Commit | Line | Data |
---|---|---|
5e28c532 | 1 | /* |
abc490a1 MD |
2 | * rculfhash.c |
3 | * | |
1475579c | 4 | * Userspace RCU library - Lock-Free Resizable RCU Hash Table |
abc490a1 MD |
5 | * |
6 | * Copyright 2010-2011 - Mathieu Desnoyers <mathieu.desnoyers@efficios.com> | |
0dcf4847 | 7 | * Copyright 2011 - Lai Jiangshan <laijs@cn.fujitsu.com> |
abc490a1 MD |
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 | |
5e28c532 MD |
22 | */ |
23 | ||
e753ff5a MD |
24 | /* |
25 | * Based on the following articles: | |
26 | * - Ori Shalev and Nir Shavit. Split-ordered lists: Lock-free | |
27 | * extensible hash tables. J. ACM 53, 3 (May 2006), 379-405. | |
28 | * - Michael, M. M. High performance dynamic lock-free hash tables | |
29 | * and list-based sets. In Proceedings of the fourteenth annual ACM | |
30 | * symposium on Parallel algorithms and architectures, ACM Press, | |
31 | * (2002), 73-82. | |
32 | * | |
1475579c | 33 | * Some specificities of this Lock-Free Resizable RCU Hash Table |
e753ff5a MD |
34 | * implementation: |
35 | * | |
36 | * - RCU read-side critical section allows readers to perform hash | |
1f67ba50 MD |
37 | * table lookups, as well as traversals, and use the returned objects |
38 | * safely by allowing memory reclaim to take place only after a grace | |
39 | * period. | |
e753ff5a MD |
40 | * - Add and remove operations are lock-free, and do not need to |
41 | * allocate memory. They need to be executed within RCU read-side | |
42 | * critical section to ensure the objects they read are valid and to | |
43 | * deal with the cmpxchg ABA problem. | |
44 | * - add and add_unique operations are supported. add_unique checks if | |
1f67ba50 MD |
45 | * the node key already exists in the hash table. It ensures not to |
46 | * populate a duplicate key if the node key already exists in the hash | |
47 | * table. | |
48 | * - The resize operation executes concurrently with | |
49 | * add/add_unique/add_replace/remove/lookup/traversal. | |
e753ff5a MD |
50 | * - Hash table nodes are contained within a split-ordered list. This |
51 | * list is ordered by incrementing reversed-bits-hash value. | |
1ee8f000 | 52 | * - An index of bucket nodes is kept. These bucket nodes are the hash |
1f67ba50 MD |
53 | * table "buckets". These buckets are internal nodes that allow to |
54 | * perform a fast hash lookup, similarly to a skip list. These | |
55 | * buckets are chained together in the split-ordered list, which | |
56 | * allows recursive expansion by inserting new buckets between the | |
57 | * existing buckets. The split-ordered list allows adding new buckets | |
58 | * between existing buckets as the table needs to grow. | |
59 | * - The resize operation for small tables only allows expanding the | |
60 | * hash table. It is triggered automatically by detecting long chains | |
61 | * in the add operation. | |
1475579c MD |
62 | * - The resize operation for larger tables (and available through an |
63 | * API) allows both expanding and shrinking the hash table. | |
4c42f1b8 | 64 | * - Split-counters are used to keep track of the number of |
1475579c | 65 | * nodes within the hash table for automatic resize triggering. |
e753ff5a MD |
66 | * - Resize operation initiated by long chain detection is executed by a |
67 | * call_rcu thread, which keeps lock-freedom of add and remove. | |
68 | * - Resize operations are protected by a mutex. | |
69 | * - The removal operation is split in two parts: first, a "removed" | |
70 | * flag is set in the next pointer within the node to remove. Then, | |
71 | * a "garbage collection" is performed in the bucket containing the | |
72 | * removed node (from the start of the bucket up to the removed node). | |
73 | * All encountered nodes with "removed" flag set in their next | |
74 | * pointers are removed from the linked-list. If the cmpxchg used for | |
75 | * removal fails (due to concurrent garbage-collection or concurrent | |
76 | * add), we retry from the beginning of the bucket. This ensures that | |
77 | * the node with "removed" flag set is removed from the hash table | |
78 | * (not visible to lookups anymore) before the RCU read-side critical | |
79 | * section held across removal ends. Furthermore, this ensures that | |
80 | * the node with "removed" flag set is removed from the linked-list | |
5c4ca589 MD |
81 | * before its memory is reclaimed. After setting the "removal" flag, |
82 | * only the thread which removal is the first to set the "removal | |
83 | * owner" flag (with an xchg) into a node's next pointer is considered | |
84 | * to have succeeded its removal (and thus owns the node to reclaim). | |
85 | * Because we garbage-collect starting from an invariant node (the | |
86 | * start-of-bucket bucket node) up to the "removed" node (or find a | |
87 | * reverse-hash that is higher), we are sure that a successful | |
88 | * traversal of the chain leads to a chain that is present in the | |
1f67ba50 | 89 | * linked-list (the start node is never removed) and that it does not |
5c4ca589 MD |
90 | * contain the "removed" node anymore, even if concurrent delete/add |
91 | * operations are changing the structure of the list concurrently. | |
1f67ba50 MD |
92 | * - The add operations perform garbage collection of buckets if they |
93 | * encounter nodes with removed flag set in the bucket where they want | |
94 | * to add their new node. This ensures lock-freedom of add operation by | |
29e669f6 MD |
95 | * helping the remover unlink nodes from the list rather than to wait |
96 | * for it do to so. | |
1f67ba50 MD |
97 | * - There are three memory backends for the hash table buckets: the |
98 | * "order table", the "chunks", and the "mmap". | |
99 | * - These bucket containers contain a compact version of the hash table | |
100 | * nodes. | |
101 | * - The RCU "order table": | |
102 | * - has a first level table indexed by log2(hash index) which is | |
103 | * copied and expanded by the resize operation. This order table | |
104 | * allows finding the "bucket node" tables. | |
105 | * - There is one bucket node table per hash index order. The size of | |
106 | * each bucket node table is half the number of hashes contained in | |
107 | * this order (except for order 0). | |
108 | * - The RCU "chunks" is best suited for close interaction with a page | |
109 | * allocator. It uses a linear array as index to "chunks" containing | |
110 | * each the same number of buckets. | |
111 | * - The RCU "mmap" memory backend uses a single memory map to hold | |
112 | * all buckets. | |
5f177b1c | 113 | * - synchronize_rcu is used to garbage-collect the old bucket node table. |
93d46c39 | 114 | * |
7f949215 | 115 | * Ordering Guarantees: |
0f5543cb | 116 | * |
7f949215 MD |
117 | * To discuss these guarantees, we first define "read" operation as any |
118 | * of the the basic cds_lfht_lookup, cds_lfht_next_duplicate, | |
119 | * cds_lfht_first, cds_lfht_next operation, as well as | |
120 | * cds_lfht_add_unique (failure). | |
121 | * | |
122 | * We define "read traversal" operation as any of the following | |
123 | * group of operations | |
0f5543cb | 124 | * - cds_lfht_lookup followed by iteration with cds_lfht_next_duplicate |
7f949215 MD |
125 | * (and/or cds_lfht_next, although less common). |
126 | * - cds_lfht_add_unique (failure) followed by iteration with | |
127 | * cds_lfht_next_duplicate (and/or cds_lfht_next, although less | |
128 | * common). | |
129 | * - cds_lfht_first followed iteration with cds_lfht_next (and/or | |
130 | * cds_lfht_next_duplicate, although less common). | |
0f5543cb MD |
131 | * |
132 | * We define "write" operations as any of cds_lfht_add, | |
7f949215 MD |
133 | * cds_lfht_add_unique (success), cds_lfht_add_replace, cds_lfht_del. |
134 | * | |
135 | * When cds_lfht_add_unique succeeds (returns the node passed as | |
136 | * parameter), it acts as a "write" operation. When cds_lfht_add_unique | |
137 | * fails (returns a node different from the one passed as parameter), it | |
138 | * acts as a "read" operation. A cds_lfht_add_unique failure is a | |
139 | * cds_lfht_lookup "read" operation, therefore, any ordering guarantee | |
140 | * referring to "lookup" imply any of "lookup" or cds_lfht_add_unique | |
141 | * (failure). | |
142 | * | |
143 | * We define "prior" and "later" node as nodes observable by reads and | |
144 | * read traversals respectively before and after a write or sequence of | |
145 | * write operations. | |
146 | * | |
147 | * Hash-table operations are often cascaded, for example, the pointer | |
148 | * returned by a cds_lfht_lookup() might be passed to a cds_lfht_next(), | |
149 | * whose return value might in turn be passed to another hash-table | |
150 | * operation. This entire cascaded series of operations must be enclosed | |
151 | * by a pair of matching rcu_read_lock() and rcu_read_unlock() | |
152 | * operations. | |
153 | * | |
154 | * The following ordering guarantees are offered by this hash table: | |
155 | * | |
156 | * A.1) "read" after "write": if there is ordering between a write and a | |
157 | * later read, then the read is guaranteed to see the write or some | |
158 | * later write. | |
159 | * A.2) "read traversal" after "write": given that there is dependency | |
160 | * ordering between reads in a "read traversal", if there is | |
161 | * ordering between a write and the first read of the traversal, | |
162 | * then the "read traversal" is guaranteed to see the write or | |
163 | * some later write. | |
164 | * B.1) "write" after "read": if there is ordering between a read and a | |
165 | * later write, then the read will never see the write. | |
166 | * B.2) "write" after "read traversal": given that there is dependency | |
167 | * ordering between reads in a "read traversal", if there is | |
168 | * ordering between the last read of the traversal and a later | |
169 | * write, then the "read traversal" will never see the write. | |
170 | * C) "write" while "read traversal": if a write occurs during a "read | |
171 | * traversal", the traversal may, or may not, see the write. | |
172 | * D.1) "write" after "write": if there is ordering between a write and | |
173 | * a later write, then the later write is guaranteed to see the | |
174 | * effects of the first write. | |
175 | * D.2) Concurrent "write" pairs: The system will assign an arbitrary | |
176 | * order to any pair of concurrent conflicting writes. | |
177 | * Non-conflicting writes (for example, to different keys) are | |
178 | * unordered. | |
179 | * E) If a grace period separates a "del" or "replace" operation | |
