Commit | Line | Data |
---|---|---|
db61f215 MD |
1 | /* |
2 | * rcuja/rcuja.c | |
3 | * | |
4 | * Userspace RCU library - RCU Judy Array | |
5 | * | |
6 | * Copyright (C) 2000 - 2002 Hewlett-Packard Company | |
7 | * Copyright 2012-2013 - Mathieu Desnoyers <mathieu.desnoyers@efficios.com> | |
8 | * | |
9 | * This library is free software; you can redistribute it and/or | |
10 | * modify it under the terms of the GNU Lesser General Public | |
11 | * License as published by the Free Software Foundation; either | |
12 | * version 2.1 of the License, or (at your option) any later version. | |
13 | * | |
14 | * This library is distributed in the hope that it will be useful, | |
15 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
16 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | |
17 | * Lesser General Public License for more details. | |
18 | * | |
19 | * You should have received a copy of the GNU Lesser General Public | |
20 | * License along with this library; if not, write to the Free Software | |
21 | * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA | |
22 | */ | |
23 | ||
24 | #define _LGPL_SOURCE | |
25 | #include <stdint.h> | |
26 | #include <errno.h> | |
27 | #include <limits.h> | |
28 | #include <string.h> | |
29 | #include <assert.h> | |
30 | #include <urcu/rcuja.h> | |
31 | #include <urcu/compiler.h> | |
32 | #include <urcu/arch.h> | |
33 | #include <urcu-pointer.h> | |
34 | #include <urcu/uatomic.h> | |
35 | ||
36 | #include "rcuja-internal.h" | |
37 | ||
38 | #ifndef abs | |
39 | #define abs_int(a) ((int) (a) > 0 ? (int) (a) : -((int) (a))) | |
40 | #endif | |
41 | ||
42 | enum cds_ja_type_class { | |
43 | RCU_JA_LINEAR = 0, /* Type A */ | |
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 */ | |
49 | RCU_JA_PIGEON = 2, /* Type C */ | |
50 | /* 32-bit: 101 to 256 children, 1024 bytes */ | |
51 | /* 64-bit: 113 to 256 children, 2048 bytes */ | |
52 | /* Leaf nodes are implicit from their height in the tree */ | |
53 | RCU_JA_NR_TYPES, | |
54 | ||
55 | RCU_JA_NULL, /* not an encoded type, but keeps code regular */ | |
56 | }; | |
57 | ||
58 | struct cds_ja_type { | |
59 | enum cds_ja_type_class type_class; | |
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 */ | |
64 | uint16_t nr_pool_order; /* number of pools */ | |
65 | uint16_t pool_size_order; /* pool size */ | |
66 | }; | |
67 | ||
68 | /* | |
69 | * Iteration on the array to find the right node size for the number of | |
70 | * children stops when it reaches .max_child == 256 (this is the largest | |
71 | * possible node size, which contains 256 children). | |
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. | |
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. | |
84 | */ | |
85 | ||
86 | #if (CAA_BITS_PER_LONG < 64) | |
87 | /* 32-bit pointers */ | |
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, | |
97 | ja_type_8_max_child = 0, /* NULL */ | |
98 | }; | |
99 | ||
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 | ||
110 | enum { | |
111 | ja_type_5_nr_pool_order = 1, | |
112 | ja_type_6_nr_pool_order = 2, | |
113 | }; | |
114 | ||
115 | const struct cds_ja_type ja_types[] = { | |
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, }, | |
121 | ||
122 | /* Pools may fill sooner than max_child */ | |
123 | /* This pool is hardcoded at index 5. See ja_node_ptr(). */ | |
124 | { .type_class = RCU_JA_POOL, .min_child = 20, .max_child = ja_type_5_max_child, .max_linear_child = ja_type_5_max_linear_child, .order = 8, .nr_pool_order = ja_type_5_nr_pool_order, .pool_size_order = 7, }, | |
125 | /* This pool is hardcoded at index 6. See ja_node_ptr(). */ | |
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, }, | |
127 | ||
128 | /* | |
129 | * Upon node removal below min_child, if child pool is filled | |
130 | * beyond capacity, we roll back to pigeon. | |
131 | */ | |
132 | { .type_class = RCU_JA_PIGEON, .min_child = 83, .max_child = ja_type_7_max_child, .order = 10, }, | |
133 | ||
134 | { .type_class = RCU_JA_NULL, .min_child = 0, .max_child = ja_type_8_max_child, }, | |
135 | }; | |
136 | #else /* !(CAA_BITS_PER_LONG < 64) */ | |
137 | /* 64-bit pointers */ | |
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, | |
147 | ja_type_8_max_child = 256, | |
148 | }; | |
149 | ||
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 | ||
160 | enum { | |
161 | ja_type_5_nr_pool_order = 1, | |
162 | ja_type_6_nr_pool_order = 2, | |
163 | }; | |
164 | ||
165 | const struct cds_ja_type ja_types[] = { | |
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, }, | |
171 | ||
172 | /* Pools may fill sooner than max_child. */ | |
173 | /* This pool is hardcoded at index 5. See ja_node_ptr(). */ | |
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, }, | |
175 | /* This pool is hardcoded at index 6. See ja_node_ptr(). */ | |
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, }, | |
177 | ||
178 | /* | |
179 | * Upon node removal below min_child, if child pool is filled | |
180 | * beyond capacity, we roll back to pigeon. | |
181 | */ | |
182 | { .type_class = RCU_JA_PIGEON, .min_child = 95, .max_child = ja_type_7_max_child, .order = 11, }, | |
183 | ||
184 | { .type_class = RCU_JA_NULL, .min_child = 0, .max_child = ja_type_8_max_child, }, | |
185 | }; | |
186 | #endif /* !(BITS_PER_LONG < 64) */ | |
187 | ||
188 | static inline __attribute__((unused)) | |
189 | void static_array_size_check(void) | |
190 | { | |
191 | CAA_BUILD_BUG_ON(CAA_ARRAY_SIZE(ja_types) < JA_TYPE_MAX_NR); | |
192 | } | |
193 | ||
194 | /* | |
195 | * The cds_ja_node contains the compressed node data needed for | |
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. | |
204 | */ | |
205 | ||
206 | #define DECLARE_LINEAR_NODE(index) \ | |
207 | struct { \ | |
208 | uint8_t nr_child; \ | |
209 | uint8_t child_value[ja_type_## index ##_max_linear_child]; \ | |
210 | struct cds_ja_inode_flag *child_ptr[ja_type_## index ##_max_linear_child]; \ | |
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]; \ | |
218 | struct cds_ja_inode_flag *child_ptr[ja_type_## index ##_max_linear_child]; \ | |
219 | } linear[1U << ja_type_## index ##_nr_pool_order]; \ | |
220 | } | |
221 | ||
222 | struct cds_ja_inode { | |
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 { | |
237 | struct cds_ja_inode_flag *child[ja_type_7_max_child]; | |
238 | } conf_7; | |
239 | /* data aliasing nodes for computed accesses */ | |
240 | uint8_t data[sizeof(struct cds_ja_inode_flag *) * ja_type_7_max_child]; | |
241 | } u; | |
242 | }; | |
243 | ||
244 | enum ja_recompact { | |
245 | JA_RECOMPACT_ADD_SAME, | |
246 | JA_RECOMPACT_ADD_NEXT, | |
247 | JA_RECOMPACT_DEL, | |
248 | }; | |
249 | ||
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 | ||
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 | ||
280 | static | |
281 | struct cds_ja_inode *alloc_cds_ja_node(struct cds_ja *ja, | |
282 | const struct cds_ja_type *ja_type) | |
283 | { | |
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); | |
293 | if (ja_debug_counters()) | |
294 | uatomic_inc(&ja->nr_nodes_allocated); | |
295 | return p; | |
296 | } | |
297 | ||
298 | void free_cds_ja_node(struct cds_ja *ja, struct cds_ja_inode *node) | |
299 | { | |
300 | free(node); | |
301 | if (ja_debug_counters() && node) | |
302 | uatomic_inc(&ja->nr_nodes_freed); | |
303 | } | |
304 | ||
305 | #define __JA_ALIGN_MASK(v, mask) (((v) + (mask)) & ~(mask)) | |
306 | #define JA_ALIGN(v, align) __JA_ALIGN_MASK(v, (typeof(v)) (align) - 1) | |
307 | #define __JA_FLOOR_MASK(v, mask) ((v) & ~(mask)) | |
308 | #define JA_FLOOR(v, align) __JA_FLOOR_MASK(v, (typeof(v)) (align) - 1) | |
309 | ||
310 | static | |
311 | uint8_t *align_ptr_size(uint8_t *ptr) | |
312 | { | |
313 | return (uint8_t *) JA_ALIGN((unsigned long) ptr, sizeof(void *)); | |
314 | } | |
315 | ||
316 | static | |
317 | uint8_t ja_linear_node_get_nr_child(const struct cds_ja_type *type, | |
318 | struct cds_ja_inode *node) | |
319 | { | |
320 | assert(type->type_class == RCU_JA_LINEAR || type->type_class == RCU_JA_POOL); | |
321 | return rcu_dereference(node->u.data[0]); | |
322 | } | |
323 | ||
324 | /* | |
325 | * The order in which values and pointers are does does not matter: if | |
326 | * a value is missing, we return NULL. If a value is there, but its | |
327 | * associated pointers is still NULL, we return NULL too. | |
328 | */ | |
329 | static | |
330 | struct cds_ja_inode_flag *ja_linear_node_get_nth(const struct cds_ja_type *type, | |
331 | struct cds_ja_inode *node, | |
332 | struct cds_ja_inode_flag ***node_flag_ptr, | |
333 | uint8_t n) | |
334 | { | |
335 | uint8_t nr_child; | |
336 | uint8_t *values; | |
337 | struct cds_ja_inode_flag **pointers; | |
338 | struct cds_ja_inode_flag *ptr; | |
339 | unsigned int i; | |
340 | ||
341 | assert(type->type_class == RCU_JA_LINEAR || type->type_class == RCU_JA_POOL); | |
342 | ||
343 | nr_child = ja_linear_node_get_nr_child(type, node); | |
344 | cmm_smp_rmb(); /* read nr_child before values and pointers */ | |
345 | assert(nr_child <= type->max_linear_child); | |
346 | assert(type->type_class != RCU_JA_LINEAR || nr_child >= type->min_child); | |
347 | ||
348 | values = &node->u.data[1]; | |
349 | for (i = 0; i < nr_child; i++) { | |
350 | if (CMM_LOAD_SHARED(values[i]) == n) | |
351 | break; | |
352 | } | |
353 | if (i >= nr_child) { | |
354 | if (caa_unlikely(node_flag_ptr)) | |
355 | *node_flag_ptr = NULL; | |
356 | return NULL; | |
357 | } | |
358 | pointers = (struct cds_ja_inode_flag **) align_ptr_size(&values[type->max_linear_child]); | |
359 | ptr = rcu_dereference(pointers[i]); | |
360 | if (caa_unlikely(node_flag_ptr)) | |
361 | *node_flag_ptr = &pointers[i]; | |
362 | return ptr; | |
363 | } | |
364 | ||
365 | static | |
366 | struct cds_ja_inode_flag *ja_linear_node_get_direction(const struct cds_ja_type *type, | |
367 | struct cds_ja_inode *node, | |
368 | int n, uint8_t *result_key, | |
369 | enum ja_direction dir) | |
370 | { | |
371 | uint8_t nr_child; | |
372 | uint8_t *values; | |
373 | struct cds_ja_inode_flag **pointers; | |
374 | struct cds_ja_inode_flag *ptr, *match_ptr = NULL; | |
375 | unsigned int i; | |
376 | int match_v; | |
377 | ||
378 | assert(type->type_class == RCU_JA_LINEAR || type->type_class == RCU_JA_POOL); | |
379 | assert(dir == JA_LEFT || dir == JA_RIGHT); | |
380 | ||
381 | if (dir == JA_LEFT) { | |
382 | match_v = -1; | |
383 | } else { | |
384 | match_v = JA_ENTRY_PER_NODE; | |
385 | } | |
386 | ||
387 | nr_child = ja_linear_node_get_nr_child(type, node); | |
388 | cmm_smp_rmb(); /* read nr_child before values and pointers */ | |
389 | assert(nr_child <= type->max_linear_child); | |
390 | assert(type->type_class != RCU_JA_LINEAR || nr_child >= type->min_child); | |
391 | ||
392 | values = &node->u.data[1]; | |
393 | pointers = (struct cds_ja_inode_flag **) align_ptr_size(&values[type->max_linear_child]); | |
394 | for (i = 0; i < nr_child; i++) { | |
395 | unsigned int v; | |
396 | ||
397 | v = CMM_LOAD_SHARED(values[i]); | |
398 | ptr = CMM_LOAD_SHARED(pointers[i]); | |
399 | if (!ptr) | |
400 | continue; | |
401 | if (dir == JA_LEFT) { | |
402 | if ((int) v < n && (int) v > match_v) { | |
403 | match_v = v; | |
404 | match_ptr = ptr; | |
405 | } | |
406 | } else { | |
407 | if ((int) v > n && (int) v < match_v) { | |
408 | match_v = v; | |
409 | match_ptr = ptr; | |
410 | } | |
411 | } | |
412 | } | |
413 | ||
414 | if (!match_ptr) { | |
415 | return NULL; | |
416 | } | |
417 | assert(match_v >= 0 && match_v < JA_ENTRY_PER_NODE); | |
418 | ||
419 | *result_key = (uint8_t) match_v; | |
420 | return rcu_dereference(match_ptr); | |
421 | } | |
422 | ||
423 | static | |
424 | void ja_linear_node_get_ith_pos(const struct cds_ja_type *type, | |
425 | struct cds_ja_inode *node, | |
426 | uint8_t i, | |
427 | uint8_t *v, | |
428 | struct cds_ja_inode_flag **iter) | |
429 | { | |
430 | uint8_t *values; | |
