint8_t n) |
630 | { | |
631 | switch (type->type_class) { | |
632 | case RCU_JA_LINEAR: | |
af3cbd45 | 633 | return ja_linear_node_clear_ptr(type, node, shadow_node, node_flag_ptr); |
2e313670 | 634 | case RCU_JA_POOL: |
af3cbd45 | 635 | return ja_pool_node_clear_ptr(type, node, shadow_node, node_flag_ptr, n); |
2e313670 | 636 | case RCU_JA_PIGEON: |
af3cbd45 | 637 | return ja_pigeon_node_clear_ptr(type, node, shadow_node, node_flag_ptr); |
2e313670 MD |
638 | case RCU_JA_NULL: |
639 | return -ENOENT; | |
640 | default: | |
641 | assert(0); | |
642 | return -EINVAL; | |
643 | } | |
644 | ||
645 | return 0; | |
646 | } | |
647 | ||
7a0b2331 MD |
648 | /* |
649 | * ja_node_recompact_add: recompact a node, adding a new child. | |
e1db2db5 | 650 | * TODO: for pool type, take selection bit(s) into account. |
2e313670 | 651 | * Return 0 on success, -EAGAIN if need to retry, or other negative |
5a9a87dd | 652 | * error value otherwise. |
7a0b2331 MD |
653 | */ |
654 | static | |
2e313670 MD |
655 | int ja_node_recompact(enum ja_recompact mode, |
656 | struct cds_ja *ja, | |
e1db2db5 | 657 | unsigned int old_type_index, |
d96bfb0d | 658 | const struct cds_ja_type *old_type, |
b4540e8a | 659 | struct cds_ja_inode *old_node, |
5a9a87dd | 660 | struct cds_ja_shadow_node *shadow_node, |
3d8fe307 | 661 | struct cds_ja_inode_flag **old_node_flag_ptr, uint8_t n, |
af3cbd45 MD |
662 | struct cds_ja_inode_flag *child_node_flag, |
663 | struct cds_ja_inode_flag **nullify_node_flag_ptr) | |
7a0b2331 | 664 | { |
e1db2db5 | 665 | unsigned int new_type_index; |
b4540e8a | 666 | struct cds_ja_inode *new_node; |
af3cbd45 | 667 | struct cds_ja_shadow_node *new_shadow_node = NULL; |
d96bfb0d | 668 | const struct cds_ja_type *new_type; |
3d8fe307 | 669 | struct cds_ja_inode_flag *new_node_flag, *old_node_flag; |
7a0b2331 | 670 | int ret; |
f07b240f | 671 | int fallback = 0; |
7a0b2331 | 672 | |
3d8fe307 MD |
673 | old_node_flag = *old_node_flag_ptr; |
674 | ||
2e313670 MD |
675 | switch (mode) { |
676 | case JA_RECOMPACT: | |
677 | new_type_index = old_type_index; | |
678 | break; | |
679 | case JA_RECOMPACT_ADD: | |
680 | if (!shadow_node || old_type_index == NODE_INDEX_NULL) { | |
681 | new_type_index = 0; | |
682 | } else { | |
683 | new_type_index = old_type_index + 1; | |
684 | } | |
685 | break; | |
686 | case JA_RECOMPACT_DEL: | |
687 | if (old_type_index == 0) { | |
688 | new_type_index = NODE_INDEX_NULL; | |
689 | } else { | |
690 | new_type_index = old_type_index - 1; | |
691 | } | |
692 | break; | |
693 | default: | |
694 | assert(0); | |
7a0b2331 | 695 | } |
a2a7ff59 | 696 | |
f07b240f | 697 | retry: /* for fallback */ |
582a6ade MD |
698 | dbg_printf("Recompact from type %d to type %d\n", |
699 | old_type_index, new_type_index); | |
7a0b2331 | 700 | new_type = &ja_types[new_type_index]; |
2e313670 MD |
701 | if (new_type_index != NODE_INDEX_NULL) { |
702 | new_node = alloc_cds_ja_node(new_type); | |
703 | if (!new_node) | |
704 | return -ENOMEM; | |
705 | new_node_flag = ja_node_flag(new_node, new_type_index); | |
706 | dbg_printf("Recompact inherit lock from %p\n", shadow_node); | |
3d8fe307 | 707 | new_shadow_node = rcuja_shadow_set(ja->ht, new_node_flag, shadow_node, ja); |
2e313670 MD |
708 | if (!new_shadow_node) { |
709 | free(new_node); | |
710 | return -ENOMEM; | |
711 | } | |
712 | if (fallback) | |
713 | new_shadow_node->fallback_removal_count = | |
714 | JA_FALLBACK_REMOVAL_COUNT; | |
715 | } else { | |
716 | new_node = NULL; | |
717 | new_node_flag = NULL; | |
e1db2db5 | 718 | } |
11c5e016 | 719 | |
2e313670 MD |
720 | assert(mode != JA_RECOMPACT_ADD || old_type->type_class != RCU_JA_PIGEON); |
721 | ||
722 | if (new_type_index == NODE_INDEX_NULL) | |
723 | goto skip_copy; | |
724 | ||
11c5e016 MD |
725 | switch (old_type->type_class) { |
726 | case RCU_JA_LINEAR: | |
727 | { | |
728 | uint8_t nr_child = | |
729 | ja_linear_node_get_nr_child(old_type, old_node); | |
730 | unsigned int i; | |
731 | ||
732 | for (i = 0; i < nr_child; i++) { | |
b4540e8a | 733 | struct cds_ja_inode_flag *iter; |
11c5e016 MD |
734 | uint8_t v; |
735 | ||
736 | ja_linear_node_get_ith_pos(old_type, old_node, i, &v, &iter); | |
737 | if (!iter) | |
738 | continue; | |
af3cbd45 | 739 | if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) |
2e313670 | 740 | continue; |
f07b240f MD |
741 | ret = _ja_node_set_nth(new_type, new_node, |
742 | new_shadow_node, | |
11c5e016 | 743 | v, iter); |
f07b240f MD |
744 | if (new_type->type_class == RCU_JA_POOL && ret) { |
745 | goto fallback_toosmall; | |
746 | } | |
11c5e016 MD |
747 | assert(!ret); |
748 | } | |
749 | break; | |
750 | } | |
751 | case RCU_JA_POOL: | |
752 | { | |
753 | unsigned int pool_nr; | |
754 | ||
755 | for (pool_nr = 0; pool_nr < (1U << old_type->nr_pool_order); pool_nr++) { | |
b4540e8a | 756 | struct cds_ja_inode *pool = |
11c5e016 MD |
757 | ja_pool_node_get_ith_pool(old_type, |
758 | old_node, pool_nr); | |
759 | uint8_t nr_child = | |
760 | ja_linear_node_get_nr_child(old_type, pool); | |
761 | unsigned int j; | |
762 | ||