180 | * and a subsequent operation, then that subsequent operation is | |
181 | * guaranteed not to see the removed item. | |
182 | * F) Uniqueness guarantee: given a hash table that does not contain | |
183 | * duplicate items for a given key, there will only be one item in | |
184 | * the hash table after an arbitrary sequence of add_unique and/or | |
185 | * add_replace operations. Note, however, that a pair of | |
186 | * concurrent read operations might well access two different items | |
187 | * with that key. | |
188 | * G.1) If a pair of lookups for a given key are ordered (e.g. by a | |
189 | * memory barrier), then the second lookup will return the same | |
190 | * node as the previous lookup, or some later node. | |
191 | * G.2) A "read traversal" that starts after the end of a prior "read | |
192 | * traversal" (ordered by memory barriers) is guaranteed to see the | |
193 | * same nodes as the previous traversal, or some later nodes. | |
194 | * G.3) Concurrent "read" pairs: concurrent reads are unordered. For | |
195 | * example, if a pair of reads to the same key run concurrently | |
196 | * with an insertion of that same key, the reads remain unordered | |
197 | * regardless of their return values. In other words, you cannot | |
198 | * rely on the values returned by the reads to deduce ordering. | |
199 | * | |
200 | * Progress guarantees: | |
201 | * | |
202 | * * Reads are wait-free. These operations always move forward in the | |
203 | * hash table linked list, and this list has no loop. | |
204 | * * Writes are lock-free. Any retry loop performed by a write operation | |
205 | * is triggered by progress made within another update operation. | |
0f5543cb | 206 | * |
1ee8f000 | 207 | * Bucket node tables: |
93d46c39 | 208 | * |
1ee8f000 LJ |
209 | * hash table hash table the last all bucket node tables |
210 | * order size bucket node 0 1 2 3 4 5 6(index) | |
93d46c39 LJ |
211 | * table size |
212 | * 0 1 1 1 | |
213 | * 1 2 1 1 1 | |
214 | * 2 4 2 1 1 2 | |
215 | * 3 8 4 1 1 2 4 | |
216 | * 4 16 8 1 1 2 4 8 | |
217 | * 5 32 16 1 1 2 4 8 16 | |
218 | * 6 64 32 1 1 2 4 8 16 32 | |
219 | * | |
1ee8f000 | 220 | * When growing/shrinking, we only focus on the last bucket node table |
93d46c39 LJ |
221 | * which size is (!order ? 1 : (1 << (order -1))). |
222 | * | |
223 | * Example for growing/shrinking: | |
1ee8f000 LJ |
224 | * grow hash table from order 5 to 6: init the index=6 bucket node table |
225 | * shrink hash table from order 6 to 5: fini the index=6 bucket node table | |
93d46c39 | 226 | * |
1475579c MD |
227 | * A bit of ascii art explanation: |
228 | * | |
1f67ba50 MD |
229 | * The order index is the off-by-one compared to the actual power of 2 |
230 | * because we use index 0 to deal with the 0 special-case. | |
1475579c MD |
231 | * |
232 | * This shows the nodes for a small table ordered by reversed bits: | |
233 | * | |
234 | * bits reverse | |
235 | * 0 000 000 | |
236 | * 4 100 001 | |
237 | * 2 010 010 | |
238 | * 6 110 011 | |
239 | * 1 001 100 | |
240 | * 5 101 101 | |
241 | * 3 011 110 | |
242 | * 7 111 111 | |
243 | * | |
244 | * This shows the nodes in order of non-reversed bits, linked by | |
245 | * reversed-bit order. | |
246 | * | |
247 | * order bits reverse | |
248 | * 0 0 000 000 | |
0adc36a8 LJ |
249 | * 1 | 1 001 100 <- |
250 | * 2 | | 2 010 010 <- | | |
f6fdd688 | 251 | * | | | 3 011 110 | <- | |
1475579c MD |
252 | * 3 -> | | | 4 100 001 | | |
253 | * -> | | 5 101 101 | | |
254 | * -> | 6 110 011 | |
255 | * -> 7 111 111 | |
e753ff5a MD |
256 | */ |
257 | ||
2ed95849 | 258 | #define _LGPL_SOURCE |
125f41db | 259 | #define _GNU_SOURCE |
2ed95849 | 260 | #include <stdlib.h> |
e0ba718a MD |
261 | #include <errno.h> |
262 | #include <assert.h> | |
263 | #include <stdio.h> | |
abc490a1 | 264 | #include <stdint.h> |
f000907d | 265 | #include <string.h> |
125f41db | 266 | #include <sched.h> |
e0ba718a | 267 | |
15cfbec7 | 268 | #include "config.h" |
2ed95849 | 269 | #include <urcu.h> |
abc490a1 | 270 | #include <urcu-call-rcu.h> |
7b17c13e | 271 | #include <urcu-flavor.h> |
a42cc659 MD |
272 | #include <urcu/arch.h> |
273 | #include <urcu/uatomic.h> | |
a42cc659 | 274 | #include <urcu/compiler.h> |
abc490a1 | 275 | #include <urcu/rculfhash.h> |
0b6aa001 | 276 | #include <rculfhash-internal.h> |
5e28c532 | 277 | #include <stdio.h> |
464a1ec9 | 278 | #include <pthread.h> |
44395fb7 | 279 | |
f8994aee | 280 | /* |
4c42f1b8 | 281 | * Split-counters lazily update the global counter each 1024 |
f8994aee MD |
282 | * addition/removal. It automatically keeps track of resize required. |
283 | * We use the bucket length as indicator for need to expand for small | |
284 | * tables and machines lacking per-cpu data suppport. | |
285 | */ | |
286 | #define COUNT_COMMIT_ORDER 10 | |
4ddbb355 | 287 | #define DEFAULT_SPLIT_COUNT_MASK 0xFUL |
6ea6bc67 MD |
288 | #define CHAIN_LEN_TARGET 1 |
289 | #define CHAIN_LEN_RESIZE_THRESHOLD 3 | |
2ed95849 | 290 | |
cd95516d | 291 | /* |
76a73da8 | 292 | * Define the minimum table size. |
cd95516d | 293 | */ |
d0d8f9aa LJ |
294 | #define MIN_TABLE_ORDER 0 |
295 | #define MIN_TABLE_SIZE (1UL << MIN_TABLE_ORDER) | |
cd95516d | 296 | |
b7d619b0 | 297 | /* |
1ee8f000 | 298 | * Minimum number of bucket nodes to touch per thread to parallelize grow/shrink. |
b7d619b0 | 299 | */ |
6083a889 MD |
300 | #define MIN_PARTITION_PER_THREAD_ORDER 12 |
301 | #define MIN_PARTITION_PER_THREAD (1UL << MIN_PARTITION_PER_THREAD_ORDER) | |
b7d619b0 | 302 | |
d95bd160 MD |
303 | /* |
304 | * The removed flag needs to be updated atomically with the pointer. | |
48ed1c18 | 305 | * It indicates that no node must attach to the node scheduled for |
b198f0fd | 306 | * removal, and that node garbage collection must be performed. |
1ee8f000 | 307 | * The bucket flag does not require to be updated atomically with the |
d95bd160 | 308 | * pointer, but it is added as a pointer low bit flag to save space. |
1f67ba50 MD |
309 | * The "removal owner" flag is used to detect which of the "del" |
310 | * operation that has set the "removed flag" gets to return the removed | |
311 | * node to its caller. Note that the replace operation does not need to | |
312 | * iteract with the "removal owner" flag, because it validates that | |
313 | * the "removed" flag is not set before performing its cmpxchg. | |
d95bd160 | 314 | */ |
d37166c6 | 315 | #define REMOVED_FLAG (1UL << 0) |
1ee8f000 | 316 | #define BUCKET_FLAG (1UL << 1) |
db00ccc3 MD |
317 | #define REMOVAL_OWNER_FLAG (1UL << 2) |
318 | #define FLAGS_MASK ((1UL << 3) - 1) | |
d37166c6 | 319 | |
bb7b2f26 | 320 | /* Value of the end pointer. Should not interact with flags. */ |
f9c80341 | 321 | #define END_VALUE NULL |
bb7b2f26 | 322 | |
7f52427b MD |
323 | /* |
324 | * ht_items_count: Split-counters counting the number of node addition | |
325 | * and removal in the table. Only used if the CDS_LFHT_ACCOUNTING flag | |
326 | * is set at hash table creation. | |
327 | * | |
328 | * These are free-running counters, never reset to zero. They count the | |
329 | * number of add/remove, and trigger every (1 << COUNT_COMMIT_ORDER) | |
330 | * operations to update the global counter. We choose a power-of-2 value | |
331 | * for the trigger to deal with 32 or 64-bit overflow of the counter. | |
332 | */ | |
df44348d | 333 | struct ht_items_count { |
860d07e8 | 334 | unsigned long add, del; |
df44348d MD |
335 | } __attribute__((aligned(CAA_CACHE_LINE_SIZE))); |
336 | ||
7f52427b MD |
337 | /* |
338 | * rcu_resize_work: Contains arguments passed to RCU worker thread | |
339 | * responsible for performing lazy resize. | |
340 | */ | |
abc490a1 MD |
341 | struct rcu_resize_work { |
342 | struct rcu_head head; | |
14044b37 | 343 | struct cds_lfht *ht; |
abc490a1 | 344 | }; |
2ed95849 | 345 | |
7f52427b MD |
346 | /* |
347 | * partition_resize_work: Contains arguments passed to worker threads | |
348 | * executing the hash table resize on partitions of the hash table | |
349 | * assigned to each processor's worker thread. | |
350 | */ | |
b7d619b0 | 351 | struct partition_resize_work { |
1af6e26e | 352 | pthread_t thread_id; |
b7d619b0 MD |
353 | struct cds_lfht *ht; |
354 | unsigned long i, start, len; | |
355 | void (*fct)(struct cds_lfht *ht, unsigned long i, | |
356 | unsigned long start, unsigned long len); | |
357 | }; | |
358 | ||
abc490a1 MD |
359 | /* |
360 | * Algorithm to reverse bits in a word by lookup table, extended to | |
361 | * 64-bit words. | |
f9830efd | 362 | * Source: |
abc490a1 | 363 | * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable |
f9830efd | 364 | * Originally from Public Domain. |
abc490a1 MD |
365 | */ |
366 | ||
367 | static const uint8_t BitReverseTable256[256] = | |
2ed95849 | 368 | { |
abc490a1 MD |
369 | #define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64 |
370 | #define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16) | |
371 | #define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 ) | |
372 | R6(0), R6(2), R6(1), R6(3) | |
373 | }; | |
374 | #undef R2 | |
375 | #undef R4 | |
376 | #undef R6 | |
2ed95849 | 377 | |
abc490a1 MD |
378 | static |
379 | uint8_t bit_reverse_u8(uint8_t v) | |
380 | { | |
381 | return BitReverseTable256[v]; | |
382 | } | |
ab7d5fc6 | 383 | |
abc490a1 MD |
384 | static __attribute__((unused)) |
385 | uint32_t bit_reverse_u32(uint32_t v) | |
386 | { | |
387 | return ((uint32_t) bit_reverse_u8(v) << 24) | | |
388 | ((uint32_t) bit_reverse_u8(v >> 8) << 16) | | |
389 | ((uint32_t) bit_reverse_u8(v >> 16) << 8) | | |
390 | ((uint32_t) bit_reverse_u8(v >> 24)); | |
2ed95849 MD |
391 | } |
392 | ||
abc490a1 MD |
393 | static __attribute__((unused)) |
394 | uint64_t bit_reverse_u64(uint64_t v) | |
2ed95849 | 395 | { |
abc490a1 MD |
396 | return ((uint64_t) bit_reverse_u8(v) << 56) | |
397 | ((uint64_t) bit_reverse_u8(v >> 8) << 48) | | |
398 | ((uint64_t) bit_reverse_u8(v >> 16) << 40) | | |
399 | ((uint64_t) bit_reverse_u8(v >> 24) << 32) | | |
400 | ((uint64_t) bit_reverse_u8(v >> 32) << 24) | | |
401 | ((uint64_t) bit_reverse_u8(v >> 40) << 16) | | |
402 | ((uint64_t) bit_reverse_u8(v >> 48) << 8) | | |
403 | ((uint64_t) bit_reverse_u8(v >> 56)); | |
404 | } | |
405 | ||
406 | static | |
407 | unsigned long bit_reverse_ulong(unsigned long v) | |
408 | { | |
409 | #if (CAA_BITS_PER_LONG == 32) | |
410 | return bit_reverse_u32(v); | |
411 | #else | |
412 | return bit_reverse_u64(v); | |
413 | #endif | |
414 | } | |
415 | ||
f9830efd | 416 | /* |
24365af7 MD |
417 | * fls: returns the position of the most significant bit. |
