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, |
b62a8d0c | 280 | struct cds_ja_inode_flag **child_node_flag_v, |
b0ca2d21 | 281 | struct cds_ja_inode_flag ***node_flag_ptr, |
8e519e3c | 282 | uint8_t n) |
d68c6810 MD |
283 | { |
284 | uint8_t nr_child; | |
285 | uint8_t *values; | |
b4540e8a MD |
286 | struct cds_ja_inode_flag **pointers; |
287 | struct cds_ja_inode_flag *ptr; | |
d68c6810 MD |
288 | unsigned int i; |
289 | ||
8e519e3c | 290 | assert(type->type_class == RCU_JA_LINEAR || type->type_class == RCU_JA_POOL); |
d68c6810 | 291 | |
11c5e016 | 292 | nr_child = ja_linear_node_get_nr_child(type, node); |
13a7f5a6 | 293 | cmm_smp_rmb(); /* read nr_child before values and pointers */ |
8e519e3c MD |
294 | assert(nr_child <= type->max_linear_child); |
295 | assert(type->type_class != RCU_JA_LINEAR || nr_child >= type->min_child); | |
d68c6810 | 296 | |
1db4943c | 297 | values = &node->u.data[1]; |
d68c6810 | 298 | for (i = 0; i < nr_child; i++) { |
13a7f5a6 | 299 | if (CMM_LOAD_SHARED(values[i]) == n) |
d68c6810 MD |
300 | break; |
301 | } | |
b0ca2d21 MD |
302 | if (i >= nr_child) { |
303 | if (caa_unlikely(node_flag_ptr)) | |
304 | *node_flag_ptr = NULL; | |
d68c6810 | 305 | return NULL; |
b0ca2d21 | 306 | } |
b4540e8a | 307 | pointers = (struct cds_ja_inode_flag **) align_ptr_size(&values[type->max_linear_child]); |
13a7f5a6 | 308 | ptr = rcu_dereference(pointers[i]); |
2e313670 MD |
309 | if (caa_unlikely(child_node_flag_ptr) && ptr) |
310 | *child_node_flag_ptr = &pointers[i]; | |
b62a8d0c MD |
311 | if (caa_unlikely(child_node_flag_v) && ptr) |
312 | *child_node_flag_v = ptr; | |
b0ca2d21 MD |
313 | if (caa_unlikely(node_flag_ptr)) |
314 | *node_flag_ptr = &pointers[i]; | |
d68c6810 MD |
315 | return ptr; |
316 | } | |
317 | ||
11c5e016 | 318 | static |
5a9a87dd | 319 | void ja_linear_node_get_ith_pos(const struct cds_ja_type *type, |
b4540e8a | 320 | struct cds_ja_inode *node, |
11c5e016 MD |
321 | uint8_t i, |
322 | uint8_t *v, | |
b4540e8a | 323 | struct cds_ja_inode_flag **iter) |
11c5e016 MD |
324 | { |
325 | uint8_t *values; | |
b4540e8a | 326 | struct cds_ja_inode_flag **pointers; |
11c5e016 MD |
327 | |
328 | assert(type->type_class == RCU_JA_LINEAR || type->type_class == RCU_JA_POOL); | |
329 | assert(i < ja_linear_node_get_nr_child(type, node)); | |
330 | ||
331 | values = &node->u.data[1]; | |
332 | *v = values[i]; | |
b4540e8a | 333 | pointers = (struct cds_ja_inode_flag **) align_ptr_size(&values[type->max_linear_child]); |
11c5e016 MD |
334 | *iter = pointers[i]; |
335 | } | |
336 | ||
d68c6810 | 337 | static |
b4540e8a MD |
338 | struct cds_ja_inode_flag *ja_pool_node_get_nth(const struct cds_ja_type *type, |
339 | struct cds_ja_inode *node, | |
5a9a87dd | 340 | struct cds_ja_inode_flag ***child_node_flag_ptr, |
b62a8d0c | 341 | struct cds_ja_inode_flag **child_node_flag_v, |
b0ca2d21 | 342 | struct cds_ja_inode_flag ***node_flag_ptr, |
8e519e3c | 343 | uint8_t n) |
d68c6810 | 344 | { |
b4540e8a | 345 | struct cds_ja_inode *linear; |
d68c6810 | 346 | |
fd800776 | 347 | assert(type->type_class == RCU_JA_POOL); |
e1db2db5 MD |
348 | /* |
349 | * TODO: currently, we select the pool by highest bits. We | |
350 | * should support various encodings. | |
351 | */ | |
b4540e8a | 352 | linear = (struct cds_ja_inode *) |
1db4943c | 353 | &node->u.data[((unsigned long) n >> (CHAR_BIT - type->nr_pool_order)) << type->pool_size_order]; |
b0ca2d21 | 354 | return ja_linear_node_get_nth(type, linear, child_node_flag_ptr, |
b62a8d0c | 355 | child_node_flag_v, node_flag_ptr, n); |
d68c6810 MD |
356 | } |
357 | ||
11c5e016 | 358 | static |
b4540e8a MD |
359 | struct cds_ja_inode *ja_pool_node_get_ith_pool(const struct cds_ja_type *type, |
360 | struct cds_ja_inode *node, | |
11c5e016 MD |
361 | uint8_t i) |
362 | { | |
363 | assert(type->type_class == RCU_JA_POOL); | |
b4540e8a | 364 | return (struct cds_ja_inode *) |
11c5e016 MD |
365 | &node->u.data[(unsigned int) i << type->pool_size_order]; |
366 | } | |
367 | ||
d68c6810 | 368 | static |
b4540e8a MD |
369 | struct cds_ja_inode_flag *ja_pigeon_node_get_nth(const struct cds_ja_type *type, |
370 | struct cds_ja_inode *node, | |
5a9a87dd | 371 | struct cds_ja_inode_flag ***child_node_flag_ptr, |
b62a8d0c | 372 | struct cds_ja_inode_flag **child_node_flag_v, |
b0ca2d21 | 373 | struct cds_ja_inode_flag ***node_flag_ptr, |
8e519e3c | 374 | uint8_t n) |
d68c6810 | 375 | { |
5a9a87dd | 376 | struct cds_ja_inode_flag **child_node_flag; |
b62a8d0c | 377 | struct cds_ja_inode_flag *child_node_flag_read; |
5a9a87dd | 378 | |
d68c6810 | 379 | assert(type->type_class == RCU_JA_PIGEON); |
5a9a87dd | 380 | child_node_flag = &((struct cds_ja_inode_flag **) node->u.data)[n]; |
b62a8d0c | 381 | child_node_flag_read = rcu_dereference(*child_node_flag); |
582a6ade MD |
382 | dbg_printf("ja_pigeon_node_get_nth child_node_flag_ptr %p\n", |
383 | child_node_flag); | |
b62a8d0c | 384 | if (caa_unlikely(child_node_flag_ptr) && child_node_flag_read) |
5a9a87dd | 385 | *child_node_flag_ptr = child_node_flag; |
b62a8d0c MD |
386 | if (caa_unlikely(child_node_flag_v) && child_node_flag_read) |
387 | *child_node_flag_v = child_node_flag_read; | |
b0ca2d21 MD |
388 | if (caa_unlikely(node_flag_ptr)) |
389 | *node_flag_ptr = child_node_flag; | |
b62a8d0c | 390 | return child_node_flag_read; |
d68c6810 MD |
391 | } |
392 | ||
2e313670 MD |
393 | static |
394 | struct cds_ja_inode_flag *ja_pigeon_node_get_ith_pos(const struct cds_ja_type *type, | |
395 | struct cds_ja_inode *node, | |
396 | uint8_t i) | |
397 | { | |
b62a8d0c | 398 | return ja_pigeon_node_get_nth(type, node, NULL, NULL, NULL, i); |
2e313670 MD |
399 | } |
400 | ||
13a7f5a6 MD |
401 | /* |
402 | * ja_node_get_nth: get nth item from a node. | |
403 | * node_flag is already rcu_dereference'd. | |
404 | */ | |
d68c6810 | 405 | static |
b62a8d0c | 406 | struct cds_ja_inode_flag *ja_node_get_nth(struct cds_ja_inode_flag *node_flag, |
5a9a87dd | 407 | struct cds_ja_inode_flag ***child_node_flag_ptr, |
b62a8d0c | 408 | struct cds_ja_inode_flag **child_node_flag, |
b0ca2d21 | 409 | struct cds_ja_inode_flag ***node_flag_ptr, |
8e519e3c | 410 | uint8_t n) |
d68c6810 MD |
411 | { |
412 | unsigned int type_index; | |
b4540e8a | 413 | struct cds_ja_inode *node; |
d96bfb0d | 414 | const struct cds_ja_type *type; |
d68c6810 | 415 | |
d68c6810 | 416 | node = ja_node_ptr(node_flag); |
5a9a87dd | 417 | assert(node != NULL); |
d68c6810 MD |
418 | type_index = ja_node_type(node_flag); |
419 | type = &ja_types[type_index]; | |
420 | ||
421 | switch (type->type_class) { | |
422 | case RCU_JA_LINEAR: | |
5a9a87dd | 423 | return ja_linear_node_get_nth(type, node, |
b62a8d0c MD |
424 | child_node_flag_ptr, child_node_flag, |
425 | node_flag_ptr, n); | |
fd800776 | 426 | case RCU_JA_POOL: |
5a9a87dd | 427 | return ja_pool_node_get_nth(type, node, |
b62a8d0c MD |
428 | child_node_flag_ptr, child_node_flag, |
429 | node_flag_ptr, n); | |
d68c6810 | 430 | case RCU_JA_PIGEON: |
5a9a87dd | 431 | return ja_pigeon_node_get_nth(type, node, |
b62a8d0c MD |
432 | child_node_flag_ptr, child_node_flag, |
433 | node_flag_ptr, n); | |
d68c6810 MD |
434 | default: |
435 | assert(0); | |
436 | return (void *) -1UL; | |
437 | } | |
438 | } | |
439 | ||
8e519e3c | 440 | static |
d96bfb0d | 441 | int ja_linear_node_set_nth(const struct cds_ja_type *type, |
b4540e8a | 442 | struct cds_ja_inode *node, |