763 | for (j = 0; j < nr_child; j++) { | |
b4540e8a | 764 | struct cds_ja_inode_flag *iter; |
11c5e016 MD |
765 | uint8_t v; |
766 | ||
767 | ja_linear_node_get_ith_pos(old_type, pool, | |
768 | j, &v, &iter); | |
769 | if (!iter) | |
770 | continue; | |
af3cbd45 | 771 | if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) |
2e313670 | 772 | continue; |
f07b240f MD |
773 | ret = _ja_node_set_nth(new_type, new_node, |
774 | new_shadow_node, | |
11c5e016 | 775 | v, iter); |
f07b240f MD |
776 | if (new_type->type_class == RCU_JA_POOL |
777 | && ret) { | |
778 | goto fallback_toosmall; | |
779 | } | |
11c5e016 MD |
780 | assert(!ret); |
781 | } | |
782 | } | |
783 | break; | |
7a0b2331 | 784 | } |
a2a7ff59 | 785 | case RCU_JA_NULL: |
2e313670 | 786 | assert(mode == JA_RECOMPACT_ADD); |
a2a7ff59 | 787 | break; |
11c5e016 | 788 | case RCU_JA_PIGEON: |
2e313670 MD |
789 | { |
790 | uint8_t nr_child; | |
791 | unsigned int i; | |
792 | ||
793 | assert(mode == JA_RECOMPACT_DEL); | |
794 | nr_child = shadow_node->nr_child; | |
795 | for (i = 0; i < nr_child; i++) { | |
796 | struct cds_ja_inode_flag *iter; | |
797 | ||
798 | iter = ja_pigeon_node_get_ith_pos(old_type, old_node, i); | |
799 | if (!iter) | |
800 | continue; | |
af3cbd45 | 801 | if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) |
2e313670 MD |
802 | continue; |
803 | ret = _ja_node_set_nth(new_type, new_node, | |
804 | new_shadow_node, | |
805 | i, iter); | |
806 | if (new_type->type_class == RCU_JA_POOL && ret) { | |
807 | goto fallback_toosmall; | |
808 | } | |
809 | assert(!ret); | |
810 | } | |
811 | break; | |
812 | } | |
11c5e016 MD |
813 | default: |
814 | assert(0); | |
5a9a87dd | 815 | ret = -EINVAL; |
f07b240f | 816 | goto end; |
11c5e016 | 817 | } |
2e313670 | 818 | skip_copy: |
11c5e016 | 819 | |
9a1c1915 | 820 | if (mode == JA_RECOMPACT_ADD) { |
2e313670 MD |
821 | /* add node */ |
822 | ret = _ja_node_set_nth(new_type, new_node, | |
823 | new_shadow_node, | |
824 | n, child_node_flag); | |
825 | assert(!ret); | |
826 | } | |
3d8fe307 MD |
827 | /* Return pointer to new recompacted node through old_node_flag_ptr */ |
828 | *old_node_flag_ptr = new_node_flag; | |
a2a7ff59 | 829 | if (old_node) { |
2e313670 MD |
830 | int flags; |
831 | ||
832 | flags = RCUJA_SHADOW_CLEAR_FREE_NODE; | |
833 | /* | |
834 | * It is OK to free the lock associated with a node | |
835 | * going to NULL, since we are holding the parent lock. | |
836 | * This synchronizes removal with re-add of that node. | |
837 | */ | |
838 | if (new_type_index == NODE_INDEX_NULL) | |
839 | flags = RCUJA_SHADOW_CLEAR_FREE_LOCK; | |
3d8fe307 | 840 | ret = rcuja_shadow_clear(ja->ht, old_node_flag, shadow_node, |
2e313670 | 841 | flags); |
a2a7ff59 MD |
842 | assert(!ret); |
843 | } | |
5a9a87dd MD |
844 | |
845 | ret = 0; | |
f07b240f | 846 | end: |
5a9a87dd | 847 | return ret; |
f07b240f MD |
848 | |
849 | fallback_toosmall: | |
850 | /* fallback if next pool is too small */ | |
af3cbd45 | 851 | assert(new_shadow_node); |
3d8fe307 | 852 | ret = rcuja_shadow_clear(ja->ht, new_node_flag, new_shadow_node, |
f07b240f MD |
853 | RCUJA_SHADOW_CLEAR_FREE_NODE); |
854 | assert(!ret); | |
855 | ||
2e313670 | 856 | /* Choose fallback type: pigeon */ |
f07b240f MD |
857 | new_type_index = (1UL << JA_TYPE_BITS) - 1; |
858 | dbg_printf("Fallback to type %d\n", new_type_index); | |
859 | uatomic_inc(&ja->nr_fallback); | |
860 | fallback = 1; | |
861 | goto retry; | |
7a0b2331 MD |
862 | } |
863 | ||
5a9a87dd | 864 | /* |
2e313670 | 865 | * Return 0 on success, -EAGAIN if need to retry, or other negative |
5a9a87dd MD |
866 | * error value otherwise. |
867 | */ | |
7a0b2331 | 868 | static |
d96bfb0d | 869 | int ja_node_set_nth(struct cds_ja *ja, |
b4540e8a | 870 | struct cds_ja_inode_flag **node_flag, uint8_t n, |
5a9a87dd MD |
871 | struct cds_ja_inode_flag *child_node_flag, |
872 | struct cds_ja_shadow_node *shadow_node) | |
7a0b2331 MD |
873 | { |
874 | int ret; | |
e1db2db5 | 875 | unsigned int type_index; |
d96bfb0d | 876 | const struct cds_ja_type *type; |
b4540e8a | 877 | struct cds_ja_inode *node; |
7a0b2331 | 878 | |
a2a7ff59 MD |
879 | dbg_printf("ja_node_set_nth for n=%u, node %p, shadow %p\n", |
880 | (unsigned int) n, ja_node_ptr(*node_flag), shadow_node); | |
881 | ||
e1db2db5 MD |
882 | node = ja_node_ptr(*node_flag); |
883 | type_index = ja_node_type(*node_flag); | |
884 | type = &ja_types[type_index]; | |
e1db2db5 MD |
885 | ret = _ja_node_set_nth(type, node, shadow_node, |
886 | n, child_node_flag); | |
2e313670 MD |
887 | switch (ret) { |
888 | case -ENOSPC: | |
e1db2db5 | 889 | /* Not enough space in node, need to recompact. */ |
2e313670 | 890 | ret = ja_node_recompact(JA_RECOMPACT_ADD, ja, type_index, type, node, |
af3cbd45 | 891 | shadow_node, node_flag, n, child_node_flag, NULL); |
2e313670 MD |
892 | break; |
893 | case -ERANGE: | |
894 | /* Node needs to be recompacted. */ | |
895 | ret = ja_node_recompact(JA_RECOMPACT, ja, type_index, type, node, | |
af3cbd45 | 896 | shadow_node, node_flag, n, child_node_flag, NULL); |
2e313670 MD |