418 | * Returns 0 if no bit is set, else returns the position of the most | |
419 | * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit). | |
f9830efd | 420 | */ |
24365af7 MD |
421 | #if defined(__i386) || defined(__x86_64) |
422 | static inline | |
423 | unsigned int fls_u32(uint32_t x) | |
f9830efd | 424 | { |
24365af7 MD |
425 | int r; |
426 | ||
427 | asm("bsrl %1,%0\n\t" | |
428 | "jnz 1f\n\t" | |
429 | "movl $-1,%0\n\t" | |
430 | "1:\n\t" | |
431 | : "=r" (r) : "rm" (x)); | |
432 | return r + 1; | |
433 | } | |
434 | #define HAS_FLS_U32 | |
435 | #endif | |
436 | ||
437 | #if defined(__x86_64) | |
438 | static inline | |
439 | unsigned int fls_u64(uint64_t x) | |
440 | { | |
441 | long r; | |
442 | ||
443 | asm("bsrq %1,%0\n\t" | |
444 | "jnz 1f\n\t" | |
445 | "movq $-1,%0\n\t" | |
446 | "1:\n\t" | |
447 | : "=r" (r) : "rm" (x)); | |
448 | return r + 1; | |
449 | } | |
450 | #define HAS_FLS_U64 | |
451 | #endif | |
452 | ||
453 | #ifndef HAS_FLS_U64 | |
454 | static __attribute__((unused)) | |
455 | unsigned int fls_u64(uint64_t x) | |
456 | { | |
457 | unsigned int r = 64; | |
458 | ||
459 | if (!x) | |
460 | return 0; | |
461 | ||
462 | if (!(x & 0xFFFFFFFF00000000ULL)) { | |
463 | x <<= 32; | |
464 | r -= 32; | |
465 | } | |
466 | if (!(x & 0xFFFF000000000000ULL)) { | |
467 | x <<= 16; | |
468 | r -= 16; | |
469 | } | |
470 | if (!(x & 0xFF00000000000000ULL)) { | |
471 | x <<= 8; | |
472 | r -= 8; | |
473 | } | |
474 | if (!(x & 0xF000000000000000ULL)) { | |
475 | x <<= 4; | |
476 | r -= 4; | |
477 | } | |
478 | if (!(x & 0xC000000000000000ULL)) { | |
479 | x <<= 2; | |
480 | r -= 2; | |
481 | } | |
482 | if (!(x & 0x8000000000000000ULL)) { | |
483 | x <<= 1; | |
484 | r -= 1; | |
485 | } | |
486 | return r; | |
487 | } | |
488 | #endif | |
489 | ||
490 | #ifndef HAS_FLS_U32 | |
491 | static __attribute__((unused)) | |
492 | unsigned int fls_u32(uint32_t x) | |
493 | { | |
494 | unsigned int r = 32; | |
f9830efd | 495 | |
24365af7 MD |
496 | if (!x) |
497 | return 0; | |
498 | if (!(x & 0xFFFF0000U)) { | |
499 | x <<= 16; | |
500 | r -= 16; | |
501 | } | |
502 | if (!(x & 0xFF000000U)) { | |
503 | x <<= 8; | |
504 | r -= 8; | |
505 | } | |
506 | if (!(x & 0xF0000000U)) { | |
507 | x <<= 4; | |
508 | r -= 4; | |
509 | } | |
510 | if (!(x & 0xC0000000U)) { | |
511 | x <<= 2; | |
512 | r -= 2; | |
513 | } | |
514 | if (!(x & 0x80000000U)) { | |
515 | x <<= 1; | |
516 | r -= 1; | |
517 | } | |
518 | return r; | |
519 | } | |
520 | #endif | |
521 | ||
5bc6b66f | 522 | unsigned int cds_lfht_fls_ulong(unsigned long x) |
f9830efd | 523 | { |
6887cc5e | 524 | #if (CAA_BITS_PER_LONG == 32) |
24365af7 MD |
525 | return fls_u32(x); |
526 | #else | |
527 | return fls_u64(x); | |
528 | #endif | |
529 | } | |
f9830efd | 530 | |
920f8ef6 LJ |
531 | /* |
532 | * Return the minimum order for which x <= (1UL << order). | |
533 | * Return -1 if x is 0. | |
534 | */ | |
5bc6b66f | 535 | int cds_lfht_get_count_order_u32(uint32_t x) |
24365af7 | 536 | { |
920f8ef6 LJ |
537 | if (!x) |
538 | return -1; | |
24365af7 | 539 | |
920f8ef6 | 540 | return fls_u32(x - 1); |
24365af7 MD |
541 | } |
542 | ||
920f8ef6 LJ |
543 | /* |
544 | * Return the minimum order for which x <= (1UL << order). | |
545 | * Return -1 if x is 0. | |
546 | */ | |
5bc6b66f | 547 | int cds_lfht_get_count_order_ulong(unsigned long x) |
24365af7 | 548 | { |
920f8ef6 LJ |
549 | if (!x) |
550 | return -1; | |
24365af7 | 551 | |
5bc6b66f | 552 | return cds_lfht_fls_ulong(x - 1); |
f9830efd MD |
553 | } |
554 | ||
555 | static | |
ab65b890 | 556 | void cds_lfht_resize_lazy_grow(struct cds_lfht *ht, unsigned long size, int growth); |
f9830efd | 557 | |
f8994aee | 558 | static |
4105056a | 559 | void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size, |
f8994aee MD |
560 | unsigned long count); |
561 | ||
df44348d | 562 | static long nr_cpus_mask = -1; |
4c42f1b8 LJ |
563 | static long split_count_mask = -1; |
564 | ||
4ddbb355 | 565 | #if defined(HAVE_SYSCONF) |
4c42f1b8 LJ |
566 | static void ht_init_nr_cpus_mask(void) |
567 | { | |
568 | long maxcpus; | |
569 | ||
570 | maxcpus = sysconf(_SC_NPROCESSORS_CONF); | |
571 | if (maxcpus <= 0) { | |
572 | nr_cpus_mask = -2; | |
573 | return; | |
574 | } | |
575 | /* | |
576 | * round up number of CPUs to next power of two, so we | |
577 | * can use & for modulo. | |
578 | */ | |
5bc6b66f | 579 | maxcpus = 1UL << cds_lfht_get_count_order_ulong(maxcpus); |
4c42f1b8 LJ |
580 | nr_cpus_mask = maxcpus - 1; |
581 | } | |
4ddbb355 LJ |
582 | #else /* #if defined(HAVE_SYSCONF) */ |
583 | static void ht_init_nr_cpus_mask(void) | |
584 | { | |
585 | nr_cpus_mask = -2; | |
586 | } | |
587 | #endif /* #else #if defined(HAVE_SYSCONF) */ | |
df44348d MD |
588 | |
589 | static | |
5afadd12 | 590 | void alloc_split_items_count(struct cds_lfht *ht) |
df44348d MD |
591 | { |
592 | struct ht_items_count *count; | |
593 | ||
4c42f1b8 LJ |
594 | if (nr_cpus_mask == -1) { |
595 | ht_init_nr_cpus_mask(); | |
4ddbb355 LJ |
596 | if (nr_cpus_mask < 0) |
597 | split_count_mask = DEFAULT_SPLIT_COUNT_MASK; | |
598 | else | |
599 | split_count_mask = nr_cpus_mask; | |
df44348d | 600 | } |
4c42f1b8 | 601 | |
4ddbb355 | 602 | assert(split_count_mask >= 0); |
5afadd12 LJ |
603 | |
604 | if (ht->flags & CDS_LFHT_ACCOUNTING) { | |
605 | ht->split_count = calloc(split_count_mask + 1, sizeof(*count)); | |
606 | assert(ht->split_count); | |
607 | } else { | |
608 | ht->split_count = NULL; | |
609 | } | |
df44348d MD |
610 | } |
611 | ||
612 | static | |
5afadd12 | 613 | void free_split_items_count(struct cds_lfht *ht) |
df44348d | 614 | { |
5afadd12 | 615 | poison_free(ht->split_count); |
df44348d MD |
616 | } |
617 | ||
14360f1c | 618 | #if defined(HAVE_SCHED_GETCPU) |
df44348d | 619 | static |
14360f1c | 620 | int ht_get_split_count_index(unsigned long hash) |
df44348d MD |
621 | { |
622 | int cpu; | |
623 | ||
4c42f1b8 | 624 | assert(split_count_mask >= 0); |
df44348d | 625 | cpu = sched_getcpu(); |
8ed51e04 | 626 | if (caa_unlikely(cpu < 0)) |
14360f1c | 627 | return hash & split_count_mask; |
df44348d | 628 | else |
4c42f1b8 | 629 | return cpu & split_count_mask; |
df44348d | 630 | } |
14360f1c LJ |
631 | #else /* #if defined(HAVE_SCHED_GETCPU) */ |
632 | static | |
633 | int ht_get_split_count_index(unsigned long hash) | |
634 | { | |
635 | return hash & split_count_mask; | |
636 | } | |
637 | #endif /* #else #if defined(HAVE_SCHED_GETCPU) */ | |
df44348d MD |
638 | |
639 | static | |
14360f1c | 640 | void ht_count_add(struct cds_lfht *ht, unsigned long size, unsigned long hash) |
df44348d | 641 | { |
4c42f1b8 LJ |
642 | unsigned long split_count; |
643 | int index; | |
314558bf | 644 | long count; |
df44348d | 645 | |
8ed51e04 | 646 | if (caa_unlikely(!ht->split_count)) |
3171717f | 647 | return; |
14360f1c | 648 | index = ht_get_split_count_index(hash); |
4c42f1b8 | 649 | split_count = uatomic_add_return(&ht->split_count[index].add, 1); |
314558bf MD |
650 | if (caa_likely(split_count & ((1UL << COUNT_COMMIT_ORDER) - 1))) |
651 | return; | |
652 | /* Only if number of add multiple of 1UL << COUNT_COMMIT_ORDER */ | |
653 | ||
654 | dbg_printf("add split count %lu\n", split_count); | |
655 | count = uatomic_add_return(&ht->count, | |
656 | 1UL << COUNT_COMMIT_ORDER); | |
4c299dcb | 657 | if (caa_likely(count & (count - 1))) |
314558bf MD |
658 | return; |
659 | /* Only if global count is power of 2 */ | |
660 | ||
661 | if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) < size) | |
662 | return; | |
663 | dbg_printf("add set global %ld\n", count); | |
664 | cds_lfht_resize_lazy_count(ht, size, | |
665 | count >> (CHAIN_LEN_TARGET - 1)); | |
df44348d MD |
666 | } |
667 | ||
668 | static | |
14360f1c | 669 | void ht_count_del(struct cds_lfht *ht, unsigned long size, unsigned long hash) |
df44348d | 670 | { |
4c42f1b8 LJ |
671 | unsigned long split_count; |
672 | int index; | |
314558bf | 673 | long count; |
df44348d | 674 | |
8ed51e04 | 675 | if (caa_unlikely(!ht->split_count)) |
3171717f | 676 | return; |
14360f1c | 677 | index = ht_get_split_count_index(hash); |
4c42f1b8 | 678 | split_count = uatomic_add_return(&ht->split_count[index].del, 1); |
314558bf MD |
679 | if (caa_likely(split_count & ((1UL << COUNT_COMMIT_ORDER) - 1))) |
680 | return; | |
681 | /* Only if number of deletes multiple of 1UL << COUNT_COMMIT_ORDER */ | |
682 | ||
683 | dbg_printf("del split count %lu\n", split_count); | |
684 | count = uatomic_add_return(&ht->count, | |
685 | -(1UL << COUNT_COMMIT_ORDER)); | |
4c299dcb | 686 | if (caa_likely(count & (count - 1))) |
314558bf MD |
687 | return; |
688 | /* Only if global count is power of 2 */ | |
689 | ||
690 | if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) >= size) | |
691 | return; | |
692 | dbg_printf("del set global %ld\n", count); | |
693 | /* | |
694 | * Don't shrink table if the number of nodes is below a | |
695 | * certain threshold. | |
696 | */ | |
697 | if (count < (1UL << COUNT_COMMIT_ORDER) * (split_count_mask + 1)) | |
698 | return; | |
699 | cds_lfht_resize_lazy_count(ht, size, | |
700 | count >> (CHAIN_LEN_TARGET - 1)); | |
df44348d MD |
701 | } |
702 | ||
f9830efd | 703 | static |
4105056a | 704 | void check_resize(struct cds_lfht *ht, unsigned long size, uint32_t chain_len) |
f9830efd | 705 | { |
f8994aee MD |
706 | unsigned long count; |
707 | ||
b8af5011 MD |
708 | if (!(ht->flags & CDS_LFHT_AUTO_RESIZE)) |
709 | return; | |
f8994aee MD |
710 | count = uatomic_read(&ht->count); |
711 | /* | |
712 | * Use bucket-local length for small table expand and for | |
713 | * environments lacking per-cpu data support. | |
714 | */ | |
715 | if (count >= (1UL << COUNT_COMMIT_ORDER)) | |
716 | return; | |
24365af7 | 717 | if (chain_len > 100) |
f0c29ed7 | 718 | dbg_printf("WARNING: large chain length: %u.\n", |
24365af7 | 719 | chain_len); |
3390d470 | 720 | if (chain_len >= CHAIN_LEN_RESIZE_THRESHOLD) |
ab65b890 | 721 | cds_lfht_resize_lazy_grow(ht, size, |
5bc6b66f | 722 | cds_lfht_get_count_order_u32(chain_len - (CHAIN_LEN_TARGET - 1))); |
f9830efd MD |
723 | } |
724 | ||
abc490a1 | 725 | static |
14044b37 | 726 | struct cds_lfht_node *clear_flag(struct cds_lfht_node *node) |
abc490a1 | 727 | { |
14044b37 | 728 | return (struct cds_lfht_node *) (((unsigned long) node) & ~FLAGS_MASK); |
abc490a1 MD |
729 | } |
730 | ||
731 | static | |
14044b37 | 732 | int is_removed(struct cds_lfht_node *node) |
abc490a1 | 733 | { |
d37166c6 | 734 | return ((unsigned long) node) & REMOVED_FLAG; |
abc490a1 MD |
735 | } |
736 | ||
f5596c94 | 737 | static |