d96bfb0d | 443 | struct cds_ja_shadow_node *shadow_node, |
8e519e3c | 444 | uint8_t n, |
b4540e8a | 445 | struct cds_ja_inode_flag *child_node_flag) |
8e519e3c MD |
446 | { |
447 | uint8_t nr_child; | |
448 | uint8_t *values, *nr_child_ptr; | |
b4540e8a | 449 | struct cds_ja_inode_flag **pointers; |
2e313670 | 450 | unsigned int i, unused = 0; |
8e519e3c MD |
451 | |
452 | assert(type->type_class == RCU_JA_LINEAR || type->type_class == RCU_JA_POOL); | |
453 | ||
454 | nr_child_ptr = &node->u.data[0]; | |
a2a7ff59 | 455 | dbg_printf("linear set nth: nr_child_ptr %p\n", nr_child_ptr); |
8e519e3c MD |
456 | nr_child = *nr_child_ptr; |
457 | assert(nr_child <= type->max_linear_child); | |
8e519e3c MD |
458 | |
459 | values = &node->u.data[1]; | |
2e313670 MD |
460 | pointers = (struct cds_ja_inode_flag **) align_ptr_size(&values[type->max_linear_child]); |
461 | /* Check if node value is already populated */ | |
8e519e3c | 462 | for (i = 0; i < nr_child; i++) { |
2e313670 MD |
463 | if (values[i] == n) { |
464 | if (pointers[i]) | |
465 | return -EEXIST; | |
466 | else | |
467 | break; | |
468 | } else { | |
469 | if (!pointers[i]) | |
470 | unused++; | |
471 | } | |
8e519e3c | 472 | } |
2e313670 MD |
473 | if (i == nr_child && nr_child >= type->max_linear_child) { |
474 | if (unused) | |
475 | return -ERANGE; /* recompact node */ | |
476 | else | |
477 | return -ENOSPC; /* No space left in this node type */ | |
478 | } | |
479 | ||
480 | assert(pointers[i] == NULL); | |
481 | rcu_assign_pointer(pointers[i], child_node_flag); | |
482 | /* If we expanded the nr_child, increment it */ | |
483 | if (i == nr_child) { | |
484 | CMM_STORE_SHARED(values[nr_child], n); | |
485 | /* write pointer and value before nr_child */ | |
486 | cmm_smp_wmb(); | |
487 | CMM_STORE_SHARED(*nr_child_ptr, nr_child + 1); | |
8e519e3c | 488 | } |
e1db2db5 | 489 | shadow_node->nr_child++; |
a2a7ff59 MD |
490 | dbg_printf("linear set nth: %u child, shadow: %u child, for node %p shadow %p\n", |
491 | (unsigned int) CMM_LOAD_SHARED(*nr_child_ptr), | |
492 | (unsigned int) shadow_node->nr_child, | |
493 | node, shadow_node); | |
494 | ||
8e519e3c MD |
495 | return 0; |
496 | } | |
497 | ||
498 | static | |
d96bfb0d | 499 | int ja_pool_node_set_nth(const struct cds_ja_type *type, |
b4540e8a | 500 | struct cds_ja_inode *node, |
d96bfb0d | 501 | struct cds_ja_shadow_node *shadow_node, |
8e519e3c | 502 | uint8_t n, |
b4540e8a | 503 | struct cds_ja_inode_flag *child_node_flag) |
8e519e3c | 504 | { |
b4540e8a | 505 | struct cds_ja_inode *linear; |
8e519e3c MD |
506 | |
507 | assert(type->type_class == RCU_JA_POOL); | |
b4540e8a | 508 | linear = (struct cds_ja_inode *) |
8e519e3c | 509 | &node->u.data[((unsigned long) n >> (CHAR_BIT - type->nr_pool_order)) << type->pool_size_order]; |
e1db2db5 MD |
510 | return ja_linear_node_set_nth(type, linear, shadow_node, |
511 | n, child_node_flag); | |
8e519e3c MD |
512 | } |
513 | ||
514 | static | |
d96bfb0d | 515 | int ja_pigeon_node_set_nth(const struct cds_ja_type *type, |
b4540e8a | 516 | struct cds_ja_inode *node, |
d96bfb0d | 517 | struct cds_ja_shadow_node *shadow_node, |
8e519e3c | 518 | uint8_t n, |
b4540e8a | 519 | struct cds_ja_inode_flag *child_node_flag) |
8e519e3c | 520 | { |
b4540e8a | 521 | struct cds_ja_inode_flag **ptr; |
8e519e3c MD |
522 | |
523 | assert(type->type_class == RCU_JA_PIGEON); | |
b4540e8a | 524 | ptr = &((struct cds_ja_inode_flag **) node->u.data)[n]; |
5a9a87dd | 525 | if (*ptr) |
8e519e3c MD |
526 | return -EEXIST; |
527 | rcu_assign_pointer(*ptr, child_node_flag); | |
e1db2db5 | 528 | shadow_node->nr_child++; |
8e519e3c MD |
529 | return 0; |
530 | } | |
531 | ||
d68c6810 | 532 | /* |
7a0b2331 | 533 | * _ja_node_set_nth: set nth item within a node. Return an error |
8e519e3c | 534 | * (negative error value) if it is already there. |
d68c6810 | 535 | */ |
8e519e3c | 536 | static |
d96bfb0d | 537 | int _ja_node_set_nth(const struct cds_ja_type *type, |
b4540e8a | 538 | struct cds_ja_inode *node, |
d96bfb0d | 539 | struct cds_ja_shadow_node *shadow_node, |
e1db2db5 | 540 | uint8_t n, |
b4540e8a | 541 | struct cds_ja_inode_flag *child_node_flag) |
8e519e3c | 542 | { |
8e519e3c MD |
543 | switch (type->type_class) { |
544 | case RCU_JA_LINEAR: | |
e1db2db5 | 545 | return ja_linear_node_set_nth(type, node, shadow_node, n, |
8e519e3c MD |
546 | child_node_flag); |
547 | case RCU_JA_POOL: | |
e1db2db5 | 548 | return ja_pool_node_set_nth(type, node, shadow_node, n, |
8e519e3c MD |
549 | child_node_flag); |
550 | case RCU_JA_PIGEON: | |
e1db2db5 | 551 | return ja_pigeon_node_set_nth(type, node, shadow_node, n, |
8e519e3c | 552 | child_node_flag); |
e1db2db5 MD |
553 | case RCU_JA_NULL: |
554 | return -ENOSPC; | |
8e519e3c MD |
555 | default: |
556 | assert(0); | |
557 | return -EINVAL; | |
558 | } | |
559 | ||
560 | return 0; | |
561 | } | |
7a0b2331 | 562 | |
2e313670 | 563 | static |
af3cbd45 | 564 | int ja_linear_node_clear_ptr(const struct cds_ja_type *type, |
2e313670 MD |
565 | struct cds_ja_inode *node, |
566 | struct cds_ja_shadow_node *shadow_node, | |
af3cbd45 | 567 | struct cds_ja_inode_flag **node_flag_ptr) |
2e313670 MD |
568 | { |
569 | uint8_t nr_child; | |
af3cbd45 | 570 | uint8_t *nr_child_ptr; |
2e313670 MD |
571 | |
572 | assert(type->type_class == RCU_JA_LINEAR || type->type_class == RCU_JA_POOL); | |
573 | ||
574 | nr_child_ptr = &node->u.data[0]; | |
af3cbd45 | 575 | dbg_printf("linear clear ptr: nr_child_ptr %p\n", nr_child_ptr); |
2e313670 MD |
576 | nr_child = *nr_child_ptr; |
577 | assert(nr_child <= type->max_linear_child); | |
578 | ||
2e313670 MD |
579 | if (shadow_node->fallback_removal_count) { |
580 | shadow_node->fallback_removal_count--; | |
581 | } else { | |
582 | if (shadow_node->nr_child <= type->min_child) { | |
583 | /* We need to try recompacting the node */ | |
584 | return -EFBIG; | |
585 | } | |
586 | } | |
af3cbd45 MD |
587 | assert(*node_flag_ptr != NULL); |
588 | rcu_assign_pointer(*node_flag_ptr, NULL); | |
2e313670 MD |
589 | /* |
590 | * Value and nr_child are never changed (would cause ABA issue). | |
591 | * Instead, we leave the pointer to NULL and recompact the node | |
592 | * once in a while. It is allowed to set a NULL pointer to a new | |
593 | * value without recompaction though. | |
594 | * Only update the shadow node accounting. | |
595 | */ | |
596 | shadow_node->nr_child--; | |
af3cbd45 | 597 | dbg_printf("linear clear ptr: %u child, shadow: %u child, for node %p shadow %p\n", |
2e313670 MD |
598 | (unsigned int) CMM_LOAD_SHARED(*nr_child_ptr), |
599 | (unsigned int) shadow_node->nr_child, | |
600 | node, shadow_node); | |
601 | ||
602 | return 0; | |
603 | } | |
604 | ||
605 | static | |
af3cbd45 | 606 | int ja_pool_node_clear_ptr(const struct cds_ja_type *type, |
2e313670 MD |
607 | struct cds_ja_inode *node, |
608 | struct cds_ja_shadow_node *shadow_node, | |
af3cbd45 | 609 | struct cds_ja_inode_flag **node_flag_ptr, |
2e313670 MD |
610 | uint8_t n) |
611 | { | |
612 | struct cds_ja_inode *linear; | |
613 | ||
614 | assert(type->type_class == RCU_JA_POOL); | |
615 | linear = (struct cds_ja_inode *) | |
616 | &node->u.data[((unsigned long) n >> (CHAR_BIT - type->nr_pool_order)) << type->pool_size_order]; | |
af3cbd45 | 617 | return ja_linear_node_clear_ptr(type, linear, shadow_node, node_flag_ptr); |
2e313670 MD |
618 | } |
619 | ||
620 | static | |
