Commit | Line | Data |
---|---|---|
61009379 MD |
1 | /* |
2 | * rcuja/rcuja.c | |
3 | * | |
4 | * Userspace RCU library - RCU Judy Array | |
5 | * | |
6 | * Copyright 2012 - Mathieu Desnoyers <mathieu.desnoyers@efficios.com> | |
7 | * | |
8 | * This library is free software; you can redistribute it and/or | |
9 | * modify it under the terms of the GNU Lesser General Public | |
10 | * License as published by the Free Software Foundation; either | |
11 | * version 2.1 of the License, or (at your option) any later version. | |
12 | * | |
13 | * This library is distributed in the hope that it will be useful, | |
14 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
15 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | |
16 | * Lesser General Public License for more details. | |
17 | * | |
18 | * You should have received a copy of the GNU Lesser General Public | |
19 | * License along with this library; if not, write to the Free Software | |
20 | * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA | |
21 | */ | |
22 | ||
195e72d3 | 23 | #define _LGPL_SOURCE |
e5227865 | 24 | #include <stdint.h> |
8e519e3c | 25 | #include <errno.h> |
d68c6810 | 26 | #include <limits.h> |
b1a90ce3 | 27 | #include <string.h> |
61009379 | 28 | #include <urcu/rcuja.h> |
d68c6810 MD |
29 | #include <urcu/compiler.h> |
30 | #include <urcu/arch.h> | |
31 | #include <assert.h> | |
8e519e3c | 32 | #include <urcu-pointer.h> |
f07b240f | 33 | #include <urcu/uatomic.h> |
b4540e8a | 34 | #include <stdint.h> |
8e519e3c | 35 | |
61009379 MD |
36 | #include "rcuja-internal.h" |
37 | ||
b1a90ce3 MD |
38 | #ifndef abs |
39 | #define abs_int(a) ((int) (a) > 0 ? (int) (a) : -((int) (a))) | |
40 | #endif | |
41 | ||
d96bfb0d | 42 | enum cds_ja_type_class { |
e5227865 | 43 | RCU_JA_LINEAR = 0, /* Type A */ |
fd800776 MD |
44 | /* 32-bit: 1 to 25 children, 8 to 128 bytes */ |
45 | /* 64-bit: 1 to 28 children, 16 to 256 bytes */ | |
46 | RCU_JA_POOL = 1, /* Type B */ | |
47 | /* 32-bit: 26 to 100 children, 256 to 512 bytes */ | |
48 | /* 64-bit: 29 to 112 children, 512 to 1024 bytes */ | |
e5227865 | 49 | RCU_JA_PIGEON = 2, /* Type C */ |
fd800776 MD |
50 | /* 32-bit: 101 to 256 children, 1024 bytes */ |
51 | /* 64-bit: 113 to 256 children, 2048 bytes */ | |
e5227865 | 52 | /* Leaf nodes are implicit from their height in the tree */ |
1db4943c | 53 | RCU_JA_NR_TYPES, |
e1db2db5 MD |
54 | |
55 | RCU_JA_NULL, /* not an encoded type, but keeps code regular */ | |
e5227865 MD |
56 | }; |
57 | ||
d96bfb0d MD |
58 | struct cds_ja_type { |
59 | enum cds_ja_type_class type_class; | |
8e519e3c MD |
60 | uint16_t min_child; /* minimum number of children: 1 to 256 */ |
61 | uint16_t max_child; /* maximum number of children: 1 to 256 */ | |
62 | uint16_t max_linear_child; /* per-pool max nr. children: 1 to 256 */ | |
63 | uint16_t order; /* node size is (1 << order), in bytes */ | |
fd800776 MD |
64 | uint16_t nr_pool_order; /* number of pools */ |
65 | uint16_t pool_size_order; /* pool size */ | |
e5227865 MD |
66 | }; |
67 | ||
68 | /* | |
69 | * Iteration on the array to find the right node size for the number of | |
d68c6810 | 70 | * children stops when it reaches .max_child == 256 (this is the largest |
e5227865 | 71 | * possible node size, which contains 256 children). |
d68c6810 MD |
72 | * The min_child overlaps with the previous max_child to provide an |
73 | * hysteresis loop to reallocation for patterns of cyclic add/removal | |
74 | * within the same node. | |
75 | * The node the index within the following arrays is represented on 3 | |
76 | * bits. It identifies the node type, min/max number of children, and | |
77 | * the size order. | |
3d45251f MD |
78 | * The max_child values for the RCU_JA_POOL below result from |
79 | * statistical approximation: over million populations, the max_child | |
80 | * covers between 97% and 99% of the populations generated. Therefore, a | |
81 | * fallback should exist to cover the rare extreme population unbalance | |
82 | * cases, but it will not have a major impact on speed nor space | |
83 | * consumption, since those are rare cases. | |
e5227865 | 84 | */ |
e5227865 | 85 | |
d68c6810 MD |
86 | #if (CAA_BITS_PER_LONG < 64) |
87 | /* 32-bit pointers */ | |
1db4943c MD |
88 | enum { |
89 | ja_type_0_max_child = 1, | |
90 | ja_type_1_max_child = 3, | |
91 | ja_type_2_max_child = 6, | |
92 | ja_type_3_max_child = 12, | |
93 | ja_type_4_max_child = 25, | |
94 | ja_type_5_max_child = 48, | |
95 | ja_type_6_max_child = 92, | |
96 | ja_type_7_max_child = 256, | |
e1db2db5 | 97 | ja_type_8_max_child = 0, /* NULL */ |
1db4943c MD |
98 | }; |
99 | ||
8e519e3c MD |
100 | enum { |
101 | ja_type_0_max_linear_child = 1, | |
102 | ja_type_1_max_linear_child = 3, | |
103 | ja_type_2_max_linear_child = 6, | |
104 | ja_type_3_max_linear_child = 12, | |
105 | ja_type_4_max_linear_child = 25, | |
106 | ja_type_5_max_linear_child = 24, | |
107 | ja_type_6_max_linear_child = 23, | |
108 | }; | |
109 | ||
1db4943c MD |
110 | enum { |
111 | ja_type_5_nr_pool_order = 1, | |
112 | ja_type_6_nr_pool_order = 2, | |
113 | }; | |
114 | ||
d96bfb0d | 115 | const struct cds_ja_type ja_types[] = { |
8e519e3c MD |
116 | { .type_class = RCU_JA_LINEAR, .min_child = 1, .max_child = ja_type_0_max_child, .max_linear_child = ja_type_0_max_linear_child, .order = 3, }, |
117 | { .type_class = RCU_JA_LINEAR, .min_child = 1, .max_child = ja_type_1_max_child, .max_linear_child = ja_type_1_max_linear_child, .order = 4, }, | |
118 | { .type_class = RCU_JA_LINEAR, .min_child = 3, .max_child = ja_type_2_max_child, .max_linear_child = ja_type_2_max_linear_child, .order = 5, }, | |
119 | { .type_class = RCU_JA_LINEAR, .min_child = 4, .max_child = ja_type_3_max_child, .max_linear_child = ja_type_3_max_linear_child, .order = 6, }, | |
120 | { .type_class = RCU_JA_LINEAR, .min_child = 10, .max_child = ja_type_4_max_child, .max_linear_child = ja_type_4_max_linear_child, .order = 7, }, | |
e5227865 | 121 | |
fd800776 | 122 | /* Pools may fill sooner than max_child */ |
1cee749c | 123 | /* This pool is hardcoded at index 5. See ja_node_ptr(). */ |
8e519e3c | 124 | { .type_class = RCU_JA_POOL, .min_child = 20, .max_child = ja_type_5_max_child, .max_linear_child = ja_type_5_max_linear_child, .order = 8, .nr_pool_order = ja_type_5_nr_pool_order, .pool_size_order = 7, }, |
1cee749c | 125 | /* This pool is hardcoded at index 6. See ja_node_ptr(). */ |
8e519e3c | 126 | { .type_class = RCU_JA_POOL, .min_child = 45, .max_child = ja_type_6_max_child, .max_linear_child = ja_type_6_max_linear_child, .order = 9, .nr_pool_order = ja_type_6_nr_pool_order, .pool_size_order = 7, }, |
3d45251f MD |
127 | |
128 | /* | |
b1a90ce3 MD |
129 | * Upon node removal below min_child, if child pool is filled |
130 | * beyond capacity, we roll back to pigeon. | |
3d45251f | 131 | */ |
58c16c03 | 132 | { .type_class = RCU_JA_PIGEON, .min_child = 83, .max_child = ja_type_7_max_child, .order = 10, }, |
e1db2db5 MD |
133 | |
134 | { .type_class = RCU_JA_NULL, .min_child = 0, .max_child = ja_type_8_max_child, }, | |
d68c6810 | 135 | }; |
d68c6810 MD |
136 | #else /* !(CAA_BITS_PER_LONG < 64) */ |
137 | /* 64-bit pointers */ | |
1db4943c MD |
138 | enum { |
139 | ja_type_0_max_child = 1, | |
140 | ja_type_1_max_child = 3, | |
141 | ja_type_2_max_child = 7, | |
142 | ja_type_3_max_child = 14, | |
143 | ja_type_4_max_child = 28, | |
144 | ja_type_5_max_child = 54, | |
145 | ja_type_6_max_child = 104, | |
146 | ja_type_7_max_child = 256, | |
e1db2db5 | 147 | ja_type_8_max_child = 256, |
1db4943c MD |
148 | }; |
149 | ||
8e519e3c MD |
150 | enum { |
151 | ja_type_0_max_linear_child = 1, | |
152 | ja_type_1_max_linear_child = 3, | |
153 | ja_type_2_max_linear_child = 7, | |
154 | ja_type_3_max_linear_child = 14, | |
155 | ja_type_4_max_linear_child = 28, | |
156 | ja_type_5_max_linear_child = 27, | |
157 | ja_type_6_max_linear_child = 26, | |
158 | }; | |
159 | ||
1db4943c MD |
160 | enum { |
161 | ja_type_5_nr_pool_order = 1, | |
162 | ja_type_6_nr_pool_order = 2, | |
163 | }; | |
164 | ||
d96bfb0d | 165 | const struct cds_ja_type ja_types[] = { |
8e519e3c MD |
166 | { .type_class = RCU_JA_LINEAR, .min_child = 1, .max_child = ja_type_0_max_child, .max_linear_child = ja_type_0_max_linear_child, .order = 4, }, |
167 | { .type_class = RCU_JA_LINEAR, .min_child = 1, .max_child = ja_type_1_max_child, .max_linear_child = ja_type_1_max_linear_child, .order = 5, }, | |
168 | { .type_class = RCU_JA_LINEAR, .min_child = 3, .max_child = ja_type_2_max_child, .max_linear_child = ja_type_2_max_linear_child, .order = 6, }, | |
169 | { .type_class = RCU_JA_LINEAR, .min_child = 5, .max_child = ja_type_3_max_child, .max_linear_child = ja_type_3_max_linear_child, .order = 7, }, | |
170 | { .type_class = RCU_JA_LINEAR, .min_child = 10, .max_child = ja_type_4_max_child, .max_linear_child = ja_type_4_max_linear_child, .order = 8, }, | |
e5227865 | 171 | |
3d45251f | 172 | /* Pools may fill sooner than max_child. */ |
1cee749c | 173 | /* This pool is hardcoded at index 5. See ja_node_ptr(). */ |
8e519e3c | 174 | { .type_class = RCU_JA_POOL, .min_child = 22, .max_child = ja_type_5_max_child, .max_linear_child = ja_type_5_max_linear_child, .order = 9, .nr_pool_order = ja_type_5_nr_pool_order, .pool_size_order = 8, }, |
1cee749c | 175 | /* This pool is hardcoded at index 6. See ja_node_ptr(). */ |
8e519e3c | 176 | { .type_class = RCU_JA_POOL, .min_child = 51, .max_child = ja_type_6_max_child, .max_linear_child = ja_type_6_max_linear_child, .order = 10, .nr_pool_order = ja_type_6_nr_pool_order, .pool_size_order = 8, }, |
e5227865 | 177 | |
3d45251f | 178 | /* |
b1a90ce3 MD |
179 | * Upon node removal below min_child, if child pool is filled |
180 | * beyond capacity, we roll back to pigeon. | |
3d45251f | 181 | */ |
64457f6c | 182 | { .type_class = RCU_JA_PIGEON, .min_child = 95, .max_child = ja_type_7_max_child, .order = 11, }, |
e1db2db5 MD |
183 | |
184 | { .type_class = RCU_JA_NULL, .min_child = 0, .max_child = ja_type_8_max_child, }, | |
e5227865 | 185 | }; |
d68c6810 | 186 | #endif /* !(BITS_PER_LONG < 64) */ |
e5227865 | 187 | |
1db4943c MD |
188 | static inline __attribute__((unused)) |
189 | void static_array_size_check(void) | |
190 | { | |
e1db2db5 | 191 | CAA_BUILD_BUG_ON(CAA_ARRAY_SIZE(ja_types) < JA_TYPE_MAX_NR); |
1db4943c MD |
192 | } |
193 | ||
e5227865 | 194 | /* |
d96bfb0d | 195 | * The cds_ja_node contains the compressed node data needed for |
1db4943c MD |
196 | * read-side. For linear and pool node configurations, it starts with a |
197 | * byte counting the number of children in the node. Then, the | |
198 | * node-specific data is placed. | |
199 | * The node mutex, if any is needed, protecting concurrent updated of | |
200 | * each node is placed in a separate hash table indexed by node address. | |
201 | * For the pigeon configuration, the number of children is also kept in | |
202 | * a separate hash table, indexed by node address, because it is only | |
203 | * required for updates. | |
e5227865 | 204 | */ |
1db4943c | 205 | |
ff38c745 MD |
206 | #define DECLARE_LINEAR_NODE(index) \ |
207 | struct { \ | |
208 | uint8_t nr_child; \ | |
209 | uint8_t child_value[ja_type_## index ##_max_linear_child]; \ | |
b4540e8a | 210 | struct cds_ja_inode_flag *child_ptr[ja_type_## index ##_max_linear_child]; \ |
ff38c745 MD |
211 | } |
212 | ||
213 | #define DECLARE_POOL_NODE(index) \ | |
214 | struct { \ | |
215 | struct { \ | |
216 | uint8_t nr_child; \ | |
217 | uint8_t child_value[ja_type_## index ##_max_linear_child]; \ | |
b4540e8a | 218 | struct cds_ja_inode_flag *child_ptr[ja_type_## index ##_max_linear_child]; \ |
ff38c745 MD |
219 | } linear[1U << ja_type_## index ##_nr_pool_order]; \ |
220 | } | |
1db4943c | 221 | |
b4540e8a | 222 | struct cds_ja_inode { |
1db4943c MD |
223 | union { |
224 | /* Linear configuration */ | |
225 | DECLARE_LINEAR_NODE(0) conf_0; | |
226 | DECLARE_LINEAR_NODE(1) conf_1; | |
227 | DECLARE_LINEAR_NODE(2) conf_2; | |
228 | DECLARE_LINEAR_NODE(3) conf_3; | |
229 | DECLARE_LINEAR_NODE(4) conf_4; | |
230 | ||
231 | /* Pool configuration */ | |
232 | DECLARE_POOL_NODE(5) conf_5; | |
233 | DECLARE_POOL_NODE(6) conf_6; | |
234 | ||
235 | /* Pigeon configuration */ | |
236 | struct { | |
b4540e8a | 237 | struct cds_ja_inode_flag *child[ja_type_7_max_child]; |
1db4943c MD |
238 | } conf_7; |
239 | /* data aliasing nodes for computed accesses */ | |
b4540e8a | 240 | uint8_t data[sizeof(struct cds_ja_inode_flag *) * ja_type_7_max_child]; |
1db4943c | 241 | } u; |
e5227865 MD |
242 | }; |
243 | ||
2e313670 | 244 | enum ja_recompact { |
19ddcd04 MD |
245 | JA_RECOMPACT_ADD_SAME, |
246 | JA_RECOMPACT_ADD_NEXT, | |
2e313670 MD |
247 | JA_RECOMPACT_DEL, |
248 | }; | |
249 | ||
b023ba9f MD |
250 | enum ja_lookup_inequality { |
251 | JA_LOOKUP_BE, | |
252 | JA_LOOKUP_AE, | |
253 | }; | |
254 | ||
255 | enum ja_direction { | |
256 | JA_LEFT, | |
257 | JA_RIGHT, | |
258 | JA_LEFTMOST, | |
259 | JA_RIGHTMOST, | |
260 | }; | |
261 | ||
b1a90ce3 MD |
262 | static |
263 | struct cds_ja_inode *_ja_node_mask_ptr(struct cds_ja_inode_flag *node) | |
264 | { | |
265 | return (struct cds_ja_inode *) (((unsigned long) node) & JA_PTR_MASK); | |
266 | } | |
267 | ||
268 | unsigned long ja_node_type(struct cds_ja_inode_flag *node) | |
269 | { | |
270 | unsigned long type; | |
271 | ||
272 | if (_ja_node_mask_ptr(node) == NULL) { | |
273 | return NODE_INDEX_NULL; | |
274 | } | |
275 | type = (unsigned int) ((unsigned long) node & JA_TYPE_MASK); | |
276 | assert(type < (1UL << JA_TYPE_BITS)); | |
277 | return type; | |
278 | } | |
279 | ||
354981c2 MD |
280 | static |
281 | struct cds_ja_inode *alloc_cds_ja_node(struct cds_ja *ja, | |
282 | const struct cds_ja_type *ja_type) | |
e5227865 | 283 | { |
b1a90ce3 MD |