1ee8f000 | 738 | int is_bucket(struct cds_lfht_node *node) |
f5596c94 | 739 | { |
1ee8f000 | 740 | return ((unsigned long) node) & BUCKET_FLAG; |
f5596c94 MD |
741 | } |
742 | ||
743 | static | |
1ee8f000 | 744 | struct cds_lfht_node *flag_bucket(struct cds_lfht_node *node) |
f5596c94 | 745 | { |
1ee8f000 | 746 | return (struct cds_lfht_node *) (((unsigned long) node) | BUCKET_FLAG); |
f5596c94 | 747 | } |
bb7b2f26 | 748 | |
db00ccc3 MD |
749 | static |
750 | int is_removal_owner(struct cds_lfht_node *node) | |
751 | { | |
752 | return ((unsigned long) node) & REMOVAL_OWNER_FLAG; | |
753 | } | |
754 | ||
755 | static | |
756 | struct cds_lfht_node *flag_removal_owner(struct cds_lfht_node *node) | |
757 | { | |
758 | return (struct cds_lfht_node *) (((unsigned long) node) | REMOVAL_OWNER_FLAG); | |
759 | } | |
760 | ||
71bb3aca MD |
761 | static |
762 | struct cds_lfht_node *flag_removed_or_removal_owner(struct cds_lfht_node *node) | |
763 | { | |
764 | return (struct cds_lfht_node *) (((unsigned long) node) | REMOVED_FLAG | REMOVAL_OWNER_FLAG); | |
765 | } | |
766 | ||
bb7b2f26 MD |
767 | static |
768 | struct cds_lfht_node *get_end(void) | |
769 | { | |
770 | return (struct cds_lfht_node *) END_VALUE; | |
771 | } | |
772 | ||
773 | static | |
774 | int is_end(struct cds_lfht_node *node) | |
775 | { | |
776 | return clear_flag(node) == (struct cds_lfht_node *) END_VALUE; | |
777 | } | |
778 | ||
abc490a1 | 779 | static |
ab65b890 LJ |
780 | unsigned long _uatomic_xchg_monotonic_increase(unsigned long *ptr, |
781 | unsigned long v) | |
abc490a1 MD |
782 | { |
783 | unsigned long old1, old2; | |
784 | ||
785 | old1 = uatomic_read(ptr); | |
786 | do { | |
787 | old2 = old1; | |
788 | if (old2 >= v) | |
f9830efd | 789 | return old2; |
abc490a1 | 790 | } while ((old1 = uatomic_cmpxchg(ptr, old2, v)) != old2); |
ab65b890 | 791 | return old2; |
abc490a1 MD |
792 | } |
793 | ||
48f1b16d LJ |
794 | static |
795 | void cds_lfht_alloc_bucket_table(struct cds_lfht *ht, unsigned long order) | |
796 | { | |
0b6aa001 | 797 | return ht->mm->alloc_bucket_table(ht, order); |
48f1b16d LJ |
798 | } |
799 | ||
800 | /* | |
801 | * cds_lfht_free_bucket_table() should be called with decreasing order. | |
802 | * When cds_lfht_free_bucket_table(0) is called, it means the whole | |
803 | * lfht is destroyed. | |
804 | */ | |
805 | static | |
806 | void cds_lfht_free_bucket_table(struct cds_lfht *ht, unsigned long order) | |
807 | { | |
0b6aa001 | 808 | return ht->mm->free_bucket_table(ht, order); |
48f1b16d LJ |
809 | } |
810 | ||
9d72a73f LJ |
811 | static inline |
812 | struct cds_lfht_node *bucket_at(struct cds_lfht *ht, unsigned long index) | |
f4a9cc0b | 813 | { |
0b6aa001 | 814 | return ht->bucket_at(ht, index); |
f4a9cc0b LJ |
815 | } |
816 | ||
9d72a73f LJ |
817 | static inline |
818 | struct cds_lfht_node *lookup_bucket(struct cds_lfht *ht, unsigned long size, | |
819 | unsigned long hash) | |
820 | { | |
821 | assert(size > 0); | |
822 | return bucket_at(ht, hash & (size - 1)); | |
823 | } | |
824 | ||
273399de MD |
825 | /* |
826 | * Remove all logically deleted nodes from a bucket up to a certain node key. | |
827 | */ | |
828 | static | |
1ee8f000 | 829 | void _cds_lfht_gc_bucket(struct cds_lfht_node *bucket, struct cds_lfht_node *node) |
273399de | 830 | { |
14044b37 | 831 | struct cds_lfht_node *iter_prev, *iter, *next, *new_next; |
273399de | 832 | |
1ee8f000 LJ |
833 | assert(!is_bucket(bucket)); |
834 | assert(!is_removed(bucket)); | |
835 | assert(!is_bucket(node)); | |
c90201ac | 836 | assert(!is_removed(node)); |
273399de | 837 | for (;;) { |
1ee8f000 LJ |
838 | iter_prev = bucket; |
839 | /* We can always skip the bucket node initially */ | |
04db56f8 | 840 | iter = rcu_dereference(iter_prev->next); |
b4cb483f | 841 | assert(!is_removed(iter)); |
04db56f8 | 842 | assert(iter_prev->reverse_hash <= node->reverse_hash); |
bd4db153 | 843 | /* |
1ee8f000 | 844 | * We should never be called with bucket (start of chain) |
bd4db153 MD |
845 | * and logically removed node (end of path compression |
846 | * marker) being the actual same node. This would be a | |
847 | * bug in the algorithm implementation. | |
848 | */ | |
1ee8f000 | 849 | assert(bucket != node); |
273399de | 850 | for (;;) { |
8ed51e04 | 851 | if (caa_unlikely(is_end(iter))) |
f9c80341 | 852 | return; |
04db56f8 | 853 | if (caa_likely(clear_flag(iter)->reverse_hash > node->reverse_hash)) |
f9c80341 | 854 | return; |
04db56f8 | 855 | next = rcu_dereference(clear_flag(iter)->next); |
8ed51e04 | 856 | if (caa_likely(is_removed(next))) |
273399de | 857 | break; |
b453eae1 | 858 | iter_prev = clear_flag(iter); |
273399de MD |
859 | iter = next; |
860 | } | |
b198f0fd | 861 | assert(!is_removed(iter)); |
1ee8f000 LJ |
862 | if (is_bucket(iter)) |
863 | new_next = flag_bucket(clear_flag(next)); | |
f5596c94 MD |
864 | else |
865 | new_next = clear_flag(next); | |
04db56f8 | 866 | (void) uatomic_cmpxchg(&iter_prev->next, iter, new_next); |
273399de MD |
867 | } |
868 | } | |
869 | ||
9357c415 MD |
870 | static |
871 | int _cds_lfht_replace(struct cds_lfht *ht, unsigned long size, | |
872 | struct cds_lfht_node *old_node, | |
3fb86f26 | 873 | struct cds_lfht_node *old_next, |
9357c415 MD |
874 | struct cds_lfht_node *new_node) |
875 | { | |
04db56f8 | 876 | struct cds_lfht_node *bucket, *ret_next; |
9357c415 MD |
877 | |
878 | if (!old_node) /* Return -ENOENT if asked to replace NULL node */ | |
7801dadd | 879 | return -ENOENT; |
9357c415 MD |
880 | |
881 | assert(!is_removed(old_node)); | |
1ee8f000 | 882 | assert(!is_bucket(old_node)); |
9357c415 | 883 | assert(!is_removed(new_node)); |
1ee8f000 | 884 | assert(!is_bucket(new_node)); |
9357c415 | 885 | assert(new_node != old_node); |
3fb86f26 | 886 | for (;;) { |
9357c415 | 887 | /* Insert after node to be replaced */ |
9357c415 MD |
888 | if (is_removed(old_next)) { |
889 | /* | |
890 | * Too late, the old node has been removed under us | |
891 | * between lookup and replace. Fail. | |
892 | */ | |
7801dadd | 893 | return -ENOENT; |
9357c415 | 894 | } |
feda2722 LJ |
895 | assert(old_next == clear_flag(old_next)); |
896 | assert(new_node != old_next); | |
71bb3aca MD |
897 | /* |
898 | * REMOVAL_OWNER flag is _NEVER_ set before the REMOVED | |
899 | * flag. It is either set atomically at the same time | |
900 | * (replace) or after (del). | |
901 | */ | |
902 | assert(!is_removal_owner(old_next)); | |
feda2722 | 903 | new_node->next = old_next; |
9357c415 MD |
904 | /* |
905 | * Here is the whole trick for lock-free replace: we add | |
906 | * the replacement node _after_ the node we want to | |
907 | * replace by atomically setting its next pointer at the | |
908 | * same time we set its removal flag. Given that | |
909 | * the lookups/get next use an iterator aware of the | |
910 | * next pointer, they will either skip the old node due | |
911 | * to the removal flag and see the new node, or use | |
912 | * the old node, but will not see the new one. | |
db00ccc3 MD |
913 | * This is a replacement of a node with another node |
914 | * that has the same value: we are therefore not | |
71bb3aca MD |
915 | * removing a value from the hash table. We set both the |
916 | * REMOVED and REMOVAL_OWNER flags atomically so we own | |
917 | * the node after successful cmpxchg. | |
9357c415 | 918 | */ |
04db56f8 | 919 | ret_next = uatomic_cmpxchg(&old_node->next, |
71bb3aca | 920 | old_next, flag_removed_or_removal_owner(new_node)); |
3fb86f26 | 921 | if (ret_next == old_next) |
7801dadd | 922 | break; /* We performed the replacement. */ |
3fb86f26 LJ |
923 | old_next = ret_next; |
924 | } | |
9357c415 | 925 | |
9357c415 MD |
926 | /* |
927 | * Ensure that the old node is not visible to readers anymore: | |
928 | * lookup for the node, and remove it (along with any other | |
929 | * logically removed node) if found. | |
930 | */ | |
04db56f8 LJ |
931 | bucket = lookup_bucket(ht, size, bit_reverse_ulong(old_node->reverse_hash)); |
932 | _cds_lfht_gc_bucket(bucket, new_node); | |
7801dadd | 933 | |
a85eff52 | 934 | assert(is_removed(CMM_LOAD_SHARED(old_node->next))); |
7801dadd | 935 | return 0; |
9357c415 MD |
936 | } |
937 | ||
83beee94 MD |
938 | /* |
939 | * A non-NULL unique_ret pointer uses the "add unique" (or uniquify) add | |
940 | * mode. A NULL unique_ret allows creation of duplicate keys. | |
941 | */ | |
abc490a1 | 942 | static |
83beee94 | 943 | void _cds_lfht_add(struct cds_lfht *ht, |
91a75cc5 | 944 | unsigned long hash, |
0422d92c | 945 | cds_lfht_match_fct match, |
996ff57c | 946 | const void *key, |
83beee94 MD |
947 | unsigned long size, |
948 | struct cds_lfht_node *node, | |
949 | struct cds_lfht_iter *unique_ret, | |
1ee8f000 | 950 | int bucket_flag) |
abc490a1 | 951 | { |
14044b37 | 952 | struct cds_lfht_node *iter_prev, *iter, *next, *new_node, *new_next, |
960c9e4f | 953 | *return_node; |
04db56f8 | 954 | struct cds_lfht_node *bucket; |
abc490a1 | 955 | |
1ee8f000 | 956 | assert(!is_bucket(node)); |
c90201ac | 957 | assert(!is_removed(node)); |
91a75cc5 | 958 | bucket = lookup_bucket(ht, size, hash); |
abc490a1 | 959 | for (;;) { |
adc0de68 | 960 | uint32_t chain_len = 0; |
abc490a1 | 961 | |
11519af6 MD |
962 | /* |
963 | * iter_prev points to the non-removed node prior to the | |
964 | * insert location. | |
11519af6 | 965 | */ |
04db56f8 | 966 | iter_prev = bucket; |
1ee8f000 | 967 | /* We can always skip the bucket node initially */ |
04db56f8 LJ |
968 | iter = rcu_dereference(iter_prev->next); |
969 | assert(iter_prev->reverse_hash <= node->reverse_hash); | |
abc490a1 | 970 | for (;;) { |
8ed51e04 | 971 | if (caa_unlikely(is_end(iter))) |
273399de | 972 | goto insert; |
04db56f8 | 973 | if (caa_likely(clear_flag(iter)->reverse_hash > node->reverse_hash)) |
273399de | 974 | goto insert; |
238cc06e | 975 | |
1ee8f000 LJ |
976 | /* bucket node is the first node of the identical-hash-value chain */ |
977 | if (bucket_flag && clear_flag(iter)->reverse_hash == node->reverse_hash) | |
194fdbd1 | 978 | goto insert; |
238cc06e | 979 | |
04db56f8 | 980 | next = rcu_dereference(clear_flag(iter)->next); |
8ed51e04 | 981 | if (caa_unlikely(is_removed(next))) |
9dba85be | 982 | goto gc_node; |
238cc06e LJ |
983 | |
984 | /* uniquely add */ | |
83beee94 | 985 | if (unique_ret |
1ee8f000 | 986 | && !is_bucket(next) |
04db56f8 | 987 | && clear_flag(iter)->reverse_hash == node->reverse_hash) { |
238cc06e LJ |
988 | struct cds_lfht_iter d_iter = { .node = node, .next = iter, }; |
989 | ||
990 | /* | |
991 | * uniquely adding inserts the node as the first | |