897 | break; |
898 | } | |
899 | return ret; | |
900 | } | |
901 | ||
902 | /* | |
903 | * Return 0 on success, -EAGAIN if need to retry, or other negative | |
904 | * error value otherwise. | |
905 | */ | |
906 | static | |
af3cbd45 MD |
907 | int ja_node_clear_ptr(struct cds_ja *ja, |
908 | struct cds_ja_inode_flag **node_flag_ptr, /* Pointer to location to nullify */ | |
909 | struct cds_ja_inode_flag **parent_node_flag_ptr, /* Address of parent ptr in its parent */ | |
910 | struct cds_ja_shadow_node *shadow_node, /* of parent */ | |
911 | uint8_t n) | |
2e313670 MD |
912 | { |
913 | int ret; | |
914 | unsigned int type_index; | |
915 | const struct cds_ja_type *type; | |
916 | struct cds_ja_inode *node; | |
917 | ||
af3cbd45 MD |
918 | dbg_printf("ja_node_clear_ptr for node %p, shadow %p, target ptr %p\n", |
919 | ja_node_ptr(*parent_node_flag_ptr), shadow_node, node_flag_ptr); | |
2e313670 | 920 | |
af3cbd45 MD |
921 | node = ja_node_ptr(*parent_node_flag_ptr); |
922 | type_index = ja_node_type(*parent_node_flag_ptr); | |
2e313670 | 923 | type = &ja_types[type_index]; |
af3cbd45 | 924 | ret = _ja_node_clear_ptr(type, node, shadow_node, node_flag_ptr, n); |
2e313670 MD |
925 | if (ret == -EFBIG) { |
926 | /* Should to try recompaction. */ | |
927 | ret = ja_node_recompact(JA_RECOMPACT_DEL, ja, type_index, type, node, | |
af3cbd45 MD |
928 | shadow_node, parent_node_flag_ptr, n, NULL, |
929 | node_flag_ptr); | |
7a0b2331 MD |
930 | } |
931 | return ret; | |
932 | } | |
be9a7474 | 933 | |
af3cbd45 | 934 | struct cds_hlist_head cds_ja_lookup(struct cds_ja *ja, uint64_t key) |
b4540e8a | 935 | { |
41975c12 MD |
936 | unsigned int tree_depth, i; |
937 | struct cds_ja_inode_flag *node_flag; | |
af3cbd45 | 938 | struct cds_hlist_head head = { NULL }; |
41975c12 MD |
939 | |
940 | if (caa_unlikely(key > ja->key_max)) | |
af3cbd45 | 941 | return head; |
41975c12 | 942 | tree_depth = ja->tree_depth; |
5a9a87dd | 943 | node_flag = rcu_dereference(ja->root); |
41975c12 | 944 | |
5a9a87dd MD |
945 | /* level 0: root node */ |
946 | if (!ja_node_ptr(node_flag)) | |
af3cbd45 | 947 | return head; |
5a9a87dd MD |
948 | |
949 | for (i = 1; i < tree_depth; i++) { | |
79b41067 MD |
950 | uint8_t iter_key; |
951 | ||
952 | iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (tree_depth - i - 1))); | |
b0ca2d21 | 953 | node_flag = ja_node_get_nth(node_flag, NULL, NULL, |
79b41067 | 954 | iter_key); |
582a6ade MD |
955 | dbg_printf("cds_ja_lookup iter key lookup %u finds node_flag %p\n", |
956 | (unsigned int) iter_key, node_flag); | |
41975c12 | 957 | if (!ja_node_ptr(node_flag)) |
af3cbd45 | 958 | return head; |
41975c12 MD |
959 | } |
960 | ||
5a9a87dd | 961 | /* Last level lookup succeded. We got an actual match. */ |
af3cbd45 MD |
962 | head.next = (struct cds_hlist_node *) node_flag; |
963 | return head; | |
5a9a87dd MD |
964 | } |
965 | ||
966 | /* | |
967 | * We reached an unpopulated node. Create it and the children we need, | |
968 | * and then attach the entire branch to the current node. This may | |
969 | * trigger recompaction of the current node. Locks needed: node lock | |
970 | * (for add), and, possibly, parent node lock (to update pointer due to | |
971 | * node recompaction). | |
972 | * | |
973 | * First take node lock, check if recompaction is needed, then take | |
974 | * parent lock (if needed). Then we can proceed to create the new | |
975 | * branch. Publish the new branch, and release locks. | |
976 | * TODO: we currently always take the parent lock even when not needed. | |
977 | */ | |
978 | static | |
979 | int ja_attach_node(struct cds_ja *ja, | |
b0ca2d21 | 980 | struct cds_ja_inode_flag **attach_node_flag_ptr, |
5a9a87dd MD |
981 | struct cds_ja_inode_flag **node_flag_ptr, |
982 | struct cds_ja_inode_flag *node_flag, | |
983 | struct cds_ja_inode_flag *parent_node_flag, | |
984 | uint64_t key, | |
79b41067 | 985 | unsigned int level, |
5a9a87dd MD |
986 | struct cds_ja_node *child_node) |
987 | { | |
988 | struct cds_ja_shadow_node *shadow_node = NULL, | |
af3cbd45 | 989 | *parent_shadow_node = NULL; |
5a9a87dd MD |
990 | struct cds_ja_inode *node = ja_node_ptr(node_flag); |
991 | struct cds_ja_inode *parent_node = ja_node_ptr(parent_node_flag); | |
992 | struct cds_hlist_head head; | |
993 | struct cds_ja_inode_flag *iter_node_flag, *iter_dest_node_flag; | |
994 | int ret, i; | |
a2a7ff59 | 995 | struct cds_ja_inode_flag *created_nodes[JA_MAX_DEPTH]; |
5a9a87dd MD |
996 | int nr_created_nodes = 0; |
997 | ||
582a6ade MD |
998 | dbg_printf("Attach node at level %u (node %p, node_flag %p)\n", |
999 | level, node, node_flag); | |
a2a7ff59 | 1000 | |
5a9a87dd | 1001 | assert(node); |
3d8fe307 | 1002 | shadow_node = rcuja_shadow_lookup_lock(ja->ht, node_flag); |
5a9a87dd | 1003 | if (!shadow_node) { |
2e313670 | 1004 | ret = -EAGAIN; |
5a9a87dd MD |
1005 | goto end; |
1006 | } | |
1007 | if (parent_node) { | |
1008 | parent_shadow_node = rcuja_shadow_lookup_lock(ja->ht, | |
3d8fe307 | 1009 | parent_node_flag); |
5a9a87dd | 1010 | if (!parent_shadow_node) { |