af3cbd45 | 621 | int ja_pigeon_node_clear_ptr(const struct cds_ja_type *type, |
2e313670 MD |
622 | struct cds_ja_inode *node, |
623 | struct cds_ja_shadow_node *shadow_node, | |
af3cbd45 | 624 | struct cds_ja_inode_flag **node_flag_ptr) |
2e313670 | 625 | { |
2e313670 | 626 | assert(type->type_class == RCU_JA_PIGEON); |
4d6ef45e | 627 | dbg_printf("ja_pigeon_node_clear_ptr: clearing ptr: %p\n", *node_flag_ptr); |
af3cbd45 | 628 | rcu_assign_pointer(*node_flag_ptr, NULL); |
2e313670 MD |
629 | shadow_node->nr_child--; |
630 | return 0; | |
631 | } | |
632 | ||
633 | /* | |
af3cbd45 | 634 | * _ja_node_clear_ptr: clear ptr item within a node. Return an error |
2e313670 MD |
635 | * (negative error value) if it is not found (-ENOENT). |
636 | */ | |
637 | static | |
af3cbd45 | 638 | int _ja_node_clear_ptr(const struct cds_ja_type *type, |
2e313670 MD |
639 | struct cds_ja_inode *node, |
640 | struct cds_ja_shadow_node *shadow_node, | |
af3cbd45 | 641 | struct cds_ja_inode_flag **node_flag_ptr, |
2e313670 MD |
642 | uint8_t n) |
643 | { | |
644 | switch (type->type_class) { | |
645 | case RCU_JA_LINEAR: | |
af3cbd45 | 646 | return ja_linear_node_clear_ptr(type, node, shadow_node, node_flag_ptr); |
2e313670 | 647 | case RCU_JA_POOL: |
af3cbd45 | 648 | return ja_pool_node_clear_ptr(type, node, shadow_node, node_flag_ptr, n); |
2e313670 | 649 | case RCU_JA_PIGEON: |
af3cbd45 | 650 | return ja_pigeon_node_clear_ptr(type, node, shadow_node, node_flag_ptr); |
2e313670 MD |
651 | case RCU_JA_NULL: |
652 | return -ENOENT; | |
653 | default: | |
654 | assert(0); | |
655 | return -EINVAL; | |
656 | } | |
657 | ||
658 | return 0; | |
659 | } | |
660 | ||
7a0b2331 MD |
661 | /* |
662 | * ja_node_recompact_add: recompact a node, adding a new child. | |
e1db2db5 | 663 | * TODO: for pool type, take selection bit(s) into account. |
2e313670 | 664 | * Return 0 on success, -EAGAIN if need to retry, or other negative |
5a9a87dd | 665 | * error value otherwise. |
7a0b2331 MD |
666 | */ |
667 | static | |
2e313670 MD |
668 | int ja_node_recompact(enum ja_recompact mode, |
669 | struct cds_ja *ja, | |
e1db2db5 | 670 | unsigned int old_type_index, |
d96bfb0d | 671 | const struct cds_ja_type *old_type, |
b4540e8a | 672 | struct cds_ja_inode *old_node, |
5a9a87dd | 673 | struct cds_ja_shadow_node *shadow_node, |
3d8fe307 | 674 | struct cds_ja_inode_flag **old_node_flag_ptr, uint8_t n, |
af3cbd45 MD |
675 | struct cds_ja_inode_flag *child_node_flag, |
676 | struct cds_ja_inode_flag **nullify_node_flag_ptr) | |
7a0b2331 | 677 | { |
e1db2db5 | 678 | unsigned int new_type_index; |
b4540e8a | 679 | struct cds_ja_inode *new_node; |
af3cbd45 | 680 | struct cds_ja_shadow_node *new_shadow_node = NULL; |
d96bfb0d | 681 | const struct cds_ja_type *new_type; |
3d8fe307 | 682 | struct cds_ja_inode_flag *new_node_flag, *old_node_flag; |
7a0b2331 | 683 | int ret; |
f07b240f | 684 | int fallback = 0; |
7a0b2331 | 685 | |
3d8fe307 MD |
686 | old_node_flag = *old_node_flag_ptr; |
687 | ||
2e313670 MD |
688 | switch (mode) { |
689 | case JA_RECOMPACT: | |
690 | new_type_index = old_type_index; | |
691 | break; | |
692 | case JA_RECOMPACT_ADD: | |
693 | if (!shadow_node || old_type_index == NODE_INDEX_NULL) { | |
694 | new_type_index = 0; | |
695 | } else { | |
696 | new_type_index = old_type_index + 1; | |
697 | } | |
698 | break; | |
699 | case JA_RECOMPACT_DEL: | |
700 | if (old_type_index == 0) { | |
701 | new_type_index = NODE_INDEX_NULL; | |
702 | } else { | |
703 | new_type_index = old_type_index - 1; | |
704 | } | |
705 | break; | |
706 | default: | |
707 | assert(0); | |
7a0b2331 | 708 | } |
a2a7ff59 | 709 | |
f07b240f | 710 | retry: /* for fallback */ |
582a6ade MD |
711 | dbg_printf("Recompact from type %d to type %d\n", |
712 | old_type_index, new_type_index); | |
7a0b2331 | 713 | new_type = &ja_types[new_type_index]; |
2e313670 MD |
714 | if (new_type_index != NODE_INDEX_NULL) { |
715 | new_node = alloc_cds_ja_node(new_type); | |
716 | if (!new_node) | |
717 | return -ENOMEM; | |
718 | new_node_flag = ja_node_flag(new_node, new_type_index); | |
719 | dbg_printf("Recompact inherit lock from %p\n", shadow_node); | |
3d8fe307 | 720 | new_shadow_node = rcuja_shadow_set(ja->ht, new_node_flag, shadow_node, ja); |
2e313670 MD |
721 | if (!new_shadow_node) { |
722 | free(new_node); | |
723 | return -ENOMEM; | |
724 | } | |
725 | if (fallback) | |
726 | new_shadow_node->fallback_removal_count = | |
727 | JA_FALLBACK_REMOVAL_COUNT; | |
728 | } else { | |
729 | new_node = NULL; | |
730 | new_node_flag = NULL; | |
e1db2db5 | 731 | } |
11c5e016 | 732 | |
2e313670 MD |
733 | assert(mode != JA_RECOMPACT_ADD || old_type->type_class != RCU_JA_PIGEON); |
734 | ||
735 | if (new_type_index == NODE_INDEX_NULL) | |
736 | goto skip_copy; | |
737 | ||
11c5e016 MD |
738 | switch (old_type->type_class) { |
739 | case RCU_JA_LINEAR: | |
740 | { | |
741 | uint8_t nr_child = | |
742 | ja_linear_node_get_nr_child(old_type, old_node); | |
743 | unsigned int i; | |
744 | ||
745 | for (i = 0; i < nr_child; i++) { | |
b4540e8a | 746 | struct cds_ja_inode_flag *iter; |
11c5e016 MD |
747 | uint8_t v; |
748 | ||
749 | ja_linear_node_get_ith_pos(old_type, old_node, i, &v, &iter); | |
750 | if (!iter) | |
751 | continue; | |
af3cbd45 | 752 | if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) |
2e313670 | 753 | continue; |
f07b240f MD |
754 | ret = _ja_node_set_nth(new_type, new_node, |
755 | new_shadow_node, | |
11c5e016 | 756 | v, iter); |
f07b240f MD |
757 | if (new_type->type_class == RCU_JA_POOL && ret) { |
758 | goto fallback_toosmall; | |
759 | } | |
11c5e016 MD |
760 | assert(!ret); |
761 | } | |
762 | break; | |
763 | } | |
764 | case RCU_JA_POOL: | |
765 | { | |
766 | unsigned int pool_nr; | |
767 | ||
768 | for (pool_nr = 0; pool_nr < (1U << old_type->nr_pool_order); pool_nr++) { | |
b4540e8a | 769 | struct cds_ja_inode *pool = |
11c5e016 MD |
770 | ja_pool_node_get_ith_pool(old_type, |
771 | old_node, pool_nr); | |
772 | uint8_t nr_child = | |
773 | ja_linear_node_get_nr_child(old_type, pool); | |
774 | unsigned int j; | |
775 | ||
776 | for (j = 0; j < nr_child; j++) { | |
b4540e8a | 777 | struct cds_ja_inode_flag *iter; |
11c5e016 MD |
778 | uint8_t v; |
779 | ||
780 | ja_linear_node_get_ith_pos(old_type, pool, | |
781 | j, &v, &iter); | |
782 | if (!iter) | |
783 | continue; | |
af3cbd45 | 784 | if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) |
2e313670 | 785 | continue; |
f07b240f MD |
786 | ret = _ja_node_set_nth(new_type, new_node, |
787 | new_shadow_node, | |
11c5e016 | 788 | v, iter); |
f07b240f MD |
789 | if (new_type->type_class == RCU_JA_POOL |
790 | && ret) { | |
791 | goto fallback_toosmall; | |
792 | } | |
11c5e016 MD |
793 | assert(!ret); |
794 | } | |
795 | } | |
796 | break; | |
7a0b2331 | 797 | } |
a2a7ff59 | 798 | case RCU_JA_NULL: |
2e313670 | 799 | assert(mode == JA_RECOMPACT_ADD); |
a2a7ff59 | 800 | break; |
11c5e016 | 801 | case RCU_JA_PIGEON: |
2e313670 MD |
802 | { |
803 | uint8_t nr_child; | |
804 | unsigned int i; | |
805 | ||
806 | assert(mode == JA_RECOMPACT_DEL); | |
807 | nr_child = shadow_node->nr_child; | |
808 | for (i = 0; i < nr_child; i++) { | |
809 | struct cds_ja_inode_flag *iter; | |
810 | ||
811 | iter = ja_pigeon_node_get_ith_pos(old_type, old_node, i); | |
812 | if (!iter) | |
813 | continue; | |
af3cbd45 | 814 | if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) |
