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 | |
74 | * before its memory is reclaimed. Only the thread which removal | |
75 | * successfully set the "removed" flag (with a cmpxchg) into a node's | |
76 | * next pointer is considered to have succeeded its removal (and thus | |
77 | * owns the node to reclaim). Because we garbage-collect starting from | |
1ee8f000 | 78 | * an invariant node (the start-of-bucket bucket node) up to the |
e753ff5a MD |
79 | * "removed" node (or find a reverse-hash that is higher), we are sure |
80 | * that a successful traversal of the chain leads to a chain that is | |
81 | * present in the linked-list (the start node is never removed) and | |
82 | * that is does not contain the "removed" node anymore, even if | |
83 | * concurrent delete/add operations are changing the structure of the | |
84 | * list concurrently. | |
29e669f6 MD |
85 | * - The add operation performs gargage collection of buckets if it |
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 | * |
1ee8f000 | 101 | * Bucket node tables: |
93d46c39 | 102 | * |
1ee8f000 LJ |
103 | * hash table hash table the last all bucket node tables |
104 | * order size bucket node 0 1 2 3 4 5 6(index) | |
93d46c39 LJ |
105 | * table size |
106 | * 0 1 1 1 | |
107 | * 1 2 1 1 1 | |
108 | * 2 4 2 1 1 2 | |
109 | * 3 8 4 1 1 2 4 | |
110 | * 4 16 8 1 1 2 4 8 | |
111 | * 5 32 16 1 1 2 4 8 16 | |
112 | * 6 64 32 1 1 2 4 8 16 32 | |
113 | * | |
1ee8f000 | 114 | * When growing/shrinking, we only focus on the last bucket node table |
93d46c39 LJ |
115 | * which size is (!order ? 1 : (1 << (order -1))). |
116 | * | |
117 | * Example for growing/shrinking: | |
1ee8f000 LJ |
118 | * grow hash table from order 5 to 6: init the index=6 bucket node table |
119 | * shrink hash table from order 6 to 5: fini the index=6 bucket node table | |
93d46c39 | 120 | * |
1475579c MD |
121 | * A bit of ascii art explanation: |
122 | * | |
123 | * Order index is the off-by-one compare to the actual power of 2 because | |
124 | * we use index 0 to deal with the 0 special-case. | |
125 | * | |
126 | * This shows the nodes for a small table ordered by reversed bits: | |
127 | * | |
128 | * bits reverse | |
129 | * 0 000 000 | |
130 | * 4 100 001 | |
131 | * 2 010 010 | |
132 | * 6 110 011 | |
133 | * 1 001 100 | |
134 | * 5 101 101 | |
135 | * 3 011 110 | |
136 | * 7 111 111 | |
137 | * | |
138 | * This shows the nodes in order of non-reversed bits, linked by | |
139 | * reversed-bit order. | |
140 | * | |
141 | * order bits reverse | |
142 | * 0 0 000 000 | |
0adc36a8 LJ |
143 | * 1 | 1 001 100 <- |
144 | * 2 | | 2 010 010 <- | | |
f6fdd688 | 145 | * | | | 3 011 110 | <- | |
1475579c MD |
146 | * 3 -> | | | 4 100 001 | | |
147 | * -> | | 5 101 101 | | |
148 | * -> | 6 110 011 | |
149 | * -> 7 111 111 | |
e753ff5a MD |
150 | */ |
151 | ||
2ed95849 MD |
152 | #define _LGPL_SOURCE |
153 | #include <stdlib.h> | |
e0ba718a MD |
154 | #include <errno.h> |
155 | #include <assert.h> | |
156 | #include <stdio.h> | |
abc490a1 | 157 | #include <stdint.h> |
f000907d | 158 | #include <string.h> |
e0ba718a | 159 | |
15cfbec7 | 160 | #include "config.h" |
2ed95849 | 161 | #include <urcu.h> |
abc490a1 | 162 | #include <urcu-call-rcu.h> |
7b17c13e | 163 | #include <urcu-flavor.h> |
a42cc659 MD |
164 | #include <urcu/arch.h> |
165 | #include <urcu/uatomic.h> | |
a42cc659 | 166 | #include <urcu/compiler.h> |
abc490a1 | 167 | #include <urcu/rculfhash.h> |
0b6aa001 | 168 | #include <rculfhash-internal.h> |
5e28c532 | 169 | #include <stdio.h> |
464a1ec9 | 170 | #include <pthread.h> |
44395fb7 | 171 | |
f8994aee | 172 | /* |
4c42f1b8 | 173 | * Split-counters lazily update the global counter each 1024 |
f8994aee MD |
174 | * addition/removal. It automatically keeps track of resize required. |
175 | * We use the bucket length as indicator for need to expand for small | |
176 | * tables and machines lacking per-cpu data suppport. | |
177 | */ | |
178 | #define COUNT_COMMIT_ORDER 10 | |
4ddbb355 | 179 | #define DEFAULT_SPLIT_COUNT_MASK 0xFUL |
6ea6bc67 MD |
180 | #define CHAIN_LEN_TARGET 1 |
181 | #define CHAIN_LEN_RESIZE_THRESHOLD 3 | |
2ed95849 | 182 | |
cd95516d | 183 | /* |
76a73da8 | 184 | * Define the minimum table size. |
cd95516d | 185 | */ |
d0d8f9aa LJ |
186 | #define MIN_TABLE_ORDER 0 |
187 | #define MIN_TABLE_SIZE (1UL << MIN_TABLE_ORDER) | |
cd95516d | 188 | |
b7d619b0 | 189 | /* |
1ee8f000 | 190 | * Minimum number of bucket nodes to touch per thread to parallelize grow/shrink. |
b7d619b0 | 191 | */ |
6083a889 MD |
192 | #define MIN_PARTITION_PER_THREAD_ORDER 12 |
193 | #define MIN_PARTITION_PER_THREAD (1UL << MIN_PARTITION_PER_THREAD_ORDER) | |
b7d619b0 | 194 | |
d95bd160 MD |
195 | /* |
196 | * The removed flag needs to be updated atomically with the pointer. | |
48ed1c18 | 197 | * It indicates that no node must attach to the node scheduled for |
b198f0fd | 198 | * removal, and that node garbage collection must be performed. |
1ee8f000 | 199 | * The bucket flag does not require to be updated atomically with the |
d95bd160 MD |
200 | * pointer, but it is added as a pointer low bit flag to save space. |
201 | */ | |
d37166c6 | 202 | #define REMOVED_FLAG (1UL << 0) |
1ee8f000 | 203 | #define BUCKET_FLAG (1UL << 1) |
b198f0fd | 204 | #define FLAGS_MASK ((1UL << 2) - 1) |
d37166c6 | 205 | |
bb7b2f26 | 206 | /* Value of the end pointer. Should not interact with flags. */ |
f9c80341 | 207 | #define END_VALUE NULL |
bb7b2f26 | 208 | |
7f52427b MD |
209 | /* |
210 | * ht_items_count: Split-counters counting the number of node addition | |
211 | * and removal in the table. Only used if the CDS_LFHT_ACCOUNTING flag | |
212 | * is set at hash table creation. | |
213 | * | |
214 | * These are free-running counters, never reset to zero. They count the | |
215 | * number of add/remove, and trigger every (1 << COUNT_COMMIT_ORDER) | |
216 | * operations to update the global counter. We choose a power-of-2 value | |
217 | * for the trigger to deal with 32 or 64-bit overflow of the counter. | |
218 | */ | |
df44348d | 219 | struct ht_items_count { |
860d07e8 | 220 | unsigned long add, del; |
df44348d MD |
221 | } __attribute__((aligned(CAA_CACHE_LINE_SIZE))); |
222 | ||
7f52427b MD |
223 | /* |
224 | * rcu_resize_work: Contains arguments passed to RCU worker thread | |
225 | * responsible for performing lazy resize. | |
226 | */ | |
abc490a1 MD |
227 | struct rcu_resize_work { |
228 | struct rcu_head head; | |
14044b37 | 229 | struct cds_lfht *ht; |
abc490a1 | 230 | }; |
2ed95849 | 231 | |
7f52427b MD |
232 | /* |
233 | * partition_resize_work: Contains arguments passed to worker threads | |
234 | * executing the hash table resize on partitions of the hash table | |
235 | * assigned to each processor's worker thread. | |
236 | */ | |
b7d619b0 | 237 | struct partition_resize_work { |
1af6e26e | 238 | pthread_t thread_id; |
b7d619b0 MD |
239 | struct cds_lfht *ht; |
240 | unsigned long i, start, len; | |
241 | void (*fct)(struct cds_lfht *ht, unsigned long i, | |
242 | unsigned long start, unsigned long len); | |
243 | }; | |
244 | ||
76a73da8 | 245 | static |
83beee94 | 246 | void _cds_lfht_add(struct cds_lfht *ht, |
0422d92c | 247 | cds_lfht_match_fct match, |
996ff57c | 248 | const void *key, |
83beee94 MD |
249 | unsigned long size, |
250 | struct cds_lfht_node *node, | |
251 | struct cds_lfht_iter *unique_ret, | |
1ee8f000 | 252 | int bucket); |
48ed1c18 | 253 | |
abc490a1 MD |
254 | /* |
255 | * Algorithm to reverse bits in a word by lookup table, extended to | |
256 | * 64-bit words. | |
f9830efd | 257 | * Source: |
abc490a1 | 258 | * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable |
f9830efd | 259 | * Originally from Public Domain. |
abc490a1 MD |
260 | */ |
261 | ||
262 | static const uint8_t BitReverseTable256[256] = | |
2ed95849 | 263 | { |
abc490a1 MD |
264 | #define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64 |
265 | #define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16) | |
266 | #define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 ) | |
267 | R6(0), R6(2), R6(1), R6(3) | |
268 | }; | |
269 | #undef R2 | |
270 | #undef R4 | |
271 | #undef R6 | |
2ed95849 | 272 | |
abc490a1 MD |
273 | static |
274 | uint8_t bit_reverse_u8(uint8_t v) | |
275 | { | |
276 | return BitReverseTable256[v]; | |
277 | } | |
ab7d5fc6 | 278 | |
abc490a1 MD |
279 | static __attribute__((unused)) |
280 | uint32_t bit_reverse_u32(uint32_t v) | |
281 | { | |
282 | return ((uint32_t) bit_reverse_u8(v) << 24) | | |
