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