284 | size_t len = 1U << ja_type->order; |
285 | void *p; | |
286 | int ret; | |
287 | ||
288 | ret = posix_memalign(&p, len, len); | |
289 | if (ret || !p) { | |
290 | return NULL; | |
291 | } | |
292 | memset(p, 0, len); | |
354981c2 | 293 | uatomic_inc(&ja->nr_nodes_allocated); |
b1a90ce3 | 294 | return p; |
e5227865 MD |
295 | } |
296 | ||
354981c2 | 297 | void free_cds_ja_node(struct cds_ja *ja, struct cds_ja_inode *node) |
e5227865 MD |
298 | { |
299 | free(node); | |
48cbe001 | 300 | if (node) |
354981c2 | 301 | uatomic_inc(&ja->nr_nodes_freed); |
e5227865 MD |
302 | } |
303 | ||
d68c6810 MD |
304 | #define __JA_ALIGN_MASK(v, mask) (((v) + (mask)) & ~(mask)) |
305 | #define JA_ALIGN(v, align) __JA_ALIGN_MASK(v, (typeof(v)) (align) - 1) | |
306 | #define __JA_FLOOR_MASK(v, mask) ((v) & ~(mask)) | |
307 | #define JA_FLOOR(v, align) __JA_FLOOR_MASK(v, (typeof(v)) (align) - 1) | |
308 | ||
309 | static | |
1db4943c | 310 | uint8_t *align_ptr_size(uint8_t *ptr) |
d68c6810 | 311 | { |
1db4943c | 312 | return (uint8_t *) JA_ALIGN((unsigned long) ptr, sizeof(void *)); |
d68c6810 MD |
313 | } |
314 | ||
11c5e016 | 315 | static |
d96bfb0d | 316 | uint8_t ja_linear_node_get_nr_child(const struct cds_ja_type *type, |
b4540e8a | 317 | struct cds_ja_inode *node) |
11c5e016 MD |
318 | { |
319 | assert(type->type_class == RCU_JA_LINEAR || type->type_class == RCU_JA_POOL); | |
2e313670 | 320 | return rcu_dereference(node->u.data[0]); |
11c5e016 MD |
321 | } |
322 | ||
13a7f5a6 MD |
323 | /* |
324 | * The order in which values and pointers are does does not matter: if | |
325 | * a value is missing, we return NULL. If a value is there, but its | |
326 | * associated pointers is still NULL, we return NULL too. | |
327 | */ | |
d68c6810 | 328 | static |
b4540e8a MD |
329 | struct cds_ja_inode_flag *ja_linear_node_get_nth(const struct cds_ja_type *type, |
330 | struct cds_ja_inode *node, | |
b0ca2d21 | 331 | struct cds_ja_inode_flag ***node_flag_ptr, |
8e519e3c | 332 | uint8_t n) |
d68c6810 MD |
333 | { |
334 | uint8_t nr_child; | |
335 | uint8_t *values; | |
b4540e8a MD |
336 | struct cds_ja_inode_flag **pointers; |
337 | struct cds_ja_inode_flag *ptr; | |
d68c6810 MD |
338 | unsigned int i; |
339 | ||
8e519e3c | 340 | assert(type->type_class == RCU_JA_LINEAR || type->type_class == RCU_JA_POOL); |
d68c6810 | 341 | |
11c5e016 | 342 | nr_child = ja_linear_node_get_nr_child(type, node); |
13a7f5a6 | 343 | cmm_smp_rmb(); /* read nr_child before values and pointers */ |
8e519e3c MD |
344 | assert(nr_child <= type->max_linear_child); |
345 | assert(type->type_class != RCU_JA_LINEAR || nr_child >= type->min_child); | |
d68c6810 | 346 | |
1db4943c | 347 | values = &node->u.data[1]; |
d68c6810 | 348 | for (i = 0; i < nr_child; i++) { |
13a7f5a6 | 349 | if (CMM_LOAD_SHARED(values[i]) == n) |
d68c6810 MD |
350 | break; |
351 | } | |
b0ca2d21 MD |
352 | if (i >= nr_child) { |
353 | if (caa_unlikely(node_flag_ptr)) | |
354 | *node_flag_ptr = NULL; | |
d68c6810 | 355 | return NULL; |
b0ca2d21 | 356 | } |
b4540e8a | 357 | pointers = (struct cds_ja_inode_flag **) align_ptr_size(&values[type->max_linear_child]); |
13a7f5a6 | 358 | ptr = rcu_dereference(pointers[i]); |
b0ca2d21 MD |
359 | if (caa_unlikely(node_flag_ptr)) |
360 | *node_flag_ptr = &pointers[i]; | |
d68c6810 MD |
361 | return ptr; |
362 | } | |
363 | ||
291b2543 | 364 | static |
b023ba9f | 365 | struct cds_ja_inode_flag *ja_linear_node_get_direction(const struct cds_ja_type *type, |
291b2543 | 366 | struct cds_ja_inode *node, |
36305a3d | 367 | int n, uint8_t *result_key, |
b023ba9f | 368 | enum ja_direction dir) |
291b2543 MD |
369 | { |
370 | uint8_t nr_child; | |
371 | uint8_t *values; | |
372 | struct cds_ja_inode_flag **pointers; | |
373 | struct cds_ja_inode_flag *ptr; | |
b023ba9f MD |
374 | unsigned int i; |
375 | int match_idx = -1, match_v; | |
291b2543 MD |
376 | |
377 | assert(type->type_class == RCU_JA_LINEAR || type->type_class == RCU_JA_POOL); | |
b023ba9f MD |
378 | assert(dir == JA_LEFT || dir == JA_RIGHT); |
379 | ||
380 | if (dir == JA_LEFT) { | |
381 | match_v = -1; | |
382 | } else { | |
383 | match_v = JA_ENTRY_PER_NODE; | |
384 | } | |
291b2543 MD |
385 | |
386 | nr_child = ja_linear_node_get_nr_child(type, node); | |
387 | cmm_smp_rmb(); /* read nr_child before values and pointers */ | |
388 | assert(nr_child <= type->max_linear_child); | |
389 | assert(type->type_class != RCU_JA_LINEAR || nr_child >= type->min_child); | |
390 | ||
391 | values = &node->u.data[1]; | |
392 | for (i = 0; i < nr_child; i++) { | |
393 | unsigned int v; | |
394 | ||
395 | v = CMM_LOAD_SHARED(values[i]); | |
b023ba9f MD |
396 | if (dir == JA_LEFT) { |
397 | if ((int) v < n && (int) v > match_v) { | |
398 | match_v = v; | |
399 | match_idx = i; | |
400 | } | |
401 | } else { | |
402 | if ((int) v > n && (int) v < match_v) { | |
403 | match_v = v; | |
404 | match_idx = i; | |
405 | } | |
291b2543 MD |
406 | } |
407 | } | |
b023ba9f MD |
408 | |
409 | if (match_idx < 0) { | |
291b2543 MD |
410 | return NULL; |
411 | } | |
b023ba9f MD |
412 | assert(match_v >= 0 && match_v < JA_ENTRY_PER_NODE); |
413 | ||
36305a3d | 414 | *result_key = (uint8_t) match_v; |
291b2543 MD |
415 | pointers = (struct cds_ja_inode_flag **) align_ptr_size(&values[type->max_linear_child]); |
416 | ptr = rcu_dereference(pointers[match_idx]); | |
417 | return ptr; | |
418 | } | |
419 | ||
11c5e016 | 420 | static |
5a9a87dd | 421 | void ja_linear_node_get_ith_pos(const struct cds_ja_type *type, |
b4540e8a | 422 | struct cds_ja_inode *node, |
11c5e016 MD |
423 | uint8_t i, |
424 | uint8_t *v, | |
b4540e8a | 425 | struct cds_ja_inode_flag **iter) |
11c5e016 MD |
426 | { |
427 | uint8_t *values; | |
b4540e8a | 428 | struct cds_ja_inode_flag **pointers; |
11c5e016 MD |
429 | |
430 | assert(type->type_class == RCU_JA_LINEAR || type->type_class == RCU_JA_POOL); | |
431 | assert(i < ja_linear_node_get_nr_child(type, node)); | |
432 | ||
433 | values = &node->u.data[1]; | |
434 | *v = values[i]; | |
b4540e8a | 435 | pointers = (struct cds_ja_inode_flag **) align_ptr_size(&values[type->max_linear_child]); |
11c5e016 MD |
436 | *iter = pointers[i]; |
437 | } | |
438 | ||
d68c6810 | 439 | static |
b4540e8a MD |
440 | struct cds_ja_inode_flag *ja_pool_node_get_nth(const struct cds_ja_type *type, |
441 | struct cds_ja_inode *node, | |
b1a90ce3 | 442 | struct cds_ja_inode_flag *node_flag, |
b0ca2d21 | 443 | struct cds_ja_inode_flag ***node_flag_ptr, |
8e519e3c | 444 | uint8_t n) |
d68c6810 | 445 | { |
b4540e8a | 446 | struct cds_ja_inode *linear; |
d68c6810 | 447 | |
fd800776 | 448 | assert(type->type_class == RCU_JA_POOL); |
b1a90ce3 MD |
449 | |
450 | switch (type->nr_pool_order) { | |
451 | case 1: | |
452 | { | |
453 | unsigned long bitsel, index; | |
454 | ||
455 | bitsel = ja_node_pool_1d_bitsel(node_flag); | |
456 | assert(bitsel < CHAR_BIT); | |
19ddcd04 | 457 | index = ((unsigned long) n >> bitsel) & 0x1; |
b1a90ce3 MD |
458 | linear = (struct cds_ja_inode *) &node->u.data[index << type->pool_size_order]; |
459 | break; | |
460 | } | |
461 | case 2: | |
462 | { | |
19ddcd04 MD |
463 | unsigned long bitsel[2], index[2], rindex; |
464 | ||
465 | ja_node_pool_2d_bitsel(node_flag, bitsel); | |
466 | assert(bitsel[0] < CHAR_BIT); | |
467 | assert(bitsel[1] < CHAR_BIT); | |
468 | index[0] = ((unsigned long) n >> bitsel[0]) & 0x1; | |
469 | index[0] <<= 1; | |
470 | index[1] = ((unsigned long) n >> bitsel[1]) & 0x1; | |
471 | rindex = index[0] | index[1]; | |
472 | linear = (struct cds_ja_inode *) &node->u.data[rindex << type->pool_size_order]; | |
b1a90ce3 MD |
473 | break; |
474 | } | |
475 | default: | |
476 | linear = NULL; | |
477 | assert(0); | |
478 | } | |
48cbe001 | 479 | return ja_linear_node_get_nth(type, linear, node_flag_ptr, n); |
d68c6810 MD |
480 | } |
481 | ||
11c5e016 | 482 | static |
b4540e8a MD |
483 | struct cds_ja_inode *ja_pool_node_get_ith_pool(const struct cds_ja_type *type, |
484 | struct cds_ja_inode *node, | |
11c5e016 MD |
485 | uint8_t i) |
486 | { | |
487 | assert(type->type_class == RCU_JA_POOL); | |
b4540e8a | 488 | return (struct cds_ja_inode *) |
11c5e016 MD |
489 | &node->u.data[(unsigned int) i << type->pool_size_order]; |
490 | } | |
491 | ||
291b2543 | 492 | static |
b023ba9f | 493 | struct cds_ja_inode_flag *ja_pool_node_get_direction(const struct cds_ja_type *type, |
291b2543 | 494 | struct cds_ja_inode *node, |
36305a3d | 495 | int n, uint8_t *result_key, |
b023ba9f | 496 | enum ja_direction dir) |
291b2543 MD |
497 | { |
498 | unsigned int pool_nr; | |
b023ba9f | 499 | int match_v; |
291b2543 MD |
500 | struct cds_ja_inode_flag *match_node_flag = NULL; |
501 | ||
502 | assert(type->type_class == RCU_JA_POOL); | |
b023ba9f MD |
503 | assert(dir == JA_LEFT || dir == JA_RIGHT); |
504 | ||
505 | if (dir == JA_LEFT) { | |
506 | match_v = -1; | |
507 | } else { | |
508 | match_v = JA_ENTRY_PER_NODE; | |
509 | } | |
291b2543 MD |
510 | |
511 | for (pool_nr = 0; pool_nr < (1U << type->nr_pool_order); pool_nr++) { | |
512 | struct cds_ja_inode *pool = | |
513 | ja_pool_node_get_ith_pool(type, | |
514 | node, pool_nr); | |
515 | uint8_t nr_child = | |
516 | ja_linear_node_get_nr_child(type, pool); | |
517 | unsigned int j; | |
518 | ||
519 | for (j = 0; j < nr_child; j++) { | |
520 | struct cds_ja_inode_flag *iter; | |
521 | uint8_t v; | |
522 | ||
523 | ja_linear_node_get_ith_pos(type, pool, | |
524 | j, &v, &iter); | |
525 | if (!iter) | |
526 | continue; | |
b023ba9f MD |
527 | if (dir == JA_LEFT) { |
528 | if ((int) v < n && (int) v > match_v) { | |
529 | match_v = v; | |
530 | match_node_flag = iter; | |
531 | } | |
532 | } else { | |
533 | if ((int) v > n && (int) v < match_v) { | |
534 | match_v = v; | |
535 | match_node_flag = iter; | |
536 | } | |
291b2543 MD |
537 | } |
538 | } | |
539 | } | |
36305a3d MD |
540 | if (match_node_flag) |
541 | *result_key = (uint8_t) match_v; | |
291b2543 MD |
542 | return match_node_flag; |
543 | } | |
544 | ||
d68c6810 | 545 | static |
b4540e8a MD |
546 | struct cds_ja_inode_flag *ja_pigeon_node_get_nth(const struct cds_ja_type *type, |
547 | struct cds_ja_inode *node, | |
b0ca2d21 | 548 | struct cds_ja_inode_flag ***node_flag_ptr, |
8e519e3c | 549 | uint8_t n) |
d68c6810 | 550 | { |
48cbe001 MD |
551 | struct cds_ja_inode_flag **child_node_flag_ptr; |
552 | struct cds_ja_inode_flag *child_node_flag; | |
5a9a87dd | 553 | |
d68c6810 | 554 | assert(type->type_class == RCU_JA_PIGEON); |
48cbe001 MD |
555 | child_node_flag_ptr = &((struct cds_ja_inode_flag **) node->u.data)[n]; |
556 | child_node_flag = rcu_dereference(*child_node_flag_ptr); | |
582a6ade | 557 | dbg_printf("ja_pigeon_node_get_nth child_node_flag_ptr %p\n", |
48cbe001 | 558 | child_node_flag_ptr); |
b0ca2d21 | 559 | if (caa_unlikely(node_flag_ptr)) |
48cbe001 MD |
560 | *node_flag_ptr = child_node_flag_ptr; |
561 | return child_node_flag; | |
d68c6810 MD |
562 | } |
563 | ||
291b2543 | 564 | static |
b023ba9f | 565 | struct cds_ja_inode_flag *ja_pigeon_node_get_direction(const struct cds_ja_type *type, |
291b2543 | 566 | struct cds_ja_inode *node, |
36305a3d | 567 | int n, uint8_t *result_key, |
b023ba9f | 568 | enum ja_direction dir) |
291b2543 MD |
569 | { |
570 | struct cds_ja_inode_flag **child_node_flag_ptr; | |
571 | struct cds_ja_inode_flag *child_node_flag; | |
572 | int i; | |
573 | ||
574 | assert(type->type_class == RCU_JA_PIGEON); | |
b023ba9f MD |
575 | assert(dir == JA_LEFT || dir == JA_RIGHT); |
576 | ||
577 | if (dir == JA_LEFT) { | |
578 | /* n - 1 is first value left of n */ | |
579 | for (i = n - 1; i >= 0; i--) { | |
580 | child_node_flag_ptr = &((struct cds_ja_inode_flag **) node->u.data)[i]; | |
581 | child_node_flag = rcu_dereference(*child_node_flag_ptr); | |
582 | if (child_node_flag) { | |
583 | dbg_printf("ja_pigeon_node_get_left child_node_flag %p\n", | |
584 | child_node_flag); | |
53f943b0 | 585 | *result_key = (uint8_t) i; |
b023ba9f MD |
586 | return child_node_flag; |
587 | } | |
588 | } | |
589 | } else { | |
590 | /* n + 1 is first value right of n */ | |
591 | for (i = n + 1; i < JA_ENTRY_PER_NODE; i++) { | |
592 | child_node_flag_ptr = &((struct cds_ja_inode_flag **) node->u.data)[i]; | |
593 | child_node_flag = rcu_dereference(*child_node_flag_ptr); | |
594 | if (child_node_flag) { | |
595 | dbg_printf("ja_pigeon_node_get_right child_node_flag %p\n", | |
596 | child_node_flag); | |
53f943b0 | 597 | *result_key = (uint8_t) i; |
b023ba9f MD |
598 | return child_node_flag; |
599 | } | |
291b2543 MD |
600 | } |
601 | } | |
602 | return NULL; | |
603 | } | |
604 | ||
2e313670 MD |
605 | static |
606 | struct cds_ja_inode_flag *ja_pigeon_node_get_ith_pos(const struct cds_ja_type *type, | |
607 | struct cds_ja_inode *node, | |
608 | uint8_t i) | |
609 | { | |
48cbe001 | 610 | return ja_pigeon_node_get_nth(type, node, NULL, i); |
2e313670 MD |
611 | } |
612 | ||
13a7f5a6 MD |
613 | /* |
614 | * ja_node_get_nth: get nth item from a node. | |
615 | * node_flag is already rcu_dereference'd. | |
616 | */ | |
d68c6810 | 617 | static |
b62a8d0c | 618 | struct cds_ja_inode_flag *ja_node_get_nth(struct cds_ja_inode_flag *node_flag, |
b0ca2d21 | 619 | struct cds_ja_inode_flag ***node_flag_ptr, |
8e519e3c | 620 | uint8_t n) |
d68c6810 MD |
621 | { |
622 | unsigned int type_index; | |
b4540e8a | 623 | struct cds_ja_inode *node; |
d96bfb0d | 624 | const struct cds_ja_type *type; |
d68c6810 | 625 | |
d68c6810 | 626 | node = ja_node_ptr(node_flag); |
5a9a87dd | 627 | assert(node != NULL); |
d68c6810 MD |
628 | type_index = ja_node_type(node_flag); |
629 | type = &ja_types[type_index]; | |
630 | ||
631 | switch (type->type_class) { | |
632 | case RCU_JA_LINEAR: | |
5a9a87dd | 633 | return ja_linear_node_get_nth(type, node, |
b62a8d0c | 634 | node_flag_ptr, n); |
fd800776 | 635 | case RCU_JA_POOL: |