2e313670 MD |
815 | continue; |
816 | ret = _ja_node_set_nth(new_type, new_node, | |
817 | new_shadow_node, | |
818 | i, iter); | |
819 | if (new_type->type_class == RCU_JA_POOL && ret) { | |
820 | goto fallback_toosmall; | |
821 | } | |
822 | assert(!ret); | |
823 | } | |
824 | break; | |
825 | } | |
11c5e016 MD |
826 | default: |
827 | assert(0); | |
5a9a87dd | 828 | ret = -EINVAL; |
f07b240f | 829 | goto end; |
11c5e016 | 830 | } |
2e313670 | 831 | skip_copy: |
11c5e016 | 832 | |
9a1c1915 | 833 | if (mode == JA_RECOMPACT_ADD) { |
2e313670 MD |
834 | /* add node */ |
835 | ret = _ja_node_set_nth(new_type, new_node, | |
836 | new_shadow_node, | |
837 | n, child_node_flag); | |
7b413155 MD |
838 | if (new_type->type_class == RCU_JA_POOL && ret) { |
839 | goto fallback_toosmall; | |
840 | } | |
2e313670 MD |
841 | assert(!ret); |
842 | } | |
3d8fe307 MD |
843 | /* Return pointer to new recompacted node through old_node_flag_ptr */ |
844 | *old_node_flag_ptr = new_node_flag; | |
a2a7ff59 | 845 | if (old_node) { |
2e313670 MD |
846 | int flags; |
847 | ||
848 | flags = RCUJA_SHADOW_CLEAR_FREE_NODE; | |
849 | /* | |
850 | * It is OK to free the lock associated with a node | |
851 | * going to NULL, since we are holding the parent lock. | |
852 | * This synchronizes removal with re-add of that node. | |
853 | */ | |
854 | if (new_type_index == NODE_INDEX_NULL) | |
855 | flags = RCUJA_SHADOW_CLEAR_FREE_LOCK; | |
3d8fe307 | 856 | ret = rcuja_shadow_clear(ja->ht, old_node_flag, shadow_node, |
2e313670 | 857 | flags); |
a2a7ff59 MD |
858 | assert(!ret); |
859 | } | |
5a9a87dd MD |
860 | |
861 | ret = 0; | |
f07b240f | 862 | end: |
5a9a87dd | 863 | return ret; |
f07b240f MD |
864 | |
865 | fallback_toosmall: | |
866 | /* fallback if next pool is too small */ | |
af3cbd45 | 867 | assert(new_shadow_node); |
3d8fe307 | 868 | ret = rcuja_shadow_clear(ja->ht, new_node_flag, new_shadow_node, |
f07b240f MD |
869 | RCUJA_SHADOW_CLEAR_FREE_NODE); |
870 | assert(!ret); | |
871 | ||
2e313670 | 872 | /* Choose fallback type: pigeon */ |
f07b240f MD |
873 | new_type_index = (1UL << JA_TYPE_BITS) - 1; |
874 | dbg_printf("Fallback to type %d\n", new_type_index); | |
875 | uatomic_inc(&ja->nr_fallback); | |
876 | fallback = 1; | |
877 | goto retry; | |
7a0b2331 MD |
878 | } |
879 | ||
5a9a87dd | 880 | /* |
2e313670 | 881 | * Return 0 on success, -EAGAIN if need to retry, or other negative |
5a9a87dd MD |
882 | * error value otherwise. |
883 | */ | |
7a0b2331 | 884 | static |
d96bfb0d | 885 | int ja_node_set_nth(struct cds_ja *ja, |
b4540e8a | 886 | struct cds_ja_inode_flag **node_flag, uint8_t n, |
5a9a87dd MD |
887 | struct cds_ja_inode_flag *child_node_flag, |
888 | struct cds_ja_shadow_node *shadow_node) | |
7a0b2331 MD |
889 | { |
890 | int ret; | |
e1db2db5 | 891 | unsigned int type_index; |
d96bfb0d | 892 | const struct cds_ja_type *type; |
b4540e8a | 893 | struct cds_ja_inode *node; |
7a0b2331 | 894 | |
a2a7ff59 MD |
895 | dbg_printf("ja_node_set_nth for n=%u, node %p, shadow %p\n", |
896 | (unsigned int) n, ja_node_ptr(*node_flag), shadow_node); | |
897 | ||
e1db2db5 MD |
898 | node = ja_node_ptr(*node_flag); |
899 | type_index = ja_node_type(*node_flag); | |
900 | type = &ja_types[type_index]; | |
e1db2db5 MD |
901 | ret = _ja_node_set_nth(type, node, shadow_node, |
902 | n, child_node_flag); | |
2e313670 MD |
903 | switch (ret) { |
904 | case -ENOSPC: | |
e1db2db5 | 905 | /* Not enough space in node, need to recompact. */ |
2e313670 | 906 | ret = ja_node_recompact(JA_RECOMPACT_ADD, ja, type_index, type, node, |
af3cbd45 | 907 | shadow_node, node_flag, n, child_node_flag, NULL); |
2e313670 MD |
908 | break; |
909 | case -ERANGE: | |
910 | /* Node needs to be recompacted. */ | |
911 | ret = ja_node_recompact(JA_RECOMPACT, ja, type_index, type, node, | |
af3cbd45 | 912 | shadow_node, node_flag, n, child_node_flag, NULL); |
2e313670 MD |
913 | break; |
914 | } | |
915 | return ret; | |
916 | } | |
917 | ||
918 | /* | |
919 | * Return 0 on success, -EAGAIN if need to retry, or other negative | |
920 | * error value otherwise. | |
921 | */ | |
922 | static | |
af3cbd45 MD |
923 | int ja_node_clear_ptr(struct cds_ja *ja, |
924 | struct cds_ja_inode_flag **node_flag_ptr, /* Pointer to location to nullify */ | |
925 | struct cds_ja_inode_flag **parent_node_flag_ptr, /* Address of parent ptr in its parent */ | |
926 | struct cds_ja_shadow_node *shadow_node, /* of parent */ | |
927 | uint8_t n) | |
2e313670 MD |
928 | { |
929 | int ret; | |
930 | unsigned int type_index; | |
931 | const struct cds_ja_type *type; | |
932 | struct cds_ja_inode *node; | |
933 | ||
af3cbd45 MD |
934 | dbg_printf("ja_node_clear_ptr for node %p, shadow %p, target ptr %p\n", |
935 | ja_node_ptr(*parent_node_flag_ptr), shadow_node, node_flag_ptr); | |
2e313670 | 936 | |
af3cbd45 MD |
937 | node = ja_node_ptr(*parent_node_flag_ptr); |
938 | type_index = ja_node_type(*parent_node_flag_ptr); | |
2e313670 | 939 | type = &ja_types[type_index]; |
af3cbd45 | 940 | ret = _ja_node_clear_ptr(type, node, shadow_node, node_flag_ptr, n); |
2e313670 MD |
941 | if (ret == -EFBIG) { |
942 | /* Should to try recompaction. */ | |
943 | ret = ja_node_recompact(JA_RECOMPACT_DEL, ja, type_index, type, node, | |
af3cbd45 MD |
944 | shadow_node, parent_node_flag_ptr, n, NULL, |
945 | node_flag_ptr); | |
7a0b2331 MD |
946 | } |
947 | return ret; | |
948 | } | |
be9a7474 | 949 | |
af3cbd45 | 950 | struct cds_hlist_head cds_ja_lookup(struct cds_ja *ja, uint64_t key) |
b4540e8a | 951 | { |
41975c12 MD |
952 | unsigned int tree_depth, i; |
953 | struct cds_ja_inode_flag *node_flag; | |
af3cbd45 | 954 | struct cds_hlist_head head = { NULL }; |
41975c12 MD |
955 | |
956 | if (caa_unlikely(key > ja->key_max)) | |
af3cbd45 | 957 | return head; |
41975c12 | 958 | tree_depth = ja->tree_depth; |
5a9a87dd | 959 | node_flag = rcu_dereference(ja->root); |
41975c12 | 960 | |
5a9a87dd MD |
961 | /* level 0: root node */ |
962 | if (!ja_node_ptr(node_flag)) | |
af3cbd45 | 963 | return head; |
5a9a87dd MD |
964 | |
965 | for (i = 1; i < tree_depth; i++) { | |
79b41067 MD |
966 | uint8_t iter_key; |
967 | ||
968 | iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (tree_depth - i - 1))); | |
b62a8d0c | 969 | node_flag = ja_node_get_nth(node_flag, NULL, NULL, NULL, |
79b41067 | 970 | iter_key); |
582a6ade MD |
971 | dbg_printf("cds_ja_lookup iter key lookup %u finds node_flag %p\n", |
972 | (unsigned int) iter_key, node_flag); | |
41975c12 | 973 | if (!ja_node_ptr(node_flag)) |
af3cbd45 | 974 | return head; |
41975c12 MD |
975 | } |
976 | ||
5a9a87dd | 977 | /* Last level lookup succeded. We got an actual match. */ |
af3cbd45 MD |
978 | head.next = (struct cds_hlist_node *) node_flag; |
979 | return head; | |
5a9a87dd MD |
980 | } |
981 | ||
982 | /* | |
983 | * We reached an unpopulated node. Create it and the children we need, | |
984 | * and then attach the entire branch to the current node. This may | |
985 | * trigger recompaction of the current node. Locks needed: node lock | |
986 | * (for add), and, possibly, parent node lock (to update pointer due to | |
987 | * node recompaction). | |
988 | * | |
989 | * First take node lock, check if recompaction is needed, then take | |
990 | * parent lock (if needed). Then we can proceed to create the new | |
991 | * branch. Publish the new branch, and release locks. | |