992 | * node of the identical-hash-value node chain. | |
993 | * | |
994 | * This semantic ensures no duplicated keys | |
995 | * should ever be observable in the table | |
1f67ba50 MD |
996 | * (including traversing the table node by |
997 | * node by forward iterations) | |
238cc06e | 998 | */ |
04db56f8 | 999 | cds_lfht_next_duplicate(ht, match, key, &d_iter); |
238cc06e LJ |
1000 | if (!d_iter.node) |
1001 | goto insert; | |
1002 | ||
1003 | *unique_ret = d_iter; | |
83beee94 | 1004 | return; |
48ed1c18 | 1005 | } |
238cc06e | 1006 | |
11519af6 | 1007 | /* Only account for identical reverse hash once */ |
04db56f8 | 1008 | if (iter_prev->reverse_hash != clear_flag(iter)->reverse_hash |
1ee8f000 | 1009 | && !is_bucket(next)) |
4105056a | 1010 | check_resize(ht, size, ++chain_len); |
11519af6 | 1011 | iter_prev = clear_flag(iter); |
273399de | 1012 | iter = next; |
abc490a1 | 1013 | } |
48ed1c18 | 1014 | |
273399de | 1015 | insert: |
7ec59d3b | 1016 | assert(node != clear_flag(iter)); |
11519af6 | 1017 | assert(!is_removed(iter_prev)); |
c90201ac | 1018 | assert(!is_removed(iter)); |
f000907d | 1019 | assert(iter_prev != node); |
1ee8f000 | 1020 | if (!bucket_flag) |
04db56f8 | 1021 | node->next = clear_flag(iter); |
f9c80341 | 1022 | else |
1ee8f000 LJ |
1023 | node->next = flag_bucket(clear_flag(iter)); |
1024 | if (is_bucket(iter)) | |
1025 | new_node = flag_bucket(node); | |
f5596c94 MD |
1026 | else |
1027 | new_node = node; | |
04db56f8 | 1028 | if (uatomic_cmpxchg(&iter_prev->next, iter, |
48ed1c18 | 1029 | new_node) != iter) { |
273399de | 1030 | continue; /* retry */ |
48ed1c18 | 1031 | } else { |
83beee94 | 1032 | return_node = node; |
960c9e4f | 1033 | goto end; |
48ed1c18 MD |
1034 | } |
1035 | ||
9dba85be MD |
1036 | gc_node: |
1037 | assert(!is_removed(iter)); | |
1ee8f000 LJ |
1038 | if (is_bucket(iter)) |
1039 | new_next = flag_bucket(clear_flag(next)); | |
f5596c94 MD |
1040 | else |
1041 | new_next = clear_flag(next); | |
04db56f8 | 1042 | (void) uatomic_cmpxchg(&iter_prev->next, iter, new_next); |
273399de | 1043 | /* retry */ |
464a1ec9 | 1044 | } |
9357c415 | 1045 | end: |
83beee94 MD |
1046 | if (unique_ret) { |
1047 | unique_ret->node = return_node; | |
1048 | /* unique_ret->next left unset, never used. */ | |
1049 | } | |
abc490a1 | 1050 | } |
464a1ec9 | 1051 | |
abc490a1 | 1052 | static |
860d07e8 | 1053 | int _cds_lfht_del(struct cds_lfht *ht, unsigned long size, |
b65ec430 | 1054 | struct cds_lfht_node *node) |
abc490a1 | 1055 | { |
db00ccc3 | 1056 | struct cds_lfht_node *bucket, *next; |
5e28c532 | 1057 | |
9357c415 | 1058 | if (!node) /* Return -ENOENT if asked to delete NULL node */ |
743f9143 | 1059 | return -ENOENT; |
9357c415 | 1060 | |
7ec59d3b | 1061 | /* logically delete the node */ |
1ee8f000 | 1062 | assert(!is_bucket(node)); |
c90201ac | 1063 | assert(!is_removed(node)); |
db00ccc3 | 1064 | assert(!is_removal_owner(node)); |
48ed1c18 | 1065 | |
db00ccc3 MD |
1066 | /* |
1067 | * We are first checking if the node had previously been | |
1068 | * logically removed (this check is not atomic with setting the | |
1069 | * logical removal flag). Return -ENOENT if the node had | |
1070 | * previously been removed. | |
1071 | */ | |
a85eff52 | 1072 | next = CMM_LOAD_SHARED(node->next); /* next is not dereferenced */ |
db00ccc3 MD |
1073 | if (caa_unlikely(is_removed(next))) |
1074 | return -ENOENT; | |
b65ec430 | 1075 | assert(!is_bucket(next)); |
196f4fab MD |
1076 | /* |
1077 | * The del operation semantic guarantees a full memory barrier | |
1078 | * before the uatomic_or atomic commit of the deletion flag. | |
1079 | */ | |
1080 | cmm_smp_mb__before_uatomic_or(); | |
db00ccc3 MD |
1081 | /* |
1082 | * We set the REMOVED_FLAG unconditionally. Note that there may | |
1083 | * be more than one concurrent thread setting this flag. | |
1084 | * Knowing which wins the race will be known after the garbage | |
1085 | * collection phase, stay tuned! | |
1086 | */ | |
1087 | uatomic_or(&node->next, REMOVED_FLAG); | |
7ec59d3b | 1088 | /* We performed the (logical) deletion. */ |
7ec59d3b MD |
1089 | |
1090 | /* | |
1091 | * Ensure that the node is not visible to readers anymore: lookup for | |
273399de MD |
1092 | * the node, and remove it (along with any other logically removed node) |
1093 | * if found. | |
11519af6 | 1094 | */ |
04db56f8 LJ |
1095 | bucket = lookup_bucket(ht, size, bit_reverse_ulong(node->reverse_hash)); |
1096 | _cds_lfht_gc_bucket(bucket, node); | |
743f9143 | 1097 | |
a85eff52 | 1098 | assert(is_removed(CMM_LOAD_SHARED(node->next))); |
db00ccc3 MD |
1099 | /* |
1100 | * Last phase: atomically exchange node->next with a version | |
1101 | * having "REMOVAL_OWNER_FLAG" set. If the returned node->next | |
1102 | * pointer did _not_ have "REMOVAL_OWNER_FLAG" set, we now own | |
1103 | * the node and win the removal race. | |
1104 | * It is interesting to note that all "add" paths are forbidden | |
1105 | * to change the next pointer starting from the point where the | |
1106 | * REMOVED_FLAG is set, so here using a read, followed by a | |
1107 | * xchg() suffice to guarantee that the xchg() will ever only | |
1108 | * set the "REMOVAL_OWNER_FLAG" (or change nothing if the flag | |
1109 | * was already set). | |
1110 | */ | |
1111 | if (!is_removal_owner(uatomic_xchg(&node->next, | |
1112 | flag_removal_owner(node->next)))) | |
1113 | return 0; | |
1114 | else | |
1115 | return -ENOENT; | |
abc490a1 | 1116 | } |
2ed95849 | 1117 | |
b7d619b0 MD |
1118 | static |
1119 | void *partition_resize_thread(void *arg) | |
1120 | { | |
1121 | struct partition_resize_work *work = arg; | |
1122 | ||
7b17c13e | 1123 | work->ht->flavor->register_thread(); |
b7d619b0 | 1124 | work->fct(work->ht, work->i, work->start, work->len); |
7b17c13e | 1125 | work->ht->flavor->unregister_thread(); |
b7d619b0 MD |
1126 | return NULL; |
1127 | } | |
1128 | ||
1129 | static | |
1130 | void partition_resize_helper(struct cds_lfht *ht, unsigned long i, | |
1131 | unsigned long len, | |
1132 | void (*fct)(struct cds_lfht *ht, unsigned long i, | |
1133 | unsigned long start, unsigned long len)) | |
1134 | { | |
1135 | unsigned long partition_len; | |
1136 | struct partition_resize_work *work; | |
6083a889 MD |
1137 | int thread, ret; |
1138 | unsigned long nr_threads; | |
b7d619b0 | 1139 | |
6083a889 MD |
1140 | /* |
1141 | * Note: nr_cpus_mask + 1 is always power of 2. | |
1142 | * We spawn just the number of threads we need to satisfy the minimum | |
1143 | * partition size, up to the number of CPUs in the system. | |
1144 | */ | |
91452a6a MD |
1145 | if (nr_cpus_mask > 0) { |
1146 | nr_threads = min(nr_cpus_mask + 1, | |
1147 | len >> MIN_PARTITION_PER_THREAD_ORDER); | |
1148 | } else { | |
1149 | nr_threads = 1; | |
1150 | } | |
5bc6b66f | 1151 | partition_len = len >> cds_lfht_get_count_order_ulong(nr_threads); |
6083a889 | 1152 | work = calloc(nr_threads, sizeof(*work)); |
b7d619b0 | 1153 | assert(work); |
6083a889 MD |
1154 | for (thread = 0; thread < nr_threads; thread++) { |
1155 | work[thread].ht = ht; | |
1156 | work[thread].i = i; | |
1157 | work[thread].len = partition_len; | |
1158 | work[thread].start = thread * partition_len; | |
1159 | work[thread].fct = fct; | |
1af6e26e | 1160 | ret = pthread_create(&(work[thread].thread_id), ht->resize_attr, |
6083a889 | 1161 | partition_resize_thread, &work[thread]); |
b7d619b0 MD |
1162 | assert(!ret); |
1163 | } | |
6083a889 | 1164 | for (thread = 0; thread < nr_threads; thread++) { |
1af6e26e | 1165 | ret = pthread_join(work[thread].thread_id, NULL); |
b7d619b0 MD |
1166 | assert(!ret); |
1167 | } | |
1168 | free(work); | |
b7d619b0 MD |
1169 | } |
1170 | ||
e8de508e MD |
1171 | /* |
1172 | * Holding RCU read lock to protect _cds_lfht_add against memory | |
1173 | * reclaim that could be performed by other call_rcu worker threads (ABA | |
1174 | * problem). | |
9ee0fc9a | 1175 | * |
b7d619b0 | 1176 | * When we reach a certain length, we can split this population phase over |
9ee0fc9a MD |
1177 | * many worker threads, based on the number of CPUs available in the system. |
1178 | * This should therefore take care of not having the expand lagging behind too | |
1179 | * many concurrent insertion threads by using the scheduler's ability to | |
1ee8f000 | 1180 | * schedule bucket node population fairly with insertions. |
e8de508e | 1181 | */ |
4105056a | 1182 | static |
b7d619b0 MD |
1183 | void init_table_populate_partition(struct cds_lfht *ht, unsigned long i, |
1184 | unsigned long start, unsigned long len) | |
4105056a | 1185 | { |
9d72a73f | 1186 | unsigned long j, size = 1UL << (i - 1); |
4105056a | 1187 | |
d0d8f9aa | 1188 | assert(i > MIN_TABLE_ORDER); |
7b17c13e | 1189 | ht->flavor->read_lock(); |
9d72a73f LJ |
1190 | for (j = size + start; j < size + start + len; j++) { |
1191 | struct cds_lfht_node *new_node = bucket_at(ht, j); | |
1192 | ||
1193 | assert(j >= size && j < (size << 1)); | |
1194 | dbg_printf("init populate: order %lu index %lu hash %lu\n", | |
1195 | i, j, j); | |
1196 | new_node->reverse_hash = bit_reverse_ulong(j); | |
91a75cc5 | 1197 | _cds_lfht_add(ht, j, NULL, NULL, size, new_node, NULL, 1); |
4105056a | 1198 | } |
7b17c13e | 1199 | ht->flavor->read_unlock(); |
b7d619b0 MD |
1200 | } |
1201 | ||
1202 | static | |
1203 | void init_table_populate(struct cds_lfht *ht, unsigned long i, | |
1204 | unsigned long len) | |
1205 | { | |
1206 | assert(nr_cpus_mask != -1); | |
6083a889 | 1207 | if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD) { |
7b17c13e | 1208 | ht->flavor->thread_online(); |
b7d619b0 | 1209 | init_table_populate_partition(ht, i, 0, len); |
7b17c13e | 1210 | ht->flavor->thread_offline(); |
b7d619b0 MD |
1211 | return; |
1212 | } | |
1213 | partition_resize_helper(ht, i, len, init_table_populate_partition); | |
4105056a MD |
1214 | } |
1215 | ||
abc490a1 | 1216 | static |
4105056a | 1217 | void init_table(struct cds_lfht *ht, |
93d46c39 | 1218 | unsigned long first_order, unsigned long last_order) |
24365af7 | 1219 | { |
93d46c39 | 1220 | unsigned long i; |
24365af7 | 1221 | |
93d46c39 LJ |
1222 | dbg_printf("init table: first_order %lu last_order %lu\n", |
1223 | first_order, last_order); | |
d0d8f9aa | 1224 | assert(first_order > MIN_TABLE_ORDER); |
93d46c39 | 1225 | for (i = first_order; i <= last_order; i++) { |
4105056a | 1226 | unsigned long len; |
24365af7 | 1227 | |
4f6e90b7 | 1228 | len = 1UL << (i - 1); |
f0c29ed7 | 1229 | dbg_printf("init order %lu len: %lu\n", i, len); |
4d676753 MD |
1230 | |
1231 | /* Stop expand if the resize target changes under us */ | |