283 | ((uint32_t) bit_reverse_u8(v >> 8) << 16) | | |
284 | ((uint32_t) bit_reverse_u8(v >> 16) << 8) | | |
285 | ((uint32_t) bit_reverse_u8(v >> 24)); | |
2ed95849 MD |
286 | } |
287 | ||
abc490a1 MD |
288 | static __attribute__((unused)) |
289 | uint64_t bit_reverse_u64(uint64_t v) | |
2ed95849 | 290 | { |
abc490a1 MD |
291 | return ((uint64_t) bit_reverse_u8(v) << 56) | |
292 | ((uint64_t) bit_reverse_u8(v >> 8) << 48) | | |
293 | ((uint64_t) bit_reverse_u8(v >> 16) << 40) | | |
294 | ((uint64_t) bit_reverse_u8(v >> 24) << 32) | | |
295 | ((uint64_t) bit_reverse_u8(v >> 32) << 24) | | |
296 | ((uint64_t) bit_reverse_u8(v >> 40) << 16) | | |
297 | ((uint64_t) bit_reverse_u8(v >> 48) << 8) | | |
298 | ((uint64_t) bit_reverse_u8(v >> 56)); | |
299 | } | |
300 | ||
301 | static | |
302 | unsigned long bit_reverse_ulong(unsigned long v) | |
303 | { | |
304 | #if (CAA_BITS_PER_LONG == 32) | |
305 | return bit_reverse_u32(v); | |
306 | #else | |
307 | return bit_reverse_u64(v); | |
308 | #endif | |
309 | } | |
310 | ||
f9830efd | 311 | /* |
24365af7 MD |
312 | * fls: returns the position of the most significant bit. |
313 | * Returns 0 if no bit is set, else returns the position of the most | |
314 | * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit). | |
f9830efd | 315 | */ |
24365af7 MD |
316 | #if defined(__i386) || defined(__x86_64) |
317 | static inline | |
318 | unsigned int fls_u32(uint32_t x) | |
f9830efd | 319 | { |
24365af7 MD |
320 | int r; |
321 | ||
322 | asm("bsrl %1,%0\n\t" | |
323 | "jnz 1f\n\t" | |
324 | "movl $-1,%0\n\t" | |
325 | "1:\n\t" | |
326 | : "=r" (r) : "rm" (x)); | |
327 | return r + 1; | |
328 | } | |
329 | #define HAS_FLS_U32 | |
330 | #endif | |
331 | ||
332 | #if defined(__x86_64) | |
333 | static inline | |
334 | unsigned int fls_u64(uint64_t x) | |
335 | { | |
336 | long r; | |
337 | ||
338 | asm("bsrq %1,%0\n\t" | |
339 | "jnz 1f\n\t" | |
340 | "movq $-1,%0\n\t" | |
341 | "1:\n\t" | |
342 | : "=r" (r) : "rm" (x)); | |
343 | return r + 1; | |
344 | } | |
345 | #define HAS_FLS_U64 | |
346 | #endif | |
347 | ||
348 | #ifndef HAS_FLS_U64 | |
349 | static __attribute__((unused)) | |
350 | unsigned int fls_u64(uint64_t x) | |
351 | { | |
352 | unsigned int r = 64; | |
353 | ||
354 | if (!x) | |
355 | return 0; | |
356 | ||
357 | if (!(x & 0xFFFFFFFF00000000ULL)) { | |
358 | x <<= 32; | |
359 | r -= 32; | |
360 | } | |
361 | if (!(x & 0xFFFF000000000000ULL)) { | |
362 | x <<= 16; | |
363 | r -= 16; | |
364 | } | |
365 | if (!(x & 0xFF00000000000000ULL)) { | |
366 | x <<= 8; | |
367 | r -= 8; | |
368 | } | |
369 | if (!(x & 0xF000000000000000ULL)) { | |
370 | x <<= 4; | |
371 | r -= 4; | |
372 | } | |
373 | if (!(x & 0xC000000000000000ULL)) { | |
374 | x <<= 2; | |
375 | r -= 2; | |
376 | } | |
377 | if (!(x & 0x8000000000000000ULL)) { | |
378 | x <<= 1; | |
379 | r -= 1; | |
380 | } | |
381 | return r; | |
382 | } | |
383 | #endif | |
384 | ||
385 | #ifndef HAS_FLS_U32 | |
386 | static __attribute__((unused)) | |
387 | unsigned int fls_u32(uint32_t x) | |
388 | { | |
389 | unsigned int r = 32; | |
f9830efd | 390 | |
24365af7 MD |
391 | if (!x) |
392 | return 0; | |
393 | if (!(x & 0xFFFF0000U)) { | |
394 | x <<= 16; | |
395 | r -= 16; | |
396 | } | |
397 | if (!(x & 0xFF000000U)) { | |
398 | x <<= 8; | |
399 | r -= 8; | |
400 | } | |
401 | if (!(x & 0xF0000000U)) { | |
402 | x <<= 4; | |
403 | r -= 4; | |
404 | } | |
405 | if (!(x & 0xC0000000U)) { | |
406 | x <<= 2; | |
407 | r -= 2; | |
408 | } | |
409 | if (!(x & 0x80000000U)) { | |
410 | x <<= 1; | |
411 | r -= 1; | |
412 | } | |
413 | return r; | |
414 | } | |
415 | #endif | |
416 | ||
417 | unsigned int fls_ulong(unsigned long x) | |
f9830efd | 418 | { |
6887cc5e | 419 | #if (CAA_BITS_PER_LONG == 32) |
24365af7 MD |
420 | return fls_u32(x); |
421 | #else | |
422 | return fls_u64(x); | |
423 | #endif | |
424 | } | |
f9830efd | 425 | |
920f8ef6 LJ |
426 | /* |
427 | * Return the minimum order for which x <= (1UL << order). | |
428 | * Return -1 if x is 0. | |
429 | */ | |
24365af7 MD |
430 | int get_count_order_u32(uint32_t x) |
431 | { | |
920f8ef6 LJ |
432 | if (!x) |
433 | return -1; | |
24365af7 | 434 | |
920f8ef6 | 435 | return fls_u32(x - 1); |
24365af7 MD |
436 | } |
437 | ||
920f8ef6 LJ |
438 | /* |
439 | * Return the minimum order for which x <= (1UL << order). | |
440 | * Return -1 if x is 0. | |
441 | */ | |
24365af7 MD |
442 | int get_count_order_ulong(unsigned long x) |
443 | { | |
920f8ef6 LJ |
444 | if (!x) |
445 | return -1; | |
24365af7 | 446 | |
920f8ef6 | 447 | return fls_ulong(x - 1); |
f9830efd MD |
448 | } |
449 | ||
450 | static | |
ab65b890 | 451 | void cds_lfht_resize_lazy_grow(struct cds_lfht *ht, unsigned long size, int growth); |
f9830efd | 452 | |
f8994aee | 453 | static |
4105056a | 454 | void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size, |
f8994aee MD |
455 | unsigned long count); |
456 | ||
df44348d | 457 | static long nr_cpus_mask = -1; |
4c42f1b8 LJ |
458 | static long split_count_mask = -1; |
459 | ||
4ddbb355 | 460 | #if defined(HAVE_SYSCONF) |
4c42f1b8 LJ |
461 | static void ht_init_nr_cpus_mask(void) |
462 | { | |
463 | long maxcpus; | |
464 | ||
465 | maxcpus = sysconf(_SC_NPROCESSORS_CONF); | |
466 | if (maxcpus <= 0) { | |
467 | nr_cpus_mask = -2; | |
468 | return; | |
469 | } | |
470 | /* | |
471 | * round up number of CPUs to next power of two, so we | |
472 | * can use & for modulo. | |
473 | */ | |
474 | maxcpus = 1UL << get_count_order_ulong(maxcpus); | |
475 | nr_cpus_mask = maxcpus - 1; | |
476 | } | |
4ddbb355 LJ |
477 | #else /* #if defined(HAVE_SYSCONF) */ |
478 | static void ht_init_nr_cpus_mask(void) | |
479 | { | |
480 | nr_cpus_mask = -2; | |
481 | } | |
482 | #endif /* #else #if defined(HAVE_SYSCONF) */ | |
df44348d MD |
483 | |
484 | static | |
5afadd12 | 485 | void alloc_split_items_count(struct cds_lfht *ht) |
df44348d MD |
486 | { |
487 | struct ht_items_count *count; | |
488 | ||
4c42f1b8 LJ |
489 | if (nr_cpus_mask == -1) { |
490 | ht_init_nr_cpus_mask(); | |
4ddbb355 LJ |
491 | if (nr_cpus_mask < 0) |
492 | split_count_mask = DEFAULT_SPLIT_COUNT_MASK; | |
493 | else | |
494 | split_count_mask = nr_cpus_mask; | |
df44348d | 495 | } |
4c42f1b8 | 496 | |
4ddbb355 | 497 | assert(split_count_mask >= 0); |
5afadd12 LJ |
498 | |
499 | if (ht->flags & CDS_LFHT_ACCOUNTING) { | |
500 | ht->split_count = calloc(split_count_mask + 1, sizeof(*count)); | |
501 | assert(ht->split_count); | |
502 | } else { | |
503 | ht->split_count = NULL; | |
504 | } | |
df44348d MD |
505 | } |
506 | ||
507 | static | |
5afadd12 | 508 | void free_split_items_count(struct cds_lfht *ht) |
df44348d | 509 | { |
5afadd12 | 510 | poison_free(ht->split_count); |
df44348d MD |
511 | } |
512 | ||
14360f1c | 513 | #if defined(HAVE_SCHED_GETCPU) |
df44348d | 514 | static |
14360f1c | 515 | int ht_get_split_count_index(unsigned long hash) |
df44348d MD |
516 | { |
517 | int cpu; | |
518 | ||
4c42f1b8 | 519 | assert(split_count_mask >= 0); |
df44348d | 520 | cpu = sched_getcpu(); |
8ed51e04 | 521 | if (caa_unlikely(cpu < 0)) |
14360f1c | 522 | return hash & split_count_mask; |
df44348d | 523 | else |
4c42f1b8 | 524 | return cpu & split_count_mask; |
df44348d | 525 | } |
14360f1c LJ |
526 | #else /* #if defined(HAVE_SCHED_GETCPU) */ |
527 | static | |
528 | int ht_get_split_count_index(unsigned long hash) | |
529 | { | |
530 | return hash & split_count_mask; | |
531 | } | |
532 | #endif /* #else #if defined(HAVE_SCHED_GETCPU) */ | |
df44348d MD |
533 | |
534 | static | |
14360f1c | 535 | void ht_count_add(struct cds_lfht *ht, unsigned long size, unsigned long hash) |
df44348d | 536 | { |
4c42f1b8 LJ |
537 | unsigned long split_count; |
538 | int index; | |
df44348d | 539 | |
8ed51e04 | 540 | if (caa_unlikely(!ht->split_count)) |
3171717f | 541 | return; |
14360f1c | 542 | index = ht_get_split_count_index(hash); |
4c42f1b8 | 543 | split_count = uatomic_add_return(&ht->split_count[index].add, 1); |
8ed51e04 | 544 | if (caa_unlikely(!(split_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) { |
e3ecefd6 | 545 | long count; |
df44348d | 546 | |
4c42f1b8 | 547 | dbg_printf("add split count %lu\n", split_count); |
df44348d MD |
548 | count = uatomic_add_return(&ht->count, |
549 | 1UL << COUNT_COMMIT_ORDER); | |
550 | /* If power of 2 */ | |
551 | if (!(count & (count - 1))) { | |
4105056a | 552 | if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) < size) |
f8994aee | 553 | return; |
e3ecefd6 | 554 | dbg_printf("add set global %ld\n", count); |