992 | * TODO: we currently always take the parent lock even when not needed. | |
993 | */ | |
994 | static | |
995 | int ja_attach_node(struct cds_ja *ja, | |
b0ca2d21 | 996 | struct cds_ja_inode_flag **attach_node_flag_ptr, |
b62a8d0c | 997 | struct cds_ja_inode_flag *attach_node_flag, |
5a9a87dd MD |
998 | struct cds_ja_inode_flag **node_flag_ptr, |
999 | struct cds_ja_inode_flag *node_flag, | |
1000 | struct cds_ja_inode_flag *parent_node_flag, | |
1001 | uint64_t key, | |
79b41067 | 1002 | unsigned int level, |
5a9a87dd MD |
1003 | struct cds_ja_node *child_node) |
1004 | { | |
1005 | struct cds_ja_shadow_node *shadow_node = NULL, | |
af3cbd45 | 1006 | *parent_shadow_node = NULL; |
5a9a87dd MD |
1007 | struct cds_ja_inode *node = ja_node_ptr(node_flag); |
1008 | struct cds_ja_inode *parent_node = ja_node_ptr(parent_node_flag); | |
1009 | struct cds_hlist_head head; | |
1010 | struct cds_ja_inode_flag *iter_node_flag, *iter_dest_node_flag; | |
1011 | int ret, i; | |
a2a7ff59 | 1012 | struct cds_ja_inode_flag *created_nodes[JA_MAX_DEPTH]; |
5a9a87dd MD |
1013 | int nr_created_nodes = 0; |
1014 | ||
582a6ade MD |
1015 | dbg_printf("Attach node at level %u (node %p, node_flag %p)\n", |
1016 | level, node, node_flag); | |
a2a7ff59 | 1017 | |
5a9a87dd | 1018 | assert(node); |
3d8fe307 | 1019 | shadow_node = rcuja_shadow_lookup_lock(ja->ht, node_flag); |
5a9a87dd | 1020 | if (!shadow_node) { |
2e313670 | 1021 | ret = -EAGAIN; |
5a9a87dd MD |
1022 | goto end; |
1023 | } | |
1024 | if (parent_node) { | |
1025 | parent_shadow_node = rcuja_shadow_lookup_lock(ja->ht, | |
3d8fe307 | 1026 | parent_node_flag); |
5a9a87dd | 1027 | if (!parent_shadow_node) { |
2e313670 | 1028 | ret = -EAGAIN; |
5a9a87dd MD |
1029 | goto unlock_shadow; |
1030 | } | |
1031 | } | |
1032 | ||
b62a8d0c | 1033 | if (node_flag_ptr && ja_node_ptr(*node_flag_ptr)) { |
b306a0fe | 1034 | /* |
c112acaa MD |
1035 | * Target node has been updated between RCU lookup and |
1036 | * lock acquisition. We need to re-try lookup and | |
1037 | * attach. | |
1038 | */ | |
1039 | ret = -EAGAIN; | |
1040 | goto unlock_parent; | |
1041 | } | |
1042 | ||
1043 | if (attach_node_flag_ptr && ja_node_ptr(*attach_node_flag_ptr) != | |
b62a8d0c | 1044 | ja_node_ptr(attach_node_flag)) { |
c112acaa MD |
1045 | /* |
1046 | * Target node has been updated between RCU lookup and | |
1047 | * lock acquisition. We need to re-try lookup and | |
1048 | * attach. | |
b306a0fe MD |
1049 | */ |
1050 | ret = -EAGAIN; | |
1051 | goto unlock_parent; | |
1052 | } | |
1053 | ||
a2a7ff59 | 1054 | /* Create new branch, starting from bottom */ |
5a9a87dd MD |
1055 | CDS_INIT_HLIST_HEAD(&head); |
1056 | cds_hlist_add_head_rcu(&child_node->list, &head); | |
a2a7ff59 | 1057 | iter_node_flag = (struct cds_ja_inode_flag *) head.next; |
5a9a87dd | 1058 | |
79b41067 MD |
1059 | for (i = ja->tree_depth; i > (int) level; i--) { |
1060 | uint8_t iter_key; | |
1061 | ||
1062 | iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (ja->tree_depth - i))); | |
1063 | dbg_printf("branch creation level %d, key %u\n", | |
1064 | i - 1, (unsigned int) iter_key); | |
5a9a87dd MD |
1065 | iter_dest_node_flag = NULL; |
1066 | ret = ja_node_set_nth(ja, &iter_dest_node_flag, | |
79b41067 | 1067 | iter_key, |
5a9a87dd MD |
1068 | iter_node_flag, |
1069 | NULL); | |
1070 | if (ret) | |
1071 | goto check_error; | |
1072 | created_nodes[nr_created_nodes++] = iter_dest_node_flag; | |
1073 | iter_node_flag = iter_dest_node_flag; | |
1074 | } | |
1075 | ||
79b41067 MD |
1076 | if (level > 1) { |
1077 | uint8_t iter_key; | |
1078 | ||
1079 | iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (ja->tree_depth - level))); | |
a2a7ff59 MD |
1080 | /* We need to use set_nth on the previous level. */ |
1081 | iter_dest_node_flag = node_flag; | |
1082 | ret = ja_node_set_nth(ja, &iter_dest_node_flag, | |
79b41067 | 1083 | iter_key, |
a2a7ff59 MD |
1084 | iter_node_flag, |
1085 | shadow_node); | |
1086 | if (ret) | |
1087 | goto check_error; | |
1088 | created_nodes[nr_created_nodes++] = iter_dest_node_flag; | |
1089 | iter_node_flag = iter_dest_node_flag; | |
1090 | } | |
1091 | ||
5a9a87dd | 1092 | /* Publish new branch */ |
a2a7ff59 | 1093 | dbg_printf("Publish branch %p, replacing %p\n", |
b0ca2d21 MD |
1094 | iter_node_flag, *attach_node_flag_ptr); |
1095 | rcu_assign_pointer(*attach_node_flag_ptr, iter_node_flag); | |
5a9a87dd MD |
1096 | |
1097 | /* Success */ | |
1098 | ret = 0; | |
1099 | ||
1100 | check_error: | |
1101 | if (ret) { | |
1102 | for (i = 0; i < nr_created_nodes; i++) { | |
1103 | int tmpret; | |
a2a7ff59 MD |
1104 | int flags; |
1105 | ||
1106 | flags = RCUJA_SHADOW_CLEAR_FREE_LOCK; | |
1107 | if (i) | |
1108 | flags |= RCUJA_SHADOW_CLEAR_FREE_NODE; | |
5a9a87dd | 1109 | tmpret = rcuja_shadow_clear(ja->ht, |
3d8fe307 | 1110 | created_nodes[i], |
a2a7ff59 MD |
1111 | NULL, |
1112 | flags); | |
5a9a87dd MD |
1113 | assert(!tmpret); |
1114 | } | |
1115 | } | |
b306a0fe | 1116 | unlock_parent: |
5a9a87dd MD |
1117 | if (parent_shadow_node) |
1118 | rcuja_shadow_unlock(parent_shadow_node); | |
1119 | unlock_shadow: | |
1120 | if (shadow_node) | |
1121 | rcuja_shadow_unlock(shadow_node); | |
1122 | end: | |
1123 | return ret; | |
1124 | } | |
1125 | ||
1126 | /* | |
af3cbd45 MD |
1127 | * Lock the parent containing the hlist head pointer, and add node to list of |
1128 | * duplicates. Failure can happen if concurrent update changes the | |
1129 | * parent before we get the lock. We return -EAGAIN in that case. | |
5a9a87dd MD |
1130 | * Return 0 on success, negative error value on failure. |
1131 | */ | |
1132 | static | |
1133 | int ja_chain_node(struct cds_ja *ja, | |
af3cbd45 | 1134 | struct cds_ja_inode_flag *parent_node_flag, |
fa112799 | 1135 | struct cds_ja_inode_flag **node_flag_ptr, |
c112acaa | 1136 | struct cds_ja_inode_flag *node_flag, |
5a9a87dd MD |
1137 | struct cds_hlist_head *head, |
1138 | struct cds_ja_node *node) | |
1139 | { | |
1140 | struct cds_ja_shadow_node *shadow_node; | |
fa112799 | 1141 | int ret = 0; |
5a9a87dd | 1142 | |
3d8fe307 | 1143 | shadow_node = rcuja_shadow_lookup_lock(ja->ht, parent_node_flag); |
b306a0fe | 1144 | if (!shadow_node) { |
2e313670 | 1145 | return -EAGAIN; |
b306a0fe | 1146 | } |
c112acaa | 1147 | if (ja_node_ptr(*node_flag_ptr) != ja_node_ptr(node_flag)) { |
fa112799 MD |
1148 | ret = -EAGAIN; |
1149 | goto end; | |
1150 | } | |
5a9a87dd | 1151 | cds_hlist_add_head_rcu(&node->list, head); |
fa112799 | 1152 | end: |
5a9a87dd | 1153 | rcuja_shadow_unlock(shadow_node); |
fa112799 | 1154 | return ret; |
5a9a87dd MD |
1155 | } |
1156 | ||
1157 | int cds_ja_add(struct cds_ja *ja, uint64_t key, | |
1158 | struct cds_ja_node *new_node) | |
1159 | { | |
1160 | unsigned int tree_depth, i; | |
b0ca2d21 MD |
1161 | struct cds_ja_inode_flag **attach_node_flag_ptr, |
1162 | **node_flag_ptr; | |
5a9a87dd MD |
1163 | struct cds_ja_inode_flag *node_flag, |
1164 | *parent_node_flag, | |
b62a8d0c MD |
1165 | *parent2_node_flag, |
1166 | *attach_node_flag; | |
5a9a87dd MD |
1167 | int ret; |
1168 | ||
b306a0fe | 1169 | if (caa_unlikely(key > ja->key_max)) { |
5a9a87dd | 1170 | return -EINVAL; |
b306a0fe | 1171 | } |
5a9a87dd MD |
1172 | tree_depth = ja->tree_depth; |
1173 | ||
1174 | retry: | |
a2a7ff59 MD |
1175 | dbg_printf("cds_ja_add attempt: key %" PRIu64 ", node %p\n", |
1176 | key, new_node); | |