2e313670 | 1011 | ret = -EAGAIN; |
5a9a87dd MD |
1012 | goto unlock_shadow; |
1013 | } | |
1014 | } | |
1015 | ||
fa112799 | 1016 | if (node_flag_ptr && ja_node_ptr(*node_flag_ptr)) { |
b306a0fe | 1017 | /* |
d810c97f | 1018 | * Target node is non-NULL: it has been updated between |
b306a0fe MD |
1019 | * RCU lookup and lock acquisition. We need to re-try |
1020 | * lookup and attach. | |
1021 | */ | |
1022 | ret = -EAGAIN; | |
1023 | goto unlock_parent; | |
1024 | } | |
1025 | ||
a2a7ff59 | 1026 | /* Create new branch, starting from bottom */ |
5a9a87dd MD |
1027 | CDS_INIT_HLIST_HEAD(&head); |
1028 | cds_hlist_add_head_rcu(&child_node->list, &head); | |
a2a7ff59 | 1029 | iter_node_flag = (struct cds_ja_inode_flag *) head.next; |
5a9a87dd | 1030 | |
79b41067 MD |
1031 | for (i = ja->tree_depth; i > (int) level; i--) { |
1032 | uint8_t iter_key; | |
1033 | ||
1034 | iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (ja->tree_depth - i))); | |
1035 | dbg_printf("branch creation level %d, key %u\n", | |
1036 | i - 1, (unsigned int) iter_key); | |
5a9a87dd MD |
1037 | iter_dest_node_flag = NULL; |
1038 | ret = ja_node_set_nth(ja, &iter_dest_node_flag, | |
79b41067 | 1039 | iter_key, |
5a9a87dd MD |
1040 | iter_node_flag, |
1041 | NULL); | |
1042 | if (ret) | |
1043 | goto check_error; | |
1044 | created_nodes[nr_created_nodes++] = iter_dest_node_flag; | |
1045 | iter_node_flag = iter_dest_node_flag; | |
1046 | } | |
1047 | ||
79b41067 MD |
1048 | if (level > 1) { |
1049 | uint8_t iter_key; | |
1050 | ||
1051 | iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (ja->tree_depth - level))); | |
a2a7ff59 MD |
1052 | /* We need to use set_nth on the previous level. */ |
1053 | iter_dest_node_flag = node_flag; | |
1054 | ret = ja_node_set_nth(ja, &iter_dest_node_flag, | |
79b41067 | 1055 | iter_key, |
a2a7ff59 MD |
1056 | iter_node_flag, |
1057 | shadow_node); | |
1058 | if (ret) | |
1059 | goto check_error; | |
1060 | created_nodes[nr_created_nodes++] = iter_dest_node_flag; | |
1061 | iter_node_flag = iter_dest_node_flag; | |
1062 | } | |
1063 | ||
5a9a87dd | 1064 | /* Publish new branch */ |
a2a7ff59 | 1065 | dbg_printf("Publish branch %p, replacing %p\n", |
b0ca2d21 MD |
1066 | iter_node_flag, *attach_node_flag_ptr); |
1067 | rcu_assign_pointer(*attach_node_flag_ptr, iter_node_flag); | |
5a9a87dd MD |
1068 | |
1069 | /* Success */ | |
1070 | ret = 0; | |
1071 | ||
1072 | check_error: | |
1073 | if (ret) { | |
1074 | for (i = 0; i < nr_created_nodes; i++) { | |
1075 | int tmpret; | |
a2a7ff59 MD |
1076 | int flags; |
1077 | ||
1078 | flags = RCUJA_SHADOW_CLEAR_FREE_LOCK; | |
1079 | if (i) | |
1080 | flags |= RCUJA_SHADOW_CLEAR_FREE_NODE; | |
5a9a87dd | 1081 | tmpret = rcuja_shadow_clear(ja->ht, |
3d8fe307 | 1082 | created_nodes[i], |
a2a7ff59 MD |
1083 | NULL, |
1084 | flags); | |
5a9a87dd MD |
1085 | assert(!tmpret); |
1086 | } | |
1087 | } | |
b306a0fe | 1088 | unlock_parent: |
5a9a87dd MD |
1089 | if (parent_shadow_node) |
1090 | rcuja_shadow_unlock(parent_shadow_node); | |
1091 | unlock_shadow: | |
1092 | if (shadow_node) | |
1093 | rcuja_shadow_unlock(shadow_node); | |
1094 | end: | |
1095 | return ret; | |
1096 | } | |
1097 | ||
1098 | /* | |
af3cbd45 MD |
1099 | * Lock the parent containing the hlist head pointer, and add node to list of |
1100 | * duplicates. Failure can happen if concurrent update changes the | |
1101 | * parent before we get the lock. We return -EAGAIN in that case. | |
5a9a87dd MD |
1102 | * Return 0 on success, negative error value on failure. |
1103 | */ | |
1104 | static | |
1105 | int ja_chain_node(struct cds_ja *ja, | |
af3cbd45 | 1106 | struct cds_ja_inode_flag *parent_node_flag, |
fa112799 | 1107 | struct cds_ja_inode_flag **node_flag_ptr, |
5a9a87dd MD |
1108 | struct cds_hlist_head *head, |
1109 | struct cds_ja_node *node) | |
1110 | { | |
1111 | struct cds_ja_shadow_node *shadow_node; | |
fa112799 | 1112 | int ret = 0; |
5a9a87dd | 1113 | |
3d8fe307 | 1114 | shadow_node = rcuja_shadow_lookup_lock(ja->ht, parent_node_flag); |
b306a0fe | 1115 | if (!shadow_node) { |
2e313670 | 1116 | return -EAGAIN; |
b306a0fe | 1117 | } |
fa112799 MD |
1118 | if (!ja_node_ptr(*node_flag_ptr)) { |
1119 | ret = -EAGAIN; | |
1120 | goto end; | |
1121 | } | |
5a9a87dd | 1122 | cds_hlist_add_head_rcu(&node->list, head); |
fa112799 | 1123 | end: |
5a9a87dd | 1124 | rcuja_shadow_unlock(shadow_node); |
fa112799 | 1125 | return ret; |
5a9a87dd MD |
1126 | } |
1127 | ||
1128 | int cds_ja_add(struct cds_ja *ja, uint64_t key, | |
1129 | struct cds_ja_node *new_node) | |
1130 | { | |
1131 | unsigned int tree_depth, i; | |
b0ca2d21 MD |
1132 | struct cds_ja_inode_flag **attach_node_flag_ptr, |
1133 | **node_flag_ptr; | |
5a9a87dd MD |
1134 | struct cds_ja_inode_flag *node_flag, |
1135 | *parent_node_flag, | |
1136 | *parent2_node_flag; | |
1137 | int ret; | |
1138 | ||
b306a0fe | 1139 | if (caa_unlikely(key > ja->key_max)) { |
5a9a87dd | 1140 | return -EINVAL; |
b306a0fe | 1141 | } |
5a9a87dd MD |
1142 | tree_depth = ja->tree_depth; |
1143 | ||
1144 | retry: | |
a2a7ff59 MD |