7b3893e4 | 1232 | if (CMM_LOAD_SHARED(ht->resize_target) < (1UL << i)) |
4d676753 MD |
1233 | break; |
1234 | ||
48f1b16d | 1235 | cds_lfht_alloc_bucket_table(ht, i); |
4105056a | 1236 | |
4105056a | 1237 | /* |
1ee8f000 LJ |
1238 | * Set all bucket nodes reverse hash values for a level and |
1239 | * link all bucket nodes into the table. | |
4105056a | 1240 | */ |
dc1da8f6 | 1241 | init_table_populate(ht, i, len); |
4105056a | 1242 | |
f9c80341 MD |
1243 | /* |
1244 | * Update table size. | |
1245 | */ | |
1246 | cmm_smp_wmb(); /* populate data before RCU size */ | |
7b3893e4 | 1247 | CMM_STORE_SHARED(ht->size, 1UL << i); |
f9c80341 | 1248 | |
4f6e90b7 | 1249 | dbg_printf("init new size: %lu\n", 1UL << i); |
4105056a MD |
1250 | if (CMM_LOAD_SHARED(ht->in_progress_destroy)) |
1251 | break; | |
1252 | } | |
1253 | } | |
1254 | ||
e8de508e MD |
1255 | /* |
1256 | * Holding RCU read lock to protect _cds_lfht_remove against memory | |
1257 | * reclaim that could be performed by other call_rcu worker threads (ABA | |
1258 | * problem). | |
1259 | * For a single level, we logically remove and garbage collect each node. | |
1260 | * | |
1261 | * As a design choice, we perform logical removal and garbage collection on a | |
1262 | * node-per-node basis to simplify this algorithm. We also assume keeping good | |
1263 | * cache locality of the operation would overweight possible performance gain | |
1264 | * that could be achieved by batching garbage collection for multiple levels. | |
1265 | * However, this would have to be justified by benchmarks. | |
1266 | * | |
1267 | * Concurrent removal and add operations are helping us perform garbage | |
1268 | * collection of logically removed nodes. We guarantee that all logically | |
1269 | * removed nodes have been garbage-collected (unlinked) before call_rcu is | |
1ee8f000 | 1270 | * invoked to free a hole level of bucket nodes (after a grace period). |
e8de508e | 1271 | * |
1f67ba50 MD |
1272 | * Logical removal and garbage collection can therefore be done in batch |
1273 | * or on a node-per-node basis, as long as the guarantee above holds. | |
9ee0fc9a | 1274 | * |
b7d619b0 MD |
1275 | * When we reach a certain length, we can split this removal over many worker |
1276 | * threads, based on the number of CPUs available in the system. This should | |
1277 | * take care of not letting resize process lag behind too many concurrent | |
9ee0fc9a | 1278 | * updater threads actively inserting into the hash table. |
e8de508e | 1279 | */ |
4105056a | 1280 | static |
b7d619b0 MD |
1281 | void remove_table_partition(struct cds_lfht *ht, unsigned long i, |
1282 | unsigned long start, unsigned long len) | |
4105056a | 1283 | { |
9d72a73f | 1284 | unsigned long j, size = 1UL << (i - 1); |
4105056a | 1285 | |
d0d8f9aa | 1286 | assert(i > MIN_TABLE_ORDER); |
7b17c13e | 1287 | ht->flavor->read_lock(); |
9d72a73f | 1288 | for (j = size + start; j < size + start + len; j++) { |
2e2ce1e9 LJ |
1289 | struct cds_lfht_node *fini_bucket = bucket_at(ht, j); |
1290 | struct cds_lfht_node *parent_bucket = bucket_at(ht, j - size); | |
9d72a73f LJ |
1291 | |
1292 | assert(j >= size && j < (size << 1)); | |
1293 | dbg_printf("remove entry: order %lu index %lu hash %lu\n", | |
1294 | i, j, j); | |
2e2ce1e9 LJ |
1295 | /* Set the REMOVED_FLAG to freeze the ->next for gc */ |
1296 | uatomic_or(&fini_bucket->next, REMOVED_FLAG); | |
1297 | _cds_lfht_gc_bucket(parent_bucket, fini_bucket); | |
abc490a1 | 1298 | } |
7b17c13e | 1299 | ht->flavor->read_unlock(); |
b7d619b0 MD |
1300 | } |
1301 | ||
1302 | static | |
1303 | void remove_table(struct cds_lfht *ht, unsigned long i, unsigned long len) | |
1304 | { | |
1305 | ||
1306 | assert(nr_cpus_mask != -1); | |
6083a889 | 1307 | if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD) { |
7b17c13e | 1308 | ht->flavor->thread_online(); |
b7d619b0 | 1309 | remove_table_partition(ht, i, 0, len); |
7b17c13e | 1310 | ht->flavor->thread_offline(); |
b7d619b0 MD |
1311 | return; |
1312 | } | |
1313 | partition_resize_helper(ht, i, len, remove_table_partition); | |
2ed95849 MD |
1314 | } |
1315 | ||
61adb337 MD |
1316 | /* |
1317 | * fini_table() is never called for first_order == 0, which is why | |
1318 | * free_by_rcu_order == 0 can be used as criterion to know if free must | |
1319 | * be called. | |
1320 | */ | |
1475579c | 1321 | static |
4105056a | 1322 | void fini_table(struct cds_lfht *ht, |
93d46c39 | 1323 | unsigned long first_order, unsigned long last_order) |
1475579c | 1324 | { |
93d46c39 | 1325 | long i; |
48f1b16d | 1326 | unsigned long free_by_rcu_order = 0; |
1475579c | 1327 | |
93d46c39 LJ |
1328 | dbg_printf("fini table: first_order %lu last_order %lu\n", |
1329 | first_order, last_order); | |
d0d8f9aa | 1330 | assert(first_order > MIN_TABLE_ORDER); |
93d46c39 | 1331 | for (i = last_order; i >= first_order; i--) { |
4105056a | 1332 | unsigned long len; |
1475579c | 1333 | |
4f6e90b7 | 1334 | len = 1UL << (i - 1); |
1475579c | 1335 | dbg_printf("fini order %lu len: %lu\n", i, len); |
4105056a | 1336 | |
4d676753 | 1337 | /* Stop shrink if the resize target changes under us */ |
7b3893e4 | 1338 | if (CMM_LOAD_SHARED(ht->resize_target) > (1UL << (i - 1))) |
4d676753 MD |
1339 | break; |
1340 | ||
1341 | cmm_smp_wmb(); /* populate data before RCU size */ | |
7b3893e4 | 1342 | CMM_STORE_SHARED(ht->size, 1UL << (i - 1)); |
4d676753 MD |
1343 | |
1344 | /* | |
1345 | * We need to wait for all add operations to reach Q.S. (and | |
1346 | * thus use the new table for lookups) before we can start | |
1ee8f000 | 1347 | * releasing the old bucket nodes. Otherwise their lookup will |
4d676753 MD |
1348 | * return a logically removed node as insert position. |
1349 | */ | |
7b17c13e | 1350 | ht->flavor->update_synchronize_rcu(); |
48f1b16d LJ |
1351 | if (free_by_rcu_order) |
1352 | cds_lfht_free_bucket_table(ht, free_by_rcu_order); | |
4d676753 | 1353 | |
21263e21 | 1354 | /* |
1ee8f000 LJ |
1355 | * Set "removed" flag in bucket nodes about to be removed. |
1356 | * Unlink all now-logically-removed bucket node pointers. | |
4105056a MD |
1357 | * Concurrent add/remove operation are helping us doing |
1358 | * the gc. | |
21263e21 | 1359 | */ |
4105056a MD |
1360 | remove_table(ht, i, len); |
1361 | ||
48f1b16d | 1362 | free_by_rcu_order = i; |
4105056a MD |
1363 | |
1364 | dbg_printf("fini new size: %lu\n", 1UL << i); | |
1475579c MD |
1365 | if (CMM_LOAD_SHARED(ht->in_progress_destroy)) |
1366 | break; | |
1367 | } | |
0d14ceb2 | 1368 | |
48f1b16d | 1369 | if (free_by_rcu_order) { |
7b17c13e | 1370 | ht->flavor->update_synchronize_rcu(); |
48f1b16d | 1371 | cds_lfht_free_bucket_table(ht, free_by_rcu_order); |
0d14ceb2 | 1372 | } |
1475579c MD |
1373 | } |
1374 | ||
ff0d69de | 1375 | static |
1ee8f000 | 1376 | void cds_lfht_create_bucket(struct cds_lfht *ht, unsigned long size) |
ff0d69de | 1377 | { |
04db56f8 | 1378 | struct cds_lfht_node *prev, *node; |
9d72a73f | 1379 | unsigned long order, len, i; |
ff0d69de | 1380 | |
48f1b16d | 1381 | cds_lfht_alloc_bucket_table(ht, 0); |
ff0d69de | 1382 | |
9d72a73f LJ |
1383 | dbg_printf("create bucket: order 0 index 0 hash 0\n"); |
1384 | node = bucket_at(ht, 0); | |
1385 | node->next = flag_bucket(get_end()); | |
1386 | node->reverse_hash = 0; | |
ff0d69de | 1387 | |
5bc6b66f | 1388 | for (order = 1; order < cds_lfht_get_count_order_ulong(size) + 1; order++) { |
ff0d69de | 1389 | len = 1UL << (order - 1); |
48f1b16d | 1390 | cds_lfht_alloc_bucket_table(ht, order); |
ff0d69de | 1391 | |
9d72a73f LJ |
1392 | for (i = 0; i < len; i++) { |
1393 | /* | |
1394 | * Now, we are trying to init the node with the | |
1395 | * hash=(len+i) (which is also a bucket with the | |
1396 | * index=(len+i)) and insert it into the hash table, | |
1397 | * so this node has to be inserted after the bucket | |
1398 | * with the index=(len+i)&(len-1)=i. And because there | |
1399 | * is no other non-bucket node nor bucket node with | |
1400 | * larger index/hash inserted, so the bucket node | |
1401 | * being inserted should be inserted directly linked | |
1402 | * after the bucket node with index=i. | |
1403 | */ | |
1404 | prev = bucket_at(ht, i); | |
1405 | node = bucket_at(ht, len + i); | |
ff0d69de | 1406 | |
1ee8f000 | 1407 | dbg_printf("create bucket: order %lu index %lu hash %lu\n", |
9d72a73f LJ |
1408 | order, len + i, len + i); |
1409 | node->reverse_hash = bit_reverse_ulong(len + i); | |
1410 | ||
1411 | /* insert after prev */ | |
1412 | assert(is_bucket(prev->next)); | |
ff0d69de | 1413 | node->next = prev->next; |
1ee8f000 | 1414 | prev->next = flag_bucket(node); |
ff0d69de LJ |
1415 | } |
1416 | } | |
1417 | } | |
1418 | ||
0422d92c | 1419 | struct cds_lfht *_cds_lfht_new(unsigned long init_size, |
0722081a | 1420 | unsigned long min_nr_alloc_buckets, |
747d725c | 1421 | unsigned long max_nr_buckets, |
b8af5011 | 1422 | int flags, |
0b6aa001 | 1423 | const struct cds_lfht_mm_type *mm, |
7b17c13e | 1424 | const struct rcu_flavor_struct *flavor, |
b7d619b0 | 1425 | pthread_attr_t *attr) |
abc490a1 | 1426 | { |
14044b37 | 1427 | struct cds_lfht *ht; |
24365af7 | 1428 | unsigned long order; |
abc490a1 | 1429 | |
0722081a LJ |
1430 | /* min_nr_alloc_buckets must be power of two */ |
1431 | if (!min_nr_alloc_buckets || (min_nr_alloc_buckets & (min_nr_alloc_buckets - 1))) | |
5488222b | 1432 | return NULL; |
747d725c | 1433 | |
8129be4e | 1434 | /* init_size must be power of two */ |
5488222b | 1435 | if (!init_size || (init_size & (init_size - 1))) |
8129be4e | 1436 | return NULL; |
747d725c | 1437 | |
c1888f3a MD |
1438 | /* |
1439 | * Memory management plugin default. | |
1440 | */ | |
1441 | if (!mm) { | |
5a2141a7 MD |
1442 | if (CAA_BITS_PER_LONG > 32 |
1443 | && max_nr_buckets | |
c1888f3a MD |
1444 | && max_nr_buckets <= (1ULL << 32)) { |
1445 | /* | |
1446 | * For 64-bit architectures, with max number of | |
1447 | * buckets small enough not to use the entire | |
1448 | * 64-bit memory mapping space (and allowing a | |
1449 | * fair number of hash table instances), use the | |
1450 | * mmap allocator, which is faster than the | |
1451 | * order allocator. | |
1452 | */ | |
1453 | mm = &cds_lfht_mm_mmap; | |
1454 | } else { | |
1455 | /* | |
1456 | * The fallback is to use the order allocator. | |
1457 | */ | |
1458 | mm = &cds_lfht_mm_order; | |
1459 | } | |
1460 | } | |
1461 | ||
0b6aa001 LJ |
1462 | /* max_nr_buckets == 0 for order based mm means infinite */ |
1463 | if (mm == &cds_lfht_mm_order && !max_nr_buckets) | |
747d725c LJ |
1464 | max_nr_buckets = 1UL << (MAX_TABLE_ORDER - 1); |