5a9a87dd | 1177 | parent2_node_flag = NULL; |
b0f74e47 MD |
1178 | parent_node_flag = |
1179 | (struct cds_ja_inode_flag *) &ja->root; /* Use root ptr address as key for mutex */ | |
b0ca2d21 | 1180 | attach_node_flag_ptr = &ja->root; |
b62a8d0c | 1181 | attach_node_flag = rcu_dereference(ja->root); |
5a9a87dd | 1182 | node_flag_ptr = &ja->root; |
35170a44 | 1183 | node_flag = rcu_dereference(ja->root); |
5a9a87dd MD |
1184 | |
1185 | /* Iterate on all internal levels */ | |
a2a7ff59 | 1186 | for (i = 1; i < tree_depth; i++) { |
79b41067 MD |
1187 | uint8_t iter_key; |
1188 | ||
b0ca2d21 | 1189 | dbg_printf("cds_ja_add iter attach_node_flag_ptr %p node_flag_ptr %p node_flag %p\n", |
c112acaa | 1190 | attach_node_flag_ptr, node_flag_ptr, node_flag); |
5a9a87dd | 1191 | if (!ja_node_ptr(node_flag)) { |
b0ca2d21 | 1192 | ret = ja_attach_node(ja, attach_node_flag_ptr, |
b62a8d0c | 1193 | attach_node_flag, |
b0ca2d21 | 1194 | node_flag_ptr, |
c112acaa MD |
1195 | parent_node_flag, |
1196 | parent2_node_flag, | |
5a9a87dd | 1197 | key, i, new_node); |
2e313670 | 1198 | if (ret == -EAGAIN || ret == -EEXIST) |
5a9a87dd MD |
1199 | goto retry; |
1200 | else | |
1201 | goto end; | |
1202 | } | |
79b41067 | 1203 | iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (tree_depth - i - 1))); |
5a9a87dd MD |
1204 | parent2_node_flag = parent_node_flag; |
1205 | parent_node_flag = node_flag; | |
1206 | node_flag = ja_node_get_nth(node_flag, | |
b0ca2d21 | 1207 | &attach_node_flag_ptr, |
b62a8d0c | 1208 | &attach_node_flag, |
5a9a87dd | 1209 | &node_flag_ptr, |
79b41067 | 1210 | iter_key); |
b0ca2d21 MD |
1211 | dbg_printf("cds_ja_add iter key lookup %u finds node_flag %p attach_node_flag_ptr %p node_flag_ptr %p\n", |
1212 | (unsigned int) iter_key, node_flag, | |
c112acaa MD |
1213 | attach_node_flag_ptr, |
1214 | node_flag_ptr); | |
5a9a87dd MD |
1215 | } |
1216 | ||
1217 | /* | |
1218 | * We reached bottom of tree, simply add node to last internal | |
1219 | * level, or chain it if key is already present. | |
1220 | */ | |
1221 | if (!ja_node_ptr(node_flag)) { | |
c112acaa MD |
1222 | dbg_printf("cds_ja_add attach_node_flag_ptr %p node_flag_ptr %p node_flag %p\n", |
1223 | attach_node_flag_ptr, node_flag_ptr, node_flag); | |
b0ca2d21 | 1224 | ret = ja_attach_node(ja, attach_node_flag_ptr, |
b62a8d0c | 1225 | attach_node_flag, |
b0ca2d21 | 1226 | node_flag_ptr, parent_node_flag, |
5a9a87dd MD |
1227 | parent2_node_flag, key, i, new_node); |
1228 | } else { | |
1229 | ret = ja_chain_node(ja, | |
af3cbd45 | 1230 | parent_node_flag, |
fa112799 | 1231 | node_flag_ptr, |
c112acaa | 1232 | node_flag, |
b0ca2d21 | 1233 | (struct cds_hlist_head *) attach_node_flag_ptr, |
5a9a87dd MD |
1234 | new_node); |
1235 | } | |
b306a0fe | 1236 | if (ret == -EAGAIN || ret == -EEXIST) |
5a9a87dd MD |
1237 | goto retry; |
1238 | end: | |
1239 | return ret; | |
b4540e8a MD |
1240 | } |
1241 | ||
af3cbd45 MD |
1242 | /* |
1243 | * Note: there is no need to lookup the pointer address associated with | |
1244 | * each node's nth item after taking the lock: it's already been done by | |
1245 | * cds_ja_del while holding the rcu read-side lock, and our node rules | |
1246 | * ensure that when a match value -> pointer is found in a node, it is | |
1247 | * _NEVER_ changed for that node without recompaction, and recompaction | |
1248 | * reallocates the node. | |
b306a0fe MD |
1249 | * However, when a child is removed from "linear" nodes, its pointer |
1250 | * is set to NULL. We therefore check, while holding the locks, if this | |
1251 | * pointer is NULL, and return -ENOENT to the caller if it is the case. | |
af3cbd45 | 1252 | */ |
35170a44 MD |
1253 | static |
1254 | int ja_detach_node(struct cds_ja *ja, | |
1255 | struct cds_ja_inode_flag **snapshot, | |
af3cbd45 MD |
1256 | struct cds_ja_inode_flag ***snapshot_ptr, |
1257 | uint8_t *snapshot_n, | |
35170a44 MD |
1258 | int nr_snapshot, |
1259 | uint64_t key, | |
1260 | struct cds_ja_node *node) | |
1261 | { | |
af3cbd45 MD |
1262 | struct cds_ja_shadow_node *shadow_nodes[JA_MAX_DEPTH]; |
1263 | struct cds_ja_inode_flag **node_flag_ptr = NULL, | |
1264 | *parent_node_flag = NULL, | |
1265 | **parent_node_flag_ptr = NULL; | |
b62a8d0c | 1266 | struct cds_ja_inode_flag *iter_node_flag; |
4d6ef45e MD |
1267 | int ret, i, nr_shadow = 0, nr_clear = 0, nr_branch = 0; |
1268 | uint8_t n = 0; | |
35170a44 | 1269 | |
4d6ef45e | 1270 | assert(nr_snapshot == ja->tree_depth + 1); |
35170a44 | 1271 | |
af3cbd45 MD |
1272 | /* |
1273 | * From the last internal level node going up, get the node | |
1274 | * lock, check if the node has only one child left. If it is the | |
1275 | * case, we continue iterating upward. When we reach a node | |
1276 | * which has more that one child left, we lock the parent, and | |
1277 | * proceed to the node deletion (removing its children too). | |
1278 | */ | |
4d6ef45e | 1279 | for (i = nr_snapshot - 2; i >= 1; i--) { |
af3cbd45 MD |
1280 | struct cds_ja_shadow_node *shadow_node; |
1281 | ||
1282 | shadow_node = rcuja_shadow_lookup_lock(ja->ht, | |
3d8fe307 | 1283 | snapshot[i]); |
af3cbd45 MD |
1284 | if (!shadow_node) { |
1285 | ret = -EAGAIN; | |
1286 | goto end; | |
1287 | } | |
af3cbd45 | 1288 | shadow_nodes[nr_shadow++] = shadow_node; |
b62a8d0c MD |
1289 | |
1290 | /* | |
1291 | * Check if node has been removed between RCU | |
1292 | * lookup and lock acquisition. | |
1293 | */ | |
1294 | assert(snapshot_ptr[i + 1]); | |
1295 | if (ja_node_ptr(*snapshot_ptr[i + 1]) | |
1296 | != ja_node_ptr(snapshot[i + 1])) { | |
1297 | ret = -ENOENT; | |
1298 | goto end; | |
1299 | } | |
1300 | ||
1301 | assert(shadow_node->nr_child > 0); | |
d810c97f | 1302 | if (shadow_node->nr_child == 1 && i > 1) |
4d6ef45e MD |
1303 | nr_clear++; |
1304 | nr_branch++; | |
af3cbd45 MD |
1305 | if (shadow_node->nr_child > 1 || i == 1) { |
1306 | /* Lock parent and break */ | |
1307 | shadow_node = rcuja_shadow_lookup_lock(ja->ht, | |
3d8fe307 | 1308 | snapshot[i - 1]); |
af3cbd45 MD |
1309 | if (!shadow_node) { |
1310 | ret = -EAGAIN; | |
1311 | goto end; | |
1312 | } | |
1313 | shadow_nodes[nr_shadow++] = shadow_node; | |
b62a8d0c | 1314 | |
c112acaa MD |
1315 | /* |
1316 | * Check if node has been removed between RCU | |
1317 | * lookup and lock acquisition. | |
1318 | */ | |
b62a8d0c MD |
1319 | assert(snapshot_ptr[i]); |
1320 | if (ja_node_ptr(*snapshot_ptr[i]) | |
1321 | != ja_node_ptr(snapshot[i])) { | |
c112acaa MD |
1322 | ret = -ENOENT; |
1323 | goto end; | |
1324 | } | |
1325 | ||
b62a8d0c | 1326 | node_flag_ptr = snapshot_ptr[i + 1]; |
4d6ef45e MD |
1327 | n = snapshot_n[i + 1]; |
1328 | parent_node_flag_ptr = snapshot_ptr[i]; | |
1329 | parent_node_flag = snapshot[i]; | |
c112acaa | 1330 | |
af3cbd45 MD |
1331 | if (i > 1) { |
1332 | /* | |
1333 | * Lock parent's parent, in case we need | |
1334 | * to recompact parent. | |
1335 | */ | |
1336 | shadow_node = rcuja_shadow_lookup_lock(ja->ht, | |
3d8fe307 | 1337 | snapshot[i - 2]); |
af3cbd45 MD |
1338 | if (!shadow_node) { |
1339 | ret = -EAGAIN; | |
1340 | goto end; | |
1341 | } | |
1342 | shadow_nodes[nr_shadow++] = shadow_node; | |
b62a8d0c MD |
1343 | |
1344 | /* | |
1345 | * Check if node has been removed between RCU | |
1346 | * lookup and lock acquisition. | |
1347 | */ | |
1348 | assert(snapshot_ptr[i - 1]); | |
1349 | if (ja_node_ptr(*snapshot_ptr[i - 1]) | |