4105056a | 555 | cds_lfht_resize_lazy_count(ht, size, |
6ea6bc67 | 556 | count >> (CHAIN_LEN_TARGET - 1)); |
df44348d MD |
557 | } |
558 | } | |
559 | } | |
560 | ||
561 | static | |
14360f1c | 562 | void ht_count_del(struct cds_lfht *ht, unsigned long size, unsigned long hash) |
df44348d | 563 | { |
4c42f1b8 LJ |
564 | unsigned long split_count; |
565 | int index; | |
df44348d | 566 | |
8ed51e04 | 567 | if (caa_unlikely(!ht->split_count)) |
3171717f | 568 | return; |
14360f1c | 569 | index = ht_get_split_count_index(hash); |
4c42f1b8 | 570 | split_count = uatomic_add_return(&ht->split_count[index].del, 1); |
8ed51e04 | 571 | if (caa_unlikely(!(split_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) { |
e3ecefd6 | 572 | long count; |
df44348d | 573 | |
4c42f1b8 | 574 | dbg_printf("del split count %lu\n", split_count); |
df44348d | 575 | count = uatomic_add_return(&ht->count, |
3171717f | 576 | -(1UL << COUNT_COMMIT_ORDER)); |
df44348d MD |
577 | /* If power of 2 */ |
578 | if (!(count & (count - 1))) { | |
4105056a | 579 | if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) >= size) |
f8994aee | 580 | return; |
e3ecefd6 MD |
581 | dbg_printf("del set global %ld\n", count); |
582 | /* | |
c941bb9e | 583 | * Don't shrink table if the number of nodes is below a |
e3ecefd6 MD |
584 | * certain threshold. |
585 | */ | |
4c42f1b8 | 586 | if (count < (1UL << COUNT_COMMIT_ORDER) * (split_count_mask + 1)) |
e3ecefd6 | 587 | return; |
4105056a | 588 | cds_lfht_resize_lazy_count(ht, size, |
6ea6bc67 | 589 | count >> (CHAIN_LEN_TARGET - 1)); |
df44348d MD |
590 | } |
591 | } | |
592 | } | |
593 | ||
f9830efd | 594 | static |
4105056a | 595 | void check_resize(struct cds_lfht *ht, unsigned long size, uint32_t chain_len) |
f9830efd | 596 | { |
f8994aee MD |
597 | unsigned long count; |
598 | ||
b8af5011 MD |
599 | if (!(ht->flags & CDS_LFHT_AUTO_RESIZE)) |
600 | return; | |
f8994aee MD |
601 | count = uatomic_read(&ht->count); |
602 | /* | |
603 | * Use bucket-local length for small table expand and for | |
604 | * environments lacking per-cpu data support. | |
605 | */ | |
606 | if (count >= (1UL << COUNT_COMMIT_ORDER)) | |
607 | return; | |
24365af7 | 608 | if (chain_len > 100) |
f0c29ed7 | 609 | dbg_printf("WARNING: large chain length: %u.\n", |
24365af7 | 610 | chain_len); |
3390d470 | 611 | if (chain_len >= CHAIN_LEN_RESIZE_THRESHOLD) |
ab65b890 | 612 | cds_lfht_resize_lazy_grow(ht, size, |
01370f0b | 613 | get_count_order_u32(chain_len - (CHAIN_LEN_TARGET - 1))); |
f9830efd MD |
614 | } |
615 | ||
abc490a1 | 616 | static |
14044b37 | 617 | struct cds_lfht_node *clear_flag(struct cds_lfht_node *node) |
abc490a1 | 618 | { |
14044b37 | 619 | return (struct cds_lfht_node *) (((unsigned long) node) & ~FLAGS_MASK); |
abc490a1 MD |
620 | } |
621 | ||
622 | static | |
14044b37 | 623 | int is_removed(struct cds_lfht_node *node) |
abc490a1 | 624 | { |
d37166c6 | 625 | return ((unsigned long) node) & REMOVED_FLAG; |
abc490a1 MD |
626 | } |
627 | ||
628 | static | |
14044b37 | 629 | struct cds_lfht_node *flag_removed(struct cds_lfht_node *node) |
abc490a1 | 630 | { |
14044b37 | 631 | return (struct cds_lfht_node *) (((unsigned long) node) | REMOVED_FLAG); |
abc490a1 MD |
632 | } |
633 | ||
f5596c94 | 634 | static |
1ee8f000 | 635 | int is_bucket(struct cds_lfht_node *node) |
f5596c94 | 636 | { |
1ee8f000 | 637 | return ((unsigned long) node) & BUCKET_FLAG; |
f5596c94 MD |
638 | } |
639 | ||
640 | static | |
1ee8f000 | 641 | struct cds_lfht_node *flag_bucket(struct cds_lfht_node *node) |
f5596c94 | 642 | { |
1ee8f000 | 643 | return (struct cds_lfht_node *) (((unsigned long) node) | BUCKET_FLAG); |
f5596c94 | 644 | } |
bb7b2f26 MD |
645 | |
646 | static | |
647 | struct cds_lfht_node *get_end(void) | |
648 | { | |
649 | return (struct cds_lfht_node *) END_VALUE; | |
650 | } | |
651 | ||
652 | static | |
653 | int is_end(struct cds_lfht_node *node) | |
654 | { | |
655 | return clear_flag(node) == (struct cds_lfht_node *) END_VALUE; | |
656 | } | |
657 | ||
abc490a1 | 658 | static |
ab65b890 LJ |
659 | unsigned long _uatomic_xchg_monotonic_increase(unsigned long *ptr, |
660 | unsigned long v) | |
abc490a1 MD |
661 | { |
662 | unsigned long old1, old2; | |
663 | ||
664 | old1 = uatomic_read(ptr); | |
665 | do { | |
666 | old2 = old1; | |
667 | if (old2 >= v) | |
f9830efd | 668 | return old2; |
abc490a1 | 669 | } while ((old1 = uatomic_cmpxchg(ptr, old2, v)) != old2); |
ab65b890 | 670 | return old2; |
abc490a1 MD |
671 | } |
672 | ||
48f1b16d LJ |
673 | static |
674 | void cds_lfht_alloc_bucket_table(struct cds_lfht *ht, unsigned long order) | |
675 | { | |
0b6aa001 | 676 | return ht->mm->alloc_bucket_table(ht, order); |
48f1b16d LJ |
677 | } |
678 | ||
679 | /* | |
680 | * cds_lfht_free_bucket_table() should be called with decreasing order. | |
681 | * When cds_lfht_free_bucket_table(0) is called, it means the whole | |
682 | * lfht is destroyed. | |
683 | */ | |
684 | static | |
685 | void cds_lfht_free_bucket_table(struct cds_lfht *ht, unsigned long order) | |
686 | { | |
0b6aa001 | 687 | return ht->mm->free_bucket_table(ht, order); |
48f1b16d LJ |
688 | } |
689 | ||
9d72a73f LJ |
690 | static inline |
691 | struct cds_lfht_node *bucket_at(struct cds_lfht *ht, unsigned long index) | |
f4a9cc0b | 692 | { |
0b6aa001 | 693 | return ht->bucket_at(ht, index); |
f4a9cc0b LJ |
694 | } |
695 | ||
9d72a73f LJ |
696 | static inline |
697 | struct cds_lfht_node *lookup_bucket(struct cds_lfht *ht, unsigned long size, | |
698 | unsigned long hash) | |
699 | { | |
700 | assert(size > 0); | |
701 | return bucket_at(ht, hash & (size - 1)); | |
702 | } | |
703 | ||
273399de MD |
704 | /* |
705 | * Remove all logically deleted nodes from a bucket up to a certain node key. | |
706 | */ | |
707 | static | |
1ee8f000 | 708 | void _cds_lfht_gc_bucket(struct cds_lfht_node *bucket, struct cds_lfht_node *node) |
273399de | 709 | { |
14044b37 | 710 | struct cds_lfht_node *iter_prev, *iter, *next, *new_next; |
273399de | 711 | |
1ee8f000 LJ |
712 | assert(!is_bucket(bucket)); |
713 | assert(!is_removed(bucket)); | |
714 | assert(!is_bucket(node)); | |
c90201ac | 715 | assert(!is_removed(node)); |
273399de | 716 | for (;;) { |
1ee8f000 LJ |
717 | iter_prev = bucket; |
718 | /* We can always skip the bucket node initially */ | |
04db56f8 | 719 | iter = rcu_dereference(iter_prev->next); |
b4cb483f | 720 | assert(!is_removed(iter)); |
04db56f8 | 721 | assert(iter_prev->reverse_hash <= node->reverse_hash); |
bd4db153 | 722 | /* |
1ee8f000 | 723 | * We should never be called with bucket (start of chain) |
bd4db153 MD |
724 | * and logically removed node (end of path compression |
725 | * marker) being the actual same node. This would be a | |
726 | * bug in the algorithm implementation. | |
727 | */ | |
1ee8f000 | 728 | assert(bucket != node); |
273399de | 729 | for (;;) { |
8ed51e04 | 730 | if (caa_unlikely(is_end(iter))) |
f9c80341 | 731 | return; |
04db56f8 | 732 | if (caa_likely(clear_flag(iter)->reverse_hash > node->reverse_hash)) |
f9c80341 | 733 | return; |
04db56f8 | 734 | next = rcu_dereference(clear_flag(iter)->next); |
8ed51e04 | 735 | if (caa_likely(is_removed(next))) |
273399de | 736 | break; |
b453eae1 | 737 | iter_prev = clear_flag(iter); |
273399de MD |
738 | iter = next; |
739 | } | |
b198f0fd | 740 | assert(!is_removed(iter)); |
1ee8f000 LJ |
741 | if (is_bucket(iter)) |
742 | new_next = flag_bucket(clear_flag(next)); | |
f5596c94 MD |
743 | else |
744 | new_next = clear_flag(next); | |
04db56f8 | 745 | (void) uatomic_cmpxchg(&iter_prev->next, iter, new_next); |
273399de | 746 | } |
f9c80341 | 747 | return; |
273399de MD |
748 | } |
749 | ||
9357c415 MD |
750 | static |
751 | int _cds_lfht_replace(struct cds_lfht *ht, unsigned long size, | |
752 | struct cds_lfht_node *old_node, | |
3fb86f26 | 753 | struct cds_lfht_node *old_next, |
9357c415 MD |
754 | struct cds_lfht_node *new_node) |
755 | { | |
04db56f8 | 756 | struct cds_lfht_node *bucket, *ret_next; |
9357c415 MD |
757 | |
758 | if (!old_node) /* Return -ENOENT if asked to replace NULL node */ | |
7801dadd | 759 | return -ENOENT; |
9357c415 MD |
760 | |
761 | assert(!is_removed(old_node)); | |
1ee8f000 | 762 | assert(!is_bucket(old_node)); |
9357c415 | 763 | assert(!is_removed(new_node)); |
1ee8f000 | 764 | assert(!is_bucket(new_node)); |
9357c415 | 765 | assert(new_node != old_node); |
3fb86f26 | 766 | for (;;) { |
9357c415 | 767 | /* Insert after node to be replaced */ |
9357c415 MD |
768 | if (is_removed(old_next)) { |
769 | /* | |