1145 | dbg_printf("cds_ja_add attempt: key %" PRIu64 ", node %p\n", |
1146 | key, new_node); | |
5a9a87dd | 1147 | parent2_node_flag = NULL; |
b0f74e47 MD |
1148 | parent_node_flag = |
1149 | (struct cds_ja_inode_flag *) &ja->root; /* Use root ptr address as key for mutex */ | |
b0ca2d21 | 1150 | attach_node_flag_ptr = &ja->root; |
5a9a87dd | 1151 | node_flag_ptr = &ja->root; |
35170a44 | 1152 | node_flag = rcu_dereference(ja->root); |
5a9a87dd MD |
1153 | |
1154 | /* Iterate on all internal levels */ | |
a2a7ff59 | 1155 | for (i = 1; i < tree_depth; i++) { |
79b41067 MD |
1156 | uint8_t iter_key; |
1157 | ||
b0ca2d21 MD |
1158 | dbg_printf("cds_ja_add iter attach_node_flag_ptr %p node_flag_ptr %p node_flag %p\n", |
1159 | *attach_node_flag_ptr, *node_flag_ptr, node_flag); | |
5a9a87dd | 1160 | if (!ja_node_ptr(node_flag)) { |
b0ca2d21 MD |
1161 | ret = ja_attach_node(ja, attach_node_flag_ptr, |
1162 | node_flag_ptr, | |
5a9a87dd MD |
1163 | parent_node_flag, parent2_node_flag, |
1164 | key, i, new_node); | |
2e313670 | 1165 | if (ret == -EAGAIN || ret == -EEXIST) |
5a9a87dd MD |
1166 | goto retry; |
1167 | else | |
1168 | goto end; | |
1169 | } | |
79b41067 | 1170 | iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (tree_depth - i - 1))); |
5a9a87dd MD |
1171 | parent2_node_flag = parent_node_flag; |
1172 | parent_node_flag = node_flag; | |
1173 | node_flag = ja_node_get_nth(node_flag, | |
b0ca2d21 | 1174 | &attach_node_flag_ptr, |
5a9a87dd | 1175 | &node_flag_ptr, |
79b41067 | 1176 | iter_key); |
b0ca2d21 MD |
1177 | dbg_printf("cds_ja_add iter key lookup %u finds node_flag %p attach_node_flag_ptr %p node_flag_ptr %p\n", |
1178 | (unsigned int) iter_key, node_flag, | |
1179 | *attach_node_flag_ptr, | |
1180 | *node_flag_ptr); | |
5a9a87dd MD |
1181 | } |
1182 | ||
1183 | /* | |
1184 | * We reached bottom of tree, simply add node to last internal | |
1185 | * level, or chain it if key is already present. | |
1186 | */ | |
1187 | if (!ja_node_ptr(node_flag)) { | |
b0ca2d21 MD |
1188 | dbg_printf("cds_ja_add last attach_node_flag_ptr %p node_flag_ptr %p node_flag %p\n", |
1189 | *attach_node_flag_ptr, *node_flag_ptr, node_flag); | |
1190 | ret = ja_attach_node(ja, attach_node_flag_ptr, | |
1191 | node_flag_ptr, parent_node_flag, | |
5a9a87dd MD |
1192 | parent2_node_flag, key, i, new_node); |
1193 | } else { | |
1194 | ret = ja_chain_node(ja, | |
af3cbd45 | 1195 | parent_node_flag, |
fa112799 | 1196 | node_flag_ptr, |
b0ca2d21 | 1197 | (struct cds_hlist_head *) attach_node_flag_ptr, |
5a9a87dd MD |
1198 | new_node); |
1199 | } | |
b306a0fe | 1200 | if (ret == -EAGAIN || ret == -EEXIST) |
5a9a87dd MD |
1201 | goto retry; |
1202 | end: | |
1203 | return ret; | |
b4540e8a MD |
1204 | } |
1205 | ||
af3cbd45 MD |
1206 | /* |
1207 | * Note: there is no need to lookup the pointer address associated with | |
1208 | * each node's nth item after taking the lock: it's already been done by | |
1209 | * cds_ja_del while holding the rcu read-side lock, and our node rules | |
1210 | * ensure that when a match value -> pointer is found in a node, it is | |
1211 | * _NEVER_ changed for that node without recompaction, and recompaction | |
1212 | * reallocates the node. | |
b306a0fe MD |
1213 | * However, when a child is removed from "linear" nodes, its pointer |
1214 | * is set to NULL. We therefore check, while holding the locks, if this | |
1215 | * pointer is NULL, and return -ENOENT to the caller if it is the case. | |
af3cbd45 | 1216 | */ |
35170a44 MD |
1217 | static |
1218 | int ja_detach_node(struct cds_ja *ja, | |
1219 | struct cds_ja_inode_flag **snapshot, | |
af3cbd45 MD |
1220 | struct cds_ja_inode_flag ***snapshot_ptr, |
1221 | uint8_t *snapshot_n, | |
35170a44 MD |
1222 | int nr_snapshot, |
1223 | uint64_t key, | |
1224 | struct cds_ja_node *node) | |
1225 | { | |
af3cbd45 MD |
1226 | struct cds_ja_shadow_node *shadow_nodes[JA_MAX_DEPTH]; |
1227 | struct cds_ja_inode_flag **node_flag_ptr = NULL, | |
1228 | *parent_node_flag = NULL, | |
1229 | **parent_node_flag_ptr = NULL; | |
1230 | struct cds_ja_inode_flag *iter_node_flag; | |
4d6ef45e MD |
1231 | int ret, i, nr_shadow = 0, nr_clear = 0, nr_branch = 0; |
1232 | uint8_t n = 0; | |
35170a44 | 1233 | |
4d6ef45e | 1234 | assert(nr_snapshot == ja->tree_depth + 1); |
35170a44 | 1235 | |
af3cbd45 MD |
1236 | /* |
1237 | * From the last internal level node going up, get the node | |
1238 | * lock, check if the node has only one child left. If it is the | |
1239 | * case, we continue iterating upward. When we reach a node | |
1240 | * which has more that one child left, we lock the parent, and | |
1241 | * proceed to the node deletion (removing its children too). | |
1242 | */ | |
4d6ef45e | 1243 | for (i = nr_snapshot - 2; i >= 1; i--) { |
af3cbd45 MD |
1244 | struct cds_ja_shadow_node *shadow_node; |
1245 | ||
1246 | shadow_node = rcuja_shadow_lookup_lock(ja->ht, | |
3d8fe307 | 1247 | snapshot[i]); |
af3cbd45 MD |
1248 | if (!shadow_node) { |
1249 | ret = -EAGAIN; | |
1250 | goto end; | |
1251 | } | |
1252 | assert(shadow_node->nr_child > 0); | |