1350 | != ja_node_ptr(snapshot[i - 1])) { | |
1351 | ret = -ENOENT; | |
1352 | goto end; | |
1353 | } | |
af3cbd45 | 1354 | } |
b62a8d0c | 1355 | |
af3cbd45 MD |
1356 | break; |
1357 | } | |
1358 | } | |
1359 | ||
1360 | /* | |
4d6ef45e MD |
1361 | * At this point, we want to delete all nodes that are about to |
1362 | * be removed from shadow_nodes (except the last one, which is | |
1363 | * either the root or the parent of the upmost node with 1 | |
b62a8d0c MD |
1364 | * child). OK to free lock here, because RCU read lock is held, |
1365 | * and free only performed in call_rcu. | |
af3cbd45 MD |
1366 | */ |
1367 | ||
1368 | for (i = 0; i < nr_clear; i++) { | |
1369 | ret = rcuja_shadow_clear(ja->ht, | |
3d8fe307 | 1370 | shadow_nodes[i]->node_flag, |
af3cbd45 MD |
1371 | shadow_nodes[i], |
1372 | RCUJA_SHADOW_CLEAR_FREE_NODE | |
1373 | | RCUJA_SHADOW_CLEAR_FREE_LOCK); | |
1374 | assert(!ret); | |
1375 | } | |
1376 | ||
1377 | iter_node_flag = parent_node_flag; | |
1378 | /* Remove from parent */ | |
1379 | ret = ja_node_clear_ptr(ja, | |
1380 | node_flag_ptr, /* Pointer to location to nullify */ | |
1381 | &iter_node_flag, /* Old new parent ptr in its parent */ | |
4d6ef45e | 1382 | shadow_nodes[nr_branch - 1], /* of parent */ |
af3cbd45 | 1383 | n); |
b306a0fe MD |
1384 | if (ret) |
1385 | goto end; | |
af3cbd45 | 1386 | |
4d6ef45e MD |
1387 | dbg_printf("ja_detach_node: publish %p instead of %p\n", |
1388 | iter_node_flag, *parent_node_flag_ptr); | |
af3cbd45 MD |
1389 | /* Update address of parent ptr in its parent */ |
1390 | rcu_assign_pointer(*parent_node_flag_ptr, iter_node_flag); | |
1391 | ||
1392 | end: | |
1393 | for (i = 0; i < nr_shadow; i++) | |
1394 | rcuja_shadow_unlock(shadow_nodes[i]); | |
35170a44 MD |
1395 | return ret; |
1396 | } | |
1397 | ||
af3cbd45 MD |
1398 | static |
1399 | int ja_unchain_node(struct cds_ja *ja, | |
1400 | struct cds_ja_inode_flag *parent_node_flag, | |
fa112799 | 1401 | struct cds_ja_inode_flag **node_flag_ptr, |
013a6083 | 1402 | struct cds_ja_inode_flag *node_flag, |
af3cbd45 MD |
1403 | struct cds_ja_node *node) |
1404 | { | |
1405 | struct cds_ja_shadow_node *shadow_node; | |
f2758d14 | 1406 | struct cds_hlist_node *hlist_node; |
013a6083 MD |
1407 | struct cds_hlist_head hlist_head; |
1408 | int ret = 0, count = 0, found = 0; | |
af3cbd45 | 1409 | |
3d8fe307 | 1410 | shadow_node = rcuja_shadow_lookup_lock(ja->ht, parent_node_flag); |
af3cbd45 MD |
1411 | if (!shadow_node) |
1412 | return -EAGAIN; | |
013a6083 | 1413 | if (ja_node_ptr(*node_flag_ptr) != ja_node_ptr(node_flag)) { |
fa112799 MD |
1414 | ret = -EAGAIN; |
1415 | goto end; | |
1416 | } | |
013a6083 | 1417 | hlist_head.next = (struct cds_hlist_node *) ja_node_ptr(node_flag); |
af3cbd45 MD |
1418 | /* |
1419 | * Retry if another thread removed all but one of duplicates | |
fa112799 | 1420 | * since check (this check was performed without lock). |
013a6083 MD |
1421 | * Ensure that the node we are about to remove is still in the |
1422 | * list (while holding lock). | |
af3cbd45 | 1423 | */ |
013a6083 | 1424 | cds_hlist_for_each_rcu(hlist_node, &hlist_head) { |
ade342cb MD |
1425 | if (count == 0) { |
1426 | /* FIXME: currently a work-around */ | |
1427 | hlist_node->prev = (struct cds_hlist_node *) node_flag_ptr; | |
1428 | } | |
f2758d14 | 1429 | count++; |
013a6083 MD |
1430 | if (hlist_node == &node->list) |
1431 | found++; | |
f2758d14 | 1432 | } |
013a6083 MD |
1433 | assert(found <= 1); |
1434 | if (!found || count == 1) { | |
af3cbd45 MD |
1435 | ret = -EAGAIN; |
1436 | goto end; | |
1437 | } | |
1438 | cds_hlist_del_rcu(&node->list); | |
ade342cb MD |
1439 | /* |
1440 | * Validate that we indeed removed the node from linked list. | |
1441 | */ | |
1442 | assert(ja_node_ptr(*node_flag_ptr) != (struct cds_ja_inode *) node); | |
af3cbd45 MD |
1443 | end: |
1444 | rcuja_shadow_unlock(shadow_node); | |
1445 | return ret; | |
1446 | } | |
1447 | ||
1448 | /* | |
1449 | * Called with RCU read lock held. | |
1450 | */ | |
35170a44 MD |
1451 | int cds_ja_del(struct cds_ja *ja, uint64_t key, |
1452 | struct cds_ja_node *node) | |
1453 | { | |
1454 | unsigned int tree_depth, i; | |
1455 | struct cds_ja_inode_flag *snapshot[JA_MAX_DEPTH]; | |
af3cbd45 MD |
1456 | struct cds_ja_inode_flag **snapshot_ptr[JA_MAX_DEPTH]; |
1457 | uint8_t snapshot_n[JA_MAX_DEPTH]; | |
35170a44 | 1458 | struct cds_ja_inode_flag *node_flag; |
fa112799 MD |
1459 | struct cds_ja_inode_flag **prev_node_flag_ptr, |
1460 | **node_flag_ptr; | |
4d6ef45e | 1461 | int nr_snapshot; |
35170a44 MD |
1462 | int ret; |
1463 | ||
1464 | if (caa_unlikely(key > ja->key_max)) | |
1465 | return -EINVAL; | |
1466 | tree_depth = ja->tree_depth; | |
1467 | ||
1468 | retry: | |
4d6ef45e | 1469 | nr_snapshot = 0; |
35170a44 MD |
1470 | dbg_printf("cds_ja_del attempt: key %" PRIu64 ", node %p\n", |
1471 | key, node); | |
1472 | ||
1473 | /* snapshot for level 0 is only for shadow node lookup */ | |
4d6ef45e MD |
1474 | snapshot_n[0] = 0; |
1475 | snapshot_n[1] = 0; | |
af3cbd45 | 1476 | snapshot_ptr[nr_snapshot] = NULL; |
35170a44 MD |
1477 | snapshot[nr_snapshot++] = (struct cds_ja_inode_flag *) &ja->root; |
1478 | node_flag = rcu_dereference(ja->root); | |
af3cbd45 | 1479 | prev_node_flag_ptr = &ja->root; |
fa112799 | 1480 | node_flag_ptr = &ja->root; |
35170a44 MD |
1481 | |
1482 | /* Iterate on all internal levels */ | |
1483 | for (i = 1; i < tree_depth; i++) { | |
1484 | uint8_t iter_key; | |
1485 | ||
1486 | dbg_printf("cds_ja_del iter node_flag %p\n", | |
1487 | node_flag); | |
1488 | if (!ja_node_ptr(node_flag)) { | |
1489 | return -ENOENT; | |
1490 | } | |
35170a44 | 1491 | iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (tree_depth - i - 1))); |
4d6ef45e | 1492 | snapshot_n[nr_snapshot + 1] = iter_key; |
af3cbd45 MD |
1493 | snapshot_ptr[nr_snapshot] = prev_node_flag_ptr; |
1494 | snapshot[nr_snapshot++] = node_flag; | |
35170a44 | 1495 | node_flag = ja_node_get_nth(node_flag, |
af3cbd45 | 1496 | &prev_node_flag_ptr, |
b62a8d0c | 1497 | NULL, |
fa112799 | 1498 | &node_flag_ptr, |
35170a44 | 1499 | iter_key); |
af3cbd45 MD |
1500 | dbg_printf("cds_ja_del iter key lookup %u finds node_flag %p, prev_node_flag_ptr %p\n", |
1501 | (unsigned int) iter_key, node_flag, | |
1502 | prev_node_flag_ptr); | |
35170a44 | 1503 | } |
35170a44 MD |
1504 | /* |
1505 | * We reached bottom of tree, try to find the node we are trying | |
1506 | * to remove. Fail if we cannot find it. | |
1507 | */ | |
1508 | if (!ja_node_ptr(node_flag)) { | |
4d6ef45e MD |
1509 | dbg_printf("cds_ja_del: no node found for key %" PRIu64 "\n", |
1510 | key); | |
35170a44 MD |
1511 | return -ENOENT; |
1512 | } else { | |
4d6ef45e | 1513 | struct cds_hlist_head hlist_head; |
35170a44 | 1514 | struct cds_hlist_node *hlist_node; |
af3cbd45 MD |
1515 | struct cds_ja_node *entry, *match = NULL; |
1516 | int count = 0; | |
35170a44 | 1517 | |
4d6ef45e MD |
1518 | hlist_head.next = |
1519 | (struct cds_hlist_node *) ja_node_ptr(node_flag); | |
af3cbd45 | 1520 | cds_hlist_for_each_entry_rcu(entry, |
35170a44 | 1521 | hlist_node, |
4d6ef45e | 1522 | &hlist_head, |
35170a44 | 1523 | list) { |
4d6ef45e | 1524 | dbg_printf("cds_ja_del: compare %p with entry %p\n", node, entry); |
af3cbd45 MD |
1525 | if (entry == node) |
1526 | match = entry; | |
1527 | count++; | |
35170a44 | 1528 | } |
4d6ef45e MD |
1529 | if (!match) { |