770 | * Too late, the old node has been removed under us | |
771 | * between lookup and replace. Fail. | |
772 | */ | |
7801dadd | 773 | return -ENOENT; |
9357c415 | 774 | } |
1ee8f000 | 775 | assert(!is_bucket(old_next)); |
9357c415 | 776 | assert(new_node != clear_flag(old_next)); |
04db56f8 | 777 | new_node->next = clear_flag(old_next); |
9357c415 MD |
778 | /* |
779 | * Here is the whole trick for lock-free replace: we add | |
780 | * the replacement node _after_ the node we want to | |
781 | * replace by atomically setting its next pointer at the | |
782 | * same time we set its removal flag. Given that | |
783 | * the lookups/get next use an iterator aware of the | |
784 | * next pointer, they will either skip the old node due | |
785 | * to the removal flag and see the new node, or use | |
786 | * the old node, but will not see the new one. | |
787 | */ | |
04db56f8 | 788 | ret_next = uatomic_cmpxchg(&old_node->next, |
9357c415 | 789 | old_next, flag_removed(new_node)); |
3fb86f26 | 790 | if (ret_next == old_next) |
7801dadd | 791 | break; /* We performed the replacement. */ |
3fb86f26 LJ |
792 | old_next = ret_next; |
793 | } | |
9357c415 | 794 | |
9357c415 MD |
795 | /* |
796 | * Ensure that the old node is not visible to readers anymore: | |
797 | * lookup for the node, and remove it (along with any other | |
798 | * logically removed node) if found. | |
799 | */ | |
04db56f8 LJ |
800 | bucket = lookup_bucket(ht, size, bit_reverse_ulong(old_node->reverse_hash)); |
801 | _cds_lfht_gc_bucket(bucket, new_node); | |
7801dadd | 802 | |
04db56f8 | 803 | assert(is_removed(rcu_dereference(old_node->next))); |
7801dadd | 804 | return 0; |
9357c415 MD |
805 | } |
806 | ||
83beee94 MD |
807 | /* |
808 | * A non-NULL unique_ret pointer uses the "add unique" (or uniquify) add | |
809 | * mode. A NULL unique_ret allows creation of duplicate keys. | |
810 | */ | |
abc490a1 | 811 | static |
83beee94 | 812 | void _cds_lfht_add(struct cds_lfht *ht, |
0422d92c | 813 | cds_lfht_match_fct match, |
996ff57c | 814 | const void *key, |
83beee94 MD |
815 | unsigned long size, |
816 | struct cds_lfht_node *node, | |
817 | struct cds_lfht_iter *unique_ret, | |
1ee8f000 | 818 | int bucket_flag) |
abc490a1 | 819 | { |
14044b37 | 820 | struct cds_lfht_node *iter_prev, *iter, *next, *new_node, *new_next, |
960c9e4f | 821 | *return_node; |
04db56f8 | 822 | struct cds_lfht_node *bucket; |
abc490a1 | 823 | |
1ee8f000 | 824 | assert(!is_bucket(node)); |
c90201ac | 825 | assert(!is_removed(node)); |
04db56f8 | 826 | bucket = lookup_bucket(ht, size, bit_reverse_ulong(node->reverse_hash)); |
abc490a1 | 827 | for (;;) { |
adc0de68 | 828 | uint32_t chain_len = 0; |
abc490a1 | 829 | |
11519af6 MD |
830 | /* |
831 | * iter_prev points to the non-removed node prior to the | |
832 | * insert location. | |
11519af6 | 833 | */ |
04db56f8 | 834 | iter_prev = bucket; |
1ee8f000 | 835 | /* We can always skip the bucket node initially */ |
04db56f8 LJ |
836 | iter = rcu_dereference(iter_prev->next); |
837 | assert(iter_prev->reverse_hash <= node->reverse_hash); | |
abc490a1 | 838 | for (;;) { |
8ed51e04 | 839 | if (caa_unlikely(is_end(iter))) |
273399de | 840 | goto insert; |
04db56f8 | 841 | if (caa_likely(clear_flag(iter)->reverse_hash > node->reverse_hash)) |
273399de | 842 | goto insert; |
238cc06e | 843 | |
1ee8f000 LJ |
844 | /* bucket node is the first node of the identical-hash-value chain */ |
845 | if (bucket_flag && clear_flag(iter)->reverse_hash == node->reverse_hash) | |
194fdbd1 | 846 | goto insert; |
238cc06e | 847 | |
04db56f8 | 848 | next = rcu_dereference(clear_flag(iter)->next); |
8ed51e04 | 849 | if (caa_unlikely(is_removed(next))) |
9dba85be | 850 | goto gc_node; |
238cc06e LJ |
851 | |
852 | /* uniquely add */ | |
83beee94 | 853 | if (unique_ret |
1ee8f000 | 854 | && !is_bucket(next) |
04db56f8 | 855 | && clear_flag(iter)->reverse_hash == node->reverse_hash) { |
238cc06e LJ |
856 | struct cds_lfht_iter d_iter = { .node = node, .next = iter, }; |
857 | ||
858 | /* | |
859 | * uniquely adding inserts the node as the first | |
860 | * node of the identical-hash-value node chain. | |
861 | * | |
862 | * This semantic ensures no duplicated keys | |
863 | * should ever be observable in the table | |
864 | * (including observe one node by one node | |
865 | * by forward iterations) | |
866 | */ | |
04db56f8 | 867 | cds_lfht_next_duplicate(ht, match, key, &d_iter); |
238cc06e LJ |
868 | if (!d_iter.node) |
869 | goto insert; | |
870 | ||
871 | *unique_ret = d_iter; | |
83beee94 | 872 | return; |
48ed1c18 | 873 | } |
238cc06e | 874 | |
11519af6 | 875 | /* Only account for identical reverse hash once */ |
04db56f8 | 876 | if (iter_prev->reverse_hash != clear_flag(iter)->reverse_hash |
1ee8f000 | 877 | && !is_bucket(next)) |
4105056a | 878 | check_resize(ht, size, ++chain_len); |
11519af6 | 879 | iter_prev = clear_flag(iter); |
273399de | 880 | iter = next; |
abc490a1 | 881 | } |
48ed1c18 | 882 | |
273399de | 883 | insert: |
7ec59d3b | 884 | assert(node != clear_flag(iter)); |
11519af6 | 885 | assert(!is_removed(iter_prev)); |
c90201ac | 886 | assert(!is_removed(iter)); |
f000907d | 887 | assert(iter_prev != node); |
1ee8f000 | 888 | if (!bucket_flag) |
04db56f8 | 889 | node->next = clear_flag(iter); |
f9c80341 | 890 | else |
1ee8f000 LJ |
891 | node->next = flag_bucket(clear_flag(iter)); |
892 | if (is_bucket(iter)) | |
893 | new_node = flag_bucket(node); | |
f5596c94 MD |
894 | else |
895 | new_node = node; | |
04db56f8 | 896 | if (uatomic_cmpxchg(&iter_prev->next, iter, |
48ed1c18 | 897 | new_node) != iter) { |
273399de | 898 | continue; /* retry */ |
48ed1c18 | 899 | } else { |
83beee94 | 900 | return_node = node; |
960c9e4f | 901 | goto end; |
48ed1c18 MD |
902 | } |
903 | ||
9dba85be MD |
904 | gc_node: |
905 | assert(!is_removed(iter)); | |
1ee8f000 LJ |
906 | if (is_bucket(iter)) |
907 | new_next = flag_bucket(clear_flag(next)); | |
f5596c94 MD |
908 | else |
909 | new_next = clear_flag(next); | |
04db56f8 | 910 | (void) uatomic_cmpxchg(&iter_prev->next, iter, new_next); |
273399de | 911 | /* retry */ |
464a1ec9 | 912 | } |
9357c415 | 913 | end: |
83beee94 MD |
914 | if (unique_ret) { |
915 | unique_ret->node = return_node; | |
916 | /* unique_ret->next left unset, never used. */ | |
917 | } | |
abc490a1 | 918 | } |
464a1ec9 | 919 | |
abc490a1 | 920 | static |
860d07e8 | 921 | int _cds_lfht_del(struct cds_lfht *ht, unsigned long size, |
4105056a | 922 | struct cds_lfht_node *node, |
1ee8f000 | 923 | int bucket_removal) |
abc490a1 | 924 | { |
04db56f8 | 925 | struct cds_lfht_node *bucket, *next, *old; |
5e28c532 | 926 | |
9357c415 | 927 | if (!node) /* Return -ENOENT if asked to delete NULL node */ |
743f9143 | 928 | return -ENOENT; |
9357c415 | 929 | |
7ec59d3b | 930 | /* logically delete the node */ |
1ee8f000 | 931 | assert(!is_bucket(node)); |
c90201ac | 932 | assert(!is_removed(node)); |
04db56f8 | 933 | old = rcu_dereference(node->next); |
7ec59d3b | 934 | do { |
48ed1c18 MD |
935 | struct cds_lfht_node *new_next; |
936 | ||
7ec59d3b | 937 | next = old; |
8ed51e04 | 938 | if (caa_unlikely(is_removed(next))) |
743f9143 | 939 | return -ENOENT; |
1ee8f000 LJ |
940 | if (bucket_removal) |
941 | assert(is_bucket(next)); | |
1475579c | 942 | else |
1ee8f000 | 943 | assert(!is_bucket(next)); |
48ed1c18 | 944 | new_next = flag_removed(next); |
04db56f8 | 945 | old = uatomic_cmpxchg(&node->next, next, new_next); |
7ec59d3b | 946 | } while (old != next); |
7ec59d3b | 947 | /* We performed the (logical) deletion. */ |
7ec59d3b MD |
948 | |
949 | /* | |
950 | * Ensure that the node is not visible to readers anymore: lookup for | |
273399de MD |
951 | * the node, and remove it (along with any other logically removed node) |
952 | * if found. | |
11519af6 | 953 | */ |
04db56f8 LJ |
954 | bucket = lookup_bucket(ht, size, bit_reverse_ulong(node->reverse_hash)); |
955 | _cds_lfht_gc_bucket(bucket, node); | |
743f9143 | 956 | |
04db56f8 | 957 | assert(is_removed(rcu_dereference(node->next))); |
743f9143 | 958 | return 0; |
abc490a1 | 959 | } |
2ed95849 | 960 | |
b7d619b0 MD |
961 | static |
962 | void *partition_resize_thread(void *arg) | |
963 | { | |
964 | struct partition_resize_work *work = arg; | |
965 | ||
7b17c13e | 966 | work->ht->flavor->register_thread(); |
b7d619b0 | 967 | work->fct(work->ht, work->i, work->start, work->len); |
7b17c13e | 968 | work->ht->flavor->unregister_thread(); |
b7d619b0 MD |
969 | return NULL; |
970 | } | |
971 | ||
972 | static | |