431 | struct cds_ja_inode_flag **pointers; | |
432 | ||
433 | assert(type->type_class == RCU_JA_LINEAR || type->type_class == RCU_JA_POOL); | |
434 | assert(i < ja_linear_node_get_nr_child(type, node)); | |
435 | ||
436 | values = &node->u.data[1]; | |
437 | *v = values[i]; | |
438 | pointers = (struct cds_ja_inode_flag **) align_ptr_size(&values[type->max_linear_child]); | |
439 | *iter = pointers[i]; | |
440 | } | |
441 | ||
442 | static | |
443 | struct cds_ja_inode_flag *ja_pool_node_get_nth(const struct cds_ja_type *type, | |
444 | struct cds_ja_inode *node, | |
445 | struct cds_ja_inode_flag *node_flag, | |
446 | struct cds_ja_inode_flag ***node_flag_ptr, | |
447 | uint8_t n) | |
448 | { | |
449 | struct cds_ja_inode *linear; | |
450 | ||
451 | assert(type->type_class == RCU_JA_POOL); | |
452 | ||
453 | switch (type->nr_pool_order) { | |
454 | case 1: | |
455 | { | |
456 | unsigned long bitsel, index; | |
457 | ||
458 | bitsel = ja_node_pool_1d_bitsel(node_flag); | |
459 | assert(bitsel < CHAR_BIT); | |
460 | index = ((unsigned long) n >> bitsel) & 0x1; | |
461 | linear = (struct cds_ja_inode *) &node->u.data[index << type->pool_size_order]; | |
462 | break; | |
463 | } | |
464 | case 2: | |
465 | { | |
466 | unsigned long bitsel[2], index[2], rindex; | |
467 | ||
468 | ja_node_pool_2d_bitsel(node_flag, bitsel); | |
469 | assert(bitsel[0] < CHAR_BIT); | |
470 | assert(bitsel[1] < CHAR_BIT); | |
471 | index[0] = ((unsigned long) n >> bitsel[0]) & 0x1; | |
472 | index[0] <<= 1; | |
473 | index[1] = ((unsigned long) n >> bitsel[1]) & 0x1; | |
474 | rindex = index[0] | index[1]; | |
475 | linear = (struct cds_ja_inode *) &node->u.data[rindex << type->pool_size_order]; | |
476 | break; | |
477 | } | |
478 | default: | |
479 | linear = NULL; | |
480 | assert(0); | |
481 | } | |
482 | return ja_linear_node_get_nth(type, linear, node_flag_ptr, n); | |
483 | } | |
484 | ||
485 | static | |
486 | struct cds_ja_inode *ja_pool_node_get_ith_pool(const struct cds_ja_type *type, | |
487 | struct cds_ja_inode *node, | |
488 | uint8_t i) | |
489 | { | |
490 | assert(type->type_class == RCU_JA_POOL); | |
491 | return (struct cds_ja_inode *) | |
492 | &node->u.data[(unsigned int) i << type->pool_size_order]; | |
493 | } | |
494 | ||
495 | static | |
496 | struct cds_ja_inode_flag *ja_pool_node_get_direction(const struct cds_ja_type *type, | |
497 | struct cds_ja_inode *node, | |
498 | int n, uint8_t *result_key, | |
499 | enum ja_direction dir) | |
500 | { | |
501 | unsigned int pool_nr; | |
502 | int match_v; | |
503 | struct cds_ja_inode_flag *match_node_flag = NULL; | |
504 | ||
505 | assert(type->type_class == RCU_JA_POOL); | |
506 | assert(dir == JA_LEFT || dir == JA_RIGHT); | |
507 | ||
508 | if (dir == JA_LEFT) { | |
509 | match_v = -1; | |
510 | } else { | |
511 | match_v = JA_ENTRY_PER_NODE; | |
512 | } | |
513 | ||
514 | for (pool_nr = 0; pool_nr < (1U << type->nr_pool_order); pool_nr++) { | |
515 | struct cds_ja_inode *pool = | |
516 | ja_pool_node_get_ith_pool(type, | |
517 | node, pool_nr); | |
518 | uint8_t nr_child = | |
519 | ja_linear_node_get_nr_child(type, pool); | |
520 | unsigned int j; | |
521 | ||
522 | for (j = 0; j < nr_child; j++) { | |
523 | struct cds_ja_inode_flag *iter; | |
524 | uint8_t v; | |
525 | ||
526 | ja_linear_node_get_ith_pos(type, pool, | |
527 | j, &v, &iter); | |
528 | if (!iter) | |
529 | continue; | |
530 | if (dir == JA_LEFT) { | |
531 | if ((int) v < n && (int) v > match_v) { | |
532 | match_v = v; | |
533 | match_node_flag = iter; | |
534 | } | |
535 | } else { | |
536 | if ((int) v > n && (int) v < match_v) { | |
537 | match_v = v; | |
538 | match_node_flag = iter; | |
539 | } | |
540 | } | |
541 | } | |
542 | } | |
543 | if (match_node_flag) | |
544 | *result_key = (uint8_t) match_v; | |
545 | return match_node_flag; | |
546 | } | |
547 | ||
548 | static | |
549 | struct cds_ja_inode_flag *ja_pigeon_node_get_nth(const struct cds_ja_type *type, | |
550 | struct cds_ja_inode *node, | |
551 | struct cds_ja_inode_flag ***node_flag_ptr, | |
552 | uint8_t n) | |
553 | { | |
554 | struct cds_ja_inode_flag **child_node_flag_ptr; | |
555 | struct cds_ja_inode_flag *child_node_flag; | |
556 | ||
557 | assert(type->type_class == RCU_JA_PIGEON); | |
558 | child_node_flag_ptr = &((struct cds_ja_inode_flag **) node->u.data)[n]; | |
559 | child_node_flag = rcu_dereference(*child_node_flag_ptr); | |
560 | dbg_printf("ja_pigeon_node_get_nth child_node_flag_ptr %p\n", | |
561 | child_node_flag_ptr); | |
562 | if (caa_unlikely(node_flag_ptr)) | |
563 | *node_flag_ptr = child_node_flag_ptr; | |
564 | return child_node_flag; | |
565 | } | |
566 | ||
567 | static | |
568 | struct cds_ja_inode_flag *ja_pigeon_node_get_direction(const struct cds_ja_type *type, | |
569 | struct cds_ja_inode *node, | |
570 | int n, uint8_t *result_key, | |
571 | enum ja_direction dir) | |
572 | { | |
573 | struct cds_ja_inode_flag **child_node_flag_ptr; | |
574 | struct cds_ja_inode_flag *child_node_flag; | |
575 | int i; | |
576 | ||
577 | assert(type->type_class == RCU_JA_PIGEON); | |
578 | assert(dir == JA_LEFT || dir == JA_RIGHT); | |
579 | ||
580 | if (dir == JA_LEFT) { | |
581 | /* n - 1 is first value left of n */ | |
582 | for (i = n - 1; i >= 0; i--) { | |
583 | child_node_flag_ptr = &((struct cds_ja_inode_flag **) node->u.data)[i]; | |
584 | child_node_flag = rcu_dereference(*child_node_flag_ptr); | |
585 | if (child_node_flag) { | |
586 | dbg_printf("ja_pigeon_node_get_left child_node_flag %p\n", | |
587 | child_node_flag); | |
588 | *result_key = (uint8_t) i; | |
589 | return child_node_flag; | |
590 | } | |
591 | } | |
592 | } else { | |
593 | /* n + 1 is first value right of n */ | |
594 | for (i = n + 1; i < JA_ENTRY_PER_NODE; i++) { | |
595 | child_node_flag_ptr = &((struct cds_ja_inode_flag **) node->u.data)[i]; | |
596 | child_node_flag = rcu_dereference(*child_node_flag_ptr); | |
597 | if (child_node_flag) { | |
598 | dbg_printf("ja_pigeon_node_get_right child_node_flag %p\n", | |
599 | child_node_flag); | |
600 | *result_key = (uint8_t) i; | |
601 | return child_node_flag; | |
602 | } | |
603 | } | |
604 | } | |
605 | return NULL; | |
606 | } | |
607 | ||
608 | static | |
609 | struct cds_ja_inode_flag *ja_pigeon_node_get_ith_pos(const struct cds_ja_type *type, | |
610 | struct cds_ja_inode *node, | |
611 | uint8_t i) | |
612 | { | |
613 | return ja_pigeon_node_get_nth(type, node, NULL, i); | |
614 | } | |
615 | ||
616 | /* | |
617 | * ja_node_get_nth: get nth item from a node. | |
618 | * node_flag is already rcu_dereference'd. | |
619 | */ | |
620 | static | |
621 | struct cds_ja_inode_flag *ja_node_get_nth(struct cds_ja_inode_flag *node_flag, | |
622 | struct cds_ja_inode_flag ***node_flag_ptr, | |
623 | uint8_t n) | |
624 | { | |
625 | unsigned int type_index; | |
626 | struct cds_ja_inode *node; | |
627 | const struct cds_ja_type *type; | |
628 | ||
629 | node = ja_node_ptr(node_flag); | |
630 | assert(node != NULL); | |
631 | type_index = ja_node_type(node_flag); | |
632 | type = &ja_types[type_index]; | |
633 | ||
634 | switch (type->type_class) { | |
635 | case RCU_JA_LINEAR: | |
636 | return ja_linear_node_get_nth(type, node, | |
637 | node_flag_ptr, n); | |
638 | case RCU_JA_POOL: | |
639 | return ja_pool_node_get_nth(type, node, node_flag, | |
640 | node_flag_ptr, n); | |
641 | case RCU_JA_PIGEON: | |
642 | return ja_pigeon_node_get_nth(type, node, | |
643 | node_flag_ptr, n); | |
644 | default: | |
645 | assert(0); | |
646 | return (void *) -1UL; | |
647 | } | |
648 | } | |
649 | ||
650 | static | |
651 | struct cds_ja_inode_flag *ja_node_get_direction(struct cds_ja_inode_flag *node_flag, | |
652 | int n, uint8_t *result_key, | |
653 | enum ja_direction dir) | |
654 | { | |
655 | unsigned int type_index; | |
656 | struct cds_ja_inode *node; | |
657 | const struct cds_ja_type *type; | |
658 | ||
659 | node = ja_node_ptr(node_flag); | |
660 | assert(node != NULL); | |
661 | type_index = ja_node_type(node_flag); | |
662 | type = &ja_types[type_index]; | |
663 | ||
664 | switch (type->type_class) { | |
665 | case RCU_JA_LINEAR: | |
666 | return ja_linear_node_get_direction(type, node, n, result_key, dir); | |
667 | case RCU_JA_POOL: | |
668 | return ja_pool_node_get_direction(type, node, n, result_key, dir); | |
669 | case RCU_JA_PIGEON: | |
670 | return ja_pigeon_node_get_direction(type, node, n, result_key, dir); | |
671 | default: | |
672 | assert(0); | |
673 | return (void *) -1UL; | |
674 | } | |
675 | } | |
676 | ||
677 | static | |
678 | struct cds_ja_inode_flag *ja_node_get_leftright(struct cds_ja_inode_flag *node_flag, | |
679 | unsigned int n, uint8_t *result_key, | |
680 | enum ja_direction dir) | |
681 | { | |
682 | return ja_node_get_direction(node_flag, n, result_key, dir); | |
683 | } | |
684 | ||
685 | static | |
686 | struct cds_ja_inode_flag *ja_node_get_minmax(struct cds_ja_inode_flag *node_flag, | |
687 | uint8_t *result_key, | |
688 | enum ja_direction dir) | |
689 | { | |
690 | switch (dir) { | |
691 | case JA_LEFTMOST: | |
692 | return ja_node_get_direction(node_flag, | |
693 | -1, result_key, JA_RIGHT); | |
694 | case JA_RIGHTMOST: | |
695 | return ja_node_get_direction(node_flag, | |
696 | JA_ENTRY_PER_NODE, result_key, JA_LEFT); | |
697 | default: | |
698 | assert(0); | |
699 | } | |
700 | } | |
701 | ||
702 | static | |
703 | int ja_linear_node_set_nth(const struct cds_ja_type *type, | |
704 | struct cds_ja_inode *node, | |
705 | struct cds_ja_shadow_node *shadow_node, | |
706 | uint8_t n, | |
707 | struct cds_ja_inode_flag *child_node_flag) | |
708 | { | |
709 | uint8_t nr_child; | |
710 | uint8_t *values, *nr_child_ptr; | |
711 | struct cds_ja_inode_flag **pointers; | |
712 | unsigned int i, unused = 0; | |
713 | ||
714 | assert(type->type_class == RCU_JA_LINEAR || type->type_class == RCU_JA_POOL); | |
715 | ||
716 | nr_child_ptr = &node->u.data[0]; | |
717 | dbg_printf("linear set nth: n %u, nr_child_ptr %p\n", | |
718 | (unsigned int) n, nr_child_ptr); | |
719 | nr_child = *nr_child_ptr; | |
720 | assert(nr_child <= type->max_linear_child); | |
721 | ||
722 | values = &node->u.data[1]; | |
723 | pointers = (struct cds_ja_inode_flag **) align_ptr_size(&values[type->max_linear_child]); | |
724 | /* Check if node value is already populated */ | |
725 | for (i = 0; i < nr_child; i++) { | |
726 | if (values[i] == n) { | |
727 | if (pointers[i]) | |
728 | return -EEXIST; | |
729 | else | |
730 | break; | |
731 | } else { | |
732 | if (!pointers[i]) | |
733 | unused++; | |
734 | } | |
735 | } | |
736 | if (i == nr_child && nr_child >= type->max_linear_child) { | |
737 | if (unused) | |
738 | return -ERANGE; /* recompact node */ | |
739 | else | |
740 | return -ENOSPC; /* No space left in this node type */ | |
741 | } | |
742 | ||
743 | assert(pointers[i] == NULL); | |
744 | rcu_assign_pointer(pointers[i], child_node_flag); | |
745 | /* If we expanded the nr_child, increment it */ | |
746 | if (i == nr_child) { | |
747 | CMM_STORE_SHARED(values[nr_child], n); | |
748 | /* write pointer and value before nr_child */ | |
749 | cmm_smp_wmb(); | |
750 | CMM_STORE_SHARED(*nr_child_ptr, nr_child + 1); | |
751 | } | |
752 | shadow_node->nr_child++; | |
753 | dbg_printf("linear set nth: %u child, shadow: %u child, for node %p shadow %p\n", | |
754 | (unsigned int) CMM_LOAD_SHARED(*nr_child_ptr), | |
755 | (unsigned int) shadow_node->nr_child, | |
756 | node, shadow_node); | |
757 | ||
758 | return 0; | |
759 | } | |
760 | ||
761 | static | |
762 | int ja_pool_node_set_nth(const struct cds_ja_type *type, | |
763 | struct cds_ja_inode *node, | |
764 | struct cds_ja_inode_flag *node_flag, | |
765 | struct cds_ja_shadow_node *shadow_node, | |
766 | uint8_t n, | |
767 | struct cds_ja_inode_flag *child_node_flag) | |
768 | { | |
769 | struct cds_ja_inode *linear; | |
770 | ||
771 | assert(type->type_class == RCU_JA_POOL); | |
772 | ||
773 | switch (type->nr_pool_order) { | |