1253 | shadow_nodes[nr_shadow++] = shadow_node; | |
d810c97f | 1254 | if (shadow_node->nr_child == 1 && i > 1) |
4d6ef45e MD |
1255 | nr_clear++; |
1256 | nr_branch++; | |
af3cbd45 MD |
1257 | if (shadow_node->nr_child > 1 || i == 1) { |
1258 | /* Lock parent and break */ | |
1259 | shadow_node = rcuja_shadow_lookup_lock(ja->ht, | |
3d8fe307 | 1260 | snapshot[i - 1]); |
af3cbd45 MD |
1261 | if (!shadow_node) { |
1262 | ret = -EAGAIN; | |
1263 | goto end; | |
1264 | } | |
1265 | shadow_nodes[nr_shadow++] = shadow_node; | |
4d6ef45e MD |
1266 | node_flag_ptr = snapshot_ptr[i + 1]; |
1267 | n = snapshot_n[i + 1]; | |
1268 | parent_node_flag_ptr = snapshot_ptr[i]; | |
1269 | parent_node_flag = snapshot[i]; | |
af3cbd45 MD |
1270 | if (i > 1) { |
1271 | /* | |
1272 | * Lock parent's parent, in case we need | |
1273 | * to recompact parent. | |
1274 | */ | |
1275 | shadow_node = rcuja_shadow_lookup_lock(ja->ht, | |
3d8fe307 | 1276 | snapshot[i - 2]); |
af3cbd45 MD |
1277 | if (!shadow_node) { |
1278 | ret = -EAGAIN; | |
1279 | goto end; | |
1280 | } | |
1281 | shadow_nodes[nr_shadow++] = shadow_node; | |
1282 | } | |
1283 | break; | |
1284 | } | |
1285 | } | |
1286 | ||
b306a0fe MD |
1287 | /* |
1288 | * Check if node has been removed between RCU lookup and lock | |
1289 | * acquisition. | |
1290 | */ | |
b0ca2d21 | 1291 | assert(node_flag_ptr); |
fa112799 | 1292 | if (!ja_node_ptr(*node_flag_ptr)) { |
b306a0fe MD |
1293 | ret = -ENOENT; |
1294 | goto end; | |
1295 | } | |
1296 | ||
af3cbd45 | 1297 | /* |
4d6ef45e MD |
1298 | * At this point, we want to delete all nodes that are about to |
1299 | * be removed from shadow_nodes (except the last one, which is | |
1300 | * either the root or the parent of the upmost node with 1 | |
1301 | * child). OK to as to free lock here, because RCU read lock is | |
1302 | * held, and free only performed in call_rcu. | |
af3cbd45 MD |
1303 | */ |
1304 | ||
1305 | for (i = 0; i < nr_clear; i++) { | |
1306 | ret = rcuja_shadow_clear(ja->ht, | |
3d8fe307 | 1307 | shadow_nodes[i]->node_flag, |
af3cbd45 MD |
1308 | shadow_nodes[i], |
1309 | RCUJA_SHADOW_CLEAR_FREE_NODE | |
1310 | | RCUJA_SHADOW_CLEAR_FREE_LOCK); | |
1311 | assert(!ret); | |
1312 | } | |
1313 | ||
1314 | iter_node_flag = parent_node_flag; | |
1315 | /* Remove from parent */ | |
1316 | ret = ja_node_clear_ptr(ja, | |
1317 | node_flag_ptr, /* Pointer to location to nullify */ | |
1318 | &iter_node_flag, /* Old new parent ptr in its parent */ | |
4d6ef45e | 1319 | shadow_nodes[nr_branch - 1], /* of parent */ |
af3cbd45 | 1320 | n); |
b306a0fe MD |
1321 | if (ret) |
1322 | goto end; | |
af3cbd45 | 1323 | |
4d6ef45e MD |
1324 | dbg_printf("ja_detach_node: publish %p instead of %p\n", |
1325 | iter_node_flag, *parent_node_flag_ptr); | |
af3cbd45 MD |
1326 | /* Update address of parent ptr in its parent */ |
1327 | rcu_assign_pointer(*parent_node_flag_ptr, iter_node_flag); | |
1328 | ||
1329 | end: | |
1330 | for (i = 0; i < nr_shadow; i++) | |
1331 | rcuja_shadow_unlock(shadow_nodes[i]); | |
35170a44 MD |
1332 | return ret; |
1333 | } | |
1334 | ||
af3cbd45 MD |
1335 | static |
1336 | int ja_unchain_node(struct cds_ja *ja, | |
1337 | struct cds_ja_inode_flag *parent_node_flag, | |
fa112799 | 1338 | struct cds_ja_inode_flag **node_flag_ptr, |
f2758d14 | 1339 | struct cds_hlist_head *head, |
af3cbd45 MD |
1340 | struct cds_ja_node *node) |
1341 | { | |
1342 | struct cds_ja_shadow_node *shadow_node; | |
f2758d14 MD |
1343 | struct cds_hlist_node *hlist_node; |
1344 | int ret = 0, count = 0; | |
af3cbd45 | 1345 | |
3d8fe307 | 1346 | shadow_node = rcuja_shadow_lookup_lock(ja->ht, parent_node_flag); |
af3cbd45 MD |
1347 | if (!shadow_node) |
1348 | return -EAGAIN; | |
fa112799 MD |
1349 | if (!ja_node_ptr(*node_flag_ptr)) { |
1350 | ret = -EAGAIN; | |
1351 | goto end; | |
1352 | } | |
af3cbd45 MD |
1353 | /* |
1354 | * Retry if another thread removed all but one of duplicates | |
fa112799 | 1355 | * since check (this check was performed without lock). |
af3cbd45 | 1356 | */ |
f2758d14 MD |
1357 | cds_hlist_for_each_rcu(hlist_node, head, list) { |
1358 | count++; | |
1359 | } | |
1360 | ||
1361 | if (count == 1) { | |
af3cbd45 MD |
1362 | ret = -EAGAIN; |
1363 | goto end; | |
1364 | } | |
1365 | cds_hlist_del_rcu(&node->list); | |
1366 | end: | |
1367 | rcuja_shadow_unlock(shadow_node); | |
1368 | return ret; | |
1369 | } | |
1370 | ||
1371 | /* | |
1372 | * Called with RCU read lock held. | |
1373 | */ | |
35170a44 MD |
1374 | int cds_ja_del(struct cds_ja *ja, uint64_t key, |
1375 | struct cds_ja_node *node) | |
1376 | { | |
1377 | unsigned int tree_depth, i; | |
1378 | struct cds_ja_inode_flag *snapshot[JA_MAX_DEPTH]; | |
af3cbd45 MD |
1379 | struct cds_ja_inode_flag **snapshot_ptr[JA_MAX_DEPTH]; |
1380 | uint8_t snapshot_n[JA_MAX_DEPTH]; | |
35170a44 | 1381 | struct cds_ja_inode_flag *node_flag; |
fa112799 MD |
1382 | struct cds_ja_inode_flag **prev_node_flag_ptr, |
1383 | **node_flag_ptr; | |
4d6ef45e | 1384 | int nr_snapshot; |
35170a44 MD |
1385 | int ret; |
1386 | ||