b1a90ce3 | 636 | return ja_pool_node_get_nth(type, node, node_flag, |
b62a8d0c | 637 | node_flag_ptr, n); |
d68c6810 | 638 | case RCU_JA_PIGEON: |
5a9a87dd | 639 | return ja_pigeon_node_get_nth(type, node, |
b62a8d0c | 640 | node_flag_ptr, n); |
d68c6810 MD |
641 | default: |
642 | assert(0); | |
643 | return (void *) -1UL; | |
644 | } | |
645 | } | |
646 | ||
291b2543 | 647 | static |
b023ba9f | 648 | struct cds_ja_inode_flag *ja_node_get_direction(struct cds_ja_inode_flag *node_flag, |
36305a3d | 649 | int n, uint8_t *result_key, |
b023ba9f | 650 | enum ja_direction dir) |
291b2543 MD |
651 | { |
652 | unsigned int type_index; | |
653 | struct cds_ja_inode *node; | |
654 | const struct cds_ja_type *type; | |
655 | ||
656 | node = ja_node_ptr(node_flag); | |
657 | assert(node != NULL); | |
658 | type_index = ja_node_type(node_flag); | |
659 | type = &ja_types[type_index]; | |
660 | ||
661 | switch (type->type_class) { | |
662 | case RCU_JA_LINEAR: | |
36305a3d | 663 | return ja_linear_node_get_direction(type, node, n, result_key, dir); |
291b2543 | 664 | case RCU_JA_POOL: |
36305a3d | 665 | return ja_pool_node_get_direction(type, node, n, result_key, dir); |
291b2543 | 666 | case RCU_JA_PIGEON: |
36305a3d | 667 | return ja_pigeon_node_get_direction(type, node, n, result_key, dir); |
291b2543 MD |
668 | default: |
669 | assert(0); | |
670 | return (void *) -1UL; | |
671 | } | |
672 | } | |
673 | ||
674 | static | |
b023ba9f | 675 | struct cds_ja_inode_flag *ja_node_get_leftright(struct cds_ja_inode_flag *node_flag, |
36305a3d | 676 | unsigned int n, uint8_t *result_key, |
b023ba9f | 677 | enum ja_direction dir) |
291b2543 | 678 | { |
36305a3d | 679 | return ja_node_get_direction(node_flag, n, result_key, dir); |
b023ba9f MD |
680 | } |
681 | ||
682 | static | |
683 | struct cds_ja_inode_flag *ja_node_get_minmax(struct cds_ja_inode_flag *node_flag, | |
36305a3d | 684 | uint8_t *result_key, |
b023ba9f MD |
685 | enum ja_direction dir) |
686 | { | |
687 | switch (dir) { | |
688 | case JA_LEFTMOST: | |
689 | return ja_node_get_direction(node_flag, | |
36305a3d | 690 | -1, result_key, JA_RIGHT); |
b023ba9f MD |
691 | case JA_RIGHTMOST: |
692 | return ja_node_get_direction(node_flag, | |
36305a3d | 693 | JA_ENTRY_PER_NODE, result_key, JA_LEFT); |
b023ba9f MD |
694 | default: |
695 | assert(0); | |
696 | } | |
291b2543 MD |
697 | } |
698 | ||
8e519e3c | 699 | static |
d96bfb0d | 700 | int ja_linear_node_set_nth(const struct cds_ja_type *type, |
b4540e8a | 701 | struct cds_ja_inode *node, |
d96bfb0d | 702 | struct cds_ja_shadow_node *shadow_node, |
8e519e3c | 703 | uint8_t n, |
b4540e8a | 704 | struct cds_ja_inode_flag *child_node_flag) |
8e519e3c MD |
705 | { |
706 | uint8_t nr_child; | |
707 | uint8_t *values, *nr_child_ptr; | |
b4540e8a | 708 | struct cds_ja_inode_flag **pointers; |
2e313670 | 709 | unsigned int i, unused = 0; |
8e519e3c MD |
710 | |
711 | assert(type->type_class == RCU_JA_LINEAR || type->type_class == RCU_JA_POOL); | |
712 | ||
713 | nr_child_ptr = &node->u.data[0]; | |
48cbe001 MD |
714 | dbg_printf("linear set nth: n %u, nr_child_ptr %p\n", |
715 | (unsigned int) n, nr_child_ptr); | |
8e519e3c MD |
716 | nr_child = *nr_child_ptr; |
717 | assert(nr_child <= type->max_linear_child); | |
8e519e3c MD |
718 | |
719 | values = &node->u.data[1]; | |
2e313670 MD |
720 | pointers = (struct cds_ja_inode_flag **) align_ptr_size(&values[type->max_linear_child]); |
721 | /* Check if node value is already populated */ | |
8e519e3c | 722 | for (i = 0; i < nr_child; i++) { |
2e313670 MD |
723 | if (values[i] == n) { |
724 | if (pointers[i]) | |
725 | return -EEXIST; | |
726 | else | |
727 | break; | |
728 | } else { | |
729 | if (!pointers[i]) | |
730 | unused++; | |
731 | } | |
8e519e3c | 732 | } |
2e313670 MD |
733 | if (i == nr_child && nr_child >= type->max_linear_child) { |
734 | if (unused) | |
735 | return -ERANGE; /* recompact node */ | |
736 | else | |
737 | return -ENOSPC; /* No space left in this node type */ | |
738 | } | |
739 | ||
740 | assert(pointers[i] == NULL); | |
741 | rcu_assign_pointer(pointers[i], child_node_flag); | |
742 | /* If we expanded the nr_child, increment it */ | |
743 | if (i == nr_child) { | |
744 | CMM_STORE_SHARED(values[nr_child], n); | |
745 | /* write pointer and value before nr_child */ | |
746 | cmm_smp_wmb(); | |
747 | CMM_STORE_SHARED(*nr_child_ptr, nr_child + 1); | |
8e519e3c | 748 | } |
e1db2db5 | 749 | shadow_node->nr_child++; |
a2a7ff59 MD |
750 | dbg_printf("linear set nth: %u child, shadow: %u child, for node %p shadow %p\n", |
751 | (unsigned int) CMM_LOAD_SHARED(*nr_child_ptr), | |
752 | (unsigned int) shadow_node->nr_child, | |
753 | node, shadow_node); | |
754 | ||
8e519e3c MD |
755 | return 0; |
756 | } | |
757 | ||
758 | static | |
d96bfb0d | 759 | int ja_pool_node_set_nth(const struct cds_ja_type *type, |
b4540e8a | 760 | struct cds_ja_inode *node, |
b1a90ce3 | 761 | struct cds_ja_inode_flag *node_flag, |
d96bfb0d | 762 | struct cds_ja_shadow_node *shadow_node, |
8e519e3c | 763 | uint8_t n, |
b4540e8a | 764 | struct cds_ja_inode_flag *child_node_flag) |
8e519e3c | 765 | { |
b4540e8a | 766 | struct cds_ja_inode *linear; |
8e519e3c MD |
767 | |
768 | assert(type->type_class == RCU_JA_POOL); | |
b1a90ce3 MD |
769 | |
770 | switch (type->nr_pool_order) { | |
771 | case 1: | |
772 | { | |
773 | unsigned long bitsel, index; | |
774 | ||
775 | bitsel = ja_node_pool_1d_bitsel(node_flag); | |
776 | assert(bitsel < CHAR_BIT); | |
19ddcd04 | 777 | index = ((unsigned long) n >> bitsel) & 0x1; |
b1a90ce3 MD |
778 | linear = (struct cds_ja_inode *) &node->u.data[index << type->pool_size_order]; |
779 | break; | |
780 | } | |
781 | case 2: | |
782 | { | |
19ddcd04 MD |
783 | unsigned long bitsel[2], index[2], rindex; |
784 | ||
785 | ja_node_pool_2d_bitsel(node_flag, bitsel); | |
786 | assert(bitsel[0] < CHAR_BIT); | |
787 | assert(bitsel[1] < CHAR_BIT); | |
788 | index[0] = ((unsigned long) n >> bitsel[0]) & 0x1; | |
789 | index[0] <<= 1; | |
790 | index[1] = ((unsigned long) n >> bitsel[1]) & 0x1; | |
791 | rindex = index[0] | index[1]; | |
792 | linear = (struct cds_ja_inode *) &node->u.data[rindex << type->pool_size_order]; | |
b1a90ce3 MD |
793 | break; |
794 | } | |
795 | default: | |
796 | linear = NULL; | |
797 | assert(0); | |
798 | } | |
799 | ||
e1db2db5 MD |
800 | return ja_linear_node_set_nth(type, linear, shadow_node, |
801 | n, child_node_flag); | |
8e519e3c MD |
802 | } |
803 | ||
804 | static | |
d96bfb0d | 805 | int ja_pigeon_node_set_nth(const struct cds_ja_type *type, |
b4540e8a | 806 | struct cds_ja_inode *node, |
d96bfb0d | 807 | struct cds_ja_shadow_node *shadow_node, |
8e519e3c | 808 | uint8_t n, |
b4540e8a | 809 | struct cds_ja_inode_flag *child_node_flag) |
8e519e3c | 810 | { |
b4540e8a | 811 | struct cds_ja_inode_flag **ptr; |
8e519e3c MD |
812 | |
813 | assert(type->type_class == RCU_JA_PIGEON); | |
b4540e8a | 814 | ptr = &((struct cds_ja_inode_flag **) node->u.data)[n]; |
5a9a87dd | 815 | if (*ptr) |
8e519e3c MD |
816 | return -EEXIST; |
817 | rcu_assign_pointer(*ptr, child_node_flag); | |
e1db2db5 | 818 | shadow_node->nr_child++; |
8e519e3c MD |
819 | return 0; |
820 | } | |
821 | ||
d68c6810 | 822 | /* |
7a0b2331 | 823 | * _ja_node_set_nth: set nth item within a node. Return an error |
8e519e3c | 824 | * (negative error value) if it is already there. |
d68c6810 | 825 | */ |
8e519e3c | 826 | static |
d96bfb0d | 827 | int _ja_node_set_nth(const struct cds_ja_type *type, |
b4540e8a | 828 | struct cds_ja_inode *node, |
b1a90ce3 | 829 | struct cds_ja_inode_flag *node_flag, |
d96bfb0d | 830 | struct cds_ja_shadow_node *shadow_node, |
e1db2db5 | 831 | uint8_t n, |
b4540e8a | 832 | struct cds_ja_inode_flag *child_node_flag) |
8e519e3c | 833 | { |
8e519e3c MD |
834 | switch (type->type_class) { |
835 | case RCU_JA_LINEAR: | |
e1db2db5 | 836 | return ja_linear_node_set_nth(type, node, shadow_node, n, |
8e519e3c MD |
837 | child_node_flag); |
838 | case RCU_JA_POOL: | |
b1a90ce3 | 839 | return ja_pool_node_set_nth(type, node, node_flag, shadow_node, n, |
8e519e3c MD |
840 | child_node_flag); |
841 | case RCU_JA_PIGEON: | |
e1db2db5 | 842 | return ja_pigeon_node_set_nth(type, node, shadow_node, n, |
8e519e3c | 843 | child_node_flag); |
e1db2db5 MD |
844 | case RCU_JA_NULL: |
845 | return -ENOSPC; | |
8e519e3c MD |
846 | default: |
847 | assert(0); | |
848 | return -EINVAL; | |
849 | } | |
850 | ||
851 | return 0; | |
852 | } | |
7a0b2331 | 853 | |
2e313670 | 854 | static |
af3cbd45 | 855 | int ja_linear_node_clear_ptr(const struct cds_ja_type *type, |
2e313670 MD |
856 | struct cds_ja_inode *node, |
857 | struct cds_ja_shadow_node *shadow_node, | |
af3cbd45 | 858 | struct cds_ja_inode_flag **node_flag_ptr) |
2e313670 MD |
859 | { |
860 | uint8_t nr_child; | |
af3cbd45 | 861 | uint8_t *nr_child_ptr; |
2e313670 MD |
862 | |
863 | assert(type->type_class == RCU_JA_LINEAR || type->type_class == RCU_JA_POOL); | |
864 | ||
865 | nr_child_ptr = &node->u.data[0]; | |
2e313670 MD |
866 | nr_child = *nr_child_ptr; |
867 | assert(nr_child <= type->max_linear_child); | |
868 | ||
48cbe001 MD |
869 | if (type->type_class == RCU_JA_LINEAR) { |
870 | assert(!shadow_node->fallback_removal_count); | |
871 | if (shadow_node->nr_child <= type->min_child) { | |
2e313670 MD |
872 | /* We need to try recompacting the node */ |
873 | return -EFBIG; | |
874 | } | |
875 | } | |
19ddcd04 | 876 | dbg_printf("linear clear ptr: nr_child_ptr %p\n", nr_child_ptr); |
af3cbd45 MD |
877 | assert(*node_flag_ptr != NULL); |
878 | rcu_assign_pointer(*node_flag_ptr, NULL); | |
2e313670 MD |
879 | /* |
880 | * Value and nr_child are never changed (would cause ABA issue). | |
881 | * Instead, we leave the pointer to NULL and recompact the node | |
882 | * once in a while. It is allowed to set a NULL pointer to a new | |
883 | * value without recompaction though. | |
884 | * Only update the shadow node accounting. | |
885 | */ | |
886 | shadow_node->nr_child--; | |
af3cbd45 | 887 | dbg_printf("linear clear ptr: %u child, shadow: %u child, for node %p shadow %p\n", |
2e313670 MD |
888 | (unsigned int) CMM_LOAD_SHARED(*nr_child_ptr), |
889 | (unsigned int) shadow_node->nr_child, | |
890 | node, shadow_node); | |
2e313670 MD |
891 | return 0; |
892 | } | |
893 | ||
894 | static | |
af3cbd45 | 895 | int ja_pool_node_clear_ptr(const struct cds_ja_type *type, |
2e313670 | 896 | struct cds_ja_inode *node, |
19ddcd04 | 897 | struct cds_ja_inode_flag *node_flag, |
2e313670 | 898 | struct cds_ja_shadow_node *shadow_node, |
af3cbd45 | 899 | struct cds_ja_inode_flag **node_flag_ptr, |
2e313670 MD |
900 | uint8_t n) |
901 | { | |
902 | struct cds_ja_inode *linear; | |
903 | ||
904 | assert(type->type_class == RCU_JA_POOL); | |
19ddcd04 MD |
905 | |
906 | if (shadow_node->fallback_removal_count) { | |
907 | shadow_node->fallback_removal_count--; | |
908 | } else { | |
909 | /* We should try recompacting the node */ | |
910 | if (shadow_node->nr_child <= type->min_child) | |
911 | return -EFBIG; | |
912 | } | |
913 | ||
914 | switch (type->nr_pool_order) { | |
915 | case 1: | |
916 | { | |
917 | unsigned long bitsel, index; | |
918 | ||
919 | bitsel = ja_node_pool_1d_bitsel(node_flag); | |
920 | assert(bitsel < CHAR_BIT); | |
921 | index = ((unsigned long) n >> bitsel) & type->nr_pool_order; | |
922 | linear = (struct cds_ja_inode *) &node->u.data[index << type->pool_size_order]; | |
923 | break; | |
924 | } | |
925 | case 2: | |
926 | { | |
927 | unsigned long bitsel[2], index[2], rindex; | |
928 | ||
929 | ja_node_pool_2d_bitsel(node_flag, bitsel); | |
930 | assert(bitsel[0] < CHAR_BIT); | |
931 | assert(bitsel[1] < CHAR_BIT); | |
932 | index[0] = ((unsigned long) n >> bitsel[0]) & 0x1; | |
933 | index[0] <<= 1; | |
934 | index[1] = ((unsigned long) n >> bitsel[1]) & 0x1; | |
935 | rindex = index[0] | index[1]; | |
936 | linear = (struct cds_ja_inode *) &node->u.data[rindex << type->pool_size_order]; | |
937 | break; | |
938 | } | |
939 | default: | |
940 | linear = NULL; | |
941 | assert(0); | |
942 | } | |
943 | ||
af3cbd45 | 944 | return ja_linear_node_clear_ptr(type, linear, shadow_node, node_flag_ptr); |
2e313670 MD |
945 | } |
946 | ||
947 | static | |
af3cbd45 | 948 | int ja_pigeon_node_clear_ptr(const struct cds_ja_type *type, |
2e313670 MD |
949 | struct cds_ja_inode *node, |
950 | struct cds_ja_shadow_node *shadow_node, | |
af3cbd45 | 951 | struct cds_ja_inode_flag **node_flag_ptr) |
2e313670 | 952 | { |
2e313670 | 953 | assert(type->type_class == RCU_JA_PIGEON); |
19ddcd04 MD |
954 | |
955 | if (shadow_node->fallback_removal_count) { | |
956 | shadow_node->fallback_removal_count--; | |
957 | } else { | |
958 | /* We should try recompacting the node */ | |
959 | if (shadow_node->nr_child <= type->min_child) | |
960 | return -EFBIG; | |
961 | } | |
4d6ef45e | 962 | dbg_printf("ja_pigeon_node_clear_ptr: clearing ptr: %p\n", *node_flag_ptr); |
af3cbd45 | 963 | rcu_assign_pointer(*node_flag_ptr, NULL); |
2e313670 MD |
964 | shadow_node->nr_child--; |
965 | return 0; | |
966 | } | |
967 | ||
968 | /* | |
af3cbd45 | 969 | * _ja_node_clear_ptr: clear ptr item within a node. Return an error |
2e313670 MD |
970 | * (negative error value) if it is not found (-ENOENT). |
971 | */ | |
972 | static | |
af3cbd45 | 973 | int _ja_node_clear_ptr(const struct cds_ja_type *type, |
2e313670 | 974 | struct cds_ja_inode *node, |
19ddcd04 | 975 | struct cds_ja_inode_flag *node_flag, |
2e313670 | 976 | struct cds_ja_shadow_node *shadow_node, |
af3cbd45 | 977 | struct cds_ja_inode_flag **node_flag_ptr, |
2e313670 MD |
978 | uint8_t n) |
979 | { | |
980 | switch (type->type_class) { | |
981 | case RCU_JA_LINEAR: | |
af3cbd45 | 982 | return ja_linear_node_clear_ptr(type, node, shadow_node, node_flag_ptr); |
2e313670 | 983 | case RCU_JA_POOL: |