774 | case 1: | |
775 | { | |
776 | unsigned long bitsel, index; | |
777 | ||
778 | bitsel = ja_node_pool_1d_bitsel(node_flag); | |
779 | assert(bitsel < CHAR_BIT); | |
780 | index = ((unsigned long) n >> bitsel) & 0x1; | |
781 | linear = (struct cds_ja_inode *) &node->u.data[index << type->pool_size_order]; | |
782 | break; | |
783 | } | |
784 | case 2: | |
785 | { | |
786 | unsigned long bitsel[2], index[2], rindex; | |
787 | ||
788 | ja_node_pool_2d_bitsel(node_flag, bitsel); | |
789 | assert(bitsel[0] < CHAR_BIT); | |
790 | assert(bitsel[1] < CHAR_BIT); | |
791 | index[0] = ((unsigned long) n >> bitsel[0]) & 0x1; | |
792 | index[0] <<= 1; | |
793 | index[1] = ((unsigned long) n >> bitsel[1]) & 0x1; | |
794 | rindex = index[0] | index[1]; | |
795 | linear = (struct cds_ja_inode *) &node->u.data[rindex << type->pool_size_order]; | |
796 | break; | |
797 | } | |
798 | default: | |
799 | linear = NULL; | |
800 | assert(0); | |
801 | } | |
802 | ||
803 | return ja_linear_node_set_nth(type, linear, shadow_node, | |
804 | n, child_node_flag); | |
805 | } | |
806 | ||
807 | static | |
808 | int ja_pigeon_node_set_nth(const struct cds_ja_type *type, | |
809 | struct cds_ja_inode *node, | |
810 | struct cds_ja_shadow_node *shadow_node, | |
811 | uint8_t n, | |
812 | struct cds_ja_inode_flag *child_node_flag) | |
813 | { | |
814 | struct cds_ja_inode_flag **ptr; | |
815 | ||
816 | assert(type->type_class == RCU_JA_PIGEON); | |
817 | ptr = &((struct cds_ja_inode_flag **) node->u.data)[n]; | |
818 | if (*ptr) | |
819 | return -EEXIST; | |
820 | rcu_assign_pointer(*ptr, child_node_flag); | |
821 | shadow_node->nr_child++; | |
822 | return 0; | |
823 | } | |
824 | ||
825 | /* | |
826 | * _ja_node_set_nth: set nth item within a node. Return an error | |
827 | * (negative error value) if it is already there. | |
828 | */ | |
829 | static | |
830 | int _ja_node_set_nth(const struct cds_ja_type *type, | |
831 | struct cds_ja_inode *node, | |
832 | struct cds_ja_inode_flag *node_flag, | |
833 | struct cds_ja_shadow_node *shadow_node, | |
834 | uint8_t n, | |
835 | struct cds_ja_inode_flag *child_node_flag) | |
836 | { | |
837 | switch (type->type_class) { | |
838 | case RCU_JA_LINEAR: | |
839 | return ja_linear_node_set_nth(type, node, shadow_node, n, | |
840 | child_node_flag); | |
841 | case RCU_JA_POOL: | |
842 | return ja_pool_node_set_nth(type, node, node_flag, shadow_node, n, | |
843 | child_node_flag); | |
844 | case RCU_JA_PIGEON: | |
845 | return ja_pigeon_node_set_nth(type, node, shadow_node, n, | |
846 | child_node_flag); | |
847 | case RCU_JA_NULL: | |
848 | return -ENOSPC; | |
849 | default: | |
850 | assert(0); | |
851 | return -EINVAL; | |
852 | } | |
853 | ||
854 | return 0; | |
855 | } | |
856 | ||
857 | static | |
858 | int ja_linear_node_clear_ptr(const struct cds_ja_type *type, | |
859 | struct cds_ja_inode *node, | |
860 | struct cds_ja_shadow_node *shadow_node, | |
861 | struct cds_ja_inode_flag **node_flag_ptr) | |
862 | { | |
863 | uint8_t nr_child; | |
864 | uint8_t *nr_child_ptr; | |
865 | ||
866 | assert(type->type_class == RCU_JA_LINEAR || type->type_class == RCU_JA_POOL); | |
867 | ||
868 | nr_child_ptr = &node->u.data[0]; | |
869 | nr_child = *nr_child_ptr; | |
870 | assert(nr_child <= type->max_linear_child); | |
871 | ||
872 | if (type->type_class == RCU_JA_LINEAR) { | |
873 | assert(!shadow_node->fallback_removal_count); | |
874 | if (shadow_node->nr_child <= type->min_child) { | |
875 | /* We need to try recompacting the node */ | |
876 | return -EFBIG; | |
877 | } | |
878 | } | |
879 | dbg_printf("linear clear ptr: nr_child_ptr %p\n", nr_child_ptr); | |
880 | assert(*node_flag_ptr != NULL); | |
881 | rcu_assign_pointer(*node_flag_ptr, NULL); | |
882 | /* | |
883 | * Value and nr_child are never changed (would cause ABA issue). | |
884 | * Instead, we leave the pointer to NULL and recompact the node | |
885 | * once in a while. It is allowed to set a NULL pointer to a new | |
886 | * value without recompaction though. | |
887 | * Only update the shadow node accounting. | |
888 | */ | |
889 | shadow_node->nr_child--; | |
890 | dbg_printf("linear clear ptr: %u child, shadow: %u child, for node %p shadow %p\n", | |
891 | (unsigned int) CMM_LOAD_SHARED(*nr_child_ptr), | |
892 | (unsigned int) shadow_node->nr_child, | |
893 | node, shadow_node); | |
894 | return 0; | |
895 | } | |
896 | ||
897 | static | |
898 | int ja_pool_node_clear_ptr(const struct cds_ja_type *type, | |
899 | struct cds_ja_inode *node, | |
900 | struct cds_ja_inode_flag *node_flag, | |
901 | struct cds_ja_shadow_node *shadow_node, | |
902 | struct cds_ja_inode_flag **node_flag_ptr, | |
903 | uint8_t n) | |
904 | { | |
905 | struct cds_ja_inode *linear; | |
906 | ||
907 | assert(type->type_class == RCU_JA_POOL); | |
908 | ||
909 | if (shadow_node->fallback_removal_count) { | |
910 | shadow_node->fallback_removal_count--; | |
911 | } else { | |
912 | /* We should try recompacting the node */ | |
913 | if (shadow_node->nr_child <= type->min_child) | |
914 | return -EFBIG; | |
915 | } | |
916 | ||
917 | switch (type->nr_pool_order) { | |
918 | case 1: | |
919 | { | |
920 | unsigned long bitsel, index; | |
921 | ||
922 | bitsel = ja_node_pool_1d_bitsel(node_flag); | |
923 | assert(bitsel < CHAR_BIT); | |
924 | index = ((unsigned long) n >> bitsel) & type->nr_pool_order; | |
925 | linear = (struct cds_ja_inode *) &node->u.data[index << type->pool_size_order]; | |
926 | break; | |
927 | } | |
928 | case 2: | |
929 | { | |
930 | unsigned long bitsel[2], index[2], rindex; | |
931 | ||
932 | ja_node_pool_2d_bitsel(node_flag, bitsel); | |
933 | assert(bitsel[0] < CHAR_BIT); | |
934 | assert(bitsel[1] < CHAR_BIT); | |
935 | index[0] = ((unsigned long) n >> bitsel[0]) & 0x1; | |
936 | index[0] <<= 1; | |
937 | index[1] = ((unsigned long) n >> bitsel[1]) & 0x1; | |
938 | rindex = index[0] | index[1]; | |
939 | linear = (struct cds_ja_inode *) &node->u.data[rindex << type->pool_size_order]; | |
940 | break; | |
941 | } | |
942 | default: | |
943 | linear = NULL; | |
944 | assert(0); | |
945 | } | |
946 | ||
947 | return ja_linear_node_clear_ptr(type, linear, shadow_node, node_flag_ptr); | |
948 | } | |
949 | ||
950 | static | |
951 | int ja_pigeon_node_clear_ptr(const struct cds_ja_type *type, | |
952 | struct cds_ja_inode *node, | |
953 | struct cds_ja_shadow_node *shadow_node, | |
954 | struct cds_ja_inode_flag **node_flag_ptr) | |
955 | { | |
956 | assert(type->type_class == RCU_JA_PIGEON); | |
957 | ||
958 | if (shadow_node->fallback_removal_count) { | |
959 | shadow_node->fallback_removal_count--; | |
960 | } else { | |
961 | /* We should try recompacting the node */ | |
962 | if (shadow_node->nr_child <= type->min_child) | |
963 | return -EFBIG; | |
964 | } | |
965 | dbg_printf("ja_pigeon_node_clear_ptr: clearing ptr: %p\n", *node_flag_ptr); | |
966 | rcu_assign_pointer(*node_flag_ptr, NULL); | |
967 | shadow_node->nr_child--; | |
968 | return 0; | |
969 | } | |
970 | ||
971 | /* | |
972 | * _ja_node_clear_ptr: clear ptr item within a node. Return an error | |
973 | * (negative error value) if it is not found (-ENOENT). | |
974 | */ | |
975 | static | |
976 | int _ja_node_clear_ptr(const struct cds_ja_type *type, | |
977 | struct cds_ja_inode *node, | |
978 | struct cds_ja_inode_flag *node_flag, | |
979 | struct cds_ja_shadow_node *shadow_node, | |
980 | struct cds_ja_inode_flag **node_flag_ptr, | |
981 | uint8_t n) | |
982 | { | |
983 | switch (type->type_class) { | |
984 | case RCU_JA_LINEAR: | |
985 | return ja_linear_node_clear_ptr(type, node, shadow_node, node_flag_ptr); | |
986 | case RCU_JA_POOL: | |
987 | return ja_pool_node_clear_ptr(type, node, node_flag, shadow_node, node_flag_ptr, n); | |
988 | case RCU_JA_PIGEON: | |
989 | return ja_pigeon_node_clear_ptr(type, node, shadow_node, node_flag_ptr); | |
990 | case RCU_JA_NULL: | |
991 | return -ENOENT; | |
992 | default: | |
993 | assert(0); | |
994 | return -EINVAL; | |
995 | } | |
996 | ||
997 | return 0; | |
998 | } | |
999 | ||
1000 | /* | |
1001 | * Calculate bit distribution. Returns the bit (0 to 7) that splits the | |
1002 | * distribution in two sub-distributions containing as much elements one | |
1003 | * compared to the other. | |
1004 | */ | |
1005 | static | |
1006 | unsigned int ja_node_sum_distribution_1d(enum ja_recompact mode, | |
1007 | struct cds_ja *ja, | |
1008 | unsigned int type_index, | |
1009 | const struct cds_ja_type *type, | |
1010 | struct cds_ja_inode *node, | |
1011 | struct cds_ja_shadow_node *shadow_node, | |
1012 | uint8_t n, | |
1013 | struct cds_ja_inode_flag *child_node_flag, | |
1014 | struct cds_ja_inode_flag **nullify_node_flag_ptr) | |
1015 | { | |
1016 | uint8_t nr_one[JA_BITS_PER_BYTE]; | |
1017 | unsigned int bitsel = 0, bit_i, overall_best_distance = UINT_MAX; | |
1018 | unsigned int distrib_nr_child = 0; | |
1019 | ||
1020 | memset(nr_one, 0, sizeof(nr_one)); | |
1021 | ||
1022 | switch (type->type_class) { | |
1023 | case RCU_JA_LINEAR: | |
1024 | { | |
1025 | uint8_t nr_child = | |
1026 | ja_linear_node_get_nr_child(type, node); | |
1027 | unsigned int i; | |
1028 | ||
1029 | for (i = 0; i < nr_child; i++) { | |
1030 | struct cds_ja_inode_flag *iter; | |
1031 | uint8_t v; | |
1032 | ||
1033 | ja_linear_node_get_ith_pos(type, node, i, &v, &iter); | |
1034 | if (!iter) | |
1035 | continue; | |
1036 | if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) | |
1037 | continue; | |
1038 | for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) { | |
1039 | if (v & (1U << bit_i)) | |
1040 | nr_one[bit_i]++; | |
1041 | } | |
1042 | distrib_nr_child++; | |
1043 | } | |
1044 | break; | |
1045 | } | |
1046 | case RCU_JA_POOL: | |
1047 | { | |
1048 | unsigned int pool_nr; | |
1049 | ||
1050 | for (pool_nr = 0; pool_nr < (1U << type->nr_pool_order); pool_nr++) { | |
1051 | struct cds_ja_inode *pool = | |
1052 | ja_pool_node_get_ith_pool(type, | |
1053 | node, pool_nr); | |
1054 | uint8_t nr_child = | |
1055 | ja_linear_node_get_nr_child(type, pool); | |
1056 | unsigned int j; | |
1057 | ||
1058 | for (j = 0; j < nr_child; j++) { | |
1059 | struct cds_ja_inode_flag *iter; | |
1060 | uint8_t v; | |
1061 | ||
1062 | ja_linear_node_get_ith_pos(type, pool, | |
1063 | j, &v, &iter); | |
1064 | if (!iter) | |
1065 | continue; | |
1066 | if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) | |
1067 | continue; | |
1068 | for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) { | |
1069 | if (v & (1U << bit_i)) | |
1070 | nr_one[bit_i]++; | |
1071 | } | |
1072 | distrib_nr_child++; | |
1073 | } | |
1074 | } | |
1075 | break; | |
1076 | } | |
1077 | case RCU_JA_PIGEON: | |
1078 | { | |
1079 | unsigned int i; | |
1080 | ||
1081 | assert(mode == JA_RECOMPACT_DEL); | |
1082 | for (i = 0; i < JA_ENTRY_PER_NODE; i++) { | |
1083 | struct cds_ja_inode_flag *iter; | |
1084 | ||
1085 | iter = ja_pigeon_node_get_ith_pos(type, node, i); | |
1086 | if (!iter) | |
1087 | continue; | |
1088 | if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) | |
1089 | continue; | |
1090 | for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) { | |
1091 | if (i & (1U << bit_i)) | |
1092 | nr_one[bit_i]++; | |
1093 | } | |
1094 | distrib_nr_child++; | |
1095 | } | |
1096 | break; | |
1097 | } | |
1098 | case RCU_JA_NULL: | |
1099 | assert(mode == JA_RECOMPACT_ADD_NEXT); | |
1100 | break; | |
1101 | default: | |
1102 | assert(0); | |
1103 | break; | |
1104 | } | |
1105 | ||
1106 | if (mode == JA_RECOMPACT_ADD_NEXT || mode == JA_RECOMPACT_ADD_SAME) { | |
1107 | for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) { | |
1108 | if (n & (1U << bit_i)) | |
1109 | nr_one[bit_i]++; | |
1110 | } | |
1111 | distrib_nr_child++; | |
1112 | } | |
1113 | ||
1114 | /* | |
1115 | * The best bit selector is that for which the number of ones is | |
1116 | * closest to half of the number of children in the | |
1117 | * distribution. We calculate the distance using the double of | |