1387 | if (caa_unlikely(key > ja->key_max)) | |
1388 | return -EINVAL; | |
1389 | tree_depth = ja->tree_depth; | |
1390 | ||
1391 | retry: | |
4d6ef45e | 1392 | nr_snapshot = 0; |
35170a44 MD |
1393 | dbg_printf("cds_ja_del attempt: key %" PRIu64 ", node %p\n", |
1394 | key, node); | |
1395 | ||
1396 | /* snapshot for level 0 is only for shadow node lookup */ | |
4d6ef45e MD |
1397 | snapshot_n[0] = 0; |
1398 | snapshot_n[1] = 0; | |
af3cbd45 | 1399 | snapshot_ptr[nr_snapshot] = NULL; |
35170a44 MD |
1400 | snapshot[nr_snapshot++] = (struct cds_ja_inode_flag *) &ja->root; |
1401 | node_flag = rcu_dereference(ja->root); | |
af3cbd45 | 1402 | prev_node_flag_ptr = &ja->root; |
fa112799 | 1403 | node_flag_ptr = &ja->root; |
35170a44 MD |
1404 | |
1405 | /* Iterate on all internal levels */ | |
1406 | for (i = 1; i < tree_depth; i++) { | |
1407 | uint8_t iter_key; | |
1408 | ||
1409 | dbg_printf("cds_ja_del iter node_flag %p\n", | |
1410 | node_flag); | |
1411 | if (!ja_node_ptr(node_flag)) { | |
1412 | return -ENOENT; | |
1413 | } | |
35170a44 | 1414 | iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (tree_depth - i - 1))); |
4d6ef45e | 1415 | snapshot_n[nr_snapshot + 1] = iter_key; |
af3cbd45 MD |
1416 | snapshot_ptr[nr_snapshot] = prev_node_flag_ptr; |
1417 | snapshot[nr_snapshot++] = node_flag; | |
35170a44 | 1418 | node_flag = ja_node_get_nth(node_flag, |
af3cbd45 | 1419 | &prev_node_flag_ptr, |
fa112799 | 1420 | &node_flag_ptr, |
35170a44 | 1421 | iter_key); |
af3cbd45 MD |
1422 | dbg_printf("cds_ja_del iter key lookup %u finds node_flag %p, prev_node_flag_ptr %p\n", |
1423 | (unsigned int) iter_key, node_flag, | |
1424 | prev_node_flag_ptr); | |
35170a44 MD |
1425 | } |
1426 | ||
1427 | /* | |
1428 | * We reached bottom of tree, try to find the node we are trying | |
1429 | * to remove. Fail if we cannot find it. | |
1430 | */ | |
1431 | if (!ja_node_ptr(node_flag)) { | |
4d6ef45e MD |
1432 | dbg_printf("cds_ja_del: no node found for key %" PRIu64 "\n", |
1433 | key); | |
35170a44 MD |
1434 | return -ENOENT; |
1435 | } else { | |
4d6ef45e | 1436 | struct cds_hlist_head hlist_head; |
35170a44 | 1437 | struct cds_hlist_node *hlist_node; |
af3cbd45 MD |
1438 | struct cds_ja_node *entry, *match = NULL; |
1439 | int count = 0; | |
35170a44 | 1440 | |
4d6ef45e MD |
1441 | hlist_head.next = |
1442 | (struct cds_hlist_node *) ja_node_ptr(node_flag); | |
af3cbd45 | 1443 | cds_hlist_for_each_entry_rcu(entry, |
35170a44 | 1444 | hlist_node, |
4d6ef45e | 1445 | &hlist_head, |
35170a44 | 1446 | list) { |
4d6ef45e | 1447 | dbg_printf("cds_ja_del: compare %p with entry %p\n", node, entry); |
af3cbd45 MD |
1448 | if (entry == node) |
1449 | match = entry; | |
1450 | count++; | |
35170a44 | 1451 | } |
4d6ef45e MD |
1452 | if (!match) { |
1453 | dbg_printf("cds_ja_del: no node match for node %p key %" PRIu64 "\n", node, key); | |
35170a44 | 1454 | return -ENOENT; |
4d6ef45e | 1455 | } |
af3cbd45 MD |
1456 | assert(count > 0); |
1457 | if (count == 1) { | |
1458 | /* | |
4d6ef45e MD |
1459 | * Removing last of duplicates. Last snapshot |
1460 | * does not have a shadow node (external leafs). | |
af3cbd45 MD |
1461 | */ |
1462 | snapshot_ptr[nr_snapshot] = prev_node_flag_ptr; | |
1463 | snapshot[nr_snapshot++] = node_flag; | |
1464 | ret = ja_detach_node(ja, snapshot, snapshot_ptr, | |
1465 | snapshot_n, nr_snapshot, key, node); | |
1466 | } else { | |
f2758d14 | 1467 | ret = ja_unchain_node(ja, snapshot[nr_snapshot - 1], |
fa112799 | 1468 | node_flag_ptr, &hlist_head, match); |
af3cbd45 | 1469 | } |
35170a44 | 1470 | } |
b306a0fe MD |
1471 | /* |
1472 | * Explanation of -ENOENT handling: caused by concurrent delete | |
1473 | * between RCU lookup and actual removal. Need to re-do the | |
1474 | * lookup and removal attempt. | |
1475 | */ | |
1476 | if (ret == -EAGAIN || ret == -ENOENT) | |
35170a44 MD |
1477 | goto retry; |
1478 | return ret; | |
1479 | } | |
1480 | ||
b4540e8a MD |
1481 | struct cds_ja *_cds_ja_new(unsigned int key_bits, |
1482 | const struct rcu_flavor_struct *flavor) | |
be9a7474 MD |
1483 | { |
1484 | struct cds_ja *ja; | |
b0f74e47 | 1485 | int ret; |
f07b240f | 1486 | struct cds_ja_shadow_node *root_shadow_node; |
be9a7474 MD |
1487 | |
1488 | ja = calloc(sizeof(*ja), 1); | |
1489 | if (!ja) | |
1490 | goto ja_error; | |
b4540e8a MD |
1491 | |
1492 | switch (key_bits) { | |
1493 | case 8: | |
b4540e8a | 1494 | case 16: |
1216b3d2 | 1495 | case 24: |
b4540e8a | 1496 | case 32: |
1216b3d2 MD |
1497 | case 40: |
1498 | case 48: | |
1499 | case 56: | |
1500 | ja->key_max = (1ULL << key_bits) - 1; | |
b4540e8a MD |
1501 | break; |
1502 | case 64: | |
1503 | ja->key_max = UINT64_MAX; | |
1504 | break; | |
1505 | default: | |
1506 | goto check_error; | |
1507 | } | |
1508 | ||
be9a7474 | 1509 | /* ja->root is NULL */ |
5a9a87dd | 1510 | /* tree_depth 0 is for pointer to root node */ |
582a6ade | 1511 | ja->tree_depth = (key_bits >> JA_LOG2_BITS_PER_BYTE) + 1; |
a2a7ff59 | 1512 | assert(ja->tree_depth <= JA_MAX_DEPTH); |
be9a7474 MD |
1513 | ja->ht = rcuja_create_ht(flavor); |
1514 | if (!ja->ht) | |