973 | void partition_resize_helper(struct cds_lfht *ht, unsigned long i, | |
974 | unsigned long len, | |
975 | void (*fct)(struct cds_lfht *ht, unsigned long i, | |
976 | unsigned long start, unsigned long len)) | |
977 | { | |
978 | unsigned long partition_len; | |
979 | struct partition_resize_work *work; | |
6083a889 MD |
980 | int thread, ret; |
981 | unsigned long nr_threads; | |
b7d619b0 | 982 | |
6083a889 MD |
983 | /* |
984 | * Note: nr_cpus_mask + 1 is always power of 2. | |
985 | * We spawn just the number of threads we need to satisfy the minimum | |
986 | * partition size, up to the number of CPUs in the system. | |
987 | */ | |
91452a6a MD |
988 | if (nr_cpus_mask > 0) { |
989 | nr_threads = min(nr_cpus_mask + 1, | |
990 | len >> MIN_PARTITION_PER_THREAD_ORDER); | |
991 | } else { | |
992 | nr_threads = 1; | |
993 | } | |
6083a889 MD |
994 | partition_len = len >> get_count_order_ulong(nr_threads); |
995 | work = calloc(nr_threads, sizeof(*work)); | |
b7d619b0 | 996 | assert(work); |
6083a889 MD |
997 | for (thread = 0; thread < nr_threads; thread++) { |
998 | work[thread].ht = ht; | |
999 | work[thread].i = i; | |
1000 | work[thread].len = partition_len; | |
1001 | work[thread].start = thread * partition_len; | |
1002 | work[thread].fct = fct; | |
1af6e26e | 1003 | ret = pthread_create(&(work[thread].thread_id), ht->resize_attr, |
6083a889 | 1004 | partition_resize_thread, &work[thread]); |
b7d619b0 MD |
1005 | assert(!ret); |
1006 | } | |
6083a889 | 1007 | for (thread = 0; thread < nr_threads; thread++) { |
1af6e26e | 1008 | ret = pthread_join(work[thread].thread_id, NULL); |
b7d619b0 MD |
1009 | assert(!ret); |
1010 | } | |
1011 | free(work); | |
b7d619b0 MD |
1012 | } |
1013 | ||
e8de508e MD |
1014 | /* |
1015 | * Holding RCU read lock to protect _cds_lfht_add against memory | |
1016 | * reclaim that could be performed by other call_rcu worker threads (ABA | |
1017 | * problem). | |
9ee0fc9a | 1018 | * |
b7d619b0 | 1019 | * When we reach a certain length, we can split this population phase over |
9ee0fc9a MD |
1020 | * many worker threads, based on the number of CPUs available in the system. |
1021 | * This should therefore take care of not having the expand lagging behind too | |
1022 | * many concurrent insertion threads by using the scheduler's ability to | |
1ee8f000 | 1023 | * schedule bucket node population fairly with insertions. |
e8de508e | 1024 | */ |
4105056a | 1025 | static |
b7d619b0 MD |
1026 | void init_table_populate_partition(struct cds_lfht *ht, unsigned long i, |
1027 | unsigned long start, unsigned long len) | |
4105056a | 1028 | { |
9d72a73f | 1029 | unsigned long j, size = 1UL << (i - 1); |
4105056a | 1030 | |
d0d8f9aa | 1031 | assert(i > MIN_TABLE_ORDER); |
7b17c13e | 1032 | ht->flavor->read_lock(); |
9d72a73f LJ |
1033 | for (j = size + start; j < size + start + len; j++) { |
1034 | struct cds_lfht_node *new_node = bucket_at(ht, j); | |
1035 | ||
1036 | assert(j >= size && j < (size << 1)); | |
1037 | dbg_printf("init populate: order %lu index %lu hash %lu\n", | |
1038 | i, j, j); | |
1039 | new_node->reverse_hash = bit_reverse_ulong(j); | |
1040 | _cds_lfht_add(ht, NULL, NULL, size, new_node, NULL, 1); | |
4105056a | 1041 | } |
7b17c13e | 1042 | ht->flavor->read_unlock(); |
b7d619b0 MD |
1043 | } |
1044 | ||
1045 | static | |
1046 | void init_table_populate(struct cds_lfht *ht, unsigned long i, | |
1047 | unsigned long len) | |
1048 | { | |
1049 | assert(nr_cpus_mask != -1); | |
6083a889 | 1050 | if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD) { |
7b17c13e | 1051 | ht->flavor->thread_online(); |
b7d619b0 | 1052 | init_table_populate_partition(ht, i, 0, len); |
7b17c13e | 1053 | ht->flavor->thread_offline(); |
b7d619b0 MD |
1054 | return; |
1055 | } | |
1056 | partition_resize_helper(ht, i, len, init_table_populate_partition); | |
4105056a MD |
1057 | } |
1058 | ||
abc490a1 | 1059 | static |
4105056a | 1060 | void init_table(struct cds_lfht *ht, |
93d46c39 | 1061 | unsigned long first_order, unsigned long last_order) |
24365af7 | 1062 | { |
93d46c39 | 1063 | unsigned long i; |
24365af7 | 1064 | |
93d46c39 LJ |
1065 | dbg_printf("init table: first_order %lu last_order %lu\n", |
1066 | first_order, last_order); | |
d0d8f9aa | 1067 | assert(first_order > MIN_TABLE_ORDER); |
93d46c39 | 1068 | for (i = first_order; i <= last_order; i++) { |
4105056a | 1069 | unsigned long len; |
24365af7 | 1070 | |
4f6e90b7 | 1071 | len = 1UL << (i - 1); |
f0c29ed7 | 1072 | dbg_printf("init order %lu len: %lu\n", i, len); |
4d676753 MD |
1073 | |
1074 | /* Stop expand if the resize target changes under us */ | |
7b3893e4 | 1075 | if (CMM_LOAD_SHARED(ht->resize_target) < (1UL << i)) |
4d676753 MD |
1076 | break; |
1077 | ||
48f1b16d | 1078 | cds_lfht_alloc_bucket_table(ht, i); |
4105056a | 1079 | |
4105056a | 1080 | /* |
1ee8f000 LJ |
1081 | * Set all bucket nodes reverse hash values for a level and |
1082 | * link all bucket nodes into the table. | |
4105056a | 1083 | */ |
dc1da8f6 | 1084 | init_table_populate(ht, i, len); |
4105056a | 1085 | |
f9c80341 MD |
1086 | /* |
1087 | * Update table size. | |
1088 | */ | |
1089 | cmm_smp_wmb(); /* populate data before RCU size */ | |
7b3893e4 | 1090 | CMM_STORE_SHARED(ht->size, 1UL << i); |
f9c80341 | 1091 | |
4f6e90b7 | 1092 | dbg_printf("init new size: %lu\n", 1UL << i); |
4105056a MD |
1093 | if (CMM_LOAD_SHARED(ht->in_progress_destroy)) |
1094 | break; | |
1095 | } | |
1096 | } | |
1097 | ||
e8de508e MD |
1098 | /* |
1099 | * Holding RCU read lock to protect _cds_lfht_remove against memory | |
1100 | * reclaim that could be performed by other call_rcu worker threads (ABA | |
1101 | * problem). | |
1102 | * For a single level, we logically remove and garbage collect each node. | |
1103 | * | |
1104 | * As a design choice, we perform logical removal and garbage collection on a | |
1105 | * node-per-node basis to simplify this algorithm. We also assume keeping good | |
1106 | * cache locality of the operation would overweight possible performance gain | |
1107 | * that could be achieved by batching garbage collection for multiple levels. | |
1108 | * However, this would have to be justified by benchmarks. | |
1109 | * | |
1110 | * Concurrent removal and add operations are helping us perform garbage | |
1111 | * collection of logically removed nodes. We guarantee that all logically | |
1112 | * removed nodes have been garbage-collected (unlinked) before call_rcu is | |
1ee8f000 | 1113 | * invoked to free a hole level of bucket nodes (after a grace period). |
e8de508e MD |
1114 | * |
1115 | * Logical removal and garbage collection can therefore be done in batch or on a | |
1116 | * node-per-node basis, as long as the guarantee above holds. | |
9ee0fc9a | 1117 | * |
b7d619b0 MD |
1118 | * When we reach a certain length, we can split this removal over many worker |
1119 | * threads, based on the number of CPUs available in the system. This should | |
1120 | * take care of not letting resize process lag behind too many concurrent | |
9ee0fc9a | 1121 | * updater threads actively inserting into the hash table. |
e8de508e | 1122 | */ |
4105056a | 1123 | static |
b7d619b0 MD |
1124 | void remove_table_partition(struct cds_lfht *ht, unsigned long i, |
1125 | unsigned long start, unsigned long len) | |
4105056a | 1126 | { |
9d72a73f | 1127 | unsigned long j, size = 1UL << (i - 1); |
4105056a | 1128 | |
d0d8f9aa | 1129 | assert(i > MIN_TABLE_ORDER); |
7b17c13e | 1130 | ht->flavor->read_lock(); |
9d72a73f LJ |
1131 | for (j = size + start; j < size + start + len; j++) { |
1132 | struct cds_lfht_node *fini_node = bucket_at(ht, j); | |
1133 | ||
1134 | assert(j >= size && j < (size << 1)); | |
1135 | dbg_printf("remove entry: order %lu index %lu hash %lu\n", | |
1136 | i, j, j); | |
1137 | fini_node->reverse_hash = bit_reverse_ulong(j); | |
1138 | (void) _cds_lfht_del(ht, size, fini_node, 1); | |
abc490a1 | 1139 | } |
7b17c13e | 1140 | ht->flavor->read_unlock(); |
b7d619b0 MD |
1141 | } |
1142 | ||
1143 | static | |
1144 | void remove_table(struct cds_lfht *ht, unsigned long i, unsigned long len) | |
1145 | { | |
1146 | ||
1147 | assert(nr_cpus_mask != -1); | |
6083a889 | 1148 | if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD) { |
7b17c13e | 1149 | ht->flavor->thread_online(); |
b7d619b0 | 1150 | remove_table_partition(ht, i, 0, len); |
7b17c13e | 1151 | ht->flavor->thread_offline(); |
b7d619b0 MD |
1152 | return; |
1153 | } | |
1154 | partition_resize_helper(ht, i, len, remove_table_partition); | |
2ed95849 MD |