1118 | * the sub-distribution sizes to eliminate truncation error. | |
1119 | */ | |
1120 | for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) { | |
1121 | unsigned int distance_to_best; | |
1122 | ||
1123 | distance_to_best = abs_int(((unsigned int) nr_one[bit_i] << 1U) - distrib_nr_child); | |
1124 | if (distance_to_best < overall_best_distance) { | |
1125 | overall_best_distance = distance_to_best; | |
1126 | bitsel = bit_i; | |
1127 | } | |
1128 | } | |
1129 | dbg_printf("1 dimension pool bit selection: (%u)\n", bitsel); | |
1130 | return bitsel; | |
1131 | } | |
1132 | ||
1133 | /* | |
1134 | * Calculate bit distribution in two dimensions. Returns the two bits | |
1135 | * (each 0 to 7) that splits the distribution in four sub-distributions | |
1136 | * containing as much elements one compared to the other. | |
1137 | */ | |
1138 | static | |
1139 | void ja_node_sum_distribution_2d(enum ja_recompact mode, | |
1140 | struct cds_ja *ja, | |
1141 | unsigned int type_index, | |
1142 | const struct cds_ja_type *type, | |
1143 | struct cds_ja_inode *node, | |
1144 | struct cds_ja_shadow_node *shadow_node, | |
1145 | uint8_t n, | |
1146 | struct cds_ja_inode_flag *child_node_flag, | |
1147 | struct cds_ja_inode_flag **nullify_node_flag_ptr, | |
1148 | unsigned int *_bitsel) | |
1149 | { | |
1150 | uint8_t nr_2d_11[JA_BITS_PER_BYTE][JA_BITS_PER_BYTE], | |
1151 | nr_2d_10[JA_BITS_PER_BYTE][JA_BITS_PER_BYTE], | |
1152 | nr_2d_01[JA_BITS_PER_BYTE][JA_BITS_PER_BYTE], | |
1153 | nr_2d_00[JA_BITS_PER_BYTE][JA_BITS_PER_BYTE]; | |
1154 | unsigned int bitsel[2] = { 0, 1 }; | |
1155 | unsigned int bit_i, bit_j; | |
1156 | int overall_best_distance = INT_MAX; | |
1157 | unsigned int distrib_nr_child = 0; | |
1158 | ||
1159 | memset(nr_2d_11, 0, sizeof(nr_2d_11)); | |
1160 | memset(nr_2d_10, 0, sizeof(nr_2d_10)); | |
1161 | memset(nr_2d_01, 0, sizeof(nr_2d_01)); | |
1162 | memset(nr_2d_00, 0, sizeof(nr_2d_00)); | |
1163 | ||
1164 | switch (type->type_class) { | |
1165 | case RCU_JA_LINEAR: | |
1166 | { | |
1167 | uint8_t nr_child = | |
1168 | ja_linear_node_get_nr_child(type, node); | |
1169 | unsigned int i; | |
1170 | ||
1171 | for (i = 0; i < nr_child; i++) { | |
1172 | struct cds_ja_inode_flag *iter; | |
1173 | uint8_t v; | |
1174 | ||
1175 | ja_linear_node_get_ith_pos(type, node, i, &v, &iter); | |
1176 | if (!iter) | |
1177 | continue; | |
1178 | if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) | |
1179 | continue; | |
1180 | for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) { | |
1181 | for (bit_j = 0; bit_j < bit_i; bit_j++) { | |
1182 | if (v & (1U << bit_i)) { | |
1183 | if (v & (1U << bit_j)) { | |
1184 | nr_2d_11[bit_i][bit_j]++; | |
1185 | } else { | |
1186 | nr_2d_10[bit_i][bit_j]++; | |
1187 | } | |
1188 | } else { | |
1189 | if (v & (1U << bit_j)) { | |
1190 | nr_2d_01[bit_i][bit_j]++; | |
1191 | } else { | |
1192 | nr_2d_00[bit_i][bit_j]++; | |
1193 | } | |
1194 | } | |
1195 | } | |
1196 | } | |
1197 | distrib_nr_child++; | |
1198 | } | |
1199 | break; | |
1200 | } | |
1201 | case RCU_JA_POOL: | |
1202 | { | |
1203 | unsigned int pool_nr; | |
1204 | ||
1205 | for (pool_nr = 0; pool_nr < (1U << type->nr_pool_order); pool_nr++) { | |
1206 | struct cds_ja_inode *pool = | |
1207 | ja_pool_node_get_ith_pool(type, | |
1208 | node, pool_nr); | |
1209 | uint8_t nr_child = | |
1210 | ja_linear_node_get_nr_child(type, pool); | |
1211 | unsigned int j; | |
1212 | ||
1213 | for (j = 0; j < nr_child; j++) { | |
1214 | struct cds_ja_inode_flag *iter; | |
1215 | uint8_t v; | |
1216 | ||
1217 | ja_linear_node_get_ith_pos(type, pool, | |
1218 | j, &v, &iter); | |
1219 | if (!iter) | |
1220 | continue; | |
1221 | if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) | |
1222 | continue; | |
1223 | for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) { | |
1224 | for (bit_j = 0; bit_j < bit_i; bit_j++) { | |
1225 | if (v & (1U << bit_i)) { | |
1226 | if (v & (1U << bit_j)) { | |
1227 | nr_2d_11[bit_i][bit_j]++; | |
1228 | } else { | |
1229 | nr_2d_10[bit_i][bit_j]++; | |
1230 | } | |
1231 | } else { | |
1232 | if (v & (1U << bit_j)) { | |
1233 | nr_2d_01[bit_i][bit_j]++; | |
1234 | } else { | |
1235 | nr_2d_00[bit_i][bit_j]++; | |
1236 | } | |
1237 | } | |
1238 | } | |
1239 | } | |
1240 | distrib_nr_child++; | |
1241 | } | |
1242 | } | |
1243 | break; | |
1244 | } | |
1245 | case RCU_JA_PIGEON: | |
1246 | { | |
1247 | unsigned int i; | |
1248 | ||
1249 | assert(mode == JA_RECOMPACT_DEL); | |
1250 | for (i = 0; i < JA_ENTRY_PER_NODE; i++) { | |
1251 | struct cds_ja_inode_flag *iter; | |
1252 | ||
1253 | iter = ja_pigeon_node_get_ith_pos(type, node, i); | |
1254 | if (!iter) | |
1255 | continue; | |
1256 | if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) | |
1257 | continue; | |
1258 | for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) { | |
1259 | for (bit_j = 0; bit_j < bit_i; bit_j++) { | |
1260 | if (i & (1U << bit_i)) { | |
1261 | if (i & (1U << bit_j)) { | |
1262 | nr_2d_11[bit_i][bit_j]++; | |
1263 | } else { | |
1264 | nr_2d_10[bit_i][bit_j]++; | |
1265 | } | |
1266 | } else { | |
1267 | if (i & (1U << bit_j)) { | |
1268 | nr_2d_01[bit_i][bit_j]++; | |
1269 | } else { | |
1270 | nr_2d_00[bit_i][bit_j]++; | |
1271 | } | |
1272 | } | |
1273 | } | |
1274 | } | |
1275 | distrib_nr_child++; | |
1276 | } | |
1277 | break; | |
1278 | } | |
1279 | case RCU_JA_NULL: | |
1280 | assert(mode == JA_RECOMPACT_ADD_NEXT); | |
1281 | break; | |
1282 | default: | |
1283 | assert(0); | |
1284 | break; | |
1285 | } | |
1286 | ||
1287 | if (mode == JA_RECOMPACT_ADD_NEXT || mode == JA_RECOMPACT_ADD_SAME) { | |
1288 | for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) { | |
1289 | for (bit_j = 0; bit_j < bit_i; bit_j++) { | |
1290 | if (n & (1U << bit_i)) { | |
1291 | if (n & (1U << bit_j)) { | |
1292 | nr_2d_11[bit_i][bit_j]++; | |
1293 | } else { | |
1294 | nr_2d_10[bit_i][bit_j]++; | |
1295 | } | |
1296 | } else { | |
1297 | if (n & (1U << bit_j)) { | |
1298 | nr_2d_01[bit_i][bit_j]++; | |
1299 | } else { | |
1300 | nr_2d_00[bit_i][bit_j]++; | |
1301 | } | |
1302 | } | |
1303 | } | |
1304 | } | |
1305 | distrib_nr_child++; | |
1306 | } | |
1307 | ||
1308 | /* | |
1309 | * The best bit selector is that for which the number of nodes | |
1310 | * in each sub-class is closest to one-fourth of the number of | |
1311 | * children in the distribution. We calculate the distance using | |
1312 | * 4 times the size of the sub-distribution to eliminate | |
1313 | * truncation error. | |
1314 | */ | |
1315 | for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) { | |
1316 | for (bit_j = 0; bit_j < bit_i; bit_j++) { | |
1317 | int distance_to_best[4]; | |
1318 | ||
1319 | distance_to_best[0] = ((unsigned int) nr_2d_11[bit_i][bit_j] << 2U) - distrib_nr_child; | |
1320 | distance_to_best[1] = ((unsigned int) nr_2d_10[bit_i][bit_j] << 2U) - distrib_nr_child; | |
1321 | distance_to_best[2] = ((unsigned int) nr_2d_01[bit_i][bit_j] << 2U) - distrib_nr_child; | |
1322 | distance_to_best[3] = ((unsigned int) nr_2d_00[bit_i][bit_j] << 2U) - distrib_nr_child; | |
1323 | ||
1324 | /* Consider worse distance above best */ | |
1325 | if (distance_to_best[1] > 0 && distance_to_best[1] > distance_to_best[0]) | |
1326 | distance_to_best[0] = distance_to_best[1]; | |
1327 | if (distance_to_best[2] > 0 && distance_to_best[2] > distance_to_best[0]) | |
1328 | distance_to_best[0] = distance_to_best[2]; | |
1329 | if (distance_to_best[3] > 0 && distance_to_best[3] > distance_to_best[0]) | |
1330 | distance_to_best[0] = distance_to_best[3]; | |
1331 | ||
1332 | /* | |
1333 | * If our worse distance is better than overall, | |
1334 | * we become new best candidate. | |
1335 | */ | |
1336 | if (distance_to_best[0] < overall_best_distance) { | |
1337 | overall_best_distance = distance_to_best[0]; | |
1338 | bitsel[0] = bit_i; | |
1339 | bitsel[1] = bit_j; | |
1340 | } | |
1341 | } | |
1342 | } | |
1343 | ||
1344 | dbg_printf("2 dimensions pool bit selection: (%u,%u)\n", bitsel[0], bitsel[1]); | |
1345 | ||
1346 | /* Return our bit selection */ | |
1347 | _bitsel[0] = bitsel[0]; | |
1348 | _bitsel[1] = bitsel[1]; | |
1349 | } | |
1350 | ||
1351 | static | |
1352 | unsigned int find_nearest_type_index(unsigned int type_index, | |
1353 | unsigned int nr_nodes) | |
1354 | { | |
1355 | const struct cds_ja_type *type; | |
1356 | ||
1357 | assert(type_index != NODE_INDEX_NULL); | |
1358 | if (nr_nodes == 0) | |
1359 | return NODE_INDEX_NULL; | |
1360 | for (;;) { | |
1361 | type = &ja_types[type_index]; | |
1362 | if (nr_nodes < type->min_child) | |
1363 | type_index--; | |
1364 | else if (nr_nodes > type->max_child) | |
1365 | type_index++; | |
1366 | else | |
1367 | break; | |
1368 | } | |
1369 | return type_index; | |
1370 | } | |
1371 | ||
1372 | /* | |
1373 | * ja_node_recompact_add: recompact a node, adding a new child. | |
1374 | * Return 0 on success, -EAGAIN if need to retry, or other negative | |
1375 | * error value otherwise. | |
1376 | */ | |
1377 | static | |
1378 | int ja_node_recompact(enum ja_recompact mode, | |
1379 | struct cds_ja *ja, | |
1380 | unsigned int old_type_index, | |
1381 | const struct cds_ja_type *old_type, | |
1382 | struct cds_ja_inode *old_node, | |
1383 | struct cds_ja_shadow_node *shadow_node, | |
1384 | struct cds_ja_inode_flag **old_node_flag_ptr, uint8_t n, | |
1385 | struct cds_ja_inode_flag *child_node_flag, | |
1386 | struct cds_ja_inode_flag **nullify_node_flag_ptr, | |
1387 | int level) | |
1388 | { | |
1389 | unsigned int new_type_index; | |
1390 | struct cds_ja_inode *new_node; | |
1391 | struct cds_ja_shadow_node *new_shadow_node = NULL; | |
1392 | const struct cds_ja_type *new_type; | |
1393 | struct cds_ja_inode_flag *new_node_flag, *old_node_flag; | |
1394 | int ret; | |
1395 | int fallback = 0; | |
1396 | ||
1397 | old_node_flag = *old_node_flag_ptr; | |
1398 | ||
1399 | /* | |
1400 | * Need to find nearest type index even for ADD_SAME, because | |
1401 | * this recompaction, when applied to linear nodes, will garbage | |
1402 | * collect dummy (NULL) entries, and can therefore cause a few | |
1403 | * linear representations to be skipped. | |
1404 | */ | |
1405 | switch (mode) { | |
1406 | case JA_RECOMPACT_ADD_SAME: | |
1407 | new_type_index = find_nearest_type_index(old_type_index, | |
1408 | shadow_node->nr_child + 1); | |
1409 | dbg_printf("Recompact for node with %u children\n", | |
1410 | shadow_node->nr_child + 1); | |
1411 | break; | |
1412 | case JA_RECOMPACT_ADD_NEXT: | |
1413 | if (!shadow_node || old_type_index == NODE_INDEX_NULL) { | |
1414 | new_type_index = 0; | |
1415 | dbg_printf("Recompact for NULL\n"); | |
1416 | } else { | |
1417 | new_type_index = find_nearest_type_index(old_type_index, | |
1418 | shadow_node->nr_child + 1); | |
1419 | dbg_printf("Recompact for node with %u children\n", | |
1420 | shadow_node->nr_child + 1); | |
1421 | } | |
1422 | break; | |
1423 | case JA_RECOMPACT_DEL: | |
1424 | new_type_index = find_nearest_type_index(old_type_index, | |
1425 | shadow_node->nr_child - 1); | |
1426 | dbg_printf("Recompact for node with %u children\n", | |
1427 | shadow_node->nr_child - 1); | |
1428 | break; | |
1429 | default: | |
1430 | assert(0); | |
1431 | } | |
1432 | ||
1433 | retry: /* for fallback */ | |
1434 | dbg_printf("Recompact from type %d to type %d\n", | |
1435 | old_type_index, new_type_index); | |
1436 | new_type = &ja_types[new_type_index]; | |
1437 | if (new_type_index != NODE_INDEX_NULL) { | |
1438 | new_node = alloc_cds_ja_node(ja, new_type); | |
1439 | if (!new_node) | |
1440 | return -ENOMEM; | |
1441 | ||
1442 | if (new_type->type_class == RCU_JA_POOL) { | |
1443 | switch (new_type->nr_pool_order) { | |
1444 | case 1: | |
1445 | { | |
1446 | unsigned int node_distrib_bitsel; | |
1447 | ||
1448 | node_distrib_bitsel = | |
1449 | ja_node_sum_distribution_1d(mode, ja, | |
1450 | old_type_index, old_type, | |