1465 | ||
1466 | /* max_nr_buckets must be power of two */ | |
1467 | if (!max_nr_buckets || (max_nr_buckets & (max_nr_buckets - 1))) | |
1468 | return NULL; | |
1469 | ||
0722081a | 1470 | min_nr_alloc_buckets = max(min_nr_alloc_buckets, MIN_TABLE_SIZE); |
d0d8f9aa | 1471 | init_size = max(init_size, MIN_TABLE_SIZE); |
747d725c LJ |
1472 | max_nr_buckets = max(max_nr_buckets, min_nr_alloc_buckets); |
1473 | init_size = min(init_size, max_nr_buckets); | |
0b6aa001 LJ |
1474 | |
1475 | ht = mm->alloc_cds_lfht(min_nr_alloc_buckets, max_nr_buckets); | |
b7d619b0 | 1476 | assert(ht); |
0b6aa001 LJ |
1477 | assert(ht->mm == mm); |
1478 | assert(ht->bucket_at == mm->bucket_at); | |
1479 | ||
b5d6b20f | 1480 | ht->flags = flags; |
7b17c13e | 1481 | ht->flavor = flavor; |
b7d619b0 | 1482 | ht->resize_attr = attr; |
5afadd12 | 1483 | alloc_split_items_count(ht); |
abc490a1 MD |
1484 | /* this mutex should not nest in read-side C.S. */ |
1485 | pthread_mutex_init(&ht->resize_mutex, NULL); | |
5bc6b66f | 1486 | order = cds_lfht_get_count_order_ulong(init_size); |
7b3893e4 | 1487 | ht->resize_target = 1UL << order; |
1ee8f000 | 1488 | cds_lfht_create_bucket(ht, 1UL << order); |
7b3893e4 | 1489 | ht->size = 1UL << order; |
abc490a1 MD |
1490 | return ht; |
1491 | } | |
1492 | ||
6f554439 | 1493 | void cds_lfht_lookup(struct cds_lfht *ht, unsigned long hash, |
996ff57c | 1494 | cds_lfht_match_fct match, const void *key, |
6f554439 | 1495 | struct cds_lfht_iter *iter) |
2ed95849 | 1496 | { |
04db56f8 | 1497 | struct cds_lfht_node *node, *next, *bucket; |
0422d92c | 1498 | unsigned long reverse_hash, size; |
2ed95849 | 1499 | |
abc490a1 | 1500 | reverse_hash = bit_reverse_ulong(hash); |
464a1ec9 | 1501 | |
7b3893e4 | 1502 | size = rcu_dereference(ht->size); |
04db56f8 | 1503 | bucket = lookup_bucket(ht, size, hash); |
1ee8f000 | 1504 | /* We can always skip the bucket node initially */ |
04db56f8 | 1505 | node = rcu_dereference(bucket->next); |
bb7b2f26 | 1506 | node = clear_flag(node); |
2ed95849 | 1507 | for (;;) { |
8ed51e04 | 1508 | if (caa_unlikely(is_end(node))) { |
96ad1112 | 1509 | node = next = NULL; |
abc490a1 | 1510 | break; |
bb7b2f26 | 1511 | } |
04db56f8 | 1512 | if (caa_unlikely(node->reverse_hash > reverse_hash)) { |
96ad1112 | 1513 | node = next = NULL; |
abc490a1 | 1514 | break; |
2ed95849 | 1515 | } |
04db56f8 | 1516 | next = rcu_dereference(node->next); |
7f52427b | 1517 | assert(node == clear_flag(node)); |
8ed51e04 | 1518 | if (caa_likely(!is_removed(next)) |
1ee8f000 | 1519 | && !is_bucket(next) |
04db56f8 | 1520 | && node->reverse_hash == reverse_hash |
0422d92c | 1521 | && caa_likely(match(node, key))) { |
273399de | 1522 | break; |
2ed95849 | 1523 | } |
1b81fe1a | 1524 | node = clear_flag(next); |
2ed95849 | 1525 | } |
a85eff52 | 1526 | assert(!node || !is_bucket(CMM_LOAD_SHARED(node->next))); |
adc0de68 MD |
1527 | iter->node = node; |
1528 | iter->next = next; | |
abc490a1 | 1529 | } |
e0ba718a | 1530 | |
0422d92c | 1531 | void cds_lfht_next_duplicate(struct cds_lfht *ht, cds_lfht_match_fct match, |
996ff57c | 1532 | const void *key, struct cds_lfht_iter *iter) |
a481e5ff | 1533 | { |
adc0de68 | 1534 | struct cds_lfht_node *node, *next; |
a481e5ff | 1535 | unsigned long reverse_hash; |
a481e5ff | 1536 | |
adc0de68 | 1537 | node = iter->node; |
04db56f8 | 1538 | reverse_hash = node->reverse_hash; |
adc0de68 | 1539 | next = iter->next; |
a481e5ff MD |
1540 | node = clear_flag(next); |
1541 | ||
1542 | for (;;) { | |
8ed51e04 | 1543 | if (caa_unlikely(is_end(node))) { |
96ad1112 | 1544 | node = next = NULL; |
a481e5ff | 1545 | break; |
bb7b2f26 | 1546 | } |
04db56f8 | 1547 | if (caa_unlikely(node->reverse_hash > reverse_hash)) { |
96ad1112 | 1548 | node = next = NULL; |
a481e5ff MD |
1549 | break; |
1550 | } | |
04db56f8 | 1551 | next = rcu_dereference(node->next); |
8ed51e04 | 1552 | if (caa_likely(!is_removed(next)) |
1ee8f000 | 1553 | && !is_bucket(next) |
04db56f8 | 1554 | && caa_likely(match(node, key))) { |
a481e5ff MD |
1555 | break; |
1556 | } | |
1557 | node = clear_flag(next); | |
1558 | } | |
a85eff52 | 1559 | assert(!node || !is_bucket(CMM_LOAD_SHARED(node->next))); |
adc0de68 MD |
1560 | iter->node = node; |
1561 | iter->next = next; | |
a481e5ff MD |
1562 | } |
1563 | ||
4e9b9fbf MD |
1564 | void cds_lfht_next(struct cds_lfht *ht, struct cds_lfht_iter *iter) |
1565 | { | |
1566 | struct cds_lfht_node *node, *next; | |
1567 | ||
853395e1 | 1568 | node = clear_flag(iter->next); |
4e9b9fbf | 1569 | for (;;) { |
8ed51e04 | 1570 | if (caa_unlikely(is_end(node))) { |
4e9b9fbf MD |
1571 | node = next = NULL; |
1572 | break; | |
1573 | } | |
04db56f8 | 1574 | next = rcu_dereference(node->next); |
8ed51e04 | 1575 | if (caa_likely(!is_removed(next)) |
1ee8f000 | 1576 | && !is_bucket(next)) { |
4e9b9fbf MD |
1577 | break; |
1578 | } | |
1579 | node = clear_flag(next); | |
1580 | } | |
a85eff52 | 1581 | assert(!node || !is_bucket(CMM_LOAD_SHARED(node->next))); |
4e9b9fbf MD |
1582 | iter->node = node; |
1583 | iter->next = next; | |
1584 | } | |
1585 | ||
1586 | void cds_lfht_first(struct cds_lfht *ht, struct cds_lfht_iter *iter) | |
1587 | { | |
4e9b9fbf | 1588 | /* |
1ee8f000 | 1589 | * Get next after first bucket node. The first bucket node is the |
4e9b9fbf MD |
1590 | * first node of the linked list. |
1591 | */ | |
9d72a73f | 1592 | iter->next = bucket_at(ht, 0)->next; |
4e9b9fbf MD |
1593 | cds_lfht_next(ht, iter); |
1594 | } | |
1595 | ||
0422d92c MD |
1596 | void cds_lfht_add(struct cds_lfht *ht, unsigned long hash, |
1597 | struct cds_lfht_node *node) | |
abc490a1 | 1598 | { |
0422d92c | 1599 | unsigned long size; |
ab7d5fc6 | 1600 | |
709bacf9 | 1601 | node->reverse_hash = bit_reverse_ulong(hash); |
7b3893e4 | 1602 | size = rcu_dereference(ht->size); |
91a75cc5 | 1603 | _cds_lfht_add(ht, hash, NULL, NULL, size, node, NULL, 0); |
14360f1c | 1604 | ht_count_add(ht, size, hash); |
3eca1b8c MD |
1605 | } |
1606 | ||
14044b37 | 1607 | struct cds_lfht_node *cds_lfht_add_unique(struct cds_lfht *ht, |
6f554439 | 1608 | unsigned long hash, |
0422d92c | 1609 | cds_lfht_match_fct match, |
996ff57c | 1610 | const void *key, |
48ed1c18 | 1611 | struct cds_lfht_node *node) |
3eca1b8c | 1612 | { |
0422d92c | 1613 | unsigned long size; |
83beee94 | 1614 | struct cds_lfht_iter iter; |
3eca1b8c | 1615 | |
709bacf9 | 1616 | node->reverse_hash = bit_reverse_ulong(hash); |
7b3893e4 | 1617 | size = rcu_dereference(ht->size); |
91a75cc5 | 1618 | _cds_lfht_add(ht, hash, match, key, size, node, &iter, 0); |
83beee94 | 1619 | if (iter.node == node) |
14360f1c | 1620 | ht_count_add(ht, size, hash); |
83beee94 | 1621 | return iter.node; |
2ed95849 MD |
1622 | } |
1623 | ||
9357c415 | 1624 | struct cds_lfht_node *cds_lfht_add_replace(struct cds_lfht *ht, |
6f554439 | 1625 | unsigned long hash, |
0422d92c | 1626 | cds_lfht_match_fct match, |
996ff57c | 1627 | const void *key, |
48ed1c18 MD |
1628 | struct cds_lfht_node *node) |
1629 | { | |
0422d92c | 1630 | unsigned long size; |
83beee94 | 1631 | struct cds_lfht_iter iter; |
48ed1c18 | 1632 | |
709bacf9 | 1633 | node->reverse_hash = bit_reverse_ulong(hash); |
7b3893e4 | 1634 | size = rcu_dereference(ht->size); |
83beee94 | 1635 | for (;;) { |
91a75cc5 | 1636 | _cds_lfht_add(ht, hash, match, key, size, node, &iter, 0); |
83beee94 | 1637 | if (iter.node == node) { |
14360f1c | 1638 | ht_count_add(ht, size, hash); |
83beee94 MD |
1639 | return NULL; |
1640 | } | |
1641 | ||
1642 | if (!_cds_lfht_replace(ht, size, iter.node, iter.next, node)) | |
1643 | return iter.node; | |
1644 | } | |
48ed1c18 MD |
1645 | } |
1646 | ||
2e79c445 MD |
1647 | int cds_lfht_replace(struct cds_lfht *ht, |
1648 | struct cds_lfht_iter *old_iter, | |
1649 | unsigned long hash, | |
1650 | cds_lfht_match_fct match, | |
1651 | const void *key, | |
9357c415 MD |
1652 | struct cds_lfht_node *new_node) |
1653 | { | |
1654 | unsigned long size; | |
1655 | ||
709bacf9 | 1656 | new_node->reverse_hash = bit_reverse_ulong(hash); |
2e79c445 MD |
1657 | if (!old_iter->node) |
1658 | return -ENOENT; | |
1659 | if (caa_unlikely(old_iter->node->reverse_hash != new_node->reverse_hash)) | |
1660 | return -EINVAL; | |
1661 | if (caa_unlikely(!match(old_iter->node, key))) | |
1662 | return -EINVAL; | |
7b3893e4 | 1663 | size = rcu_dereference(ht->size); |
9357c415 MD |
1664 | return _cds_lfht_replace(ht, size, old_iter->node, old_iter->next, |
1665 | new_node); | |
1666 | } | |
1667 | ||
bc8c3c74 | 1668 | int cds_lfht_del(struct cds_lfht *ht, struct cds_lfht_node *node) |
2ed95849 | 1669 | { |
14360f1c | 1670 | unsigned long size, hash; |
df44348d | 1671 | int ret; |
abc490a1 | 1672 | |
7b3893e4 | 1673 | size = rcu_dereference(ht->size); |
bc8c3c74 | 1674 | ret = _cds_lfht_del(ht, size, node); |
14360f1c | 1675 | if (!ret) { |
bc8c3c74 | 1676 | hash = bit_reverse_ulong(node->reverse_hash); |
14360f1c LJ |
1677 | ht_count_del(ht, size, hash); |
1678 | } | |
df44348d | 1679 | return ret; |
2ed95849 | 1680 | } |
ab7d5fc6 | 1681 | |
df55172a MD |
1682 | int cds_lfht_is_node_deleted(struct cds_lfht_node *node) |
1683 | { | |
a85eff52 | 1684 | return is_removed(CMM_LOAD_SHARED(node->next)); |
df55172a MD |
1685 | } |
1686 | ||
abc490a1 | 1687 | static |
1ee8f000 | 1688 | int cds_lfht_delete_bucket(struct cds_lfht *ht) |
674f7a69 | 1689 | { |
14044b37 | 1690 | struct cds_lfht_node *node; |
4105056a | 1691 | unsigned long order, i, size; |
674f7a69 | 1692 | |
abc490a1 | 1693 | /* Check that the table is empty */ |
9d72a73f | 1694 | node = bucket_at(ht, 0); |
abc490a1 | 1695 | do { |
04db56f8 | 1696 | node = clear_flag(node)->next; |
1ee8f000 | 1697 | if (!is_bucket(node)) |
abc490a1 | 1698 | return -EPERM; |
273399de | 1699 | assert(!is_removed(node)); |
bb7b2f26 | 1700 | } while (!is_end(node)); |
4105056a MD |
1701 | /* |
1702 | * size accessed without rcu_dereference because hash table is | |
1703 | * being destroyed. | |
1704 | */ | |
7b3893e4 | 1705 | size = ht->size; |
1f67ba50 | 1706 | /* Internal sanity check: all nodes left should be buckets */ |
48f1b16d LJ |
1707 | for (i = 0; i < size; i++) { |
1708 | node = bucket_at(ht, i); | |
1709 | dbg_printf("delete bucket: index %lu expected hash %lu hash %lu\n", | |
1710 | i, i, bit_reverse_ulong(node->reverse_hash)); | |
1711 | assert(is_bucket(node->next)); | |
1712 | } | |
24365af7 | 1713 | |
5bc6b66f | 1714 | for (order = cds_lfht_get_count_order_ulong(size); (long)order >= 0; order--) |
48f1b16d | 1715 | cds_lfht_free_bucket_table(ht, order); |
5488222b | 1716 | |
abc490a1 | 1717 | return 0; |
674f7a69 MD |
1718 | } |
1719 | ||