19ddcd04 | 984 | return ja_pool_node_clear_ptr(type, node, node_flag, shadow_node, node_flag_ptr, n); |
2e313670 | 985 | case RCU_JA_PIGEON: |
af3cbd45 | 986 | return ja_pigeon_node_clear_ptr(type, node, shadow_node, node_flag_ptr); |
2e313670 MD |
987 | case RCU_JA_NULL: |
988 | return -ENOENT; | |
989 | default: | |
990 | assert(0); | |
991 | return -EINVAL; | |
992 | } | |
993 | ||
994 | return 0; | |
995 | } | |
996 | ||
b1a90ce3 MD |
997 | /* |
998 | * Calculate bit distribution. Returns the bit (0 to 7) that splits the | |
999 | * distribution in two sub-distributions containing as much elements one | |
1000 | * compared to the other. | |
1001 | */ | |
1002 | static | |
1003 | unsigned int ja_node_sum_distribution_1d(enum ja_recompact mode, | |
1004 | struct cds_ja *ja, | |
1005 | unsigned int type_index, | |
1006 | const struct cds_ja_type *type, | |
1007 | struct cds_ja_inode *node, | |
1008 | struct cds_ja_shadow_node *shadow_node, | |
1009 | uint8_t n, | |
1010 | struct cds_ja_inode_flag *child_node_flag, | |
1011 | struct cds_ja_inode_flag **nullify_node_flag_ptr) | |
1012 | { | |
1013 | uint8_t nr_one[JA_BITS_PER_BYTE]; | |
1014 | unsigned int bitsel = 0, bit_i, overall_best_distance = UINT_MAX; | |
1015 | unsigned int distrib_nr_child = 0; | |
1016 | ||
1017 | memset(nr_one, 0, sizeof(nr_one)); | |
1018 | ||
1019 | switch (type->type_class) { | |
1020 | case RCU_JA_LINEAR: | |
1021 | { | |
1022 | uint8_t nr_child = | |
1023 | ja_linear_node_get_nr_child(type, node); | |
1024 | unsigned int i; | |
1025 | ||
1026 | for (i = 0; i < nr_child; i++) { | |
1027 | struct cds_ja_inode_flag *iter; | |
b1a90ce3 MD |
1028 | uint8_t v; |
1029 | ||
1030 | ja_linear_node_get_ith_pos(type, node, i, &v, &iter); | |
1031 | if (!iter) | |
1032 | continue; | |
1033 | if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) | |
1034 | continue; | |
f5531dd9 MD |
1035 | for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) { |
1036 | if (v & (1U << bit_i)) | |
1037 | nr_one[bit_i]++; | |
b1a90ce3 MD |
1038 | } |
1039 | distrib_nr_child++; | |
1040 | } | |
1041 | break; | |
1042 | } | |
1043 | case RCU_JA_POOL: | |
1044 | { | |
1045 | unsigned int pool_nr; | |
1046 | ||
1047 | for (pool_nr = 0; pool_nr < (1U << type->nr_pool_order); pool_nr++) { | |
1048 | struct cds_ja_inode *pool = | |
1049 | ja_pool_node_get_ith_pool(type, | |
1050 | node, pool_nr); | |
1051 | uint8_t nr_child = | |
1052 | ja_linear_node_get_nr_child(type, pool); | |
1053 | unsigned int j; | |
1054 | ||
1055 | for (j = 0; j < nr_child; j++) { | |
1056 | struct cds_ja_inode_flag *iter; | |
b1a90ce3 MD |
1057 | uint8_t v; |
1058 | ||
1059 | ja_linear_node_get_ith_pos(type, pool, | |
1060 | j, &v, &iter); | |
1061 | if (!iter) | |
1062 | continue; | |
1063 | if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) | |
1064 | continue; | |
f5531dd9 MD |
1065 | for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) { |
1066 | if (v & (1U << bit_i)) | |
1067 | nr_one[bit_i]++; | |
b1a90ce3 MD |
1068 | } |
1069 | distrib_nr_child++; | |
1070 | } | |
1071 | } | |
1072 | break; | |
1073 | } | |
1074 | case RCU_JA_PIGEON: | |
1075 | { | |
b1a90ce3 MD |
1076 | unsigned int i; |
1077 | ||
1078 | assert(mode == JA_RECOMPACT_DEL); | |
48cbe001 | 1079 | for (i = 0; i < JA_ENTRY_PER_NODE; i++) { |
b1a90ce3 | 1080 | struct cds_ja_inode_flag *iter; |
b1a90ce3 MD |
1081 | |
1082 | iter = ja_pigeon_node_get_ith_pos(type, node, i); | |
1083 | if (!iter) | |
1084 | continue; | |
1085 | if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) | |
1086 | continue; | |
f5531dd9 MD |
1087 | for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) { |
1088 | if (i & (1U << bit_i)) | |
1089 | nr_one[bit_i]++; | |
b1a90ce3 MD |
1090 | } |
1091 | distrib_nr_child++; | |
1092 | } | |
1093 | break; | |
1094 | } | |
1095 | case RCU_JA_NULL: | |
19ddcd04 | 1096 | assert(mode == JA_RECOMPACT_ADD_NEXT); |
b1a90ce3 MD |
1097 | break; |
1098 | default: | |
1099 | assert(0); | |
1100 | break; | |
1101 | } | |
1102 | ||
19ddcd04 | 1103 | if (mode == JA_RECOMPACT_ADD_NEXT || mode == JA_RECOMPACT_ADD_SAME) { |
f5531dd9 MD |
1104 | for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) { |
1105 | if (n & (1U << bit_i)) | |
1106 | nr_one[bit_i]++; | |
b1a90ce3 MD |
1107 | } |
1108 | distrib_nr_child++; | |
1109 | } | |
1110 | ||
1111 | /* | |
1112 | * The best bit selector is that for which the number of ones is | |
1113 | * closest to half of the number of children in the | |
f5531dd9 MD |
1114 | * distribution. We calculate the distance using the double of |
1115 | * the sub-distribution sizes to eliminate truncation error. | |
b1a90ce3 MD |
1116 | */ |
1117 | for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) { | |
1118 | unsigned int distance_to_best; | |
1119 | ||
1b34283b | 1120 | distance_to_best = abs_int(((unsigned int) nr_one[bit_i] << 1U) - distrib_nr_child); |
b1a90ce3 MD |
1121 | if (distance_to_best < overall_best_distance) { |
1122 | overall_best_distance = distance_to_best; | |
1123 | bitsel = bit_i; | |
1124 | } | |
1125 | } | |
1126 | dbg_printf("1 dimension pool bit selection: (%u)\n", bitsel); | |
1127 | return bitsel; | |
1128 | } | |
1129 | ||
19ddcd04 MD |
1130 | /* |
1131 | * Calculate bit distribution in two dimensions. Returns the two bits | |
1132 | * (each 0 to 7) that splits the distribution in four sub-distributions | |
1133 | * containing as much elements one compared to the other. | |
1134 | */ | |
1135 | static | |
1136 | void ja_node_sum_distribution_2d(enum ja_recompact mode, | |
1137 | struct cds_ja *ja, | |
1138 | unsigned int type_index, | |
1139 | const struct cds_ja_type *type, | |
1140 | struct cds_ja_inode *node, | |
1141 | struct cds_ja_shadow_node *shadow_node, | |
1142 | uint8_t n, | |
1143 | struct cds_ja_inode_flag *child_node_flag, | |
1144 | struct cds_ja_inode_flag **nullify_node_flag_ptr, | |
1145 | unsigned int *_bitsel) | |
1146 | { | |
1147 | uint8_t nr_2d_11[JA_BITS_PER_BYTE][JA_BITS_PER_BYTE], | |
1148 | nr_2d_10[JA_BITS_PER_BYTE][JA_BITS_PER_BYTE], | |
1149 | nr_2d_01[JA_BITS_PER_BYTE][JA_BITS_PER_BYTE], | |
1150 | nr_2d_00[JA_BITS_PER_BYTE][JA_BITS_PER_BYTE]; | |
1151 | unsigned int bitsel[2] = { 0, 1 }; | |
4a073c53 MD |
1152 | unsigned int bit_i, bit_j; |
1153 | int overall_best_distance = INT_MAX; | |
19ddcd04 MD |
1154 | unsigned int distrib_nr_child = 0; |
1155 | ||
1156 | memset(nr_2d_11, 0, sizeof(nr_2d_11)); | |
1157 | memset(nr_2d_10, 0, sizeof(nr_2d_10)); | |
4a073c53 MD |
1158 | memset(nr_2d_01, 0, sizeof(nr_2d_01)); |
1159 | memset(nr_2d_00, 0, sizeof(nr_2d_00)); | |
19ddcd04 MD |
1160 | |
1161 | switch (type->type_class) { | |
1162 | case RCU_JA_LINEAR: | |
1163 | { | |
1164 | uint8_t nr_child = | |
1165 | ja_linear_node_get_nr_child(type, node); | |
1166 | unsigned int i; | |
1167 | ||
1168 | for (i = 0; i < nr_child; i++) { | |
1169 | struct cds_ja_inode_flag *iter; | |
1170 | uint8_t v; | |
1171 | ||
1172 | ja_linear_node_get_ith_pos(type, node, i, &v, &iter); | |
1173 | if (!iter) | |
1174 | continue; | |
1175 | if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) | |
1176 | continue; | |
1177 | for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) { | |
1178 | for (bit_j = 0; bit_j < bit_i; bit_j++) { | |
7f14b43a MD |
1179 | if (v & (1U << bit_i)) { |
1180 | if (v & (1U << bit_j)) { | |
1181 | nr_2d_11[bit_i][bit_j]++; | |
1182 | } else { | |
1183 | nr_2d_10[bit_i][bit_j]++; | |
1184 | } | |
1185 | } else { | |
1186 | if (v & (1U << bit_j)) { | |
1187 | nr_2d_01[bit_i][bit_j]++; | |
1188 | } else { | |
1189 | nr_2d_00[bit_i][bit_j]++; | |
1190 | } | |
19ddcd04 MD |
1191 | } |
1192 | } | |
1193 | } | |
1194 | distrib_nr_child++; | |
1195 | } | |
1196 | break; | |
1197 | } | |
1198 | case RCU_JA_POOL: | |
1199 | { | |
1200 | unsigned int pool_nr; | |
1201 | ||
1202 | for (pool_nr = 0; pool_nr < (1U << type->nr_pool_order); pool_nr++) { | |
1203 | struct cds_ja_inode *pool = | |
1204 | ja_pool_node_get_ith_pool(type, | |
1205 | node, pool_nr); | |
1206 | uint8_t nr_child = | |
1207 | ja_linear_node_get_nr_child(type, pool); | |
1208 | unsigned int j; | |
1209 | ||
1210 | for (j = 0; j < nr_child; j++) { | |
1211 | struct cds_ja_inode_flag *iter; | |
1212 | uint8_t v; | |
1213 | ||
1214 | ja_linear_node_get_ith_pos(type, pool, | |
1215 | j, &v, &iter); | |
1216 | if (!iter) | |
1217 | continue; | |
1218 | if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) | |
1219 | continue; | |
1220 | for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) { | |
1221 | for (bit_j = 0; bit_j < bit_i; bit_j++) { | |
7f14b43a MD |
1222 | if (v & (1U << bit_i)) { |
1223 | if (v & (1U << bit_j)) { | |
1224 | nr_2d_11[bit_i][bit_j]++; | |
1225 | } else { | |
1226 | nr_2d_10[bit_i][bit_j]++; | |
1227 | } | |
1228 | } else { | |
1229 | if (v & (1U << bit_j)) { | |
1230 | nr_2d_01[bit_i][bit_j]++; | |
1231 | } else { | |
1232 | nr_2d_00[bit_i][bit_j]++; | |
1233 | } | |
19ddcd04 MD |
1234 | } |
1235 | } | |
1236 | } | |
1237 | distrib_nr_child++; | |
1238 | } | |
1239 | } | |
1240 | break; | |
1241 | } | |
1242 | case RCU_JA_PIGEON: | |
1243 | { | |
19ddcd04 MD |
1244 | unsigned int i; |
1245 | ||
1246 | assert(mode == JA_RECOMPACT_DEL); | |
48cbe001 | 1247 | for (i = 0; i < JA_ENTRY_PER_NODE; i++) { |
19ddcd04 MD |
1248 | struct cds_ja_inode_flag *iter; |
1249 | ||
1250 | iter = ja_pigeon_node_get_ith_pos(type, node, i); | |
1251 | if (!iter) | |
1252 | continue; | |
1253 | if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) | |
1254 | continue; | |
1255 | for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) { | |
1256 | for (bit_j = 0; bit_j < bit_i; bit_j++) { | |
7f14b43a MD |
1257 | if (i & (1U << bit_i)) { |
1258 | if (i & (1U << bit_j)) { | |
1259 | nr_2d_11[bit_i][bit_j]++; | |
1260 | } else { | |
1261 | nr_2d_10[bit_i][bit_j]++; | |
1262 | } | |
1263 | } else { | |
1264 | if (i & (1U << bit_j)) { | |
1265 | nr_2d_01[bit_i][bit_j]++; | |
1266 | } else { | |
1267 | nr_2d_00[bit_i][bit_j]++; | |
1268 | } | |
19ddcd04 MD |
1269 | } |
1270 | } | |
1271 | } | |
1272 | distrib_nr_child++; | |
1273 | } | |
1274 | break; | |
1275 | } | |
1276 | case RCU_JA_NULL: | |
1277 | assert(mode == JA_RECOMPACT_ADD_NEXT); | |
1278 | break; | |
1279 | default: | |
1280 | assert(0); | |
1281 | break; | |
1282 | } | |
1283 | ||
1284 | if (mode == JA_RECOMPACT_ADD_NEXT || mode == JA_RECOMPACT_ADD_SAME) { | |
1285 | for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) { | |
1286 | for (bit_j = 0; bit_j < bit_i; bit_j++) { | |
7f14b43a MD |
1287 | if (n & (1U << bit_i)) { |
1288 | if (n & (1U << bit_j)) { | |
1289 | nr_2d_11[bit_i][bit_j]++; | |
1290 | } else { | |
1291 | nr_2d_10[bit_i][bit_j]++; | |
1292 | } | |
1293 | } else { | |
1294 | if (n & (1U << bit_j)) { | |
1295 | nr_2d_01[bit_i][bit_j]++; | |
1296 | } else { | |
1297 | nr_2d_00[bit_i][bit_j]++; | |
1298 | } | |
19ddcd04 MD |
1299 | } |
1300 | } | |
1301 | } | |
1302 | distrib_nr_child++; | |
1303 | } | |
1304 | ||
1305 | /* | |
1306 | * The best bit selector is that for which the number of nodes | |
1307 | * in each sub-class is closest to one-fourth of the number of | |
1308 | * children in the distribution. We calculate the distance using | |
1309 | * 4 times the size of the sub-distribution to eliminate | |
1310 | * truncation error. | |
1311 | */ | |
1312 | for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) { | |
1313 | for (bit_j = 0; bit_j < bit_i; bit_j++) { | |
4a073c53 | 1314 | int distance_to_best[4]; |
19ddcd04 | 1315 | |
1b34283b MD |
1316 | distance_to_best[0] = ((unsigned int) nr_2d_11[bit_i][bit_j] << 2U) - distrib_nr_child; |
1317 | distance_to_best[1] = ((unsigned int) nr_2d_10[bit_i][bit_j] << 2U) - distrib_nr_child; | |
1318 | distance_to_best[2] = ((unsigned int) nr_2d_01[bit_i][bit_j] << 2U) - distrib_nr_child; | |
1319 | distance_to_best[3] = ((unsigned int) nr_2d_00[bit_i][bit_j] << 2U) - distrib_nr_child; | |
19ddcd04 | 1320 | |
4a073c53 MD |
1321 | /* Consider worse distance above best */ |
1322 | if (distance_to_best[1] > 0 && distance_to_best[1] > distance_to_best[0]) | |
19ddcd04 | 1323 | distance_to_best[0] = distance_to_best[1]; |
4a073c53 | 1324 | if (distance_to_best[2] > 0 && distance_to_best[2] > distance_to_best[0]) |
19ddcd04 | 1325 | distance_to_best[0] = distance_to_best[2]; |
4a073c53 | 1326 | if (distance_to_best[3] > 0 && distance_to_best[3] > distance_to_best[0]) |
19ddcd04 | 1327 | distance_to_best[0] = distance_to_best[3]; |
4a073c53 | 1328 | |
19ddcd04 MD |
1329 | /* |
1330 | * If our worse distance is better than overall, | |
1331 | * we become new best candidate. | |
1332 | */ | |
1333 | if (distance_to_best[0] < overall_best_distance) { | |
1334 | overall_best_distance = distance_to_best[0]; | |
1335 | bitsel[0] = bit_i; | |
1336 | bitsel[1] = bit_j; | |
1337 | } | |
1338 | } | |
1339 | } | |
1340 | ||
1341 | dbg_printf("2 dimensions pool bit selection: (%u,%u)\n", bitsel[0], bitsel[1]); | |
1342 | ||
1343 | /* Return our bit selection */ | |
1344 | _bitsel[0] = bitsel[0]; | |
1345 | _bitsel[1] = bitsel[1]; | |
1346 | } | |
1347 | ||
48cbe001 MD |
1348 | static |
1349 | unsigned int find_nearest_type_index(unsigned int type_index, | |
1350 | unsigned int nr_nodes) | |
1351 | { | |
1352 | const struct cds_ja_type *type; | |
1353 | ||
1354 | assert(type_index != NODE_INDEX_NULL); | |
1355 | if (nr_nodes == 0) | |
1356 | return NODE_INDEX_NULL; | |
1357 | for (;;) { | |
1358 | type = &ja_types[type_index]; | |
1359 | if (nr_nodes < type->min_child) | |
1360 | type_index--; | |
1361 | else if (nr_nodes > type->max_child) | |
1362 | type_index++; | |
1363 | else | |
1364 | break; | |
1365 | } | |
1366 | return type_index; | |
1367 | } | |
1368 | ||
7a0b2331 MD |
1369 | /* |
1370 | * ja_node_recompact_add: recompact a node, adding a new child. | |