1515 | goto ht_error; | |
b0f74e47 MD |
1516 | |
1517 | /* | |
1518 | * Note: we should not free this node until judy array destroy. | |
1519 | */ | |
f07b240f | 1520 | root_shadow_node = rcuja_shadow_set(ja->ht, |
3d8fe307 MD |
1521 | (struct cds_ja_inode_flag *) &ja->root, |
1522 | NULL, ja); | |
f07b240f MD |
1523 | if (!root_shadow_node) { |
1524 | ret = -ENOMEM; | |
b0f74e47 | 1525 | goto ht_node_error; |
f07b240f | 1526 | } |
3d8fe307 | 1527 | root_shadow_node->level = 0; |
b0f74e47 | 1528 | |
be9a7474 MD |
1529 | return ja; |
1530 | ||
b0f74e47 MD |
1531 | ht_node_error: |
1532 | ret = rcuja_delete_ht(ja->ht); | |
1533 | assert(!ret); | |
be9a7474 | 1534 | ht_error: |
b4540e8a | 1535 | check_error: |
be9a7474 MD |
1536 | free(ja); |
1537 | ja_error: | |
1538 | return NULL; | |
1539 | } | |
1540 | ||
3d8fe307 MD |
1541 | /* |
1542 | * Called from RCU read-side CS. | |
1543 | */ | |
1544 | __attribute__((visibility("protected"))) | |
1545 | void rcuja_free_all_children(struct cds_ja_shadow_node *shadow_node, | |
1546 | struct cds_ja_inode_flag *node_flag, | |
1547 | void (*free_node_cb)(struct rcu_head *head)) | |
1548 | { | |
1549 | const struct rcu_flavor_struct *flavor; | |
1550 | unsigned int type_index; | |
1551 | struct cds_ja_inode *node; | |
1552 | const struct cds_ja_type *type; | |
1553 | ||
1554 | flavor = cds_lfht_rcu_flavor(shadow_node->ja->ht); | |
1555 | node = ja_node_ptr(node_flag); | |
1556 | assert(node != NULL); | |
1557 | type_index = ja_node_type(node_flag); | |
1558 | type = &ja_types[type_index]; | |
1559 | ||
1560 | switch (type->type_class) { | |
1561 | case RCU_JA_LINEAR: | |
1562 | { | |
1563 | uint8_t nr_child = | |
1564 | ja_linear_node_get_nr_child(type, node); | |
1565 | unsigned int i; | |
1566 | ||
1567 | for (i = 0; i < nr_child; i++) { | |
1568 | struct cds_ja_inode_flag *iter; | |
1569 | struct cds_hlist_head head; | |
1570 | struct cds_ja_node *entry; | |
1571 | struct cds_hlist_node *pos; | |
1572 | uint8_t v; | |
1573 | ||
1574 | ja_linear_node_get_ith_pos(type, node, i, &v, &iter); | |
1575 | if (!iter) | |
1576 | continue; | |
1577 | head.next = (struct cds_hlist_node *) iter; | |
1578 | cds_hlist_for_each_entry_rcu(entry, pos, &head, list) { | |
1579 | flavor->update_call_rcu(&entry->head, free_node_cb); | |
1580 | } | |
1581 | } | |
1582 | break; | |
1583 | } | |
1584 | case RCU_JA_POOL: | |
1585 | { | |
1586 | unsigned int pool_nr; | |
1587 | ||
1588 | for (pool_nr = 0; pool_nr < (1U << type->nr_pool_order); pool_nr++) { | |
1589 | struct cds_ja_inode *pool = | |
1590 | ja_pool_node_get_ith_pool(type, node, pool_nr); | |
1591 | uint8_t nr_child = | |
1592 | ja_linear_node_get_nr_child(type, pool); | |
1593 | unsigned int j; | |
1594 | ||
1595 | for (j = 0; j < nr_child; j++) { | |
1596 | struct cds_ja_inode_flag *iter; | |
1597 | struct cds_hlist_head head; | |
1598 | struct cds_ja_node *entry; | |
1599 | struct cds_hlist_node *pos; | |
1600 | uint8_t v; | |
1601 | ||
1602 | ja_linear_node_get_ith_pos(type, node, j, &v, &iter); | |
1603 | if (!iter) | |
1604 | continue; | |
1605 | head.next = (struct cds_hlist_node *) iter; | |
1606 | cds_hlist_for_each_entry_rcu(entry, pos, &head, list) { | |
1607 | flavor->update_call_rcu(&entry->head, free_node_cb); | |
1608 | } | |
1609 | } | |
1610 | } | |
1611 | break; | |
1612 | } | |
1613 | case RCU_JA_NULL: | |
1614 | break; | |
1615 | case RCU_JA_PIGEON: | |
1616 | { | |
1617 | uint8_t nr_child; | |
1618 | unsigned int i; | |
1619 | ||
1620 | nr_child = shadow_node->nr_child; | |
1621 | for (i = 0; i < nr_child; i++) { | |
1622 | struct cds_ja_inode_flag *iter; | |
1623 | struct cds_hlist_head head; | |
1624 | struct cds_ja_node *entry; | |
1625 | struct cds_hlist_node *pos; | |
1626 | ||
1627 | iter = ja_pigeon_node_get_ith_pos(type, node, i); | |
1628 | if (!iter) | |
1629 | continue; | |
1630 | head.next = (struct cds_hlist_node *) iter; | |
1631 | cds_hlist_for_each_entry_rcu(entry, pos, &head, list) { | |
1632 | flavor->update_call_rcu(&entry->head, free_node_cb); | |
1633 | } | |
1634 | } | |
1635 | break; | |
1636 | } | |
1637 | default: | |
1638 | assert(0); | |
1639 | } | |
1640 | } | |
1641 | ||
be9a7474 MD |
1642 | /* |
1643 | * There should be no more concurrent add to the judy array while it is | |
1644 | * being destroyed (ensured by the caller). | |
1645 | */ | |
3d8fe307 MD |
1646 | int cds_ja_destroy(struct cds_ja *ja, |
1647 | void (*free_node_cb)(struct rcu_head *head)) | |
be9a7474 | 1648 | { |
b4540e8a MD |
1649 | int ret; |
1650 | ||
be9a7474 | 1651 | rcuja_shadow_prune(ja->ht, |
3d8fe307 MD |
1652 | RCUJA_SHADOW_CLEAR_FREE_NODE | RCUJA_SHADOW_CLEAR_FREE_LOCK, |
1653 | free_node_cb); | |
b4540e8a MD |
1654 | ret = rcuja_delete_ht(ja->ht); |
1655 | if (ret) | |
1656 | return ret; | |
f07b240f MD |
1657 | if (uatomic_read(&ja->nr_fallback)) |
1658 | fprintf(stderr, | |
1659 | "[warning] RCU Judy Array used %lu fallback node(s)\n", | |
1660 | uatomic_read(&ja->nr_fallback)); | |
b4540e8a | 1661 | free(ja); |
41975c12 | 1662 | return 0; |
be9a7474 | 1663 | } |