1720 | /* | |
1721 | * Should only be called when no more concurrent readers nor writers can | |
1722 | * possibly access the table. | |
1723 | */ | |
b7d619b0 | 1724 | int cds_lfht_destroy(struct cds_lfht *ht, pthread_attr_t **attr) |
674f7a69 | 1725 | { |
5e28c532 MD |
1726 | int ret; |
1727 | ||
848d4088 | 1728 | /* Wait for in-flight resize operations to complete */ |
24953e08 MD |
1729 | _CMM_STORE_SHARED(ht->in_progress_destroy, 1); |
1730 | cmm_smp_mb(); /* Store destroy before load resize */ | |
285b4481 | 1731 | ht->flavor->thread_offline(); |
848d4088 MD |
1732 | while (uatomic_read(&ht->in_progress_resize)) |
1733 | poll(NULL, 0, 100); /* wait for 100ms */ | |
285b4481 | 1734 | ht->flavor->thread_online(); |
1ee8f000 | 1735 | ret = cds_lfht_delete_bucket(ht); |
abc490a1 MD |
1736 | if (ret) |
1737 | return ret; | |
5afadd12 | 1738 | free_split_items_count(ht); |
b7d619b0 MD |
1739 | if (attr) |
1740 | *attr = ht->resize_attr; | |
98808fb1 | 1741 | poison_free(ht); |
5e28c532 | 1742 | return ret; |
674f7a69 MD |
1743 | } |
1744 | ||
14044b37 | 1745 | void cds_lfht_count_nodes(struct cds_lfht *ht, |
d933dd0e | 1746 | long *approx_before, |
273399de | 1747 | unsigned long *count, |
d933dd0e | 1748 | long *approx_after) |
273399de | 1749 | { |
14044b37 | 1750 | struct cds_lfht_node *node, *next; |
caf3653d | 1751 | unsigned long nr_bucket = 0, nr_removed = 0; |
273399de | 1752 | |
7ed7682f | 1753 | *approx_before = 0; |
5afadd12 | 1754 | if (ht->split_count) { |
973e5e1b MD |
1755 | int i; |
1756 | ||
4c42f1b8 LJ |
1757 | for (i = 0; i < split_count_mask + 1; i++) { |
1758 | *approx_before += uatomic_read(&ht->split_count[i].add); | |
1759 | *approx_before -= uatomic_read(&ht->split_count[i].del); | |
973e5e1b MD |
1760 | } |
1761 | } | |
1762 | ||
273399de | 1763 | *count = 0; |
273399de | 1764 | |
1ee8f000 | 1765 | /* Count non-bucket nodes in the table */ |
9d72a73f | 1766 | node = bucket_at(ht, 0); |
273399de | 1767 | do { |
04db56f8 | 1768 | next = rcu_dereference(node->next); |
b198f0fd | 1769 | if (is_removed(next)) { |
1ee8f000 | 1770 | if (!is_bucket(next)) |
caf3653d | 1771 | (nr_removed)++; |
973e5e1b | 1772 | else |
1ee8f000 LJ |
1773 | (nr_bucket)++; |
1774 | } else if (!is_bucket(next)) | |
273399de | 1775 | (*count)++; |
24365af7 | 1776 | else |
1ee8f000 | 1777 | (nr_bucket)++; |
273399de | 1778 | node = clear_flag(next); |
bb7b2f26 | 1779 | } while (!is_end(node)); |
caf3653d | 1780 | dbg_printf("number of logically removed nodes: %lu\n", nr_removed); |
1ee8f000 | 1781 | dbg_printf("number of bucket nodes: %lu\n", nr_bucket); |
7ed7682f | 1782 | *approx_after = 0; |
5afadd12 | 1783 | if (ht->split_count) { |
973e5e1b MD |
1784 | int i; |
1785 | ||
4c42f1b8 LJ |
1786 | for (i = 0; i < split_count_mask + 1; i++) { |
1787 | *approx_after += uatomic_read(&ht->split_count[i].add); | |
1788 | *approx_after -= uatomic_read(&ht->split_count[i].del); | |
973e5e1b MD |
1789 | } |
1790 | } | |
273399de MD |
1791 | } |
1792 | ||
1475579c | 1793 | /* called with resize mutex held */ |
abc490a1 | 1794 | static |
4105056a | 1795 | void _do_cds_lfht_grow(struct cds_lfht *ht, |
1475579c | 1796 | unsigned long old_size, unsigned long new_size) |
abc490a1 | 1797 | { |
1475579c | 1798 | unsigned long old_order, new_order; |
1475579c | 1799 | |
5bc6b66f MD |
1800 | old_order = cds_lfht_get_count_order_ulong(old_size); |
1801 | new_order = cds_lfht_get_count_order_ulong(new_size); | |
1a401918 LJ |
1802 | dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n", |
1803 | old_size, old_order, new_size, new_order); | |
1475579c | 1804 | assert(new_size > old_size); |
93d46c39 | 1805 | init_table(ht, old_order + 1, new_order); |
abc490a1 MD |
1806 | } |
1807 | ||
1808 | /* called with resize mutex held */ | |
1809 | static | |
4105056a | 1810 | void _do_cds_lfht_shrink(struct cds_lfht *ht, |
1475579c | 1811 | unsigned long old_size, unsigned long new_size) |
464a1ec9 | 1812 | { |
1475579c | 1813 | unsigned long old_order, new_order; |
464a1ec9 | 1814 | |
d0d8f9aa | 1815 | new_size = max(new_size, MIN_TABLE_SIZE); |
5bc6b66f MD |
1816 | old_order = cds_lfht_get_count_order_ulong(old_size); |
1817 | new_order = cds_lfht_get_count_order_ulong(new_size); | |
1a401918 LJ |
1818 | dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n", |
1819 | old_size, old_order, new_size, new_order); | |
1475579c | 1820 | assert(new_size < old_size); |
1475579c | 1821 | |
1ee8f000 | 1822 | /* Remove and unlink all bucket nodes to remove. */ |
93d46c39 | 1823 | fini_table(ht, new_order + 1, old_order); |
464a1ec9 MD |
1824 | } |
1825 | ||
1475579c MD |
1826 | |
1827 | /* called with resize mutex held */ | |
1828 | static | |
1829 | void _do_cds_lfht_resize(struct cds_lfht *ht) | |
1830 | { | |
1831 | unsigned long new_size, old_size; | |
4105056a MD |
1832 | |
1833 | /* | |
1834 | * Resize table, re-do if the target size has changed under us. | |
1835 | */ | |
1836 | do { | |
d2be3620 MD |
1837 | assert(uatomic_read(&ht->in_progress_resize)); |
1838 | if (CMM_LOAD_SHARED(ht->in_progress_destroy)) | |
1839 | break; | |
7b3893e4 LJ |
1840 | ht->resize_initiated = 1; |
1841 | old_size = ht->size; | |
1842 | new_size = CMM_LOAD_SHARED(ht->resize_target); | |
4105056a MD |
1843 | if (old_size < new_size) |
1844 | _do_cds_lfht_grow(ht, old_size, new_size); | |
1845 | else if (old_size > new_size) | |
1846 | _do_cds_lfht_shrink(ht, old_size, new_size); | |
7b3893e4 | 1847 | ht->resize_initiated = 0; |
4105056a MD |
1848 | /* write resize_initiated before read resize_target */ |
1849 | cmm_smp_mb(); | |
7b3893e4 | 1850 | } while (ht->size != CMM_LOAD_SHARED(ht->resize_target)); |
1475579c MD |
1851 | } |
1852 | ||
abc490a1 | 1853 | static |
ab65b890 | 1854 | unsigned long resize_target_grow(struct cds_lfht *ht, unsigned long new_size) |
464a1ec9 | 1855 | { |
7b3893e4 | 1856 | return _uatomic_xchg_monotonic_increase(&ht->resize_target, new_size); |
464a1ec9 MD |
1857 | } |
1858 | ||
1475579c | 1859 | static |
4105056a | 1860 | void resize_target_update_count(struct cds_lfht *ht, |
b8af5011 | 1861 | unsigned long count) |
1475579c | 1862 | { |
d0d8f9aa | 1863 | count = max(count, MIN_TABLE_SIZE); |
747d725c | 1864 | count = min(count, ht->max_nr_buckets); |
7b3893e4 | 1865 | uatomic_set(&ht->resize_target, count); |
1475579c MD |
1866 | } |
1867 | ||
1868 | void cds_lfht_resize(struct cds_lfht *ht, unsigned long new_size) | |
464a1ec9 | 1869 | { |
4105056a | 1870 | resize_target_update_count(ht, new_size); |
7b3893e4 | 1871 | CMM_STORE_SHARED(ht->resize_initiated, 1); |
7b17c13e | 1872 | ht->flavor->thread_offline(); |
1475579c MD |
1873 | pthread_mutex_lock(&ht->resize_mutex); |
1874 | _do_cds_lfht_resize(ht); | |
1875 | pthread_mutex_unlock(&ht->resize_mutex); | |
7b17c13e | 1876 | ht->flavor->thread_online(); |
abc490a1 | 1877 | } |
464a1ec9 | 1878 | |
abc490a1 MD |
1879 | static |
1880 | void do_resize_cb(struct rcu_head *head) | |
1881 | { | |
1882 | struct rcu_resize_work *work = | |
1883 | caa_container_of(head, struct rcu_resize_work, head); | |
14044b37 | 1884 | struct cds_lfht *ht = work->ht; |
abc490a1 | 1885 | |
7b17c13e | 1886 | ht->flavor->thread_offline(); |
abc490a1 | 1887 | pthread_mutex_lock(&ht->resize_mutex); |
14044b37 | 1888 | _do_cds_lfht_resize(ht); |
abc490a1 | 1889 | pthread_mutex_unlock(&ht->resize_mutex); |
7b17c13e | 1890 | ht->flavor->thread_online(); |
98808fb1 | 1891 | poison_free(work); |
848d4088 MD |
1892 | cmm_smp_mb(); /* finish resize before decrement */ |
1893 | uatomic_dec(&ht->in_progress_resize); | |
464a1ec9 MD |
1894 | } |
1895 | ||
abc490a1 | 1896 | static |
f1f119ee | 1897 | void __cds_lfht_resize_lazy_launch(struct cds_lfht *ht) |
ab7d5fc6 | 1898 | { |
abc490a1 MD |
1899 | struct rcu_resize_work *work; |
1900 | ||
4105056a MD |
1901 | /* Store resize_target before read resize_initiated */ |
1902 | cmm_smp_mb(); | |
7b3893e4 | 1903 | if (!CMM_LOAD_SHARED(ht->resize_initiated)) { |
848d4088 | 1904 | uatomic_inc(&ht->in_progress_resize); |
59290e9d | 1905 | cmm_smp_mb(); /* increment resize count before load destroy */ |
ed35e6d8 MD |
1906 | if (CMM_LOAD_SHARED(ht->in_progress_destroy)) { |
1907 | uatomic_dec(&ht->in_progress_resize); | |
59290e9d | 1908 | return; |
ed35e6d8 | 1909 | } |
f9830efd | 1910 | work = malloc(sizeof(*work)); |
741f378e MD |
1911 | if (work == NULL) { |
1912 | dbg_printf("error allocating resize work, bailing out\n"); | |
1913 | uatomic_dec(&ht->in_progress_resize); | |
1914 | return; | |
1915 | } | |
f9830efd | 1916 | work->ht = ht; |
7b17c13e | 1917 | ht->flavor->update_call_rcu(&work->head, do_resize_cb); |
7b3893e4 | 1918 | CMM_STORE_SHARED(ht->resize_initiated, 1); |
f9830efd | 1919 | } |
ab7d5fc6 | 1920 | } |
3171717f | 1921 | |
f1f119ee LJ |
1922 | static |
1923 | void cds_lfht_resize_lazy_grow(struct cds_lfht *ht, unsigned long size, int growth) | |
1924 | { | |
1925 | unsigned long target_size = size << growth; | |
1926 | ||
747d725c | 1927 | target_size = min(target_size, ht->max_nr_buckets); |
f1f119ee LJ |
1928 | if (resize_target_grow(ht, target_size) >= target_size) |
1929 | return; | |
1930 | ||
1931 | __cds_lfht_resize_lazy_launch(ht); | |
1932 | } | |
1933 | ||
89bb121d LJ |
1934 | /* |
1935 | * We favor grow operations over shrink. A shrink operation never occurs | |
1936 | * if a grow operation is queued for lazy execution. A grow operation | |
1937 | * cancels any pending shrink lazy execution. | |
1938 | */ | |
3171717f | 1939 | static |
4105056a | 1940 | void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size, |
3171717f MD |
1941 | unsigned long count) |
1942 | { | |
b8af5011 MD |
1943 | if (!(ht->flags & CDS_LFHT_AUTO_RESIZE)) |
1944 | return; | |
d0d8f9aa | 1945 | count = max(count, MIN_TABLE_SIZE); |
747d725c | 1946 | count = min(count, ht->max_nr_buckets); |
89bb121d LJ |
1947 | if (count == size) |
1948 | return; /* Already the right size, no resize needed */ | |
1949 | if (count > size) { /* lazy grow */ | |
1950 | if (resize_target_grow(ht, count) >= count) | |
1951 | return; | |
1952 | } else { /* lazy shrink */ | |
1953 | for (;;) { | |
1954 | unsigned long s; | |
1955 | ||
7b3893e4 | 1956 | s = uatomic_cmpxchg(&ht->resize_target, size, count); |
89bb121d LJ |
1957 | if (s == size) |
1958 | break; /* no resize needed */ | |
1959 | if (s > size) | |
1960 | return; /* growing is/(was just) in progress */ | |
1961 | if (s <= count) | |
1962 | return; /* some other thread do shrink */ | |
1963 | size = s; | |
1964 | } | |
1965 | } | |
f1f119ee | 1966 | __cds_lfht_resize_lazy_launch(ht); |
3171717f | 1967 | } |