2e313670 | 1371 | * Return 0 on success, -EAGAIN if need to retry, or other negative |
5a9a87dd | 1372 | * error value otherwise. |
7a0b2331 MD |
1373 | */ |
1374 | static | |
2e313670 MD |
1375 | int ja_node_recompact(enum ja_recompact mode, |
1376 | struct cds_ja *ja, | |
e1db2db5 | 1377 | unsigned int old_type_index, |
d96bfb0d | 1378 | const struct cds_ja_type *old_type, |
b4540e8a | 1379 | struct cds_ja_inode *old_node, |
5a9a87dd | 1380 | struct cds_ja_shadow_node *shadow_node, |
3d8fe307 | 1381 | struct cds_ja_inode_flag **old_node_flag_ptr, uint8_t n, |
af3cbd45 | 1382 | struct cds_ja_inode_flag *child_node_flag, |
48cbe001 MD |
1383 | struct cds_ja_inode_flag **nullify_node_flag_ptr, |
1384 | int level) | |
7a0b2331 | 1385 | { |
e1db2db5 | 1386 | unsigned int new_type_index; |
b4540e8a | 1387 | struct cds_ja_inode *new_node; |
af3cbd45 | 1388 | struct cds_ja_shadow_node *new_shadow_node = NULL; |
d96bfb0d | 1389 | const struct cds_ja_type *new_type; |
3d8fe307 | 1390 | struct cds_ja_inode_flag *new_node_flag, *old_node_flag; |
7a0b2331 | 1391 | int ret; |
f07b240f | 1392 | int fallback = 0; |
7a0b2331 | 1393 | |
3d8fe307 MD |
1394 | old_node_flag = *old_node_flag_ptr; |
1395 | ||
48cbe001 MD |
1396 | /* |
1397 | * Need to find nearest type index even for ADD_SAME, because | |
1398 | * this recompaction, when applied to linear nodes, will garbage | |
1399 | * collect dummy (NULL) entries, and can therefore cause a few | |
1400 | * linear representations to be skipped. | |
1401 | */ | |
2e313670 | 1402 | switch (mode) { |
19ddcd04 | 1403 | case JA_RECOMPACT_ADD_SAME: |
48cbe001 MD |
1404 | new_type_index = find_nearest_type_index(old_type_index, |
1405 | shadow_node->nr_child + 1); | |
1406 | dbg_printf("Recompact for node with %u children\n", | |
1407 | shadow_node->nr_child + 1); | |
2e313670 | 1408 | break; |
19ddcd04 | 1409 | case JA_RECOMPACT_ADD_NEXT: |
2e313670 MD |
1410 | if (!shadow_node || old_type_index == NODE_INDEX_NULL) { |
1411 | new_type_index = 0; | |
48cbe001 | 1412 | dbg_printf("Recompact for NULL\n"); |
2e313670 | 1413 | } else { |
48cbe001 MD |
1414 | new_type_index = find_nearest_type_index(old_type_index, |
1415 | shadow_node->nr_child + 1); | |
1416 | dbg_printf("Recompact for node with %u children\n", | |
1417 | shadow_node->nr_child + 1); | |
2e313670 MD |
1418 | } |
1419 | break; | |
1420 | case JA_RECOMPACT_DEL: | |
48cbe001 MD |
1421 | new_type_index = find_nearest_type_index(old_type_index, |
1422 | shadow_node->nr_child - 1); | |
1423 | dbg_printf("Recompact for node with %u children\n", | |
1424 | shadow_node->nr_child - 1); | |
2e313670 MD |
1425 | break; |
1426 | default: | |
1427 | assert(0); | |
7a0b2331 | 1428 | } |
a2a7ff59 | 1429 | |
f07b240f | 1430 | retry: /* for fallback */ |
582a6ade MD |
1431 | dbg_printf("Recompact from type %d to type %d\n", |
1432 | old_type_index, new_type_index); | |
7a0b2331 | 1433 | new_type = &ja_types[new_type_index]; |
2e313670 | 1434 | if (new_type_index != NODE_INDEX_NULL) { |
354981c2 | 1435 | new_node = alloc_cds_ja_node(ja, new_type); |
2e313670 MD |
1436 | if (!new_node) |
1437 | return -ENOMEM; | |
b1a90ce3 MD |
1438 | |
1439 | if (new_type->type_class == RCU_JA_POOL) { | |
1440 | switch (new_type->nr_pool_order) { | |
1441 | case 1: | |
1442 | { | |
19ddcd04 MD |
1443 | unsigned int node_distrib_bitsel; |
1444 | ||
b1a90ce3 MD |
1445 | node_distrib_bitsel = |
1446 | ja_node_sum_distribution_1d(mode, ja, | |
1447 | old_type_index, old_type, | |
1448 | old_node, shadow_node, | |
1449 | n, child_node_flag, | |
1450 | nullify_node_flag_ptr); | |
1451 | assert(!((unsigned long) new_node & JA_POOL_1D_MASK)); | |
1452 | new_node_flag = ja_node_flag_pool_1d(new_node, | |
1453 | new_type_index, node_distrib_bitsel); | |
1454 | break; | |
1455 | } | |
1456 | case 2: | |
1457 | { | |
19ddcd04 MD |
1458 | unsigned int node_distrib_bitsel[2]; |
1459 | ||
1460 | ja_node_sum_distribution_2d(mode, ja, | |
1461 | old_type_index, old_type, | |
1462 | old_node, shadow_node, | |
1463 | n, child_node_flag, | |
1464 | nullify_node_flag_ptr, | |
1465 | node_distrib_bitsel); | |
b1a90ce3 MD |
1466 | assert(!((unsigned long) new_node & JA_POOL_1D_MASK)); |
1467 | assert(!((unsigned long) new_node & JA_POOL_2D_MASK)); | |
19ddcd04 MD |
1468 | new_node_flag = ja_node_flag_pool_2d(new_node, |
1469 | new_type_index, node_distrib_bitsel); | |
b1a90ce3 MD |
1470 | break; |
1471 | } | |
1472 | default: | |
1473 | assert(0); | |
1474 | } | |
1475 | } else { | |
1476 | new_node_flag = ja_node_flag(new_node, new_type_index); | |
1477 | } | |
1478 | ||
2e313670 | 1479 | dbg_printf("Recompact inherit lock from %p\n", shadow_node); |
48cbe001 | 1480 | new_shadow_node = rcuja_shadow_set(ja->ht, new_node_flag, shadow_node, ja, level); |
2e313670 | 1481 | if (!new_shadow_node) { |
354981c2 | 1482 | free_cds_ja_node(ja, new_node); |
2e313670 MD |
1483 | return -ENOMEM; |
1484 | } | |
1485 | if (fallback) | |
1486 | new_shadow_node->fallback_removal_count = | |
1487 | JA_FALLBACK_REMOVAL_COUNT; | |
1488 | } else { | |
1489 | new_node = NULL; | |
1490 | new_node_flag = NULL; | |
e1db2db5 | 1491 | } |
11c5e016 | 1492 | |
19ddcd04 | 1493 | assert(mode != JA_RECOMPACT_ADD_NEXT || old_type->type_class != RCU_JA_PIGEON); |
2e313670 MD |
1494 | |
1495 | if (new_type_index == NODE_INDEX_NULL) | |
1496 | goto skip_copy; | |
1497 | ||
11c5e016 MD |
1498 | switch (old_type->type_class) { |
1499 | case RCU_JA_LINEAR: | |
1500 | { | |
1501 | uint8_t nr_child = | |
1502 | ja_linear_node_get_nr_child(old_type, old_node); | |
1503 | unsigned int i; | |
1504 | ||
1505 | for (i = 0; i < nr_child; i++) { | |
b4540e8a | 1506 | struct cds_ja_inode_flag *iter; |
11c5e016 MD |
1507 | uint8_t v; |
1508 | ||
1509 | ja_linear_node_get_ith_pos(old_type, old_node, i, &v, &iter); | |
1510 | if (!iter) | |
1511 | continue; | |
af3cbd45 | 1512 | if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) |
2e313670 | 1513 | continue; |
b1a90ce3 | 1514 | ret = _ja_node_set_nth(new_type, new_node, new_node_flag, |
f07b240f | 1515 | new_shadow_node, |
11c5e016 | 1516 | v, iter); |
f07b240f MD |
1517 | if (new_type->type_class == RCU_JA_POOL && ret) { |
1518 | goto fallback_toosmall; | |
1519 | } | |
11c5e016 MD |
1520 | assert(!ret); |
1521 | } | |
1522 | break; | |
1523 | } | |
1524 | case RCU_JA_POOL: | |
1525 | { | |
1526 | unsigned int pool_nr; | |
1527 | ||
1528 | for (pool_nr = 0; pool_nr < (1U << old_type->nr_pool_order); pool_nr++) { | |
b4540e8a | 1529 | struct cds_ja_inode *pool = |
11c5e016 MD |
1530 | ja_pool_node_get_ith_pool(old_type, |
1531 | old_node, pool_nr); | |
1532 | uint8_t nr_child = | |
1533 | ja_linear_node_get_nr_child(old_type, pool); | |
1534 | unsigned int j; | |
1535 | ||
1536 | for (j = 0; j < nr_child; j++) { | |
b4540e8a | 1537 | struct cds_ja_inode_flag *iter; |
11c5e016 MD |
1538 | uint8_t v; |
1539 | ||
1540 | ja_linear_node_get_ith_pos(old_type, pool, | |
1541 | j, &v, &iter); | |
1542 | if (!iter) | |
1543 | continue; | |
af3cbd45 | 1544 | if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) |
2e313670 | 1545 | continue; |
b1a90ce3 | 1546 | ret = _ja_node_set_nth(new_type, new_node, new_node_flag, |
f07b240f | 1547 | new_shadow_node, |
11c5e016 | 1548 | v, iter); |
f07b240f MD |
1549 | if (new_type->type_class == RCU_JA_POOL |
1550 | && ret) { | |
1551 | goto fallback_toosmall; | |
1552 | } | |
11c5e016 MD |
1553 | assert(!ret); |
1554 | } | |
1555 | } | |
1556 | break; | |
7a0b2331 | 1557 | } |
a2a7ff59 | 1558 | case RCU_JA_NULL: |
19ddcd04 | 1559 | assert(mode == JA_RECOMPACT_ADD_NEXT); |
a2a7ff59 | 1560 | break; |
11c5e016 | 1561 | case RCU_JA_PIGEON: |
2e313670 | 1562 | { |
2e313670 MD |
1563 | unsigned int i; |
1564 | ||
1565 | assert(mode == JA_RECOMPACT_DEL); | |
48cbe001 | 1566 | for (i = 0; i < JA_ENTRY_PER_NODE; i++) { |
2e313670 MD |
1567 | struct cds_ja_inode_flag *iter; |
1568 | ||
1569 | iter = ja_pigeon_node_get_ith_pos(old_type, old_node, i); | |
1570 | if (!iter) | |
1571 | continue; | |
af3cbd45 | 1572 | if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) |
2e313670 | 1573 | continue; |
b1a90ce3 | 1574 | ret = _ja_node_set_nth(new_type, new_node, new_node_flag, |
2e313670 MD |
1575 | new_shadow_node, |
1576 | i, iter); | |
1577 | if (new_type->type_class == RCU_JA_POOL && ret) { | |
1578 | goto fallback_toosmall; | |
1579 | } | |
1580 | assert(!ret); | |
1581 | } | |
1582 | break; | |
1583 | } | |
11c5e016 MD |
1584 | default: |
1585 | assert(0); | |
5a9a87dd | 1586 | ret = -EINVAL; |
f07b240f | 1587 | goto end; |
11c5e016 | 1588 | } |
2e313670 | 1589 | skip_copy: |
11c5e016 | 1590 | |
19ddcd04 | 1591 | if (mode == JA_RECOMPACT_ADD_NEXT || mode == JA_RECOMPACT_ADD_SAME) { |
2e313670 | 1592 | /* add node */ |
b1a90ce3 | 1593 | ret = _ja_node_set_nth(new_type, new_node, new_node_flag, |
2e313670 MD |
1594 | new_shadow_node, |
1595 | n, child_node_flag); | |
7b413155 MD |
1596 | if (new_type->type_class == RCU_JA_POOL && ret) { |
1597 | goto fallback_toosmall; | |
1598 | } | |
2e313670 MD |
1599 | assert(!ret); |
1600 | } | |
19ddcd04 MD |
1601 | |
1602 | if (fallback) { | |
1603 | dbg_printf("Using fallback for %u children, node type index: %u, mode %s\n", | |
1604 | new_shadow_node->nr_child, old_type_index, mode == JA_RECOMPACT_ADD_NEXT ? "add_next" : | |
1605 | (mode == JA_RECOMPACT_DEL ? "del" : "add_same")); | |
354981c2 | 1606 | uatomic_inc(&ja->node_fallback_count_distribution[new_shadow_node->nr_child]); |
19ddcd04 MD |
1607 | } |
1608 | ||
3d8fe307 MD |
1609 | /* Return pointer to new recompacted node through old_node_flag_ptr */ |
1610 | *old_node_flag_ptr = new_node_flag; | |
a2a7ff59 | 1611 | if (old_node) { |
2e313670 MD |
1612 | int flags; |
1613 | ||
1614 | flags = RCUJA_SHADOW_CLEAR_FREE_NODE; | |
1615 | /* | |
1616 | * It is OK to free the lock associated with a node | |
1617 | * going to NULL, since we are holding the parent lock. | |
1618 | * This synchronizes removal with re-add of that node. | |
1619 | */ | |
1620 | if (new_type_index == NODE_INDEX_NULL) | |
48cbe001 | 1621 | flags |= RCUJA_SHADOW_CLEAR_FREE_LOCK; |
3d8fe307 | 1622 | ret = rcuja_shadow_clear(ja->ht, old_node_flag, shadow_node, |
2e313670 | 1623 | flags); |
a2a7ff59 MD |
1624 | assert(!ret); |
1625 | } | |
5a9a87dd MD |
1626 | |
1627 | ret = 0; | |
f07b240f | 1628 | end: |
5a9a87dd | 1629 | return ret; |
f07b240f MD |
1630 | |
1631 | fallback_toosmall: | |
1632 | /* fallback if next pool is too small */ | |
af3cbd45 | 1633 | assert(new_shadow_node); |
3d8fe307 | 1634 | ret = rcuja_shadow_clear(ja->ht, new_node_flag, new_shadow_node, |
f07b240f MD |
1635 | RCUJA_SHADOW_CLEAR_FREE_NODE); |
1636 | assert(!ret); | |
1637 | ||
19ddcd04 MD |
1638 | switch (mode) { |
1639 | case JA_RECOMPACT_ADD_SAME: | |
1640 | /* | |
1641 | * JA_RECOMPACT_ADD_SAME is only triggered if a linear | |
1642 | * node within a pool has unused entries. It should | |
1643 | * therefore _never_ be too small. | |
1644 | */ | |
4a073c53 | 1645 | assert(0); |
4cde8267 MD |
1646 | |
1647 | /* Fall-through */ | |
19ddcd04 MD |
1648 | case JA_RECOMPACT_ADD_NEXT: |
1649 | { | |
1650 | const struct cds_ja_type *next_type; | |
1651 | ||
1652 | /* | |
1653 | * Recompaction attempt on add failed. Should only | |
1654 | * happen if target node type is pool. Caused by | |
1655 | * hard-to-split distribution. Recompact using the next | |
1656 | * distribution size. | |
1657 | */ | |
1658 | assert(new_type->type_class == RCU_JA_POOL); | |
1659 | next_type = &ja_types[new_type_index + 1]; | |
1660 | /* | |
1661 | * Try going to the next pool size if our population | |
1662 | * fits within its range. This is not flagged as a | |
1663 | * fallback. | |
1664 | */ | |
1665 | if (shadow_node->nr_child + 1 >= next_type->min_child | |
1666 | && shadow_node->nr_child + 1 <= next_type->max_child) { | |
1667 | new_type_index++; | |
1668 | goto retry; | |
1669 | } else { | |
1670 | new_type_index++; | |
1671 | dbg_printf("Add fallback to type %d\n", new_type_index); | |
1672 | uatomic_inc(&ja->nr_fallback); | |
1673 | fallback = 1; | |
1674 | goto retry; | |
1675 | } | |
1676 | break; | |
1677 | } | |
1678 | case JA_RECOMPACT_DEL: | |
1679 | /* | |
1680 | * Recompaction attempt on delete failed. Should only | |
1681 | * happen if target node type is pool. This is caused by | |
1682 | * a hard-to-split distribution. Recompact on same node | |
1683 | * size, but flag current node as "fallback" to ensure | |
1684 | * we don't attempt recompaction before some activity | |
1685 | * has reshuffled our node. | |
1686 | */ | |
1687 | assert(new_type->type_class == RCU_JA_POOL); | |
1688 | new_type_index = old_type_index; | |
1689 | dbg_printf("Delete fallback keeping type %d\n", new_type_index); | |
1690 | uatomic_inc(&ja->nr_fallback); | |
1691 | fallback = 1; | |
1692 | goto retry; | |
1693 | default: | |
1694 | assert(0); | |
1695 | return -EINVAL; | |
1696 | } | |
1697 | ||
1698 | /* | |
1699 | * Last resort fallback: pigeon. | |
1700 | */ | |
f07b240f MD |
1701 | new_type_index = (1UL << JA_TYPE_BITS) - 1; |
1702 | dbg_printf("Fallback to type %d\n", new_type_index); | |
1703 | uatomic_inc(&ja->nr_fallback); | |
1704 | fallback = 1; | |
1705 | goto retry; | |
7a0b2331 MD |
1706 | } |
1707 | ||
5a9a87dd | 1708 | /* |
2e313670 | 1709 | * Return 0 on success, -EAGAIN if need to retry, or other negative |
5a9a87dd MD |
1710 | * error value otherwise. |
1711 | */ | |
7a0b2331 | 1712 | static |
d96bfb0d | 1713 | int ja_node_set_nth(struct cds_ja *ja, |
b4540e8a | 1714 | struct cds_ja_inode_flag **node_flag, uint8_t n, |
5a9a87dd | 1715 | struct cds_ja_inode_flag *child_node_flag, |
48cbe001 MD |
1716 | struct cds_ja_shadow_node *shadow_node, |
1717 | int level) | |
7a0b2331 MD |
1718 | { |
1719 | int ret; | |