1451 | old_node, shadow_node, | |
1452 | n, child_node_flag, | |
1453 | nullify_node_flag_ptr); | |
1454 | assert(!((unsigned long) new_node & JA_POOL_1D_MASK)); | |
1455 | new_node_flag = ja_node_flag_pool_1d(new_node, | |
1456 | new_type_index, node_distrib_bitsel); | |
1457 | break; | |
1458 | } | |
1459 | case 2: | |
1460 | { | |
1461 | unsigned int node_distrib_bitsel[2]; | |
1462 | ||
1463 | ja_node_sum_distribution_2d(mode, ja, | |
1464 | old_type_index, old_type, | |
1465 | old_node, shadow_node, | |
1466 | n, child_node_flag, | |
1467 | nullify_node_flag_ptr, | |
1468 | node_distrib_bitsel); | |
1469 | assert(!((unsigned long) new_node & JA_POOL_1D_MASK)); | |
1470 | assert(!((unsigned long) new_node & JA_POOL_2D_MASK)); | |
1471 | new_node_flag = ja_node_flag_pool_2d(new_node, | |
1472 | new_type_index, node_distrib_bitsel); | |
1473 | break; | |
1474 | } | |
1475 | default: | |
1476 | assert(0); | |
1477 | } | |
1478 | } else { | |
1479 | new_node_flag = ja_node_flag(new_node, new_type_index); | |
1480 | } | |
1481 | ||
1482 | dbg_printf("Recompact inherit lock from %p\n", shadow_node); | |
1483 | new_shadow_node = rcuja_shadow_set(ja->ht, new_node_flag, shadow_node, ja, level); | |
1484 | if (!new_shadow_node) { | |
1485 | free_cds_ja_node(ja, new_node); | |
1486 | return -ENOMEM; | |
1487 | } | |
1488 | if (fallback) | |
1489 | new_shadow_node->fallback_removal_count = | |
1490 | JA_FALLBACK_REMOVAL_COUNT; | |
1491 | } else { | |
1492 | new_node = NULL; | |
1493 | new_node_flag = NULL; | |
1494 | } | |
1495 | ||
1496 | assert(mode != JA_RECOMPACT_ADD_NEXT || old_type->type_class != RCU_JA_PIGEON); | |
1497 | ||
1498 | if (new_type_index == NODE_INDEX_NULL) | |
1499 | goto skip_copy; | |
1500 | ||
1501 | switch (old_type->type_class) { | |
1502 | case RCU_JA_LINEAR: | |
1503 | { | |
1504 | uint8_t nr_child = | |
1505 | ja_linear_node_get_nr_child(old_type, old_node); | |
1506 | unsigned int i; | |
1507 | ||
1508 | for (i = 0; i < nr_child; i++) { | |
1509 | struct cds_ja_inode_flag *iter; | |
1510 | uint8_t v; | |
1511 | ||
1512 | ja_linear_node_get_ith_pos(old_type, old_node, i, &v, &iter); | |
1513 | if (!iter) | |
1514 | continue; | |
1515 | if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) | |
1516 | continue; | |
1517 | ret = _ja_node_set_nth(new_type, new_node, new_node_flag, | |
1518 | new_shadow_node, | |
1519 | v, iter); | |
1520 | if (new_type->type_class == RCU_JA_POOL && ret) { | |
1521 | goto fallback_toosmall; | |
1522 | } | |
1523 | assert(!ret); | |
1524 | } | |
1525 | break; | |
1526 | } | |
1527 | case RCU_JA_POOL: | |
1528 | { | |
1529 | unsigned int pool_nr; | |
1530 | ||
1531 | for (pool_nr = 0; pool_nr < (1U << old_type->nr_pool_order); pool_nr++) { | |
1532 | struct cds_ja_inode *pool = | |
1533 | ja_pool_node_get_ith_pool(old_type, | |
1534 | old_node, pool_nr); | |
1535 | uint8_t nr_child = | |
1536 | ja_linear_node_get_nr_child(old_type, pool); | |
1537 | unsigned int j; | |
1538 | ||
1539 | for (j = 0; j < nr_child; j++) { | |
1540 | struct cds_ja_inode_flag *iter; | |
1541 | uint8_t v; | |
1542 | ||
1543 | ja_linear_node_get_ith_pos(old_type, pool, | |
1544 | j, &v, &iter); | |
1545 | if (!iter) | |
1546 | continue; | |
1547 | if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) | |
1548 | continue; | |
1549 | ret = _ja_node_set_nth(new_type, new_node, new_node_flag, | |
1550 | new_shadow_node, | |
1551 | v, iter); | |
1552 | if (new_type->type_class == RCU_JA_POOL | |
1553 | && ret) { | |
1554 | goto fallback_toosmall; | |
1555 | } | |
1556 | assert(!ret); | |
1557 | } | |
1558 | } | |
1559 | break; | |
1560 | } | |
1561 | case RCU_JA_NULL: | |
1562 | assert(mode == JA_RECOMPACT_ADD_NEXT); | |
1563 | break; | |
1564 | case RCU_JA_PIGEON: | |
1565 | { | |
1566 | unsigned int i; | |
1567 | ||
1568 | assert(mode == JA_RECOMPACT_DEL); | |
1569 | for (i = 0; i < JA_ENTRY_PER_NODE; i++) { | |
1570 | struct cds_ja_inode_flag *iter; | |
1571 | ||
1572 | iter = ja_pigeon_node_get_ith_pos(old_type, old_node, i); | |
1573 | if (!iter) | |
1574 | continue; | |
1575 | if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) | |
1576 | continue; | |
1577 | ret = _ja_node_set_nth(new_type, new_node, new_node_flag, | |
1578 | new_shadow_node, | |
1579 | i, iter); | |
1580 | if (new_type->type_class == RCU_JA_POOL && ret) { | |
1581 | goto fallback_toosmall; | |
1582 | } | |
1583 | assert(!ret); | |
1584 | } | |
1585 | break; | |
1586 | } | |
1587 | default: | |
1588 | assert(0); | |
1589 | ret = -EINVAL; | |
1590 | goto end; | |
1591 | } | |
1592 | skip_copy: | |
1593 | ||
1594 | if (mode == JA_RECOMPACT_ADD_NEXT || mode == JA_RECOMPACT_ADD_SAME) { | |
1595 | /* add node */ | |
1596 | ret = _ja_node_set_nth(new_type, new_node, new_node_flag, | |
1597 | new_shadow_node, | |
1598 | n, child_node_flag); | |
1599 | if (new_type->type_class == RCU_JA_POOL && ret) { | |
1600 | goto fallback_toosmall; | |
1601 | } | |
1602 | assert(!ret); | |
1603 | } | |
1604 | ||
1605 | if (fallback) { | |
1606 | dbg_printf("Using fallback for %u children, node type index: %u, mode %s\n", | |
1607 | new_shadow_node->nr_child, old_type_index, mode == JA_RECOMPACT_ADD_NEXT ? "add_next" : | |
1608 | (mode == JA_RECOMPACT_DEL ? "del" : "add_same")); | |
1609 | if (ja_debug_counters()) | |
1610 | uatomic_inc(&ja->node_fallback_count_distribution[new_shadow_node->nr_child]); | |
1611 | } | |
1612 | ||
1613 | /* Return pointer to new recompacted node through old_node_flag_ptr */ | |
1614 | *old_node_flag_ptr = new_node_flag; | |
1615 | if (old_node) { | |
1616 | int flags; | |
1617 | ||
1618 | flags = RCUJA_SHADOW_CLEAR_FREE_NODE; | |
1619 | /* | |
1620 | * It is OK to free the lock associated with a node | |
1621 | * going to NULL, since we are holding the parent lock. | |
1622 | * This synchronizes removal with re-add of that node. | |
1623 | */ | |
1624 | if (new_type_index == NODE_INDEX_NULL) | |
1625 | flags |= RCUJA_SHADOW_CLEAR_FREE_LOCK; | |
1626 | ret = rcuja_shadow_clear(ja->ht, old_node_flag, shadow_node, | |
1627 | flags); | |
1628 | assert(!ret); | |
1629 | } | |
1630 | ||
1631 | ret = 0; | |
1632 | end: | |
1633 | return ret; | |
1634 | ||
1635 | fallback_toosmall: | |
1636 | /* fallback if next pool is too small */ | |
1637 | assert(new_shadow_node); | |
1638 | ret = rcuja_shadow_clear(ja->ht, new_node_flag, new_shadow_node, | |
1639 | RCUJA_SHADOW_CLEAR_FREE_NODE); | |
1640 | assert(!ret); | |
1641 | ||
1642 | switch (mode) { | |
1643 | case JA_RECOMPACT_ADD_SAME: | |
1644 | /* | |
1645 | * JA_RECOMPACT_ADD_SAME is only triggered if a linear | |
1646 | * node within a pool has unused entries. It should | |
1647 | * therefore _never_ be too small. | |
1648 | */ | |
1649 | assert(0); | |
1650 | ||
1651 | /* Fall-through */ | |
1652 | case JA_RECOMPACT_ADD_NEXT: | |
1653 | { | |
1654 | const struct cds_ja_type *next_type; | |
1655 | ||
1656 | /* | |
1657 | * Recompaction attempt on add failed. Should only | |
1658 | * happen if target node type is pool. Caused by | |
1659 | * hard-to-split distribution. Recompact using the next | |
1660 | * distribution size. | |
1661 | */ | |
1662 | assert(new_type->type_class == RCU_JA_POOL); | |
1663 | next_type = &ja_types[new_type_index + 1]; | |
1664 | /* | |
1665 | * Try going to the next pool size if our population | |
1666 | * fits within its range. This is not flagged as a | |
1667 | * fallback. | |
1668 | */ | |
1669 | if (shadow_node->nr_child + 1 >= next_type->min_child | |
1670 | && shadow_node->nr_child + 1 <= next_type->max_child) { | |
1671 | new_type_index++; | |
1672 | goto retry; | |
1673 | } else { | |
1674 | new_type_index++; | |
1675 | dbg_printf("Add fallback to type %d\n", new_type_index); | |
1676 | if (ja_debug_counters()) | |
1677 | uatomic_inc(&ja->nr_fallback); | |
1678 | fallback = 1; | |
1679 | goto retry; | |
1680 | } | |
1681 | break; | |
1682 | } | |
1683 | case JA_RECOMPACT_DEL: | |
1684 | /* | |
1685 | * Recompaction attempt on delete failed. Should only | |
1686 | * happen if target node type is pool. This is caused by | |
1687 | * a hard-to-split distribution. Recompact on same node | |
1688 | * size, but flag current node as "fallback" to ensure | |
1689 | * we don't attempt recompaction before some activity | |
1690 | * has reshuffled our node. | |
1691 | */ | |
1692 | assert(new_type->type_class == RCU_JA_POOL); | |
1693 | new_type_index = old_type_index; | |
1694 | dbg_printf("Delete fallback keeping type %d\n", new_type_index); | |
1695 | uatomic_inc(&ja->nr_fallback); | |
1696 | fallback = 1; | |
1697 | goto retry; | |
1698 | default: | |
1699 | assert(0); | |
1700 | return -EINVAL; | |
1701 | } | |
1702 | ||
1703 | /* | |
1704 | * Last resort fallback: pigeon. | |
1705 | */ | |
1706 | new_type_index = (1UL << JA_TYPE_BITS) - 1; | |
1707 | dbg_printf("Fallback to type %d\n", new_type_index); | |
1708 | uatomic_inc(&ja->nr_fallback); | |
1709 | fallback = 1; | |
1710 | goto retry; | |
1711 | } | |
1712 | ||
1713 | /* | |
1714 | * Return 0 on success, -EAGAIN if need to retry, or other negative | |
1715 | * error value otherwise. | |
1716 | */ | |
1717 | static | |
1718 | int ja_node_set_nth(struct cds_ja *ja, | |
1719 | struct cds_ja_inode_flag **node_flag, uint8_t n, | |
1720 | struct cds_ja_inode_flag *child_node_flag, | |
1721 | struct cds_ja_shadow_node *shadow_node, | |
1722 | int level) | |
1723 | { | |
1724 | int ret; | |
1725 | unsigned int type_index; | |
1726 | const struct cds_ja_type *type; | |
1727 | struct cds_ja_inode *node; | |
1728 | ||
1729 | dbg_printf("ja_node_set_nth for n=%u, node %p, shadow %p\n", | |
1730 | (unsigned int) n, ja_node_ptr(*node_flag), shadow_node); | |
1731 | ||
1732 | node = ja_node_ptr(*node_flag); | |
1733 | type_index = ja_node_type(*node_flag); | |
1734 | type = &ja_types[type_index]; | |
1735 | ret = _ja_node_set_nth(type, node, *node_flag, shadow_node, | |
1736 | n, child_node_flag); | |
1737 | switch (ret) { | |
1738 | case -ENOSPC: | |
1739 | /* Not enough space in node, need to recompact to next type. */ | |
1740 | ret = ja_node_recompact(JA_RECOMPACT_ADD_NEXT, ja, type_index, type, node, | |
1741 | shadow_node, node_flag, n, child_node_flag, NULL, level); | |
1742 | break; | |
1743 | case -ERANGE: | |
1744 | /* Node needs to be recompacted. */ | |
1745 | ret = ja_node_recompact(JA_RECOMPACT_ADD_SAME, ja, type_index, type, node, | |
1746 | shadow_node, node_flag, n, child_node_flag, NULL, level); | |
1747 | break; | |
1748 | } | |
1749 | return ret; | |
1750 | } | |
1751 | ||
1752 | /* | |
1753 | * Return 0 on success, -EAGAIN if need to retry, or other negative | |
1754 | * error value otherwise. | |
1755 | */ | |
1756 | static | |
1757 | int ja_node_clear_ptr(struct cds_ja *ja, | |
1758 | struct cds_ja_inode_flag **node_flag_ptr, /* Pointer to location to nullify */ | |
1759 | struct cds_ja_inode_flag **parent_node_flag_ptr, /* Address of parent ptr in its parent */ | |
1760 | struct cds_ja_shadow_node *shadow_node, /* of parent */ | |
1761 | uint8_t n, int level) | |
1762 | { | |
1763 | int ret; | |
1764 | unsigned int type_index; | |
1765 | const struct cds_ja_type *type; | |
1766 | struct cds_ja_inode *node; | |
1767 | ||
1768 | dbg_printf("ja_node_clear_ptr for node %p, shadow %p, target ptr %p\n", | |
1769 | ja_node_ptr(*parent_node_flag_ptr), shadow_node, node_flag_ptr); | |
1770 | ||
1771 | node = ja_node_ptr(*parent_node_flag_ptr); | |
1772 | type_index = ja_node_type(*parent_node_flag_ptr); | |
1773 | type = &ja_types[type_index]; | |
1774 | ret = _ja_node_clear_ptr(type, node, *parent_node_flag_ptr, shadow_node, node_flag_ptr, n); | |
1775 | if (ret == -EFBIG) { | |
1776 | /* Should try recompaction. */ | |
1777 | ret = ja_node_recompact(JA_RECOMPACT_DEL, ja, type_index, type, node, | |
1778 | shadow_node, parent_node_flag_ptr, n, NULL, | |
1779 | node_flag_ptr, level); | |
1780 | } | |
1781 | return ret; | |
1782 | } | |