1155 | } |
1156 | ||
1475579c | 1157 | static |
4105056a | 1158 | void fini_table(struct cds_lfht *ht, |
93d46c39 | 1159 | unsigned long first_order, unsigned long last_order) |
1475579c | 1160 | { |
93d46c39 | 1161 | long i; |
48f1b16d | 1162 | unsigned long free_by_rcu_order = 0; |
1475579c | 1163 | |
93d46c39 LJ |
1164 | dbg_printf("fini table: first_order %lu last_order %lu\n", |
1165 | first_order, last_order); | |
d0d8f9aa | 1166 | assert(first_order > MIN_TABLE_ORDER); |
93d46c39 | 1167 | for (i = last_order; i >= first_order; i--) { |
4105056a | 1168 | unsigned long len; |
1475579c | 1169 | |
4f6e90b7 | 1170 | len = 1UL << (i - 1); |
1475579c | 1171 | dbg_printf("fini order %lu len: %lu\n", i, len); |
4105056a | 1172 | |
4d676753 | 1173 | /* Stop shrink if the resize target changes under us */ |
7b3893e4 | 1174 | if (CMM_LOAD_SHARED(ht->resize_target) > (1UL << (i - 1))) |
4d676753 MD |
1175 | break; |
1176 | ||
1177 | cmm_smp_wmb(); /* populate data before RCU size */ | |
7b3893e4 | 1178 | CMM_STORE_SHARED(ht->size, 1UL << (i - 1)); |
4d676753 MD |
1179 | |
1180 | /* | |
1181 | * We need to wait for all add operations to reach Q.S. (and | |
1182 | * thus use the new table for lookups) before we can start | |
1ee8f000 | 1183 | * releasing the old bucket nodes. Otherwise their lookup will |
4d676753 MD |
1184 | * return a logically removed node as insert position. |
1185 | */ | |
7b17c13e | 1186 | ht->flavor->update_synchronize_rcu(); |
48f1b16d LJ |
1187 | if (free_by_rcu_order) |
1188 | cds_lfht_free_bucket_table(ht, free_by_rcu_order); | |
4d676753 | 1189 | |
21263e21 | 1190 | /* |
1ee8f000 LJ |
1191 | * Set "removed" flag in bucket nodes about to be removed. |
1192 | * Unlink all now-logically-removed bucket node pointers. | |
4105056a MD |
1193 | * Concurrent add/remove operation are helping us doing |
1194 | * the gc. | |
21263e21 | 1195 | */ |
4105056a MD |
1196 | remove_table(ht, i, len); |
1197 | ||
48f1b16d | 1198 | free_by_rcu_order = i; |
4105056a MD |
1199 | |
1200 | dbg_printf("fini new size: %lu\n", 1UL << i); | |
1475579c MD |
1201 | if (CMM_LOAD_SHARED(ht->in_progress_destroy)) |
1202 | break; | |
1203 | } | |
0d14ceb2 | 1204 | |
48f1b16d | 1205 | if (free_by_rcu_order) { |
7b17c13e | 1206 | ht->flavor->update_synchronize_rcu(); |
48f1b16d | 1207 | cds_lfht_free_bucket_table(ht, free_by_rcu_order); |
0d14ceb2 | 1208 | } |
1475579c MD |
1209 | } |
1210 | ||
ff0d69de | 1211 | static |
1ee8f000 | 1212 | void cds_lfht_create_bucket(struct cds_lfht *ht, unsigned long size) |
ff0d69de | 1213 | { |
04db56f8 | 1214 | struct cds_lfht_node *prev, *node; |
9d72a73f | 1215 | unsigned long order, len, i; |
ff0d69de | 1216 | |
48f1b16d | 1217 | cds_lfht_alloc_bucket_table(ht, 0); |
ff0d69de | 1218 | |
9d72a73f LJ |
1219 | dbg_printf("create bucket: order 0 index 0 hash 0\n"); |
1220 | node = bucket_at(ht, 0); | |
1221 | node->next = flag_bucket(get_end()); | |
1222 | node->reverse_hash = 0; | |
ff0d69de LJ |
1223 | |
1224 | for (order = 1; order < get_count_order_ulong(size) + 1; order++) { | |
1225 | len = 1UL << (order - 1); | |
48f1b16d | 1226 | cds_lfht_alloc_bucket_table(ht, order); |
ff0d69de | 1227 | |
9d72a73f LJ |
1228 | for (i = 0; i < len; i++) { |
1229 | /* | |
1230 | * Now, we are trying to init the node with the | |
1231 | * hash=(len+i) (which is also a bucket with the | |
1232 | * index=(len+i)) and insert it into the hash table, | |
1233 | * so this node has to be inserted after the bucket | |
1234 | * with the index=(len+i)&(len-1)=i. And because there | |
1235 | * is no other non-bucket node nor bucket node with | |
1236 | * larger index/hash inserted, so the bucket node | |
1237 | * being inserted should be inserted directly linked | |
1238 | * after the bucket node with index=i. | |
1239 | */ | |
1240 | prev = bucket_at(ht, i); | |
1241 | node = bucket_at(ht, len + i); | |
ff0d69de | 1242 | |
1ee8f000 | 1243 | dbg_printf("create bucket: order %lu index %lu hash %lu\n", |
9d72a73f LJ |
1244 | order, len + i, len + i); |
1245 | node->reverse_hash = bit_reverse_ulong(len + i); | |
1246 | ||
1247 | /* insert after prev */ | |
1248 | assert(is_bucket(prev->next)); | |
ff0d69de | 1249 | node->next = prev->next; |
1ee8f000 | 1250 | prev->next = flag_bucket(node); |
ff0d69de LJ |
1251 | } |
1252 | } | |
1253 | } | |
1254 | ||
0422d92c | 1255 | struct cds_lfht *_cds_lfht_new(unsigned long init_size, |
0722081a | 1256 | unsigned long min_nr_alloc_buckets, |
747d725c | 1257 | unsigned long max_nr_buckets, |
b8af5011 | 1258 | int flags, |
0b6aa001 | 1259 | const struct cds_lfht_mm_type *mm, |
7b17c13e | 1260 | const struct rcu_flavor_struct *flavor, |
b7d619b0 | 1261 | pthread_attr_t *attr) |
abc490a1 | 1262 | { |
14044b37 | 1263 | struct cds_lfht *ht; |
24365af7 | 1264 | unsigned long order; |
abc490a1 | 1265 | |
0722081a LJ |
1266 | /* min_nr_alloc_buckets must be power of two */ |
1267 | if (!min_nr_alloc_buckets || (min_nr_alloc_buckets & (min_nr_alloc_buckets - 1))) | |
5488222b | 1268 | return NULL; |
747d725c | 1269 | |
8129be4e | 1270 | /* init_size must be power of two */ |
5488222b | 1271 | if (!init_size || (init_size & (init_size - 1))) |
8129be4e | 1272 | return NULL; |
747d725c | 1273 | |
0b6aa001 LJ |
1274 | /* max_nr_buckets == 0 for order based mm means infinite */ |
1275 | if (mm == &cds_lfht_mm_order && !max_nr_buckets) | |
747d725c LJ |
1276 | max_nr_buckets = 1UL << (MAX_TABLE_ORDER - 1); |
1277 | ||
1278 | /* max_nr_buckets must be power of two */ | |
1279 | if (!max_nr_buckets || (max_nr_buckets & (max_nr_buckets - 1))) | |
1280 | return NULL; | |
1281 | ||
0722081a | 1282 | min_nr_alloc_buckets = max(min_nr_alloc_buckets, MIN_TABLE_SIZE); |
d0d8f9aa | 1283 | init_size = max(init_size, MIN_TABLE_SIZE); |
747d725c LJ |
1284 | max_nr_buckets = max(max_nr_buckets, min_nr_alloc_buckets); |
1285 | init_size = min(init_size, max_nr_buckets); | |
0b6aa001 LJ |
1286 | |
1287 | ht = mm->alloc_cds_lfht(min_nr_alloc_buckets, max_nr_buckets); | |
b7d619b0 | 1288 | assert(ht); |
0b6aa001 LJ |
1289 | assert(ht->mm == mm); |
1290 | assert(ht->bucket_at == mm->bucket_at); | |
1291 | ||
b5d6b20f | 1292 | ht->flags = flags; |
7b17c13e | 1293 | ht->flavor = flavor; |
b7d619b0 | 1294 | ht->resize_attr = attr; |
5afadd12 | 1295 | alloc_split_items_count(ht); |
abc490a1 MD |
1296 | /* this mutex should not nest in read-side C.S. */ |
1297 | pthread_mutex_init(&ht->resize_mutex, NULL); | |
5488222b | 1298 | order = get_count_order_ulong(init_size); |
7b3893e4 | 1299 | ht->resize_target = 1UL << order; |
1ee8f000 | 1300 | cds_lfht_create_bucket(ht, 1UL << order); |
7b3893e4 | 1301 | ht->size = 1UL << order; |
abc490a1 MD |
1302 | return ht; |
1303 | } | |
1304 | ||
6f554439 | 1305 | void cds_lfht_lookup(struct cds_lfht *ht, unsigned long hash, |
996ff57c | 1306 | cds_lfht_match_fct match, const void *key, |
6f554439 | 1307 | struct cds_lfht_iter *iter) |
2ed95849 | 1308 | { |
04db56f8 | 1309 | struct cds_lfht_node *node, *next, *bucket; |
0422d92c | 1310 | unsigned long reverse_hash, size; |
2ed95849 | 1311 | |
abc490a1 | 1312 | reverse_hash = bit_reverse_ulong(hash); |
464a1ec9 | 1313 | |
7b3893e4 | 1314 | size = rcu_dereference(ht->size); |
04db56f8 | 1315 | bucket = lookup_bucket(ht, size, hash); |
1ee8f000 | 1316 | /* We can always skip the bucket node initially */ |
04db56f8 | 1317 | node = rcu_dereference(bucket->next); |
bb7b2f26 | 1318 | node = clear_flag(node); |
2ed95849 | 1319 | for (;;) { |
8ed51e04 | 1320 | if (caa_unlikely(is_end(node))) { |
96ad1112 | 1321 | node = next = NULL; |
abc490a1 | 1322 | break; |
bb7b2f26 | 1323 | } |
04db56f8 | 1324 | if (caa_unlikely(node->reverse_hash > reverse_hash)) { |
96ad1112 | 1325 | node = next = NULL; |
abc490a1 | 1326 | break; |
2ed95849 | 1327 | } |
04db56f8 | 1328 | next = rcu_dereference(node->next); |
7f52427b | 1329 | assert(node == clear_flag(node)); |
8ed51e04 | 1330 | if (caa_likely(!is_removed(next)) |
1ee8f000 | 1331 | && !is_bucket(next) |
04db56f8 | 1332 | && node->reverse_hash == reverse_hash |
0422d92c | 1333 | && caa_likely(match(node, key))) { |
273399de | 1334 | break; |
2ed95849 | 1335 | } |
1b81fe1a | 1336 | node = clear_flag(next); |
2ed95849 | 1337 | } |
1ee8f000 | 1338 | assert(!node || !is_bucket(rcu_dereference(node->next))); |
adc0de68 MD |
1339 | iter->node = node; |
1340 | iter->next = next; | |
abc490a1 | 1341 | } |
e0ba718a | 1342 | |
0422d92c | 1343 | void cds_lfht_next_duplicate(struct cds_lfht *ht, cds_lfht_match_fct match, |