e1db2db5 | 1720 | unsigned int type_index; |
d96bfb0d | 1721 | const struct cds_ja_type *type; |
b4540e8a | 1722 | struct cds_ja_inode *node; |
7a0b2331 | 1723 | |
a2a7ff59 MD |
1724 | dbg_printf("ja_node_set_nth for n=%u, node %p, shadow %p\n", |
1725 | (unsigned int) n, ja_node_ptr(*node_flag), shadow_node); | |
1726 | ||
e1db2db5 MD |
1727 | node = ja_node_ptr(*node_flag); |
1728 | type_index = ja_node_type(*node_flag); | |
1729 | type = &ja_types[type_index]; | |
b1a90ce3 | 1730 | ret = _ja_node_set_nth(type, node, *node_flag, shadow_node, |
e1db2db5 | 1731 | n, child_node_flag); |
2e313670 MD |
1732 | switch (ret) { |
1733 | case -ENOSPC: | |
19ddcd04 MD |
1734 | /* Not enough space in node, need to recompact to next type. */ |
1735 | ret = ja_node_recompact(JA_RECOMPACT_ADD_NEXT, ja, type_index, type, node, | |
48cbe001 | 1736 | shadow_node, node_flag, n, child_node_flag, NULL, level); |
2e313670 MD |
1737 | break; |
1738 | case -ERANGE: | |
1739 | /* Node needs to be recompacted. */ | |
19ddcd04 | 1740 | ret = ja_node_recompact(JA_RECOMPACT_ADD_SAME, ja, type_index, type, node, |
48cbe001 | 1741 | shadow_node, node_flag, n, child_node_flag, NULL, level); |
2e313670 MD |
1742 | break; |
1743 | } | |
1744 | return ret; | |
1745 | } | |
1746 | ||
1747 | /* | |
1748 | * Return 0 on success, -EAGAIN if need to retry, or other negative | |
1749 | * error value otherwise. | |
1750 | */ | |
1751 | static | |
af3cbd45 MD |
1752 | int ja_node_clear_ptr(struct cds_ja *ja, |
1753 | struct cds_ja_inode_flag **node_flag_ptr, /* Pointer to location to nullify */ | |
1754 | struct cds_ja_inode_flag **parent_node_flag_ptr, /* Address of parent ptr in its parent */ | |
1755 | struct cds_ja_shadow_node *shadow_node, /* of parent */ | |
48cbe001 | 1756 | uint8_t n, int level) |
2e313670 MD |
1757 | { |
1758 | int ret; | |
1759 | unsigned int type_index; | |
1760 | const struct cds_ja_type *type; | |
1761 | struct cds_ja_inode *node; | |
1762 | ||
af3cbd45 MD |
1763 | dbg_printf("ja_node_clear_ptr for node %p, shadow %p, target ptr %p\n", |
1764 | ja_node_ptr(*parent_node_flag_ptr), shadow_node, node_flag_ptr); | |
2e313670 | 1765 | |
af3cbd45 MD |
1766 | node = ja_node_ptr(*parent_node_flag_ptr); |
1767 | type_index = ja_node_type(*parent_node_flag_ptr); | |
2e313670 | 1768 | type = &ja_types[type_index]; |
19ddcd04 | 1769 | ret = _ja_node_clear_ptr(type, node, *parent_node_flag_ptr, shadow_node, node_flag_ptr, n); |
2e313670 | 1770 | if (ret == -EFBIG) { |
19ddcd04 | 1771 | /* Should try recompaction. */ |
2e313670 | 1772 | ret = ja_node_recompact(JA_RECOMPACT_DEL, ja, type_index, type, node, |
af3cbd45 | 1773 | shadow_node, parent_node_flag_ptr, n, NULL, |
48cbe001 | 1774 | node_flag_ptr, level); |
7a0b2331 MD |
1775 | } |
1776 | return ret; | |
1777 | } | |
be9a7474 | 1778 | |
03ec1aeb | 1779 | struct cds_ja_node *cds_ja_lookup(struct cds_ja *ja, uint64_t key) |
b4540e8a | 1780 | { |
41975c12 MD |
1781 | unsigned int tree_depth, i; |
1782 | struct cds_ja_inode_flag *node_flag; | |
1783 | ||
1784 | if (caa_unlikely(key > ja->key_max)) | |
03ec1aeb | 1785 | return NULL; |
41975c12 | 1786 | tree_depth = ja->tree_depth; |
5a9a87dd | 1787 | node_flag = rcu_dereference(ja->root); |
41975c12 | 1788 | |
5a9a87dd MD |
1789 | /* level 0: root node */ |
1790 | if (!ja_node_ptr(node_flag)) | |
03ec1aeb | 1791 | return NULL; |
5a9a87dd MD |
1792 | |
1793 | for (i = 1; i < tree_depth; i++) { | |
79b41067 MD |
1794 | uint8_t iter_key; |
1795 | ||
1796 | iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (tree_depth - i - 1))); | |
48cbe001 | 1797 | node_flag = ja_node_get_nth(node_flag, NULL, iter_key); |
582a6ade MD |
1798 | dbg_printf("cds_ja_lookup iter key lookup %u finds node_flag %p\n", |
1799 | (unsigned int) iter_key, node_flag); | |
41975c12 | 1800 | if (!ja_node_ptr(node_flag)) |
03ec1aeb | 1801 | return NULL; |
41975c12 MD |
1802 | } |
1803 | ||
5a9a87dd | 1804 | /* Last level lookup succeded. We got an actual match. */ |
03ec1aeb | 1805 | return (struct cds_ja_node *) node_flag; |
5a9a87dd MD |
1806 | } |
1807 | ||
b023ba9f MD |
1808 | static |
1809 | struct cds_ja_node *cds_ja_lookup_inequality(struct cds_ja *ja, uint64_t key, | |
36305a3d | 1810 | uint64_t *result_key, enum ja_lookup_inequality mode) |
291b2543 MD |
1811 | { |
1812 | int tree_depth, level; | |
1813 | struct cds_ja_inode_flag *node_flag, *cur_node_depth[JA_MAX_DEPTH]; | |
36305a3d MD |
1814 | uint8_t cur_key[JA_MAX_DEPTH]; |
1815 | uint64_t _result_key = 0; | |
b023ba9f | 1816 | enum ja_direction dir; |
291b2543 | 1817 | |
b023ba9f MD |
1818 | switch (mode) { |
1819 | case JA_LOOKUP_BE: | |
1820 | if (caa_unlikely(key > ja->key_max || key == 0)) | |
1821 | return NULL; | |
1822 | break; | |
1823 | case JA_LOOKUP_AE: | |
1824 | if (caa_unlikely(key >= ja->key_max)) | |
1825 | return NULL; | |
1826 | break; | |
1827 | default: | |
03ec1aeb | 1828 | return NULL; |
b023ba9f | 1829 | } |
291b2543 MD |
1830 | |
1831 | memset(cur_node_depth, 0, sizeof(cur_node_depth)); | |
36305a3d | 1832 | memset(cur_key, 0, sizeof(cur_key)); |
291b2543 MD |
1833 | tree_depth = ja->tree_depth; |
1834 | node_flag = rcu_dereference(ja->root); | |
1835 | cur_node_depth[0] = node_flag; | |
1836 | ||
1837 | /* level 0: root node */ | |
1838 | if (!ja_node_ptr(node_flag)) | |
03ec1aeb | 1839 | return NULL; |
291b2543 MD |
1840 | |
1841 | for (level = 1; level < tree_depth; level++) { | |
1842 | uint8_t iter_key; | |
1843 | ||
1844 | iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (tree_depth - level - 1))); | |
1845 | node_flag = ja_node_get_nth(node_flag, NULL, iter_key); | |
1846 | if (!ja_node_ptr(node_flag)) | |
1847 | break; | |
36305a3d | 1848 | cur_key[level - 1] = iter_key; |
291b2543 | 1849 | cur_node_depth[level] = node_flag; |
b023ba9f | 1850 | dbg_printf("cds_ja_lookup_inequality iter key lookup %u finds node_flag %p\n", |
291b2543 MD |
1851 | (unsigned int) iter_key, node_flag); |
1852 | } | |
1853 | ||
1854 | if (level == tree_depth) { | |
1855 | /* Last level lookup succeded. We got an equal match. */ | |
36305a3d MD |
1856 | if (result_key) |
1857 | *result_key = key; | |
03ec1aeb | 1858 | return (struct cds_ja_node *) node_flag; |
291b2543 MD |
1859 | } |
1860 | ||
1861 | /* | |
b023ba9f | 1862 | * Find highest value left/right of current node. |
291b2543 | 1863 | * Current node is cur_node_depth[level]. |
b023ba9f MD |
1864 | * Start at current level. If we cannot find any key left/right |
1865 | * of ours, go one level up, seek highest value left/right of | |
1866 | * current (recursively), and when we find one, get the | |
1867 | * rightmost/leftmost child of its rightmost/leftmost child | |
1868 | * (recursively). | |
291b2543 | 1869 | */ |
b023ba9f MD |
1870 | switch (mode) { |
1871 | case JA_LOOKUP_BE: | |
1872 | dir = JA_LEFT; | |
1873 | break; | |
1874 | case JA_LOOKUP_AE: | |
1875 | dir = JA_RIGHT; | |
1876 | break; | |
1877 | default: | |
1878 | assert(0); | |
1879 | } | |
291b2543 MD |
1880 | for (; level > 0; level--) { |
1881 | uint8_t iter_key; | |
1882 | ||
1883 | iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (tree_depth - level - 1))); | |
b023ba9f | 1884 | node_flag = ja_node_get_leftright(cur_node_depth[level - 1], |
36305a3d MD |
1885 | iter_key, &cur_key[level - 1], dir); |
1886 | /* If found left/right sibling, find rightmost/leftmost child. */ | |
291b2543 MD |
1887 | if (ja_node_ptr(node_flag)) |
1888 | break; | |
1889 | } | |
1890 | ||
1891 | if (!level) { | |
b023ba9f | 1892 | /* Reached the root and could not find a left/right sibling. */ |
03ec1aeb | 1893 | return NULL; |
291b2543 MD |
1894 | } |
1895 | ||
1896 | level++; | |
3c52f0f9 MD |
1897 | |
1898 | /* | |
4cef6f97 | 1899 | * From this point, we are guaranteed to be able to find a |
b023ba9f MD |
1900 | * "below than"/"above than" match. ja_attach_node() and |
1901 | * ja_detach_node() both guarantee that it is not possible for a | |
1902 | * lookup to reach a dead-end. | |
3c52f0f9 MD |
1903 | */ |
1904 | ||
b023ba9f MD |
1905 | /* |
1906 | * Find rightmost/leftmost child of rightmost/leftmost child | |
1907 | * (recursively). | |
1908 | */ | |
1909 | switch (mode) { | |
1910 | case JA_LOOKUP_BE: | |
1911 | dir = JA_RIGHTMOST; | |
1912 | break; | |
1913 | case JA_LOOKUP_AE: | |
1914 | dir = JA_LEFTMOST; | |
1915 | break; | |
1916 | default: | |
1917 | assert(0); | |
1918 | } | |
291b2543 | 1919 | for (; level < tree_depth; level++) { |
36305a3d | 1920 | node_flag = ja_node_get_minmax(node_flag, &cur_key[level - 1], dir); |
291b2543 MD |
1921 | if (!ja_node_ptr(node_flag)) |
1922 | break; | |
1923 | } | |
1924 | ||
4cef6f97 | 1925 | assert(level == tree_depth); |
291b2543 | 1926 | |
36305a3d MD |
1927 | if (result_key) { |
1928 | for (level = 1; level < tree_depth; level++) { | |
1929 | _result_key |= ((uint64_t) cur_key[level - 1]) | |
1930 | << (JA_BITS_PER_BYTE * (tree_depth - level - 1)); | |
1931 | } | |
1932 | *result_key = _result_key; | |
1933 | } | |
03ec1aeb | 1934 | return (struct cds_ja_node *) node_flag; |
291b2543 MD |
1935 | } |
1936 | ||
36305a3d MD |
1937 | struct cds_ja_node *cds_ja_lookup_below_equal(struct cds_ja *ja, |
1938 | uint64_t key, uint64_t *result_key) | |
b023ba9f | 1939 | { |
36305a3d | 1940 | return cds_ja_lookup_inequality(ja, key, result_key, JA_LOOKUP_BE); |
b023ba9f MD |
1941 | } |
1942 | ||
36305a3d MD |
1943 | struct cds_ja_node *cds_ja_lookup_above_equal(struct cds_ja *ja, |
1944 | uint64_t key, uint64_t *result_key) | |
b023ba9f | 1945 | { |
36305a3d | 1946 | return cds_ja_lookup_inequality(ja, key, result_key, JA_LOOKUP_AE); |
b023ba9f MD |
1947 | } |
1948 | ||
5a9a87dd MD |
1949 | /* |
1950 | * We reached an unpopulated node. Create it and the children we need, | |
1951 | * and then attach the entire branch to the current node. This may | |
1952 | * trigger recompaction of the current node. Locks needed: node lock | |
1953 | * (for add), and, possibly, parent node lock (to update pointer due to | |
1954 | * node recompaction). | |
1955 | * | |
1956 | * First take node lock, check if recompaction is needed, then take | |
1957 | * parent lock (if needed). Then we can proceed to create the new | |
1958 | * branch. Publish the new branch, and release locks. | |
1959 | * TODO: we currently always take the parent lock even when not needed. | |
47d2eab3 MD |
1960 | * |
1961 | * ja_attach_node() ensures that a lookup will _never_ see a branch that | |
1962 | * leads to a dead-end: before attaching a branch, the entire content of | |
1963 | * the new branch is populated, thus creating a cluster, before | |
1964 | * attaching the cluster to the rest of the tree, thus making it visible | |
1965 | * to lookups. | |
5a9a87dd MD |
1966 | */ |
1967 | static | |
1968 | int ja_attach_node(struct cds_ja *ja, | |
b0ca2d21 | 1969 | struct cds_ja_inode_flag **attach_node_flag_ptr, |
b62a8d0c | 1970 | struct cds_ja_inode_flag *attach_node_flag, |
48cbe001 MD |
1971 | struct cds_ja_inode_flag *parent_attach_node_flag, |
1972 | struct cds_ja_inode_flag **old_node_flag_ptr, | |
1973 | struct cds_ja_inode_flag *old_node_flag, | |
5a9a87dd | 1974 | uint64_t key, |
79b41067 | 1975 | unsigned int level, |
5a9a87dd MD |
1976 | struct cds_ja_node *child_node) |
1977 | { | |
1978 | struct cds_ja_shadow_node *shadow_node = NULL, | |
af3cbd45 | 1979 | *parent_shadow_node = NULL; |
5a9a87dd MD |
1980 | struct cds_ja_inode_flag *iter_node_flag, *iter_dest_node_flag; |
1981 | int ret, i; | |
a2a7ff59 | 1982 | struct cds_ja_inode_flag *created_nodes[JA_MAX_DEPTH]; |
5a9a87dd MD |
1983 | int nr_created_nodes = 0; |
1984 | ||
48cbe001 MD |
1985 | dbg_printf("Attach node at level %u (old_node_flag %p, attach_node_flag_ptr %p attach_node_flag %p, parent_attach_node_flag %p)\n", |
1986 | level, old_node_flag, attach_node_flag_ptr, attach_node_flag, parent_attach_node_flag); | |
a2a7ff59 | 1987 | |
48cbe001 MD |
1988 | assert(!old_node_flag); |
1989 | if (attach_node_flag) { | |
1990 | shadow_node = rcuja_shadow_lookup_lock(ja->ht, attach_node_flag); | |
1991 | if (!shadow_node) { | |
1992 | ret = -EAGAIN; | |
1993 | goto end; | |
1994 | } | |
5a9a87dd | 1995 | } |
48cbe001 | 1996 | if (parent_attach_node_flag) { |
5a9a87dd | 1997 | parent_shadow_node = rcuja_shadow_lookup_lock(ja->ht, |
48cbe001 | 1998 | parent_attach_node_flag); |
5a9a87dd | 1999 | if (!parent_shadow_node) { |
2e313670 | 2000 | ret = -EAGAIN; |
5a9a87dd MD |
2001 | goto unlock_shadow; |
2002 | } | |
2003 | } | |
2004 | ||
48cbe001 | 2005 | if (old_node_flag_ptr && ja_node_ptr(*old_node_flag_ptr)) { |
b306a0fe | 2006 | /* |
c112acaa MD |
2007 | * Target node has been updated between RCU lookup and |
2008 | * lock acquisition. We need to re-try lookup and | |
2009 | * attach. | |
2010 | */ | |
2011 | ret = -EAGAIN; | |
2012 | goto unlock_parent; | |
2013 | } | |
2014 | ||
9be99d4a MD |
2015 | /* |
2016 | * Perform a lookup query to handle the case where | |
2017 | * old_node_flag_ptr is NULL. We cannot use it to check if the | |
2018 | * node has been populated between RCU lookup and mutex | |
2019 | * acquisition. | |
2020 | */ | |
2021 | if (!old_node_flag_ptr) { | |
2022 | uint8_t iter_key; | |
2023 | struct cds_ja_inode_flag *lookup_node_flag; | |
2024 | struct cds_ja_inode_flag **lookup_node_flag_ptr; | |
2025 | ||
2026 | iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (ja->tree_depth - level))); | |
2027 | lookup_node_flag = ja_node_get_nth(attach_node_flag, | |
2028 | &lookup_node_flag_ptr, | |
2029 | iter_key); | |
2030 | if (lookup_node_flag) { | |
2031 | ret = -EEXIST; | |
2032 | goto unlock_parent; | |
2033 | } | |
2034 | } | |
2035 | ||
c112acaa | 2036 | if (attach_node_flag_ptr && ja_node_ptr(*attach_node_flag_ptr) != |
b62a8d0c | 2037 | ja_node_ptr(attach_node_flag)) { |
c112acaa MD |
2038 | /* |