1783 | ||
1784 | struct cds_ja_node *cds_ja_lookup(struct cds_ja *ja, uint64_t key) | |
1785 | { | |
1786 | unsigned int tree_depth, i; | |
1787 | struct cds_ja_inode_flag *node_flag; | |
1788 | ||
1789 | if (caa_unlikely(key > ja->key_max || key == UINT64_MAX)) | |
1790 | return NULL; | |
1791 | tree_depth = ja->tree_depth; | |
1792 | node_flag = rcu_dereference(ja->root); | |
1793 | ||
1794 | /* level 0: root node */ | |
1795 | if (!ja_node_ptr(node_flag)) | |
1796 | return NULL; | |
1797 | ||
1798 | for (i = 1; i < tree_depth; i++) { | |
1799 | uint8_t iter_key; | |
1800 | ||
1801 | iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (tree_depth - i - 1))); | |
1802 | node_flag = ja_node_get_nth(node_flag, NULL, iter_key); | |
1803 | dbg_printf("cds_ja_lookup iter key lookup %u finds node_flag %p\n", | |
1804 | (unsigned int) iter_key, node_flag); | |
1805 | if (!ja_node_ptr(node_flag)) | |
1806 | return NULL; | |
1807 | } | |
1808 | ||
1809 | /* Last level lookup succeded. We got an actual match. */ | |
1810 | return (struct cds_ja_node *) node_flag; | |
1811 | } | |
1812 | ||
1813 | static | |
1814 | struct cds_ja_node *cds_ja_lookup_inequality(struct cds_ja *ja, uint64_t key, | |
1815 | uint64_t *result_key, enum ja_lookup_inequality mode) | |
1816 | { | |
1817 | int tree_depth, level; | |
1818 | struct cds_ja_inode_flag *node_flag, *cur_node_depth[JA_MAX_DEPTH]; | |
1819 | uint8_t cur_key[JA_MAX_DEPTH]; | |
1820 | uint64_t _result_key = 0; | |
1821 | enum ja_direction dir; | |
1822 | ||
1823 | switch (mode) { | |
1824 | case JA_LOOKUP_BE: | |
1825 | case JA_LOOKUP_AE: | |
1826 | if (caa_unlikely(key > ja->key_max || key == UINT64_MAX)) | |
1827 | return NULL; | |
1828 | break; | |
1829 | default: | |
1830 | return NULL; | |
1831 | } | |
1832 | ||
1833 | memset(cur_node_depth, 0, sizeof(cur_node_depth)); | |
1834 | memset(cur_key, 0, sizeof(cur_key)); | |
1835 | tree_depth = ja->tree_depth; | |
1836 | node_flag = rcu_dereference(ja->root); | |
1837 | cur_node_depth[0] = node_flag; | |
1838 | ||
1839 | /* level 0: root node */ | |
1840 | if (!ja_node_ptr(node_flag)) | |
1841 | return NULL; | |
1842 | ||
1843 | for (level = 1; level < tree_depth; level++) { | |
1844 | uint8_t iter_key; | |
1845 | ||
1846 | iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (tree_depth - level - 1))); | |
1847 | node_flag = ja_node_get_nth(node_flag, NULL, iter_key); | |
1848 | if (!ja_node_ptr(node_flag)) | |
1849 | break; | |
1850 | cur_key[level - 1] = iter_key; | |
1851 | cur_node_depth[level] = node_flag; | |
1852 | dbg_printf("cds_ja_lookup_inequality iter key lookup %u finds node_flag %p\n", | |
1853 | (unsigned int) iter_key, node_flag); | |
1854 | } | |
1855 | ||
1856 | if (level == tree_depth) { | |
1857 | /* Last level lookup succeded. We got an equal match. */ | |
1858 | if (result_key) | |
1859 | *result_key = key; | |
1860 | return (struct cds_ja_node *) node_flag; | |
1861 | } | |
1862 | ||
1863 | /* | |
1864 | * Find highest value left/right of current node. | |
1865 | * Current node is cur_node_depth[level]. | |
1866 | * Start at current level. If we cannot find any key left/right | |
1867 | * of ours, go one level up, seek highest value left/right of | |
1868 | * current (recursively), and when we find one, get the | |
1869 | * rightmost/leftmost child of its rightmost/leftmost child | |
1870 | * (recursively). | |
1871 | */ | |
1872 | switch (mode) { | |
1873 | case JA_LOOKUP_BE: | |
1874 | dir = JA_LEFT; | |
1875 | break; | |
1876 | case JA_LOOKUP_AE: | |
1877 | dir = JA_RIGHT; | |
1878 | break; | |
1879 | default: | |
1880 | assert(0); | |
1881 | } | |
1882 | for (; level > 0; level--) { | |
1883 | uint8_t iter_key; | |
1884 | ||
1885 | iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (tree_depth - level - 1))); | |
1886 | node_flag = ja_node_get_leftright(cur_node_depth[level - 1], | |
1887 | iter_key, &cur_key[level - 1], dir); | |
1888 | dbg_printf("cds_ja_lookup_inequality find sibling from %u at %u finds node_flag %p\n", | |
1889 | (unsigned int) iter_key, (unsigned int) cur_key[level - 1], | |
1890 | node_flag); | |
1891 | /* If found left/right sibling, find rightmost/leftmost child. */ | |
1892 | if (ja_node_ptr(node_flag)) | |
1893 | break; | |
1894 | } | |
1895 | ||
1896 | if (!level) { | |
1897 | /* Reached the root and could not find a left/right sibling. */ | |
1898 | return NULL; | |
1899 | } | |
1900 | ||
1901 | level++; | |
1902 | ||
1903 | /* | |
1904 | * From this point, we are guaranteed to be able to find a | |
1905 | * "below than"/"above than" match. ja_attach_node() and | |
1906 | * ja_detach_node() both guarantee that it is not possible for a | |
1907 | * lookup to reach a dead-end. | |
1908 | */ | |
1909 | ||
1910 | /* | |
1911 | * Find rightmost/leftmost child of rightmost/leftmost child | |
1912 | * (recursively). | |
1913 | */ | |
1914 | switch (mode) { | |
1915 | case JA_LOOKUP_BE: | |
1916 | dir = JA_RIGHTMOST; | |
1917 | break; | |
1918 | case JA_LOOKUP_AE: | |
1919 | dir = JA_LEFTMOST; | |
1920 | break; | |
1921 | default: | |
1922 | assert(0); | |
1923 | } | |
1924 | for (; level < tree_depth; level++) { | |
1925 | node_flag = ja_node_get_minmax(node_flag, &cur_key[level - 1], dir); | |
1926 | dbg_printf("cds_ja_lookup_inequality find minmax at %u finds node_flag %p\n", | |
1927 | (unsigned int) cur_key[level - 1], | |
1928 | node_flag); | |
1929 | if (!ja_node_ptr(node_flag)) | |
1930 | break; | |
1931 | } | |
1932 | ||
1933 | assert(level == tree_depth); | |
1934 | ||
1935 | if (result_key) { | |
1936 | for (level = 1; level < tree_depth; level++) { | |
1937 | _result_key |= ((uint64_t) cur_key[level - 1]) | |
1938 | << (JA_BITS_PER_BYTE * (tree_depth - level - 1)); | |
1939 | } | |
1940 | *result_key = _result_key; | |
1941 | } | |
1942 | return (struct cds_ja_node *) node_flag; | |
1943 | } | |
1944 | ||
1945 | struct cds_ja_node *cds_ja_lookup_below_equal(struct cds_ja *ja, | |
1946 | uint64_t key, uint64_t *result_key) | |
1947 | { | |
1948 | dbg_printf("cds_ja_lookup_below_equal key %" PRIu64 "\n", key); | |
1949 | return cds_ja_lookup_inequality(ja, key, result_key, JA_LOOKUP_BE); | |
1950 | } | |
1951 | ||
1952 | struct cds_ja_node *cds_ja_lookup_above_equal(struct cds_ja *ja, | |
1953 | uint64_t key, uint64_t *result_key) | |
1954 | { | |
1955 | dbg_printf("cds_ja_lookup_above_equal key %" PRIu64 "\n", key); | |
1956 | return cds_ja_lookup_inequality(ja, key, result_key, JA_LOOKUP_AE); | |
1957 | } | |
1958 | ||
1959 | /* | |
1960 | * We reached an unpopulated node. Create it and the children we need, | |
1961 | * and then attach the entire branch to the current node. This may | |
1962 | * trigger recompaction of the current node. Locks needed: node lock | |
1963 | * (for add), and, possibly, parent node lock (to update pointer due to | |
1964 | * node recompaction). | |
1965 | * | |
1966 | * First take node lock, check if recompaction is needed, then take | |
1967 | * parent lock (if needed). Then we can proceed to create the new | |
1968 | * branch. Publish the new branch, and release locks. | |
1969 | * TODO: we currently always take the parent lock even when not needed. | |
1970 | * | |
1971 | * ja_attach_node() ensures that a lookup will _never_ see a branch that | |
1972 | * leads to a dead-end: before attaching a branch, the entire content of | |
1973 | * the new branch is populated, thus creating a cluster, before | |
1974 | * attaching the cluster to the rest of the tree, thus making it visible | |
1975 | * to lookups. | |
1976 | */ | |
1977 | static | |
1978 | int ja_attach_node(struct cds_ja *ja, | |
1979 | struct cds_ja_inode_flag **attach_node_flag_ptr, | |
1980 | struct cds_ja_inode_flag *attach_node_flag, | |
1981 | struct cds_ja_inode_flag *parent_attach_node_flag, | |
1982 | struct cds_ja_inode_flag **old_node_flag_ptr, | |
1983 | struct cds_ja_inode_flag *old_node_flag, | |
1984 | uint64_t key, | |
1985 | unsigned int level, | |
1986 | struct cds_ja_node *child_node) | |
1987 | { | |
1988 | struct cds_ja_shadow_node *shadow_node = NULL, | |
1989 | *parent_shadow_node = NULL; | |
1990 | struct cds_ja_inode_flag *iter_node_flag, *iter_dest_node_flag; | |
1991 | int ret, i; | |
1992 | struct cds_ja_inode_flag *created_nodes[JA_MAX_DEPTH]; | |
1993 | int nr_created_nodes = 0; | |
1994 | ||
1995 | 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", | |
1996 | level, old_node_flag, attach_node_flag_ptr, attach_node_flag, parent_attach_node_flag); | |
1997 | ||
1998 | assert(!old_node_flag); | |
1999 | if (attach_node_flag) { | |
2000 | shadow_node = rcuja_shadow_lookup_lock(ja->ht, attach_node_flag); | |
2001 | if (!shadow_node) { | |
2002 | ret = -EAGAIN; | |
2003 | goto end; | |
2004 | } | |
2005 | } | |
2006 | if (parent_attach_node_flag) { | |
2007 | parent_shadow_node = rcuja_shadow_lookup_lock(ja->ht, | |
2008 | parent_attach_node_flag); | |
2009 | if (!parent_shadow_node) { | |
2010 | ret = -EAGAIN; | |
2011 | goto unlock_shadow; | |
2012 | } | |
2013 | } | |
2014 | ||
2015 | if (old_node_flag_ptr && ja_node_ptr(*old_node_flag_ptr)) { | |
2016 | /* | |
2017 | * Target node has been updated between RCU lookup and | |
2018 | * lock acquisition. We need to re-try lookup and | |
2019 | * attach. | |
2020 | */ | |
2021 | ret = -EAGAIN; | |
2022 | goto unlock_parent; | |
2023 | } | |
2024 | ||
2025 | /* | |
2026 | * Perform a lookup query to handle the case where | |
2027 | * old_node_flag_ptr is NULL. We cannot use it to check if the | |
2028 | * node has been populated between RCU lookup and mutex | |
2029 | * acquisition. | |
2030 | */ | |
2031 | if (!old_node_flag_ptr) { | |
2032 | uint8_t iter_key; | |
2033 | struct cds_ja_inode_flag *lookup_node_flag; | |
2034 | struct cds_ja_inode_flag **lookup_node_flag_ptr; | |
2035 | ||
2036 | iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (ja->tree_depth - level))); | |
2037 | lookup_node_flag = ja_node_get_nth(attach_node_flag, | |
2038 | &lookup_node_flag_ptr, | |
2039 | iter_key); | |
2040 | if (lookup_node_flag) { | |
2041 | ret = -EEXIST; | |
2042 | goto unlock_parent; | |
2043 | } | |
2044 | } | |
2045 | ||
2046 | if (attach_node_flag_ptr && ja_node_ptr(*attach_node_flag_ptr) != | |
2047 | ja_node_ptr(attach_node_flag)) { | |
2048 | /* | |
2049 | * Target node has been updated between RCU lookup and | |
2050 | * lock acquisition. We need to re-try lookup and | |
2051 | * attach. | |
2052 | */ | |
2053 | ret = -EAGAIN; | |
2054 | goto unlock_parent; | |
2055 | } | |
2056 | ||
2057 | /* Create new branch, starting from bottom */ | |
2058 | iter_node_flag = (struct cds_ja_inode_flag *) child_node; | |
2059 | ||
2060 | for (i = ja->tree_depth - 1; i >= (int) level; i--) { | |
2061 | uint8_t iter_key; | |
2062 | ||
2063 | iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (ja->tree_depth - i - 1))); | |
2064 | dbg_printf("branch creation level %d, key %u\n", | |
2065 | i, (unsigned int) iter_key); | |
2066 | iter_dest_node_flag = NULL; | |
2067 | ret = ja_node_set_nth(ja, &iter_dest_node_flag, | |
2068 | iter_key, | |
2069 | iter_node_flag, | |
2070 | NULL, i); | |
2071 | if (ret) { | |
2072 | dbg_printf("branch creation error %d\n", ret); | |
2073 | goto check_error; | |
2074 | } | |
2075 | created_nodes[nr_created_nodes++] = iter_dest_node_flag; | |
2076 | iter_node_flag = iter_dest_node_flag; | |
2077 | } | |
2078 | assert(level > 0); | |
2079 | ||
2080 | /* Publish branch */ | |
2081 | if (level == 1) { | |
2082 | /* | |
2083 | * Attaching to root node. | |
2084 | */ | |
2085 | rcu_assign_pointer(ja->root, iter_node_flag); | |
2086 | } else { | |
2087 | uint8_t iter_key; | |
2088 | ||
2089 | iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (ja->tree_depth - level))); | |
2090 | dbg_printf("publish branch at level %d, key %u\n", | |
2091 | level - 1, (unsigned int) iter_key); | |
2092 | /* We need to use set_nth on the previous level. */ | |