996ff57c | 1344 | const void *key, struct cds_lfht_iter *iter) |
a481e5ff | 1345 | { |
adc0de68 | 1346 | struct cds_lfht_node *node, *next; |
a481e5ff | 1347 | unsigned long reverse_hash; |
a481e5ff | 1348 | |
adc0de68 | 1349 | node = iter->node; |
04db56f8 | 1350 | reverse_hash = node->reverse_hash; |
adc0de68 | 1351 | next = iter->next; |
a481e5ff MD |
1352 | node = clear_flag(next); |
1353 | ||
1354 | for (;;) { | |
8ed51e04 | 1355 | if (caa_unlikely(is_end(node))) { |
96ad1112 | 1356 | node = next = NULL; |
a481e5ff | 1357 | break; |
bb7b2f26 | 1358 | } |
04db56f8 | 1359 | if (caa_unlikely(node->reverse_hash > reverse_hash)) { |
96ad1112 | 1360 | node = next = NULL; |
a481e5ff MD |
1361 | break; |
1362 | } | |
04db56f8 | 1363 | next = rcu_dereference(node->next); |
8ed51e04 | 1364 | if (caa_likely(!is_removed(next)) |
1ee8f000 | 1365 | && !is_bucket(next) |
04db56f8 | 1366 | && caa_likely(match(node, key))) { |
a481e5ff MD |
1367 | break; |
1368 | } | |
1369 | node = clear_flag(next); | |
1370 | } | |
1ee8f000 | 1371 | assert(!node || !is_bucket(rcu_dereference(node->next))); |
adc0de68 MD |
1372 | iter->node = node; |
1373 | iter->next = next; | |
a481e5ff MD |
1374 | } |
1375 | ||
4e9b9fbf MD |
1376 | void cds_lfht_next(struct cds_lfht *ht, struct cds_lfht_iter *iter) |
1377 | { | |
1378 | struct cds_lfht_node *node, *next; | |
1379 | ||
853395e1 | 1380 | node = clear_flag(iter->next); |
4e9b9fbf | 1381 | for (;;) { |
8ed51e04 | 1382 | if (caa_unlikely(is_end(node))) { |
4e9b9fbf MD |
1383 | node = next = NULL; |
1384 | break; | |
1385 | } | |
04db56f8 | 1386 | next = rcu_dereference(node->next); |
8ed51e04 | 1387 | if (caa_likely(!is_removed(next)) |
1ee8f000 | 1388 | && !is_bucket(next)) { |
4e9b9fbf MD |
1389 | break; |
1390 | } | |
1391 | node = clear_flag(next); | |
1392 | } | |
1ee8f000 | 1393 | assert(!node || !is_bucket(rcu_dereference(node->next))); |
4e9b9fbf MD |
1394 | iter->node = node; |
1395 | iter->next = next; | |
1396 | } | |
1397 | ||
1398 | void cds_lfht_first(struct cds_lfht *ht, struct cds_lfht_iter *iter) | |
1399 | { | |
4e9b9fbf | 1400 | /* |
1ee8f000 | 1401 | * Get next after first bucket node. The first bucket node is the |
4e9b9fbf MD |
1402 | * first node of the linked list. |
1403 | */ | |
9d72a73f | 1404 | iter->next = bucket_at(ht, 0)->next; |
4e9b9fbf MD |
1405 | cds_lfht_next(ht, iter); |
1406 | } | |
1407 | ||
0422d92c MD |
1408 | void cds_lfht_add(struct cds_lfht *ht, unsigned long hash, |
1409 | struct cds_lfht_node *node) | |
abc490a1 | 1410 | { |
0422d92c | 1411 | unsigned long size; |
ab7d5fc6 | 1412 | |
04db56f8 | 1413 | node->reverse_hash = bit_reverse_ulong((unsigned long) hash); |
7b3893e4 | 1414 | size = rcu_dereference(ht->size); |
04db56f8 | 1415 | _cds_lfht_add(ht, NULL, NULL, size, node, NULL, 0); |
14360f1c | 1416 | ht_count_add(ht, size, hash); |
3eca1b8c MD |
1417 | } |
1418 | ||
14044b37 | 1419 | struct cds_lfht_node *cds_lfht_add_unique(struct cds_lfht *ht, |
6f554439 | 1420 | unsigned long hash, |
0422d92c | 1421 | cds_lfht_match_fct match, |
996ff57c | 1422 | const void *key, |
48ed1c18 | 1423 | struct cds_lfht_node *node) |
3eca1b8c | 1424 | { |
0422d92c | 1425 | unsigned long size; |
83beee94 | 1426 | struct cds_lfht_iter iter; |
3eca1b8c | 1427 | |
04db56f8 | 1428 | node->reverse_hash = bit_reverse_ulong((unsigned long) hash); |
7b3893e4 | 1429 | size = rcu_dereference(ht->size); |
04db56f8 | 1430 | _cds_lfht_add(ht, match, key, size, node, &iter, 0); |
83beee94 | 1431 | if (iter.node == node) |
14360f1c | 1432 | ht_count_add(ht, size, hash); |
83beee94 | 1433 | return iter.node; |
2ed95849 MD |
1434 | } |
1435 | ||
9357c415 | 1436 | struct cds_lfht_node *cds_lfht_add_replace(struct cds_lfht *ht, |
6f554439 | 1437 | unsigned long hash, |
0422d92c | 1438 | cds_lfht_match_fct match, |
996ff57c | 1439 | const void *key, |
48ed1c18 MD |
1440 | struct cds_lfht_node *node) |
1441 | { | |
0422d92c | 1442 | unsigned long size; |
83beee94 | 1443 | struct cds_lfht_iter iter; |
48ed1c18 | 1444 | |
04db56f8 | 1445 | node->reverse_hash = bit_reverse_ulong((unsigned long) hash); |
7b3893e4 | 1446 | size = rcu_dereference(ht->size); |
83beee94 | 1447 | for (;;) { |
04db56f8 | 1448 | _cds_lfht_add(ht, match, key, size, node, &iter, 0); |
83beee94 | 1449 | if (iter.node == node) { |
14360f1c | 1450 | ht_count_add(ht, size, hash); |
83beee94 MD |
1451 | return NULL; |
1452 | } | |
1453 | ||
1454 | if (!_cds_lfht_replace(ht, size, iter.node, iter.next, node)) | |
1455 | return iter.node; | |
1456 | } | |
48ed1c18 MD |
1457 | } |
1458 | ||
9357c415 MD |
1459 | int cds_lfht_replace(struct cds_lfht *ht, struct cds_lfht_iter *old_iter, |
1460 | struct cds_lfht_node *new_node) | |
1461 | { | |
1462 | unsigned long size; | |
1463 | ||
7b3893e4 | 1464 | size = rcu_dereference(ht->size); |
9357c415 MD |
1465 | return _cds_lfht_replace(ht, size, old_iter->node, old_iter->next, |
1466 | new_node); | |
1467 | } | |
1468 | ||
1469 | int cds_lfht_del(struct cds_lfht *ht, struct cds_lfht_iter *iter) | |
2ed95849 | 1470 | { |
14360f1c | 1471 | unsigned long size, hash; |
df44348d | 1472 | int ret; |
abc490a1 | 1473 | |
7b3893e4 | 1474 | size = rcu_dereference(ht->size); |
9357c415 | 1475 | ret = _cds_lfht_del(ht, size, iter->node, 0); |
14360f1c | 1476 | if (!ret) { |
04db56f8 | 1477 | hash = bit_reverse_ulong(iter->node->reverse_hash); |
14360f1c LJ |
1478 | ht_count_del(ht, size, hash); |
1479 | } | |
df44348d | 1480 | return ret; |
2ed95849 | 1481 | } |
ab7d5fc6 | 1482 | |
abc490a1 | 1483 | static |
1ee8f000 | 1484 | int cds_lfht_delete_bucket(struct cds_lfht *ht) |
674f7a69 | 1485 | { |
14044b37 | 1486 | struct cds_lfht_node *node; |
4105056a | 1487 | unsigned long order, i, size; |
674f7a69 | 1488 | |
abc490a1 | 1489 | /* Check that the table is empty */ |
9d72a73f | 1490 | node = bucket_at(ht, 0); |
abc490a1 | 1491 | do { |
04db56f8 | 1492 | node = clear_flag(node)->next; |
1ee8f000 | 1493 | if (!is_bucket(node)) |
abc490a1 | 1494 | return -EPERM; |
273399de | 1495 | assert(!is_removed(node)); |
bb7b2f26 | 1496 | } while (!is_end(node)); |
4105056a MD |
1497 | /* |
1498 | * size accessed without rcu_dereference because hash table is | |
1499 | * being destroyed. | |
1500 | */ | |
7b3893e4 | 1501 | size = ht->size; |
1ee8f000 | 1502 | /* Internal sanity check: all nodes left should be bucket */ |
48f1b16d LJ |
1503 | for (i = 0; i < size; i++) { |
1504 | node = bucket_at(ht, i); | |
1505 | dbg_printf("delete bucket: index %lu expected hash %lu hash %lu\n", | |
1506 | i, i, bit_reverse_ulong(node->reverse_hash)); | |
1507 | assert(is_bucket(node->next)); | |
1508 | } | |
24365af7 | 1509 | |
48f1b16d LJ |
1510 | for (order = get_count_order_ulong(size); (long)order >= 0; order--) |
1511 | cds_lfht_free_bucket_table(ht, order); | |
5488222b | 1512 | |
abc490a1 | 1513 | return 0; |
674f7a69 MD |
1514 | } |
1515 | ||
1516 | /* | |
1517 | * Should only be called when no more concurrent readers nor writers can | |
1518 | * possibly access the table. | |
1519 | */ | |
b7d619b0 | 1520 | int cds_lfht_destroy(struct cds_lfht *ht, pthread_attr_t **attr) |
674f7a69 | 1521 | { |
5e28c532 MD |
1522 | int ret; |
1523 | ||
848d4088 | 1524 | /* Wait for in-flight resize operations to complete */ |
24953e08 MD |
1525 | _CMM_STORE_SHARED(ht->in_progress_destroy, 1); |
1526 | cmm_smp_mb(); /* Store destroy before load resize */ | |
848d4088 MD |
1527 | while (uatomic_read(&ht->in_progress_resize)) |
1528 | poll(NULL, 0, 100); /* wait for 100ms */ | |
1ee8f000 | 1529 | ret = cds_lfht_delete_bucket(ht); |
abc490a1 MD |
1530 | if (ret) |
1531 | return ret; | |
5afadd12 | 1532 | free_split_items_count(ht); |
b7d619b0 MD |
1533 | if (attr) |
1534 | *attr = ht->resize_attr; | |
98808fb1 | 1535 | poison_free(ht); |
5e28c532 | 1536 | return ret; |
674f7a69 MD |
1537 | } |
1538 | ||
14044b37 | 1539 | void cds_lfht_count_nodes(struct cds_lfht *ht, |
d933dd0e | 1540 | long *approx_before, |
273399de | 1541 | unsigned long *count, |
973e5e1b | 1542 | unsigned long *removed, |
d933dd0e | 1543 | long *approx_after) |
273399de | 1544 | { |
14044b37 | 1545 | struct cds_lfht_node *node, *next; |
1ee8f000 | 1546 | unsigned long nr_bucket = 0; |
273399de | 1547 | |
7ed7682f | 1548 | *approx_before = 0; |
5afadd12 | 1549 | if (ht->split_count) { |
973e5e1b MD |
1550 | int i; |
1551 | ||
4c42f1b8 LJ |
1552 | for (i = 0; i < split_count_mask + 1; i++) { |
1553 | *approx_before += uatomic_read(&ht->split_count[i].add); | |
1554 | *approx_before -= uatomic_read(&ht->split_count[i].del); | |
973e5e1b MD |