2039 | * Target node has been updated between RCU lookup and | |
2040 | * lock acquisition. We need to re-try lookup and | |
2041 | * attach. | |
b306a0fe MD |
2042 | */ |
2043 | ret = -EAGAIN; | |
2044 | goto unlock_parent; | |
2045 | } | |
2046 | ||
a2a7ff59 | 2047 | /* Create new branch, starting from bottom */ |
03ec1aeb | 2048 | iter_node_flag = (struct cds_ja_inode_flag *) child_node; |
5a9a87dd | 2049 | |
48cbe001 | 2050 | for (i = ja->tree_depth - 1; i >= (int) level; i--) { |
79b41067 MD |
2051 | uint8_t iter_key; |
2052 | ||
48cbe001 | 2053 | iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (ja->tree_depth - i - 1))); |
79b41067 | 2054 | dbg_printf("branch creation level %d, key %u\n", |
48cbe001 | 2055 | i, (unsigned int) iter_key); |
5a9a87dd MD |
2056 | iter_dest_node_flag = NULL; |
2057 | ret = ja_node_set_nth(ja, &iter_dest_node_flag, | |
79b41067 | 2058 | iter_key, |
5a9a87dd | 2059 | iter_node_flag, |
48cbe001 | 2060 | NULL, i); |
9be99d4a MD |
2061 | if (ret) { |
2062 | dbg_printf("branch creation error %d\n", ret); | |
5a9a87dd | 2063 | goto check_error; |
9be99d4a | 2064 | } |
5a9a87dd MD |
2065 | created_nodes[nr_created_nodes++] = iter_dest_node_flag; |
2066 | iter_node_flag = iter_dest_node_flag; | |
2067 | } | |
48cbe001 | 2068 | assert(level > 0); |
5a9a87dd | 2069 | |
48cbe001 MD |
2070 | /* Publish branch */ |
2071 | if (level == 1) { | |
2072 | /* | |
2073 | * Attaching to root node. | |
2074 | */ | |
2075 | rcu_assign_pointer(ja->root, iter_node_flag); | |
2076 | } else { | |
79b41067 MD |
2077 | uint8_t iter_key; |
2078 | ||
2079 | iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (ja->tree_depth - level))); | |
48cbe001 MD |
2080 | dbg_printf("publish branch at level %d, key %u\n", |
2081 | level - 1, (unsigned int) iter_key); | |
a2a7ff59 | 2082 | /* We need to use set_nth on the previous level. */ |
48cbe001 | 2083 | iter_dest_node_flag = attach_node_flag; |
a2a7ff59 | 2084 | ret = ja_node_set_nth(ja, &iter_dest_node_flag, |
79b41067 | 2085 | iter_key, |
a2a7ff59 | 2086 | iter_node_flag, |
48cbe001 | 2087 | shadow_node, level - 1); |
9be99d4a MD |
2088 | if (ret) { |
2089 | dbg_printf("branch publish error %d\n", ret); | |
a2a7ff59 | 2090 | goto check_error; |
9be99d4a | 2091 | } |
48cbe001 MD |
2092 | /* |
2093 | * Attach branch | |
2094 | */ | |
2095 | rcu_assign_pointer(*attach_node_flag_ptr, iter_dest_node_flag); | |
a2a7ff59 MD |
2096 | } |
2097 | ||
5a9a87dd MD |
2098 | /* Success */ |
2099 | ret = 0; | |
2100 | ||
2101 | check_error: | |
2102 | if (ret) { | |
2103 | for (i = 0; i < nr_created_nodes; i++) { | |
2104 | int tmpret; | |
a2a7ff59 MD |
2105 | int flags; |
2106 | ||
2107 | flags = RCUJA_SHADOW_CLEAR_FREE_LOCK; | |
2108 | if (i) | |
2109 | flags |= RCUJA_SHADOW_CLEAR_FREE_NODE; | |
5a9a87dd | 2110 | tmpret = rcuja_shadow_clear(ja->ht, |
3d8fe307 | 2111 | created_nodes[i], |
a2a7ff59 MD |
2112 | NULL, |
2113 | flags); | |
5a9a87dd MD |
2114 | assert(!tmpret); |
2115 | } | |
2116 | } | |
b306a0fe | 2117 | unlock_parent: |
5a9a87dd MD |
2118 | if (parent_shadow_node) |
2119 | rcuja_shadow_unlock(parent_shadow_node); | |
2120 | unlock_shadow: | |
2121 | if (shadow_node) | |
2122 | rcuja_shadow_unlock(shadow_node); | |
2123 | end: | |
2124 | return ret; | |
2125 | } | |
2126 | ||
2127 | /* | |
03ec1aeb MD |
2128 | * Lock the parent containing the pointer to list of duplicates, and add |
2129 | * node to this list. Failure can happen if concurrent update changes | |
2130 | * the parent before we get the lock. We return -EAGAIN in that case. | |
5a9a87dd MD |
2131 | * Return 0 on success, negative error value on failure. |
2132 | */ | |
2133 | static | |
2134 | int ja_chain_node(struct cds_ja *ja, | |
af3cbd45 | 2135 | struct cds_ja_inode_flag *parent_node_flag, |
fa112799 | 2136 | struct cds_ja_inode_flag **node_flag_ptr, |
c112acaa | 2137 | struct cds_ja_inode_flag *node_flag, |
5a9a87dd MD |
2138 | struct cds_ja_node *node) |
2139 | { | |
2140 | struct cds_ja_shadow_node *shadow_node; | |
fa112799 | 2141 | int ret = 0; |
5a9a87dd | 2142 | |
3d8fe307 | 2143 | shadow_node = rcuja_shadow_lookup_lock(ja->ht, parent_node_flag); |
b306a0fe | 2144 | if (!shadow_node) { |
2e313670 | 2145 | return -EAGAIN; |
b306a0fe | 2146 | } |
c112acaa | 2147 | if (ja_node_ptr(*node_flag_ptr) != ja_node_ptr(node_flag)) { |
fa112799 MD |
2148 | ret = -EAGAIN; |
2149 | goto end; | |
2150 | } | |
03ec1aeb MD |
2151 | /* |
2152 | * Add node to head of list. Safe against concurrent RCU read | |
2153 | * traversals. | |
2154 | */ | |
2155 | node->next = (struct cds_ja_node *) node_flag; | |
2156 | rcu_assign_pointer(*node_flag_ptr, (struct cds_ja_inode_flag *) node); | |
fa112799 | 2157 | end: |
5a9a87dd | 2158 | rcuja_shadow_unlock(shadow_node); |
fa112799 | 2159 | return ret; |
5a9a87dd MD |
2160 | } |
2161 | ||
75d573aa MD |
2162 | static |
2163 | int _cds_ja_add(struct cds_ja *ja, uint64_t key, | |
6475613c | 2164 | struct cds_ja_node *node, |
75d573aa | 2165 | struct cds_ja_node **unique_node_ret) |
5a9a87dd MD |
2166 | { |
2167 | unsigned int tree_depth, i; | |
48cbe001 | 2168 | struct cds_ja_inode_flag *attach_node_flag, |
5a9a87dd | 2169 | *parent_node_flag, |
b62a8d0c | 2170 | *parent2_node_flag, |
48cbe001 MD |
2171 | *node_flag, |
2172 | *parent_attach_node_flag; | |
2173 | struct cds_ja_inode_flag **attach_node_flag_ptr, | |
2174 | **parent_node_flag_ptr, | |
2175 | **node_flag_ptr; | |
5a9a87dd MD |
2176 | int ret; |
2177 | ||
b306a0fe | 2178 | if (caa_unlikely(key > ja->key_max)) { |
5a9a87dd | 2179 | return -EINVAL; |
b306a0fe | 2180 | } |
5a9a87dd MD |
2181 | tree_depth = ja->tree_depth; |
2182 | ||
2183 | retry: | |
a2a7ff59 | 2184 | dbg_printf("cds_ja_add attempt: key %" PRIu64 ", node %p\n", |
6475613c | 2185 | key, node); |
5a9a87dd | 2186 | parent2_node_flag = NULL; |
b0f74e47 MD |
2187 | parent_node_flag = |
2188 | (struct cds_ja_inode_flag *) &ja->root; /* Use root ptr address as key for mutex */ | |
48cbe001 | 2189 | parent_node_flag_ptr = NULL; |
35170a44 | 2190 | node_flag = rcu_dereference(ja->root); |
48cbe001 | 2191 | node_flag_ptr = &ja->root; |
5a9a87dd MD |
2192 | |
2193 | /* Iterate on all internal levels */ | |
a2a7ff59 | 2194 | for (i = 1; i < tree_depth; i++) { |
79b41067 MD |
2195 | uint8_t iter_key; |
2196 | ||
48cbe001 MD |
2197 | if (!ja_node_ptr(node_flag)) |
2198 | break; | |
2199 | dbg_printf("cds_ja_add iter parent2_node_flag %p parent_node_flag %p node_flag_ptr %p node_flag %p\n", | |
2200 | parent2_node_flag, parent_node_flag, node_flag_ptr, node_flag); | |
79b41067 | 2201 | iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (tree_depth - i - 1))); |
5a9a87dd MD |
2202 | parent2_node_flag = parent_node_flag; |
2203 | parent_node_flag = node_flag; | |
48cbe001 | 2204 | parent_node_flag_ptr = node_flag_ptr; |
5a9a87dd MD |
2205 | node_flag = ja_node_get_nth(node_flag, |
2206 | &node_flag_ptr, | |
79b41067 | 2207 | iter_key); |
5a9a87dd MD |
2208 | } |
2209 | ||
2210 | /* | |
48cbe001 MD |
2211 | * We reached either bottom of tree or internal NULL node, |
2212 | * simply add node to last internal level, or chain it if key is | |
2213 | * already present. | |
5a9a87dd MD |
2214 | */ |
2215 | if (!ja_node_ptr(node_flag)) { | |
48cbe001 MD |
2216 | dbg_printf("cds_ja_add NULL parent2_node_flag %p parent_node_flag %p node_flag_ptr %p node_flag %p\n", |
2217 | parent2_node_flag, parent_node_flag, node_flag_ptr, node_flag); | |
75d573aa | 2218 | |
48cbe001 MD |
2219 | attach_node_flag = parent_node_flag; |
2220 | attach_node_flag_ptr = parent_node_flag_ptr; | |
2221 | parent_attach_node_flag = parent2_node_flag; | |
2222 | ||
b0ca2d21 | 2223 | ret = ja_attach_node(ja, attach_node_flag_ptr, |
b62a8d0c | 2224 | attach_node_flag, |
48cbe001 MD |
2225 | parent_attach_node_flag, |
2226 | node_flag_ptr, | |
2227 | node_flag, | |
6475613c | 2228 | key, i, node); |
5a9a87dd | 2229 | } else { |
75d573aa MD |
2230 | if (unique_node_ret) { |
2231 | *unique_node_ret = (struct cds_ja_node *) ja_node_ptr(node_flag); | |
2232 | return -EEXIST; | |
2233 | } | |
2234 | ||
48cbe001 MD |
2235 | dbg_printf("cds_ja_add duplicate parent2_node_flag %p parent_node_flag %p node_flag_ptr %p node_flag %p\n", |
2236 | parent2_node_flag, parent_node_flag, node_flag_ptr, node_flag); | |
75d573aa | 2237 | |
48cbe001 MD |
2238 | attach_node_flag = node_flag; |
2239 | attach_node_flag_ptr = node_flag_ptr; | |
2240 | parent_attach_node_flag = parent_node_flag; | |
2241 | ||
5a9a87dd | 2242 | ret = ja_chain_node(ja, |
48cbe001 MD |
2243 | parent_attach_node_flag, |
2244 | attach_node_flag_ptr, | |
2245 | attach_node_flag, | |
6475613c | 2246 | node); |
5a9a87dd | 2247 | } |
b306a0fe | 2248 | if (ret == -EAGAIN || ret == -EEXIST) |
5a9a87dd | 2249 | goto retry; |
48cbe001 | 2250 | |
5a9a87dd | 2251 | return ret; |
b4540e8a MD |
2252 | } |
2253 | ||
75d573aa | 2254 | int cds_ja_add(struct cds_ja *ja, uint64_t key, |
6475613c | 2255 | struct cds_ja_node *node) |
75d573aa | 2256 | { |
6475613c | 2257 | return _cds_ja_add(ja, key, node, NULL); |
75d573aa MD |
2258 | } |
2259 | ||
2260 | struct cds_ja_node *cds_ja_add_unique(struct cds_ja *ja, uint64_t key, | |
6475613c | 2261 | struct cds_ja_node *node) |
75d573aa MD |
2262 | { |
2263 | int ret; | |
2264 | struct cds_ja_node *ret_node; | |
2265 | ||
6475613c | 2266 | ret = _cds_ja_add(ja, key, node, &ret_node); |
75d573aa MD |
2267 | if (ret == -EEXIST) |
2268 | return ret_node; | |
2269 | else | |
6475613c | 2270 | return node; |
75d573aa MD |
2271 | } |
2272 | ||
af3cbd45 MD |
2273 | /* |
2274 | * Note: there is no need to lookup the pointer address associated with | |
2275 | * each node's nth item after taking the lock: it's already been done by | |
2276 | * cds_ja_del while holding the rcu read-side lock, and our node rules | |
2277 | * ensure that when a match value -> pointer is found in a node, it is | |
2278 | * _NEVER_ changed for that node without recompaction, and recompaction | |
2279 | * reallocates the node. | |
b306a0fe MD |
2280 | * However, when a child is removed from "linear" nodes, its pointer |
2281 | * is set to NULL. We therefore check, while holding the locks, if this | |
2282 | * pointer is NULL, and return -ENOENT to the caller if it is the case. | |
4cef6f97 MD |
2283 | * |
2284 | * ja_detach_node() ensures that a lookup will _never_ see a branch that | |
2285 | * leads to a dead-end: when removing branch, it makes sure to perform | |
2286 | * the "cut" at the highest node that has only one child, effectively | |
2287 | * replacing it with a NULL pointer. | |
af3cbd45 | 2288 | */ |
35170a44 MD |
2289 | static |
2290 | int ja_detach_node(struct cds_ja *ja, | |
2291 | struct cds_ja_inode_flag **snapshot, | |
af3cbd45 MD |
2292 | struct cds_ja_inode_flag ***snapshot_ptr, |
2293 | uint8_t *snapshot_n, | |
35170a44 MD |
2294 | int nr_snapshot, |
2295 | uint64_t key, | |
2296 | struct cds_ja_node *node) | |
2297 | { | |
af3cbd45 MD |
2298 | struct cds_ja_shadow_node *shadow_nodes[JA_MAX_DEPTH]; |
2299 | struct cds_ja_inode_flag **node_flag_ptr = NULL, | |
2300 | *parent_node_flag = NULL, | |
2301 | **parent_node_flag_ptr = NULL; | |
b62a8d0c | 2302 | struct cds_ja_inode_flag *iter_node_flag; |
4d6ef45e MD |
2303 | int ret, i, nr_shadow = 0, nr_clear = 0, nr_branch = 0; |
2304 | uint8_t n = 0; | |
35170a44 | 2305 | |
4d6ef45e | 2306 | assert(nr_snapshot == ja->tree_depth + 1); |
35170a44 | 2307 | |
af3cbd45 MD |
2308 | /* |
2309 | * From the last internal level node going up, get the node | |
2310 | * lock, check if the node has only one child left. If it is the | |
2311 | * case, we continue iterating upward. When we reach a node | |
2312 | * which has more that one child left, we lock the parent, and | |
2313 | * proceed to the node deletion (removing its children too). | |
2314 | */ | |
4d6ef45e | 2315 | for (i = nr_snapshot - 2; i >= 1; i--) { |
af3cbd45 MD |
2316 | struct cds_ja_shadow_node *shadow_node; |
2317 | ||
2318 | shadow_node = rcuja_shadow_lookup_lock(ja->ht, | |
3d8fe307 | 2319 | snapshot[i]); |
af3cbd45 MD |
2320 | if (!shadow_node) { |
2321 | ret = -EAGAIN; | |
2322 | goto end; | |
2323 | } | |
af3cbd45 | 2324 | shadow_nodes[nr_shadow++] = shadow_node; |
b62a8d0c MD |
2325 | |
2326 | /* | |
2327 | * Check if node has been removed between RCU | |
2328 | * lookup and lock acquisition. | |
2329 | */ | |
2330 | assert(snapshot_ptr[i + 1]); | |
2331 | if (ja_node_ptr(*snapshot_ptr[i + 1]) | |
2332 | != ja_node_ptr(snapshot[i + 1])) { | |
2333 | ret = -ENOENT; | |
2334 | goto end; | |
2335 | } | |
2336 | ||
2337 | assert(shadow_node->nr_child > 0); | |
d810c97f | 2338 | if (shadow_node->nr_child == 1 && i > 1) |
4d6ef45e MD |
2339 | nr_clear++; |
2340 | nr_branch++; | |
af3cbd45 MD |
2341 | if (shadow_node->nr_child > 1 || i == 1) { |
2342 | /* Lock parent and break */ | |
2343 | shadow_node = rcuja_shadow_lookup_lock(ja->ht, | |
3d8fe307 | 2344 | snapshot[i - 1]); |
af3cbd45 MD |
2345 | if (!shadow_node) { |
2346 | ret = -EAGAIN; | |
2347 | goto end; | |
2348 | } | |
2349 | shadow_nodes[nr_shadow++] = shadow_node; | |
b62a8d0c | 2350 | |
c112acaa MD |
2351 | /* |
2352 | * Check if node has been removed between RCU | |
2353 | * lookup and lock acquisition. | |
2354 | */ | |
b62a8d0c MD |
2355 | assert(snapshot_ptr[i]); |
2356 | if (ja_node_ptr(*snapshot_ptr[i]) | |
2357 | != ja_node_ptr(snapshot[i])) { | |
c112acaa MD |
2358 | ret = -ENOENT; |
2359 | goto end; | |
2360 | } | |
2361 | ||
b62a8d0c | 2362 | node_flag_ptr = snapshot_ptr[i + 1]; |
4d6ef45e MD |
2363 | n = snapshot_n[i + 1]; |
2364 | parent_node_flag_ptr = snapshot_ptr[i]; | |
2365 | parent_node_flag = snapshot[i]; | |
c112acaa | 2366 | |
af3cbd45 MD |
2367 | if (i > 1) { |
2368 | /* | |
2369 | * Lock parent's parent, in case we need | |
2370 | * to recompact parent. | |
2371 | */ | |