2093 | iter_dest_node_flag = attach_node_flag; | |
2094 | ret = ja_node_set_nth(ja, &iter_dest_node_flag, | |
2095 | iter_key, | |
2096 | iter_node_flag, | |
2097 | shadow_node, level - 1); | |
2098 | if (ret) { | |
2099 | dbg_printf("branch publish error %d\n", ret); | |
2100 | goto check_error; | |
2101 | } | |
2102 | /* | |
2103 | * Attach branch | |
2104 | */ | |
2105 | rcu_assign_pointer(*attach_node_flag_ptr, iter_dest_node_flag); | |
2106 | } | |
2107 | ||
2108 | /* Success */ | |
2109 | ret = 0; | |
2110 | ||
2111 | check_error: | |
2112 | if (ret) { | |
2113 | for (i = 0; i < nr_created_nodes; i++) { | |
2114 | int tmpret; | |
2115 | int flags; | |
2116 | ||
2117 | flags = RCUJA_SHADOW_CLEAR_FREE_LOCK; | |
2118 | if (i) | |
2119 | flags |= RCUJA_SHADOW_CLEAR_FREE_NODE; | |
2120 | tmpret = rcuja_shadow_clear(ja->ht, | |
2121 | created_nodes[i], | |
2122 | NULL, | |
2123 | flags); | |
2124 | assert(!tmpret); | |
2125 | } | |
2126 | } | |
2127 | unlock_parent: | |
2128 | if (parent_shadow_node) | |
2129 | rcuja_shadow_unlock(parent_shadow_node); | |
2130 | unlock_shadow: | |
2131 | if (shadow_node) | |
2132 | rcuja_shadow_unlock(shadow_node); | |
2133 | end: | |
2134 | return ret; | |
2135 | } | |
2136 | ||
2137 | /* | |
2138 | * Lock the parent containing the pointer to list of duplicates, and add | |
2139 | * node to this list. Failure can happen if concurrent update changes | |
2140 | * the parent before we get the lock. We return -EAGAIN in that case. | |
2141 | * Return 0 on success, negative error value on failure. | |
2142 | */ | |
2143 | static | |
2144 | int ja_chain_node(struct cds_ja *ja, | |
2145 | struct cds_ja_inode_flag *parent_node_flag, | |
2146 | struct cds_ja_inode_flag **node_flag_ptr, | |
2147 | struct cds_ja_inode_flag *node_flag, | |
2148 | struct cds_ja_node *last_node, | |
2149 | struct cds_ja_node *node) | |
2150 | { | |
2151 | struct cds_ja_shadow_node *shadow_node; | |
2152 | struct cds_ja_node *iter_node; | |
2153 | int ret = 0, found = 0; | |
2154 | ||
2155 | shadow_node = rcuja_shadow_lookup_lock(ja->ht, parent_node_flag); | |
2156 | if (!shadow_node) { | |
2157 | return -EAGAIN; | |
2158 | } | |
2159 | /* | |
2160 | * Ensure that previous node is still there at end of list. | |
2161 | */ | |
2162 | iter_node = (struct cds_ja_node *) ja_node_ptr(node_flag); | |
2163 | if ((struct cds_ja_node *) ja_node_ptr(*node_flag_ptr) != iter_node) { | |
2164 | ret = -EAGAIN; | |
2165 | goto end; | |
2166 | } | |
2167 | cds_ja_for_each_duplicate(iter_node) { | |
2168 | if (iter_node == last_node) | |
2169 | found = 1; | |
2170 | } | |
2171 | if (!found) { | |
2172 | ret = -EAGAIN; | |
2173 | goto end; | |
2174 | } | |
2175 | /* | |
2176 | * Add node to tail of list to ensure that RCU traversals will | |
2177 | * always see either the prior node or the newly added if | |
2178 | * executed concurrently with a sequence of add followed by del | |
2179 | * on the same key. Safe against concurrent RCU read traversals. | |
2180 | */ | |
2181 | node->next = NULL; | |
2182 | rcu_assign_pointer(last_node->next, node); | |
2183 | end: | |
2184 | rcuja_shadow_unlock(shadow_node); | |
2185 | return ret; | |
2186 | } | |
2187 | ||
2188 | static | |
2189 | int _cds_ja_add(struct cds_ja *ja, uint64_t key, | |
2190 | struct cds_ja_node *node, | |
2191 | struct cds_ja_node **unique_node_ret) | |
2192 | { | |
2193 | unsigned int tree_depth, i; | |
2194 | struct cds_ja_inode_flag *attach_node_flag, | |
2195 | *parent_node_flag, | |
2196 | *parent2_node_flag, | |
2197 | *node_flag, | |
2198 | *parent_attach_node_flag; | |
2199 | struct cds_ja_inode_flag **attach_node_flag_ptr, | |
2200 | **parent_node_flag_ptr, | |
2201 | **node_flag_ptr; | |
2202 | int ret; | |
2203 | ||
2204 | if (caa_unlikely(key > ja->key_max || key == UINT64_MAX)) { | |
2205 | return -EINVAL; | |
2206 | } | |
2207 | tree_depth = ja->tree_depth; | |
2208 | ||
2209 | retry: | |
2210 | dbg_printf("cds_ja_add attempt: key %" PRIu64 ", node %p\n", | |
2211 | key, node); | |
2212 | parent2_node_flag = NULL; | |
2213 | parent_node_flag = | |
2214 | (struct cds_ja_inode_flag *) &ja->root; /* Use root ptr address as key for mutex */ | |
2215 | parent_node_flag_ptr = NULL; | |
2216 | node_flag = rcu_dereference(ja->root); | |
2217 | node_flag_ptr = &ja->root; | |
2218 | ||
2219 | /* Iterate on all internal levels */ | |
2220 | for (i = 1; i < tree_depth; i++) { | |
2221 | uint8_t iter_key; | |
2222 | ||
2223 | if (!ja_node_ptr(node_flag)) | |
2224 | break; | |
2225 | dbg_printf("cds_ja_add iter parent2_node_flag %p parent_node_flag %p node_flag_ptr %p node_flag %p\n", | |
2226 | parent2_node_flag, parent_node_flag, node_flag_ptr, node_flag); | |
2227 | iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (tree_depth - i - 1))); | |
2228 | parent2_node_flag = parent_node_flag; | |
2229 | parent_node_flag = node_flag; | |
2230 | parent_node_flag_ptr = node_flag_ptr; | |
2231 | node_flag = ja_node_get_nth(node_flag, | |
2232 | &node_flag_ptr, | |
2233 | iter_key); | |
2234 | } | |
2235 | ||
2236 | /* | |
2237 | * We reached either bottom of tree or internal NULL node, | |
2238 | * simply add node to last internal level, or chain it if key is | |
2239 | * already present. | |
2240 | */ | |
2241 | if (!ja_node_ptr(node_flag)) { | |
2242 | dbg_printf("cds_ja_add NULL parent2_node_flag %p parent_node_flag %p node_flag_ptr %p node_flag %p\n", | |
2243 | parent2_node_flag, parent_node_flag, node_flag_ptr, node_flag); | |
2244 | ||
2245 | attach_node_flag = parent_node_flag; | |
2246 | attach_node_flag_ptr = parent_node_flag_ptr; | |
2247 | parent_attach_node_flag = parent2_node_flag; | |
2248 | ||
2249 | ret = ja_attach_node(ja, attach_node_flag_ptr, | |
2250 | attach_node_flag, | |
2251 | parent_attach_node_flag, | |
2252 | node_flag_ptr, | |
2253 | node_flag, | |
2254 | key, i, node); | |
2255 | } else { | |
2256 | struct cds_ja_node *iter_node, *last_node = NULL; | |
2257 | ||
2258 | if (unique_node_ret) { | |
2259 | *unique_node_ret = (struct cds_ja_node *) ja_node_ptr(node_flag); | |
2260 | return -EEXIST; | |
2261 | } | |
2262 | ||
2263 | /* Find last duplicate */ | |
2264 | iter_node = (struct cds_ja_node *) ja_node_ptr(node_flag); | |
2265 | cds_ja_for_each_duplicate_rcu(iter_node) | |
2266 | last_node = iter_node; | |
2267 | ||
2268 | dbg_printf("cds_ja_add duplicate parent2_node_flag %p parent_node_flag %p node_flag_ptr %p node_flag %p\n", | |
2269 | parent2_node_flag, parent_node_flag, node_flag_ptr, node_flag); | |
2270 | ||
2271 | attach_node_flag = node_flag; | |
2272 | attach_node_flag_ptr = node_flag_ptr; | |
2273 | parent_attach_node_flag = parent_node_flag; | |
2274 | ||
2275 | ret = ja_chain_node(ja, | |
2276 | parent_attach_node_flag, | |
2277 | attach_node_flag_ptr, | |
2278 | attach_node_flag, | |
2279 | last_node, | |
2280 | node); | |
2281 | } | |
2282 | if (ret == -EAGAIN || ret == -EEXIST) | |
2283 | goto retry; | |
2284 | ||
2285 | return ret; | |
2286 | } | |
2287 | ||
2288 | int cds_ja_add(struct cds_ja *ja, uint64_t key, | |
2289 | struct cds_ja_node *node) | |
2290 | { | |
2291 | return _cds_ja_add(ja, key, node, NULL); | |
2292 | } | |
2293 | ||
2294 | struct cds_ja_node *cds_ja_add_unique(struct cds_ja *ja, uint64_t key, | |
2295 | struct cds_ja_node *node) | |
2296 | { | |
2297 | int ret; | |
2298 | struct cds_ja_node *ret_node; | |
2299 | ||
2300 | ret = _cds_ja_add(ja, key, node, &ret_node); | |
2301 | if (ret == -EEXIST) | |
2302 | return ret_node; | |
2303 | else | |
2304 | return node; | |
2305 | } | |
2306 | ||
2307 | /* | |
2308 | * Note: there is no need to lookup the pointer address associated with | |
2309 | * each node's nth item after taking the lock: it's already been done by | |
2310 | * cds_ja_del while holding the rcu read-side lock, and our node rules | |
2311 | * ensure that when a match value -> pointer is found in a node, it is | |
2312 | * _NEVER_ changed for that node without recompaction, and recompaction | |
2313 | * reallocates the node. | |
2314 | * However, when a child is removed from "linear" nodes, its pointer | |
2315 | * is set to NULL. We therefore check, while holding the locks, if this | |
2316 | * pointer is NULL, and return -ENOENT to the caller if it is the case. | |
2317 | * | |
2318 | * ja_detach_node() ensures that a lookup will _never_ see a branch that | |
2319 | * leads to a dead-end: when removing branch, it makes sure to perform | |
2320 | * the "cut" at the highest node that has only one child, effectively | |
2321 | * replacing it with a NULL pointer. | |
2322 | */ | |
2323 | static | |
2324 | int ja_detach_node(struct cds_ja *ja, | |
2325 | struct cds_ja_inode_flag **snapshot, | |
2326 | struct cds_ja_inode_flag ***snapshot_ptr, | |
2327 | uint8_t *snapshot_n, | |
2328 | int nr_snapshot, | |
2329 | uint64_t key, | |
2330 | struct cds_ja_node *node) | |
2331 | { | |
2332 | struct cds_ja_shadow_node *shadow_nodes[JA_MAX_DEPTH]; | |
2333 | struct cds_ja_inode_flag **node_flag_ptr = NULL, | |
2334 | *parent_node_flag = NULL, | |
2335 | **parent_node_flag_ptr = NULL; | |
2336 | struct cds_ja_inode_flag *iter_node_flag; | |
2337 | int ret, i, nr_shadow = 0, nr_clear = 0, nr_branch = 0; | |
2338 | uint8_t n = 0; | |
2339 | ||
2340 | assert(nr_snapshot == ja->tree_depth + 1); | |
2341 | ||
2342 | /* | |
2343 | * From the last internal level node going up, get the node | |
2344 | * lock, check if the node has only one child left. If it is the | |
2345 | * case, we continue iterating upward. When we reach a node | |
2346 | * which has more that one child left, we lock the parent, and | |
2347 | * proceed to the node deletion (removing its children too). | |
2348 | */ | |
2349 | for (i = nr_snapshot - 2; i >= 1; i--) { | |
2350 | struct cds_ja_shadow_node *shadow_node; | |
2351 | ||
2352 | shadow_node = rcuja_shadow_lookup_lock(ja->ht, | |
2353 | snapshot[i]); | |
2354 | if (!shadow_node) { | |
2355 | ret = -EAGAIN; | |
2356 | goto end; | |
2357 | } | |
2358 | shadow_nodes[nr_shadow++] = shadow_node; | |
2359 | ||
2360 | /* | |
2361 | * Check if node has been removed between RCU | |
2362 | * lookup and lock acquisition. | |
2363 | */ | |
2364 | assert(snapshot_ptr[i + 1]); | |
2365 | if (ja_node_ptr(*snapshot_ptr[i + 1]) | |
2366 | != ja_node_ptr(snapshot[i + 1])) { | |
2367 | ret = -ENOENT; | |
2368 | goto end; | |
2369 | } | |
2370 | ||
2371 | assert(shadow_node->nr_child > 0); | |
2372 | if (shadow_node->nr_child == 1 && i > 1) | |
2373 | nr_clear++; | |
2374 | nr_branch++; | |
2375 | if (shadow_node->nr_child > 1 || i == 1) { | |
2376 | /* Lock parent and break */ | |
2377 | shadow_node = rcuja_shadow_lookup_lock(ja->ht, | |
2378 | snapshot[i - 1]); | |
2379 | if (!shadow_node) { | |
2380 | ret = -EAGAIN; | |
2381 | goto end; | |
2382 | } | |
2383 | shadow_nodes[nr_shadow++] = shadow_node; | |
2384 | ||
2385 | /* | |
2386 | * Check if node has been removed between RCU | |
2387 | * lookup and lock acquisition. | |
2388 | */ | |
2389 | assert(snapshot_ptr[i]); | |
2390 | if (ja_node_ptr(*snapshot_ptr[i]) | |
2391 | != ja_node_ptr(snapshot[i])) { | |
2392 | ret = -ENOENT; | |
2393 | goto end; | |
2394 | } | |
2395 | ||
2396 | node_flag_ptr = snapshot_ptr[i + 1]; | |
2397 | n = snapshot_n[i + 1]; | |
2398 | parent_node_flag_ptr = snapshot_ptr[i]; | |
2399 | parent_node_flag = snapshot[i]; | |
2400 | ||
2401 | if (i > 1) { | |
2402 | /* | |
2403 | * Lock parent's parent, in case we need | |
2404 | * to recompact parent. | |
2405 | */ | |
2406 | shadow_node = rcuja_shadow_lookup_lock(ja->ht, | |
2407 | snapshot[i - 2]); | |
2408 | if (!shadow_node) { | |
2409 | ret = -EAGAIN; | |
2410 | goto end; | |
2411 | } | |
2412 | shadow_nodes[nr_shadow++] = shadow_node; | |
2413 | ||
2414 | /* | |
2415 | * Check if node has been removed between RCU | |
2416 | * lookup and lock acquisition. | |
2417 | */ | |
2418 | assert(snapshot_ptr[i - 1]); | |
2419 | if (ja_node_ptr(*snapshot_ptr[i - 1]) | |