1555 | } |
1556 | } | |
1557 | ||
273399de MD |
1558 | *count = 0; |
1559 | *removed = 0; | |
1560 | ||
1ee8f000 | 1561 | /* Count non-bucket nodes in the table */ |
9d72a73f | 1562 | node = bucket_at(ht, 0); |
273399de | 1563 | do { |
04db56f8 | 1564 | next = rcu_dereference(node->next); |
b198f0fd | 1565 | if (is_removed(next)) { |
1ee8f000 | 1566 | if (!is_bucket(next)) |
973e5e1b MD |
1567 | (*removed)++; |
1568 | else | |
1ee8f000 LJ |
1569 | (nr_bucket)++; |
1570 | } else if (!is_bucket(next)) | |
273399de | 1571 | (*count)++; |
24365af7 | 1572 | else |
1ee8f000 | 1573 | (nr_bucket)++; |
273399de | 1574 | node = clear_flag(next); |
bb7b2f26 | 1575 | } while (!is_end(node)); |
1ee8f000 | 1576 | dbg_printf("number of bucket nodes: %lu\n", nr_bucket); |
7ed7682f | 1577 | *approx_after = 0; |
5afadd12 | 1578 | if (ht->split_count) { |
973e5e1b MD |
1579 | int i; |
1580 | ||
4c42f1b8 LJ |
1581 | for (i = 0; i < split_count_mask + 1; i++) { |
1582 | *approx_after += uatomic_read(&ht->split_count[i].add); | |
1583 | *approx_after -= uatomic_read(&ht->split_count[i].del); | |
973e5e1b MD |
1584 | } |
1585 | } | |
273399de MD |
1586 | } |
1587 | ||
1475579c | 1588 | /* called with resize mutex held */ |
abc490a1 | 1589 | static |
4105056a | 1590 | void _do_cds_lfht_grow(struct cds_lfht *ht, |
1475579c | 1591 | unsigned long old_size, unsigned long new_size) |
abc490a1 | 1592 | { |
1475579c | 1593 | unsigned long old_order, new_order; |
1475579c | 1594 | |
93d46c39 LJ |
1595 | old_order = get_count_order_ulong(old_size); |
1596 | new_order = get_count_order_ulong(new_size); | |
1a401918 LJ |
1597 | dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n", |
1598 | old_size, old_order, new_size, new_order); | |
1475579c | 1599 | assert(new_size > old_size); |
93d46c39 | 1600 | init_table(ht, old_order + 1, new_order); |
abc490a1 MD |
1601 | } |
1602 | ||
1603 | /* called with resize mutex held */ | |
1604 | static | |
4105056a | 1605 | void _do_cds_lfht_shrink(struct cds_lfht *ht, |
1475579c | 1606 | unsigned long old_size, unsigned long new_size) |
464a1ec9 | 1607 | { |
1475579c | 1608 | unsigned long old_order, new_order; |
464a1ec9 | 1609 | |
d0d8f9aa | 1610 | new_size = max(new_size, MIN_TABLE_SIZE); |
93d46c39 LJ |
1611 | old_order = get_count_order_ulong(old_size); |
1612 | new_order = get_count_order_ulong(new_size); | |
1a401918 LJ |
1613 | dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n", |
1614 | old_size, old_order, new_size, new_order); | |
1475579c | 1615 | assert(new_size < old_size); |
1475579c | 1616 | |
1ee8f000 | 1617 | /* Remove and unlink all bucket nodes to remove. */ |
93d46c39 | 1618 | fini_table(ht, new_order + 1, old_order); |
464a1ec9 MD |
1619 | } |
1620 | ||
1475579c MD |
1621 | |
1622 | /* called with resize mutex held */ | |
1623 | static | |
1624 | void _do_cds_lfht_resize(struct cds_lfht *ht) | |
1625 | { | |
1626 | unsigned long new_size, old_size; | |
4105056a MD |
1627 | |
1628 | /* | |
1629 | * Resize table, re-do if the target size has changed under us. | |
1630 | */ | |
1631 | do { | |
d2be3620 MD |
1632 | assert(uatomic_read(&ht->in_progress_resize)); |
1633 | if (CMM_LOAD_SHARED(ht->in_progress_destroy)) | |
1634 | break; | |
7b3893e4 LJ |
1635 | ht->resize_initiated = 1; |
1636 | old_size = ht->size; | |
1637 | new_size = CMM_LOAD_SHARED(ht->resize_target); | |
4105056a MD |
1638 | if (old_size < new_size) |
1639 | _do_cds_lfht_grow(ht, old_size, new_size); | |
1640 | else if (old_size > new_size) | |
1641 | _do_cds_lfht_shrink(ht, old_size, new_size); | |
7b3893e4 | 1642 | ht->resize_initiated = 0; |
4105056a MD |
1643 | /* write resize_initiated before read resize_target */ |
1644 | cmm_smp_mb(); | |
7b3893e4 | 1645 | } while (ht->size != CMM_LOAD_SHARED(ht->resize_target)); |
1475579c MD |
1646 | } |
1647 | ||
abc490a1 | 1648 | static |
ab65b890 | 1649 | unsigned long resize_target_grow(struct cds_lfht *ht, unsigned long new_size) |
464a1ec9 | 1650 | { |
7b3893e4 | 1651 | return _uatomic_xchg_monotonic_increase(&ht->resize_target, new_size); |
464a1ec9 MD |
1652 | } |
1653 | ||
1475579c | 1654 | static |
4105056a | 1655 | void resize_target_update_count(struct cds_lfht *ht, |
b8af5011 | 1656 | unsigned long count) |
1475579c | 1657 | { |
d0d8f9aa | 1658 | count = max(count, MIN_TABLE_SIZE); |
747d725c | 1659 | count = min(count, ht->max_nr_buckets); |
7b3893e4 | 1660 | uatomic_set(&ht->resize_target, count); |
1475579c MD |
1661 | } |
1662 | ||
1663 | void cds_lfht_resize(struct cds_lfht *ht, unsigned long new_size) | |
464a1ec9 | 1664 | { |
4105056a | 1665 | resize_target_update_count(ht, new_size); |
7b3893e4 | 1666 | CMM_STORE_SHARED(ht->resize_initiated, 1); |
7b17c13e | 1667 | ht->flavor->thread_offline(); |
1475579c MD |
1668 | pthread_mutex_lock(&ht->resize_mutex); |
1669 | _do_cds_lfht_resize(ht); | |
1670 | pthread_mutex_unlock(&ht->resize_mutex); | |
7b17c13e | 1671 | ht->flavor->thread_online(); |
abc490a1 | 1672 | } |
464a1ec9 | 1673 | |
abc490a1 MD |
1674 | static |
1675 | void do_resize_cb(struct rcu_head *head) | |
1676 | { | |
1677 | struct rcu_resize_work *work = | |
1678 | caa_container_of(head, struct rcu_resize_work, head); | |
14044b37 | 1679 | struct cds_lfht *ht = work->ht; |
abc490a1 | 1680 | |
7b17c13e | 1681 | ht->flavor->thread_offline(); |
abc490a1 | 1682 | pthread_mutex_lock(&ht->resize_mutex); |
14044b37 | 1683 | _do_cds_lfht_resize(ht); |
abc490a1 | 1684 | pthread_mutex_unlock(&ht->resize_mutex); |
7b17c13e | 1685 | ht->flavor->thread_online(); |
98808fb1 | 1686 | poison_free(work); |
848d4088 MD |
1687 | cmm_smp_mb(); /* finish resize before decrement */ |
1688 | uatomic_dec(&ht->in_progress_resize); | |
464a1ec9 MD |
1689 | } |
1690 | ||
abc490a1 | 1691 | static |
f1f119ee | 1692 | void __cds_lfht_resize_lazy_launch(struct cds_lfht *ht) |
ab7d5fc6 | 1693 | { |
abc490a1 MD |
1694 | struct rcu_resize_work *work; |
1695 | ||
4105056a MD |
1696 | /* Store resize_target before read resize_initiated */ |
1697 | cmm_smp_mb(); | |
7b3893e4 | 1698 | if (!CMM_LOAD_SHARED(ht->resize_initiated)) { |
848d4088 | 1699 | uatomic_inc(&ht->in_progress_resize); |
59290e9d | 1700 | cmm_smp_mb(); /* increment resize count before load destroy */ |
ed35e6d8 MD |
1701 | if (CMM_LOAD_SHARED(ht->in_progress_destroy)) { |
1702 | uatomic_dec(&ht->in_progress_resize); | |
59290e9d | 1703 | return; |
ed35e6d8 | 1704 | } |
f9830efd MD |
1705 | work = malloc(sizeof(*work)); |
1706 | work->ht = ht; | |
7b17c13e | 1707 | ht->flavor->update_call_rcu(&work->head, do_resize_cb); |
7b3893e4 | 1708 | CMM_STORE_SHARED(ht->resize_initiated, 1); |
f9830efd | 1709 | } |
ab7d5fc6 | 1710 | } |
3171717f | 1711 | |
f1f119ee LJ |
1712 | static |
1713 | void cds_lfht_resize_lazy_grow(struct cds_lfht *ht, unsigned long size, int growth) | |
1714 | { | |
1715 | unsigned long target_size = size << growth; | |
1716 | ||
747d725c | 1717 | target_size = min(target_size, ht->max_nr_buckets); |
f1f119ee LJ |
1718 | if (resize_target_grow(ht, target_size) >= target_size) |
1719 | return; | |
1720 | ||
1721 | __cds_lfht_resize_lazy_launch(ht); | |
1722 | } | |
1723 | ||
89bb121d LJ |
1724 | /* |
1725 | * We favor grow operations over shrink. A shrink operation never occurs | |
1726 | * if a grow operation is queued for lazy execution. A grow operation | |
1727 | * cancels any pending shrink lazy execution. | |
1728 | */ | |
3171717f | 1729 | static |
4105056a | 1730 | void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size, |
3171717f MD |
1731 | unsigned long count) |
1732 | { | |
b8af5011 MD |
1733 | if (!(ht->flags & CDS_LFHT_AUTO_RESIZE)) |
1734 | return; | |
d0d8f9aa | 1735 | count = max(count, MIN_TABLE_SIZE); |
747d725c | 1736 | count = min(count, ht->max_nr_buckets); |
89bb121d LJ |
1737 | if (count == size) |
1738 | return; /* Already the right size, no resize needed */ | |
1739 | if (count > size) { /* lazy grow */ | |
1740 | if (resize_target_grow(ht, count) >= count) | |
1741 | return; | |
1742 | } else { /* lazy shrink */ | |
1743 | for (;;) { | |
1744 | unsigned long s; | |
1745 | ||
7b3893e4 | 1746 | s = uatomic_cmpxchg(&ht->resize_target, size, count); |
89bb121d LJ |
1747 | if (s == size) |
1748 | break; /* no resize needed */ | |
1749 | if (s > size) | |
1750 | return; /* growing is/(was just) in progress */ | |
1751 | if (s <= count) | |
1752 | return; /* some other thread do shrink */ | |
1753 | size = s; | |
1754 | } | |
1755 | } | |
f1f119ee | 1756 | __cds_lfht_resize_lazy_launch(ht); |
3171717f | 1757 | } |