2372 | shadow_node = rcuja_shadow_lookup_lock(ja->ht, | |
3d8fe307 | 2373 | snapshot[i - 2]); |
af3cbd45 MD |
2374 | if (!shadow_node) { |
2375 | ret = -EAGAIN; | |
2376 | goto end; | |
2377 | } | |
2378 | shadow_nodes[nr_shadow++] = shadow_node; | |
b62a8d0c MD |
2379 | |
2380 | /* | |
2381 | * Check if node has been removed between RCU | |
2382 | * lookup and lock acquisition. | |
2383 | */ | |
2384 | assert(snapshot_ptr[i - 1]); | |
2385 | if (ja_node_ptr(*snapshot_ptr[i - 1]) | |
2386 | != ja_node_ptr(snapshot[i - 1])) { | |
2387 | ret = -ENOENT; | |
2388 | goto end; | |
2389 | } | |
af3cbd45 | 2390 | } |
b62a8d0c | 2391 | |
af3cbd45 MD |
2392 | break; |
2393 | } | |
2394 | } | |
2395 | ||
2396 | /* | |
4d6ef45e MD |
2397 | * At this point, we want to delete all nodes that are about to |
2398 | * be removed from shadow_nodes (except the last one, which is | |
2399 | * either the root or the parent of the upmost node with 1 | |
b62a8d0c MD |
2400 | * child). OK to free lock here, because RCU read lock is held, |
2401 | * and free only performed in call_rcu. | |
af3cbd45 MD |
2402 | */ |
2403 | ||
2404 | for (i = 0; i < nr_clear; i++) { | |
2405 | ret = rcuja_shadow_clear(ja->ht, | |
3d8fe307 | 2406 | shadow_nodes[i]->node_flag, |
af3cbd45 MD |
2407 | shadow_nodes[i], |
2408 | RCUJA_SHADOW_CLEAR_FREE_NODE | |
2409 | | RCUJA_SHADOW_CLEAR_FREE_LOCK); | |
2410 | assert(!ret); | |
2411 | } | |
2412 | ||
2413 | iter_node_flag = parent_node_flag; | |
2414 | /* Remove from parent */ | |
2415 | ret = ja_node_clear_ptr(ja, | |
2416 | node_flag_ptr, /* Pointer to location to nullify */ | |
2417 | &iter_node_flag, /* Old new parent ptr in its parent */ | |
4d6ef45e | 2418 | shadow_nodes[nr_branch - 1], /* of parent */ |
48cbe001 | 2419 | n, nr_branch - 1); |
b306a0fe MD |
2420 | if (ret) |
2421 | goto end; | |
af3cbd45 | 2422 | |
4d6ef45e MD |
2423 | dbg_printf("ja_detach_node: publish %p instead of %p\n", |
2424 | iter_node_flag, *parent_node_flag_ptr); | |
af3cbd45 MD |
2425 | /* Update address of parent ptr in its parent */ |
2426 | rcu_assign_pointer(*parent_node_flag_ptr, iter_node_flag); | |
2427 | ||
2428 | end: | |
2429 | for (i = 0; i < nr_shadow; i++) | |
2430 | rcuja_shadow_unlock(shadow_nodes[i]); | |
35170a44 MD |
2431 | return ret; |
2432 | } | |
2433 | ||
af3cbd45 MD |
2434 | static |
2435 | int ja_unchain_node(struct cds_ja *ja, | |
2436 | struct cds_ja_inode_flag *parent_node_flag, | |
fa112799 | 2437 | struct cds_ja_inode_flag **node_flag_ptr, |
013a6083 | 2438 | struct cds_ja_inode_flag *node_flag, |
af3cbd45 MD |
2439 | struct cds_ja_node *node) |
2440 | { | |
2441 | struct cds_ja_shadow_node *shadow_node; | |
03ec1aeb | 2442 | struct cds_ja_node *iter_node, **iter_node_ptr, **prev_node_ptr = NULL; |
013a6083 | 2443 | int ret = 0, count = 0, found = 0; |
af3cbd45 | 2444 | |
3d8fe307 | 2445 | shadow_node = rcuja_shadow_lookup_lock(ja->ht, parent_node_flag); |
af3cbd45 MD |
2446 | if (!shadow_node) |
2447 | return -EAGAIN; | |
013a6083 | 2448 | if (ja_node_ptr(*node_flag_ptr) != ja_node_ptr(node_flag)) { |
fa112799 MD |
2449 | ret = -EAGAIN; |
2450 | goto end; | |
2451 | } | |
af3cbd45 | 2452 | /* |
03ec1aeb MD |
2453 | * Find the previous node's next pointer pointing to our node, |
2454 | * so we can update it. Retry if another thread removed all but | |
2455 | * one of duplicates since check (this check was performed | |
2456 | * without lock). Ensure that the node we are about to remove is | |
2457 | * still in the list (while holding lock). No need for RCU | |
2458 | * traversal here since we hold the lock on the parent. | |
af3cbd45 | 2459 | */ |
03ec1aeb MD |
2460 | iter_node_ptr = (struct cds_ja_node **) node_flag_ptr; |
2461 | iter_node = (struct cds_ja_node *) ja_node_ptr(node_flag); | |
2462 | cds_ja_for_each_duplicate(iter_node) { | |
f2758d14 | 2463 | count++; |
03ec1aeb MD |
2464 | if (iter_node == node) { |
2465 | prev_node_ptr = iter_node_ptr; | |
013a6083 | 2466 | found++; |
03ec1aeb MD |
2467 | } |
2468 | iter_node_ptr = &iter_node->next; | |
f2758d14 | 2469 | } |
013a6083 MD |
2470 | assert(found <= 1); |
2471 | if (!found || count == 1) { | |
af3cbd45 MD |
2472 | ret = -EAGAIN; |
2473 | goto end; | |
2474 | } | |
03ec1aeb | 2475 | CMM_STORE_SHARED(*prev_node_ptr, node->next); |
ade342cb MD |
2476 | /* |
2477 | * Validate that we indeed removed the node from linked list. | |
2478 | */ | |
2479 | assert(ja_node_ptr(*node_flag_ptr) != (struct cds_ja_inode *) node); | |
af3cbd45 MD |
2480 | end: |
2481 | rcuja_shadow_unlock(shadow_node); | |
2482 | return ret; | |
2483 | } | |
2484 | ||
2485 | /* | |
2486 | * Called with RCU read lock held. | |
2487 | */ | |
35170a44 MD |
2488 | int cds_ja_del(struct cds_ja *ja, uint64_t key, |
2489 | struct cds_ja_node *node) | |
2490 | { | |
2491 | unsigned int tree_depth, i; | |
2492 | struct cds_ja_inode_flag *snapshot[JA_MAX_DEPTH]; | |
af3cbd45 MD |
2493 | struct cds_ja_inode_flag **snapshot_ptr[JA_MAX_DEPTH]; |
2494 | uint8_t snapshot_n[JA_MAX_DEPTH]; | |
35170a44 | 2495 | struct cds_ja_inode_flag *node_flag; |
fa112799 MD |
2496 | struct cds_ja_inode_flag **prev_node_flag_ptr, |
2497 | **node_flag_ptr; | |
4d6ef45e | 2498 | int nr_snapshot; |
35170a44 MD |
2499 | int ret; |
2500 | ||
2501 | if (caa_unlikely(key > ja->key_max)) | |
2502 | return -EINVAL; | |
2503 | tree_depth = ja->tree_depth; | |
2504 | ||
2505 | retry: | |
4d6ef45e | 2506 | nr_snapshot = 0; |
35170a44 MD |
2507 | dbg_printf("cds_ja_del attempt: key %" PRIu64 ", node %p\n", |
2508 | key, node); | |
2509 | ||
2510 | /* snapshot for level 0 is only for shadow node lookup */ | |
4d6ef45e MD |
2511 | snapshot_n[0] = 0; |
2512 | snapshot_n[1] = 0; | |
af3cbd45 | 2513 | snapshot_ptr[nr_snapshot] = NULL; |
35170a44 MD |
2514 | snapshot[nr_snapshot++] = (struct cds_ja_inode_flag *) &ja->root; |
2515 | node_flag = rcu_dereference(ja->root); | |
af3cbd45 | 2516 | prev_node_flag_ptr = &ja->root; |
fa112799 | 2517 | node_flag_ptr = &ja->root; |
35170a44 MD |
2518 | |
2519 | /* Iterate on all internal levels */ | |
2520 | for (i = 1; i < tree_depth; i++) { | |
2521 | uint8_t iter_key; | |
2522 | ||
2523 | dbg_printf("cds_ja_del iter node_flag %p\n", | |
2524 | node_flag); | |
2525 | if (!ja_node_ptr(node_flag)) { | |
2526 | return -ENOENT; | |
2527 | } | |
35170a44 | 2528 | iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (tree_depth - i - 1))); |
4d6ef45e | 2529 | snapshot_n[nr_snapshot + 1] = iter_key; |
af3cbd45 MD |
2530 | snapshot_ptr[nr_snapshot] = prev_node_flag_ptr; |
2531 | snapshot[nr_snapshot++] = node_flag; | |
35170a44 | 2532 | node_flag = ja_node_get_nth(node_flag, |
fa112799 | 2533 | &node_flag_ptr, |
35170a44 | 2534 | iter_key); |
48cbe001 MD |
2535 | if (node_flag) |
2536 | prev_node_flag_ptr = node_flag_ptr; | |
af3cbd45 MD |
2537 | dbg_printf("cds_ja_del iter key lookup %u finds node_flag %p, prev_node_flag_ptr %p\n", |
2538 | (unsigned int) iter_key, node_flag, | |
2539 | prev_node_flag_ptr); | |
35170a44 | 2540 | } |
35170a44 MD |
2541 | /* |
2542 | * We reached bottom of tree, try to find the node we are trying | |
2543 | * to remove. Fail if we cannot find it. | |
2544 | */ | |
2545 | if (!ja_node_ptr(node_flag)) { | |
4d6ef45e MD |
2546 | dbg_printf("cds_ja_del: no node found for key %" PRIu64 "\n", |
2547 | key); | |
35170a44 MD |
2548 | return -ENOENT; |
2549 | } else { | |
03ec1aeb | 2550 | struct cds_ja_node *iter_node, *match = NULL; |
af3cbd45 | 2551 | int count = 0; |
35170a44 | 2552 | |
03ec1aeb MD |
2553 | iter_node = (struct cds_ja_node *) ja_node_ptr(node_flag); |
2554 | cds_ja_for_each_duplicate_rcu(iter_node) { | |
2555 | dbg_printf("cds_ja_del: compare %p with iter_node %p\n", node, iter_node); | |
2556 | if (iter_node == node) | |
2557 | match = iter_node; | |
af3cbd45 | 2558 | count++; |
35170a44 | 2559 | } |
03ec1aeb | 2560 | |
4d6ef45e MD |
2561 | if (!match) { |
2562 | dbg_printf("cds_ja_del: no node match for node %p key %" PRIu64 "\n", node, key); | |
35170a44 | 2563 | return -ENOENT; |
4d6ef45e | 2564 | } |
af3cbd45 MD |
2565 | assert(count > 0); |
2566 | if (count == 1) { | |
2567 | /* | |
4d6ef45e MD |
2568 | * Removing last of duplicates. Last snapshot |
2569 | * does not have a shadow node (external leafs). | |
af3cbd45 MD |
2570 | */ |
2571 | snapshot_ptr[nr_snapshot] = prev_node_flag_ptr; | |
2572 | snapshot[nr_snapshot++] = node_flag; | |
2573 | ret = ja_detach_node(ja, snapshot, snapshot_ptr, | |
2574 | snapshot_n, nr_snapshot, key, node); | |
2575 | } else { | |
f2758d14 | 2576 | ret = ja_unchain_node(ja, snapshot[nr_snapshot - 1], |
013a6083 | 2577 | node_flag_ptr, node_flag, match); |
af3cbd45 | 2578 | } |
35170a44 | 2579 | } |
b306a0fe MD |
2580 | /* |
2581 | * Explanation of -ENOENT handling: caused by concurrent delete | |
2582 | * between RCU lookup and actual removal. Need to re-do the | |
2583 | * lookup and removal attempt. | |
2584 | */ | |
2585 | if (ret == -EAGAIN || ret == -ENOENT) | |
35170a44 MD |
2586 | goto retry; |
2587 | return ret; | |
2588 | } | |
2589 | ||
b4540e8a MD |
2590 | struct cds_ja *_cds_ja_new(unsigned int key_bits, |
2591 | const struct rcu_flavor_struct *flavor) | |
be9a7474 MD |
2592 | { |
2593 | struct cds_ja *ja; | |
b0f74e47 | 2594 | int ret; |
f07b240f | 2595 | struct cds_ja_shadow_node *root_shadow_node; |
be9a7474 MD |
2596 | |
2597 | ja = calloc(sizeof(*ja), 1); | |
2598 | if (!ja) | |
2599 | goto ja_error; | |
b4540e8a MD |
2600 | |
2601 | switch (key_bits) { | |
2602 | case 8: | |
b4540e8a | 2603 | case 16: |
1216b3d2 | 2604 | case 24: |
b4540e8a | 2605 | case 32: |
1216b3d2 MD |
2606 | case 40: |
2607 | case 48: | |
2608 | case 56: | |
2609 | ja->key_max = (1ULL << key_bits) - 1; | |
b4540e8a MD |
2610 | break; |
2611 | case 64: | |
2612 | ja->key_max = UINT64_MAX; | |
2613 | break; | |
2614 | default: | |
2615 | goto check_error; | |
2616 | } | |
2617 | ||
be9a7474 | 2618 | /* ja->root is NULL */ |
5a9a87dd | 2619 | /* tree_depth 0 is for pointer to root node */ |
582a6ade | 2620 | ja->tree_depth = (key_bits >> JA_LOG2_BITS_PER_BYTE) + 1; |
a2a7ff59 | 2621 | assert(ja->tree_depth <= JA_MAX_DEPTH); |
be9a7474 MD |
2622 | ja->ht = rcuja_create_ht(flavor); |
2623 | if (!ja->ht) | |
2624 | goto ht_error; | |
b0f74e47 MD |
2625 | |
2626 | /* | |
2627 | * Note: we should not free this node until judy array destroy. | |
2628 | */ | |
f07b240f | 2629 | root_shadow_node = rcuja_shadow_set(ja->ht, |
3d8fe307 | 2630 | (struct cds_ja_inode_flag *) &ja->root, |
48cbe001 | 2631 | NULL, ja, 0); |
f07b240f MD |
2632 | if (!root_shadow_node) { |
2633 | ret = -ENOMEM; | |
b0f74e47 | 2634 | goto ht_node_error; |
f07b240f | 2635 | } |
b0f74e47 | 2636 | |
be9a7474 MD |
2637 | return ja; |
2638 | ||
b0f74e47 MD |
2639 | ht_node_error: |
2640 | ret = rcuja_delete_ht(ja->ht); | |
2641 | assert(!ret); | |
be9a7474 | 2642 | ht_error: |
b4540e8a | 2643 | check_error: |
be9a7474 MD |
2644 | free(ja); |
2645 | ja_error: | |
2646 | return NULL; | |
2647 | } | |
2648 | ||
19ddcd04 | 2649 | static |
354981c2 | 2650 | void print_debug_fallback_distribution(struct cds_ja *ja) |
19ddcd04 MD |
2651 | { |
2652 | int i; | |
2653 | ||
2654 | fprintf(stderr, "Fallback node distribution:\n"); | |
2655 | for (i = 0; i < JA_ENTRY_PER_NODE; i++) { | |
354981c2 | 2656 | if (!ja->node_fallback_count_distribution[i]) |
19ddcd04 MD |
2657 | continue; |
2658 | fprintf(stderr, " %3u: %4lu\n", | |
354981c2 | 2659 | i, ja->node_fallback_count_distribution[i]); |
19ddcd04 MD |
2660 | } |
2661 | } | |
2662 | ||
021c72c0 | 2663 | static |
19a748d9 | 2664 | int ja_final_checks(struct cds_ja *ja) |
021c72c0 MD |
2665 | { |
2666 | double fallback_ratio; | |
2667 | unsigned long na, nf, nr_fallback; | |
19a748d9 | 2668 | int ret = 0; |
021c72c0 MD |
2669 | |
2670 | fallback_ratio = (double) uatomic_read(&ja->nr_fallback); | |
2671 | fallback_ratio /= (double) uatomic_read(&ja->nr_nodes_allocated); | |
2672 | nr_fallback = uatomic_read(&ja->nr_fallback); | |
2673 | if (nr_fallback) | |
2674 | fprintf(stderr, | |
2675 | "[warning] RCU Judy Array used %lu fallback node(s) (ratio: %g)\n", | |
2676 | uatomic_read(&ja->nr_fallback), | |
2677 | fallback_ratio); | |
2678 | ||
2679 | na = uatomic_read(&ja->nr_nodes_allocated); | |
2680 | nf = uatomic_read(&ja->nr_nodes_freed); | |
19a748d9 MD |
2681 | dbg_printf("Nodes allocated: %lu, Nodes freed: %lu.\n", na, nf); |
2682 | if (nr_fallback) | |
2683 | print_debug_fallback_distribution(ja); | |
2684 | ||
021c72c0 MD |
2685 | if (na != nf) { |
2686 | fprintf(stderr, "[error] Judy array leaked %ld nodes. Allocated: %lu, freed: %lu.\n", | |
2687 | (long) na - nf, na, nf); | |
19a748d9 | 2688 | ret = -1; |
021c72c0 | 2689 | } |
19a748d9 | 2690 | return ret; |
021c72c0 MD |
2691 | } |
2692 | ||
be9a7474 | 2693 | /* |
dc0e9798 MD |
2694 | * There should be no more concurrent add, delete, nor look-up performed |
2695 | * on the Judy array while it is being destroyed (ensured by the | |
2696 | * caller). | |
be9a7474 | 2697 | */ |
99e6e3dc | 2698 | int cds_ja_destroy(struct cds_ja *ja) |
be9a7474 | 2699 | { |
48cbe001 | 2700 | const struct rcu_flavor_struct *flavor; |
b4540e8a MD |
2701 | int ret; |
2702 | ||
48cbe001 | 2703 | flavor = cds_lfht_rcu_flavor(ja->ht); |
be9a7474 | 2704 | rcuja_shadow_prune(ja->ht, |
99e6e3dc | 2705 | RCUJA_SHADOW_CLEAR_FREE_NODE | RCUJA_SHADOW_CLEAR_FREE_LOCK); |
48cbe001 | 2706 | flavor->thread_offline(); |
b4540e8a MD |
2707 | ret = rcuja_delete_ht(ja->ht); |
2708 | if (ret) | |
2709 | return ret; | |
f2ae7af7 MD |
2710 | |
2711 | /* Wait for in-flight call_rcu free to complete. */ | |
2712 | flavor->barrier(); | |
2713 | ||
48cbe001 | 2714 | flavor->thread_online(); |
19a748d9 | 2715 | ret = ja_final_checks(ja); |
b4540e8a | 2716 | free(ja); |
19a748d9 | 2717 | return ret; |
be9a7474 | 2718 | } |