2420 | != ja_node_ptr(snapshot[i - 1])) { | |
2421 | ret = -ENOENT; | |
2422 | goto end; | |
2423 | } | |
2424 | } | |
2425 | ||
2426 | break; | |
2427 | } | |
2428 | } | |
2429 | ||
2430 | /* | |
2431 | * At this point, we want to delete all nodes that are about to | |
2432 | * be removed from shadow_nodes (except the last one, which is | |
2433 | * either the root or the parent of the upmost node with 1 | |
2434 | * child). OK to free lock here, because RCU read lock is held, | |
2435 | * and free only performed in call_rcu. | |
2436 | */ | |
2437 | ||
2438 | for (i = 0; i < nr_clear; i++) { | |
2439 | ret = rcuja_shadow_clear(ja->ht, | |
2440 | shadow_nodes[i]->node_flag, | |
2441 | shadow_nodes[i], | |
2442 | RCUJA_SHADOW_CLEAR_FREE_NODE | |
2443 | | RCUJA_SHADOW_CLEAR_FREE_LOCK); | |
2444 | assert(!ret); | |
2445 | } | |
2446 | ||
2447 | iter_node_flag = parent_node_flag; | |
2448 | /* Remove from parent */ | |
2449 | if (nr_branch < 1) | |
2450 | abort(); /* Should never happen. */ | |
2451 | ret = ja_node_clear_ptr(ja, | |
2452 | node_flag_ptr, /* Pointer to location to nullify */ | |
2453 | &iter_node_flag, /* Old new parent ptr in its parent */ | |
2454 | shadow_nodes[nr_branch - 1], /* of parent */ | |
2455 | n, nr_branch - 1); | |
2456 | if (ret) | |
2457 | goto end; | |
2458 | ||
2459 | dbg_printf("ja_detach_node: publish %p instead of %p\n", | |
2460 | iter_node_flag, *parent_node_flag_ptr); | |
2461 | /* Update address of parent ptr in its parent */ | |
2462 | rcu_assign_pointer(*parent_node_flag_ptr, iter_node_flag); | |
2463 | ||
2464 | end: | |
2465 | for (i = 0; i < nr_shadow; i++) | |
2466 | rcuja_shadow_unlock(shadow_nodes[i]); | |
2467 | return ret; | |
2468 | } | |
2469 | ||
2470 | static | |
2471 | int ja_unchain_node(struct cds_ja *ja, | |
2472 | struct cds_ja_inode_flag *parent_node_flag, | |
2473 | struct cds_ja_inode_flag **node_flag_ptr, | |
2474 | struct cds_ja_inode_flag *node_flag, | |
2475 | struct cds_ja_node *node) | |
2476 | { | |
2477 | struct cds_ja_shadow_node *shadow_node; | |
2478 | struct cds_ja_node *iter_node, **iter_node_ptr, **prev_node_ptr = NULL; | |
2479 | int ret = 0, count = 0, found = 0; | |
2480 | ||
2481 | shadow_node = rcuja_shadow_lookup_lock(ja->ht, parent_node_flag); | |
2482 | if (!shadow_node) | |
2483 | return -EAGAIN; | |
2484 | if (ja_node_ptr(*node_flag_ptr) != ja_node_ptr(node_flag)) { | |
2485 | ret = -EAGAIN; | |
2486 | goto end; | |
2487 | } | |
2488 | /* | |
2489 | * Find the previous node's next pointer pointing to our node, | |
2490 | * so we can update it. Retry if another thread removed all but | |
2491 | * one of duplicates since check (this check was performed | |
2492 | * without lock). Ensure that the node we are about to remove is | |
2493 | * still in the list (while holding lock). No need for RCU | |
2494 | * traversal here since we hold the lock on the parent. | |
2495 | */ | |
2496 | iter_node_ptr = (struct cds_ja_node **) node_flag_ptr; | |
2497 | iter_node = (struct cds_ja_node *) ja_node_ptr(node_flag); | |
2498 | cds_ja_for_each_duplicate(iter_node) { | |
2499 | count++; | |
2500 | if (iter_node == node) { | |
2501 | prev_node_ptr = iter_node_ptr; | |
2502 | found++; | |
2503 | } | |
2504 | iter_node_ptr = &iter_node->next; | |
2505 | } | |
2506 | assert(found <= 1); | |
2507 | if (!found || count == 1) { | |
2508 | ret = -EAGAIN; | |
2509 | goto end; | |
2510 | } | |
2511 | CMM_STORE_SHARED(*prev_node_ptr, node->next); | |
2512 | /* | |
2513 | * Validate that we indeed removed the node from linked list. | |
2514 | */ | |
2515 | assert(ja_node_ptr(*node_flag_ptr) != (struct cds_ja_inode *) node); | |
2516 | end: | |
2517 | rcuja_shadow_unlock(shadow_node); | |
2518 | return ret; | |
2519 | } | |
2520 | ||
2521 | /* | |
2522 | * Called with RCU read lock held. | |
2523 | */ | |
2524 | int cds_ja_del(struct cds_ja *ja, uint64_t key, | |
2525 | struct cds_ja_node *node) | |
2526 | { | |
2527 | unsigned int tree_depth, i; | |
2528 | struct cds_ja_inode_flag *snapshot[JA_MAX_DEPTH]; | |
2529 | struct cds_ja_inode_flag **snapshot_ptr[JA_MAX_DEPTH]; | |
2530 | uint8_t snapshot_n[JA_MAX_DEPTH]; | |
2531 | struct cds_ja_inode_flag *node_flag; | |
2532 | struct cds_ja_inode_flag **prev_node_flag_ptr, | |
2533 | **node_flag_ptr; | |
2534 | int nr_snapshot; | |
2535 | int ret; | |
2536 | ||
2537 | if (caa_unlikely(key > ja->key_max || key == UINT64_MAX)) | |
2538 | return -EINVAL; | |
2539 | tree_depth = ja->tree_depth; | |
2540 | ||
2541 | retry: | |
2542 | nr_snapshot = 0; | |
2543 | dbg_printf("cds_ja_del attempt: key %" PRIu64 ", node %p\n", | |
2544 | key, node); | |
2545 | ||
2546 | /* snapshot for level 0 is only for shadow node lookup */ | |
2547 | snapshot_n[0] = 0; | |
2548 | snapshot_n[1] = 0; | |
2549 | snapshot_ptr[nr_snapshot] = NULL; | |
2550 | snapshot[nr_snapshot++] = (struct cds_ja_inode_flag *) &ja->root; | |
2551 | node_flag = rcu_dereference(ja->root); | |
2552 | prev_node_flag_ptr = &ja->root; | |
2553 | node_flag_ptr = &ja->root; | |
2554 | ||
2555 | /* Iterate on all internal levels */ | |
2556 | for (i = 1; i < tree_depth; i++) { | |
2557 | uint8_t iter_key; | |
2558 | ||
2559 | dbg_printf("cds_ja_del iter node_flag %p\n", | |
2560 | node_flag); | |
2561 | if (!ja_node_ptr(node_flag)) { | |
2562 | return -ENOENT; | |
2563 | } | |
2564 | iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (tree_depth - i - 1))); | |
2565 | snapshot_n[nr_snapshot + 1] = iter_key; | |
2566 | snapshot_ptr[nr_snapshot] = prev_node_flag_ptr; | |
2567 | snapshot[nr_snapshot++] = node_flag; | |
2568 | node_flag = ja_node_get_nth(node_flag, | |
2569 | &node_flag_ptr, | |
2570 | iter_key); | |
2571 | if (node_flag) | |
2572 | prev_node_flag_ptr = node_flag_ptr; | |
2573 | dbg_printf("cds_ja_del iter key lookup %u finds node_flag %p, prev_node_flag_ptr %p\n", | |
2574 | (unsigned int) iter_key, node_flag, | |
2575 | prev_node_flag_ptr); | |
2576 | } | |
2577 | /* | |
2578 | * We reached bottom of tree, try to find the node we are trying | |
2579 | * to remove. Fail if we cannot find it. | |
2580 | */ | |
2581 | if (!ja_node_ptr(node_flag)) { | |
2582 | dbg_printf("cds_ja_del: no node found for key %" PRIu64 "\n", | |
2583 | key); | |
2584 | return -ENOENT; | |
2585 | } else { | |
2586 | struct cds_ja_node *iter_node, *match = NULL; | |
2587 | int count = 0; | |
2588 | ||
2589 | iter_node = (struct cds_ja_node *) ja_node_ptr(node_flag); | |
2590 | cds_ja_for_each_duplicate_rcu(iter_node) { | |
2591 | dbg_printf("cds_ja_del: compare %p with iter_node %p\n", node, iter_node); | |
2592 | if (iter_node == node) | |
2593 | match = iter_node; | |
2594 | count++; | |
2595 | } | |
2596 | ||
2597 | if (!match) { | |
2598 | dbg_printf("cds_ja_del: no node match for node %p key %" PRIu64 "\n", node, key); | |
2599 | return -ENOENT; | |
2600 | } | |
2601 | assert(count > 0); | |
2602 | if (count == 1) { | |
2603 | /* | |
2604 | * Removing last of duplicates. Last snapshot | |
2605 | * does not have a shadow node (external leafs). | |
2606 | */ | |
2607 | snapshot_ptr[nr_snapshot] = prev_node_flag_ptr; | |
2608 | snapshot[nr_snapshot++] = node_flag; | |
2609 | ret = ja_detach_node(ja, snapshot, snapshot_ptr, | |
2610 | snapshot_n, nr_snapshot, key, node); | |
2611 | } else { | |
2612 | ret = ja_unchain_node(ja, snapshot[nr_snapshot - 1], | |
2613 | node_flag_ptr, node_flag, match); | |
2614 | } | |
2615 | } | |
2616 | /* | |
2617 | * Explanation of -ENOENT handling: caused by concurrent delete | |
2618 | * between RCU lookup and actual removal. Need to re-do the | |
2619 | * lookup and removal attempt. | |
2620 | */ | |
2621 | if (ret == -EAGAIN || ret == -ENOENT) | |
2622 | goto retry; | |
2623 | return ret; | |
2624 | } | |
2625 | ||
2626 | struct cds_ja *_cds_ja_new(unsigned int key_bits, | |
2627 | const struct rcu_flavor_struct *flavor) | |
2628 | { | |
2629 | struct cds_ja *ja; | |
2630 | int ret; | |
2631 | struct cds_ja_shadow_node *root_shadow_node; | |
2632 | ||
2633 | ja = calloc(sizeof(*ja), 1); | |
2634 | if (!ja) | |
2635 | goto ja_error; | |
2636 | ||
2637 | switch (key_bits) { | |
2638 | case 8: | |
2639 | case 16: | |
2640 | case 24: | |
2641 | case 32: | |
2642 | case 40: | |
2643 | case 48: | |
2644 | case 56: | |
2645 | ja->key_max = (1ULL << key_bits) - 1; | |
2646 | break; | |
2647 | case 64: | |
2648 | ja->key_max = UINT64_MAX; | |
2649 | break; | |
2650 | default: | |
2651 | goto check_error; | |
2652 | } | |
2653 | ||
2654 | /* ja->root is NULL */ | |
2655 | /* tree_depth 0 is for pointer to root node */ | |
2656 | ja->tree_depth = (key_bits >> JA_LOG2_BITS_PER_BYTE) + 1; | |
2657 | assert(ja->tree_depth <= JA_MAX_DEPTH); | |
2658 | ja->ht = rcuja_create_ht(flavor); | |
2659 | if (!ja->ht) | |
2660 | goto ht_error; | |
2661 | ||
2662 | /* | |
2663 | * Note: we should not free this node until judy array destroy. | |
2664 | */ | |
2665 | root_shadow_node = rcuja_shadow_set(ja->ht, | |
2666 | (struct cds_ja_inode_flag *) &ja->root, | |
2667 | NULL, ja, 0); | |
2668 | if (!root_shadow_node) { | |
2669 | ret = -ENOMEM; | |
2670 | goto ht_node_error; | |
2671 | } | |
2672 | ||
2673 | return ja; | |
2674 | ||
2675 | ht_node_error: | |
2676 | ret = rcuja_delete_ht(ja->ht); | |
2677 | assert(!ret); | |
2678 | ht_error: | |
2679 | check_error: | |
2680 | free(ja); | |
2681 | ja_error: | |
2682 | return NULL; | |
2683 | } | |
2684 | ||
2685 | static | |
2686 | void print_debug_fallback_distribution(struct cds_ja *ja) | |
2687 | { | |
2688 | int i; | |
2689 | ||
2690 | fprintf(stderr, "Fallback node distribution:\n"); | |
2691 | for (i = 0; i < JA_ENTRY_PER_NODE; i++) { | |
2692 | if (!ja->node_fallback_count_distribution[i]) | |
2693 | continue; | |
2694 | fprintf(stderr, " %3u: %4lu\n", | |
2695 | i, ja->node_fallback_count_distribution[i]); | |
2696 | } | |
2697 | } | |
2698 | ||
2699 | static | |
2700 | int ja_final_checks(struct cds_ja *ja) | |
2701 | { | |
2702 | double fallback_ratio; | |
2703 | unsigned long na, nf, nr_fallback; | |
2704 | int ret = 0; | |
2705 | ||
2706 | if (!ja_debug_counters()) | |
2707 | return 0; | |
2708 | ||
2709 | fallback_ratio = (double) uatomic_read(&ja->nr_fallback); | |
2710 | fallback_ratio /= (double) uatomic_read(&ja->nr_nodes_allocated); | |
2711 | nr_fallback = uatomic_read(&ja->nr_fallback); | |
2712 | if (nr_fallback) | |
2713 | fprintf(stderr, | |
2714 | "[warning] RCU Judy Array used %lu fallback node(s) (ratio: %g)\n", | |
2715 | uatomic_read(&ja->nr_fallback), | |
2716 | fallback_ratio); | |
2717 | ||
2718 | na = uatomic_read(&ja->nr_nodes_allocated); | |
2719 | nf = uatomic_read(&ja->nr_nodes_freed); | |
2720 | dbg_printf("Nodes allocated: %lu, Nodes freed: %lu.\n", na, nf); | |
2721 | if (nr_fallback) | |
2722 | print_debug_fallback_distribution(ja); | |
2723 | ||
2724 | if (na != nf) { | |
2725 | fprintf(stderr, "[error] Judy array leaked %ld nodes. Allocated: %lu, freed: %lu.\n", | |
2726 | (long) na - nf, na, nf); | |
2727 | ret = -1; | |
2728 | } | |
2729 | return ret; | |
2730 | } | |
2731 | ||
2732 | /* | |
2733 | * There should be no more concurrent add, delete, nor look-up performed | |
2734 | * on the Judy array while it is being destroyed (ensured by the | |
2735 | * caller). | |
2736 | */ | |
2737 | int cds_ja_destroy(struct cds_ja *ja) | |
2738 | { | |
2739 | const struct rcu_flavor_struct *flavor; | |
2740 | int ret; | |
2741 | ||
2742 | flavor = cds_lfht_rcu_flavor(ja->ht); | |
2743 | rcuja_shadow_prune(ja->ht, | |
2744 | RCUJA_SHADOW_CLEAR_FREE_NODE | RCUJA_SHADOW_CLEAR_FREE_LOCK); | |
2745 | flavor->thread_offline(); | |
2746 | ret = rcuja_delete_ht(ja->ht); | |
2747 | if (ret) | |
2748 | return ret; | |
2749 | ||
2750 | /* Wait for in-flight call_rcu free to complete. */ | |
2751 | flavor->barrier(); | |
2752 | ||
2753 | flavor->thread_online(); | |
2754 | ret = ja_final_checks(ja); | |
2755 | free(ja); | |
2756 | return ret; | |
2757 | } |