1530 | dbg_printf("cds_ja_del: no node match for node %p key %" PRIu64 "\n", node, key); | |
35170a44 | 1531 | return -ENOENT; |
4d6ef45e | 1532 | } |
af3cbd45 MD |
1533 | assert(count > 0); |
1534 | if (count == 1) { | |
1535 | /* | |
4d6ef45e MD |
1536 | * Removing last of duplicates. Last snapshot |
1537 | * does not have a shadow node (external leafs). | |
af3cbd45 MD |
1538 | */ |
1539 | snapshot_ptr[nr_snapshot] = prev_node_flag_ptr; | |
1540 | snapshot[nr_snapshot++] = node_flag; | |
1541 | ret = ja_detach_node(ja, snapshot, snapshot_ptr, | |
1542 | snapshot_n, nr_snapshot, key, node); | |
1543 | } else { | |
f2758d14 | 1544 | ret = ja_unchain_node(ja, snapshot[nr_snapshot - 1], |
013a6083 | 1545 | node_flag_ptr, node_flag, match); |
af3cbd45 | 1546 | } |
35170a44 | 1547 | } |
b306a0fe MD |
1548 | /* |
1549 | * Explanation of -ENOENT handling: caused by concurrent delete | |
1550 | * between RCU lookup and actual removal. Need to re-do the | |
1551 | * lookup and removal attempt. | |
1552 | */ | |
1553 | if (ret == -EAGAIN || ret == -ENOENT) | |
35170a44 MD |
1554 | goto retry; |
1555 | return ret; | |
1556 | } | |
1557 | ||
b4540e8a MD |
1558 | struct cds_ja *_cds_ja_new(unsigned int key_bits, |
1559 | const struct rcu_flavor_struct *flavor) | |
be9a7474 MD |
1560 | { |
1561 | struct cds_ja *ja; | |
b0f74e47 | 1562 | int ret; |
f07b240f | 1563 | struct cds_ja_shadow_node *root_shadow_node; |
be9a7474 MD |
1564 | |
1565 | ja = calloc(sizeof(*ja), 1); | |
1566 | if (!ja) | |
1567 | goto ja_error; | |
b4540e8a MD |
1568 | |
1569 | switch (key_bits) { | |
1570 | case 8: | |
b4540e8a | 1571 | case 16: |
1216b3d2 | 1572 | case 24: |
b4540e8a | 1573 | case 32: |
1216b3d2 MD |
1574 | case 40: |
1575 | case 48: | |
1576 | case 56: | |
1577 | ja->key_max = (1ULL << key_bits) - 1; | |
b4540e8a MD |
1578 | break; |
1579 | case 64: | |
1580 | ja->key_max = UINT64_MAX; | |
1581 | break; | |
1582 | default: | |
1583 | goto check_error; | |
1584 | } | |
1585 | ||
be9a7474 | 1586 | /* ja->root is NULL */ |
5a9a87dd | 1587 | /* tree_depth 0 is for pointer to root node */ |
582a6ade | 1588 | ja->tree_depth = (key_bits >> JA_LOG2_BITS_PER_BYTE) + 1; |
a2a7ff59 | 1589 | assert(ja->tree_depth <= JA_MAX_DEPTH); |
be9a7474 MD |
1590 | ja->ht = rcuja_create_ht(flavor); |
1591 | if (!ja->ht) | |
1592 | goto ht_error; | |
b0f74e47 MD |
1593 | |
1594 | /* | |
1595 | * Note: we should not free this node until judy array destroy. | |
1596 | */ | |
f07b240f | 1597 | root_shadow_node = rcuja_shadow_set(ja->ht, |
3d8fe307 MD |
1598 | (struct cds_ja_inode_flag *) &ja->root, |
1599 | NULL, ja); | |
f07b240f MD |
1600 | if (!root_shadow_node) { |
1601 | ret = -ENOMEM; | |
b0f74e47 | 1602 | goto ht_node_error; |
f07b240f | 1603 | } |
3d8fe307 | 1604 | root_shadow_node->level = 0; |
b0f74e47 | 1605 | |
be9a7474 MD |
1606 | return ja; |
1607 | ||
b0f74e47 MD |
1608 | ht_node_error: |
1609 | ret = rcuja_delete_ht(ja->ht); | |
1610 | assert(!ret); | |
be9a7474 | 1611 | ht_error: |
b4540e8a | 1612 | check_error: |
be9a7474 MD |
1613 | free(ja); |
1614 | ja_error: | |
1615 | return NULL; | |
1616 | } | |
1617 | ||
3d8fe307 MD |
1618 | /* |
1619 | * Called from RCU read-side CS. | |
1620 | */ | |
1621 | __attribute__((visibility("protected"))) | |
1622 | void rcuja_free_all_children(struct cds_ja_shadow_node *shadow_node, | |
1623 | struct cds_ja_inode_flag *node_flag, | |
1624 | void (*free_node_cb)(struct rcu_head *head)) | |
1625 | { | |
1626 | const struct rcu_flavor_struct *flavor; | |
1627 | unsigned int type_index; | |
1628 | struct cds_ja_inode *node; | |
1629 | const struct cds_ja_type *type; | |
1630 | ||
1631 | flavor = cds_lfht_rcu_flavor(shadow_node->ja->ht); | |
1632 | node = ja_node_ptr(node_flag); | |
1633 | assert(node != NULL); | |
1634 | type_index = ja_node_type(node_flag); | |
1635 | type = &ja_types[type_index]; | |
1636 | ||
1637 | switch (type->type_class) { | |
1638 | case RCU_JA_LINEAR: | |
1639 | { | |
1640 | uint8_t nr_child = | |
1641 | ja_linear_node_get_nr_child(type, node); | |
1642 | unsigned int i; | |
1643 | ||
1644 | for (i = 0; i < nr_child; i++) { | |
1645 | struct cds_ja_inode_flag *iter; | |
1646 | struct cds_hlist_head head; | |
1647 | struct cds_ja_node *entry; | |
1648 | struct cds_hlist_node *pos; | |
1649 | uint8_t v; | |
1650 | ||
1651 | ja_linear_node_get_ith_pos(type, node, i, &v, &iter); | |
1652 | if (!iter) | |
1653 | continue; | |
1654 | head.next = (struct cds_hlist_node *) iter; | |
1655 | cds_hlist_for_each_entry_rcu(entry, pos, &head, list) { | |
1656 | flavor->update_call_rcu(&entry->head, free_node_cb); | |
1657 | } | |
1658 | } | |
1659 | break; | |
1660 | } | |
1661 | case RCU_JA_POOL: | |
1662 | { | |
1663 | unsigned int pool_nr; | |
1664 | ||
1665 | for (pool_nr = 0; pool_nr < (1U << type->nr_pool_order); pool_nr++) { | |
1666 | struct cds_ja_inode *pool = | |
1667 | ja_pool_node_get_ith_pool(type, node, pool_nr); | |
1668 | uint8_t nr_child = | |
1669 | ja_linear_node_get_nr_child(type, pool); | |
1670 | unsigned int j; | |
1671 | ||
1672 | for (j = 0; j < nr_child; j++) { | |
1673 | struct cds_ja_inode_flag *iter; | |
1674 | struct cds_hlist_head head; | |
1675 | struct cds_ja_node *entry; | |
1676 | struct cds_hlist_node *pos; | |
1677 | uint8_t v; | |
1678 | ||
1679 | ja_linear_node_get_ith_pos(type, node, j, &v, &iter); | |
1680 | if (!iter) | |
1681 | continue; | |
1682 | head.next = (struct cds_hlist_node *) iter; | |
1683 | cds_hlist_for_each_entry_rcu(entry, pos, &head, list) { | |
1684 | flavor->update_call_rcu(&entry->head, free_node_cb); | |
1685 | } | |
1686 | } | |
1687 | } | |
1688 | break; | |
1689 | } | |
1690 | case RCU_JA_NULL: | |
1691 | break; | |
1692 | case RCU_JA_PIGEON: | |
1693 | { | |
1694 | uint8_t nr_child; | |
1695 | unsigned int i; | |
1696 | ||
1697 | nr_child = shadow_node->nr_child; | |
1698 | for (i = 0; i < nr_child; i++) { | |
1699 | struct cds_ja_inode_flag *iter; | |
1700 | struct cds_hlist_head head; | |
1701 | struct cds_ja_node *entry; | |
1702 | struct cds_hlist_node *pos; | |
1703 | ||
1704 | iter = ja_pigeon_node_get_ith_pos(type, node, i); | |
1705 | if (!iter) | |
1706 | continue; | |
1707 | head.next = (struct cds_hlist_node *) iter; | |
1708 | cds_hlist_for_each_entry_rcu(entry, pos, &head, list) { | |
1709 | flavor->update_call_rcu(&entry->head, free_node_cb); | |
1710 | } | |
1711 | } | |
1712 | break; | |
1713 | } | |
1714 | default: | |
1715 | assert(0); | |
1716 | } | |
1717 | } | |
1718 | ||
be9a7474 MD |
1719 | /* |
1720 | * There should be no more concurrent add to the judy array while it is | |
1721 | * being destroyed (ensured by the caller). | |
1722 | */ | |
3d8fe307 MD |
1723 | int cds_ja_destroy(struct cds_ja *ja, |
1724 | void (*free_node_cb)(struct rcu_head *head)) | |
be9a7474 | 1725 | { |
b4540e8a MD |
1726 | int ret; |
1727 | ||
be9a7474 | 1728 | rcuja_shadow_prune(ja->ht, |
3d8fe307 MD |
1729 | RCUJA_SHADOW_CLEAR_FREE_NODE | RCUJA_SHADOW_CLEAR_FREE_LOCK, |
1730 | free_node_cb); | |
b4540e8a MD |
1731 | ret = rcuja_delete_ht(ja->ht); |
1732 | if (ret) | |
1733 | return ret; | |
f07b240f MD |
1734 | if (uatomic_read(&ja->nr_fallback)) |
1735 | fprintf(stderr, | |
1736 | "[warning] RCU Judy Array used %lu fallback node(s)\n", | |
1737 | uatomic_read(&ja->nr_fallback)); | |
b4540e8a | 1738 | free(ja); |
41975c12 | 1739 | return 0; |
be9a7474 | 1740 | } |