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