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> | |
7 | * | |
8 | * This library is free software; you can redistribute it and/or | |
9 | * modify it under the terms of the GNU Lesser General Public | |
10 | * License as published by the Free Software Foundation; either | |
11 | * version 2.1 of the License, or (at your option) any later version. | |
12 | * | |
13 | * This library is distributed in the hope that it will be useful, | |
14 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
15 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | |
16 | * Lesser General Public License for more details. | |
17 | * | |
18 | * You should have received a copy of the GNU Lesser General Public | |
19 | * License along with this library; if not, write to the Free Software | |
20 | * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA | |
5e28c532 MD |
21 | */ |
22 | ||
e753ff5a MD |
23 | /* |
24 | * Based on the following articles: | |
25 | * - Ori Shalev and Nir Shavit. Split-ordered lists: Lock-free | |
26 | * extensible hash tables. J. ACM 53, 3 (May 2006), 379-405. | |
27 | * - Michael, M. M. High performance dynamic lock-free hash tables | |
28 | * and list-based sets. In Proceedings of the fourteenth annual ACM | |
29 | * symposium on Parallel algorithms and architectures, ACM Press, | |
30 | * (2002), 73-82. | |
31 | * | |
1475579c | 32 | * Some specificities of this Lock-Free Resizable RCU Hash Table |
e753ff5a MD |
33 | * implementation: |
34 | * | |
35 | * - RCU read-side critical section allows readers to perform hash | |
36 | * table lookups and use the returned objects safely by delaying | |
37 | * memory reclaim of a grace period. | |
38 | * - Add and remove operations are lock-free, and do not need to | |
39 | * allocate memory. They need to be executed within RCU read-side | |
40 | * critical section to ensure the objects they read are valid and to | |
41 | * deal with the cmpxchg ABA problem. | |
42 | * - add and add_unique operations are supported. add_unique checks if | |
43 | * the node key already exists in the hash table. It ensures no key | |
44 | * duplicata exists. | |
45 | * - The resize operation executes concurrently with add/remove/lookup. | |
46 | * - Hash table nodes are contained within a split-ordered list. This | |
47 | * list is ordered by incrementing reversed-bits-hash value. | |
48 | * - An index of dummy nodes is kept. These dummy nodes are the hash | |
49 | * table "buckets", and they are also chained together in the | |
50 | * split-ordered list, which allows recursive expansion. | |
1475579c MD |
51 | * - The resize operation for small tables only allows expanding the hash table. |
52 | * It is triggered automatically by detecting long chains in the add | |
53 | * operation. | |
54 | * - The resize operation for larger tables (and available through an | |
55 | * API) allows both expanding and shrinking the hash table. | |
56 | * - Per-CPU Split-counters are used to keep track of the number of | |
57 | * nodes within the hash table for automatic resize triggering. | |
e753ff5a MD |
58 | * - Resize operation initiated by long chain detection is executed by a |
59 | * call_rcu thread, which keeps lock-freedom of add and remove. | |
60 | * - Resize operations are protected by a mutex. | |
61 | * - The removal operation is split in two parts: first, a "removed" | |
62 | * flag is set in the next pointer within the node to remove. Then, | |
63 | * a "garbage collection" is performed in the bucket containing the | |
64 | * removed node (from the start of the bucket up to the removed node). | |
65 | * All encountered nodes with "removed" flag set in their next | |
66 | * pointers are removed from the linked-list. If the cmpxchg used for | |
67 | * removal fails (due to concurrent garbage-collection or concurrent | |
68 | * add), we retry from the beginning of the bucket. This ensures that | |
69 | * the node with "removed" flag set is removed from the hash table | |
70 | * (not visible to lookups anymore) before the RCU read-side critical | |
71 | * section held across removal ends. Furthermore, this ensures that | |
72 | * the node with "removed" flag set is removed from the linked-list | |
73 | * before its memory is reclaimed. Only the thread which removal | |
74 | * successfully set the "removed" flag (with a cmpxchg) into a node's | |
75 | * next pointer is considered to have succeeded its removal (and thus | |
76 | * owns the node to reclaim). Because we garbage-collect starting from | |
77 | * an invariant node (the start-of-bucket dummy node) up to the | |
78 | * "removed" node (or find a reverse-hash that is higher), we are sure | |
79 | * that a successful traversal of the chain leads to a chain that is | |
80 | * present in the linked-list (the start node is never removed) and | |
81 | * that is does not contain the "removed" node anymore, even if | |
82 | * concurrent delete/add operations are changing the structure of the | |
83 | * list concurrently. | |
29e669f6 MD |
84 | * - The add operation performs gargage collection of buckets if it |
85 | * encounters nodes with removed flag set in the bucket where it wants | |
86 | * to add its new node. This ensures lock-freedom of add operation by | |
87 | * helping the remover unlink nodes from the list rather than to wait | |
88 | * for it do to so. | |
e753ff5a MD |
89 | * - A RCU "order table" indexed by log2(hash index) is copied and |
90 | * expanded by the resize operation. This order table allows finding | |
91 | * the "dummy node" tables. | |
92 | * - There is one dummy node table per hash index order. The size of | |
93 | * each dummy node table is half the number of hashes contained in | |
94 | * this order. | |
95 | * - call_rcu is used to garbage-collect the old order table. | |
96 | * - The per-order dummy node tables contain a compact version of the | |
97 | * hash table nodes. These tables are invariant after they are | |
98 | * populated into the hash table. | |
1475579c MD |
99 | * |
100 | * A bit of ascii art explanation: | |
101 | * | |
102 | * Order index is the off-by-one compare to the actual power of 2 because | |
103 | * we use index 0 to deal with the 0 special-case. | |
104 | * | |
105 | * This shows the nodes for a small table ordered by reversed bits: | |
106 | * | |
107 | * bits reverse | |
108 | * 0 000 000 | |
109 | * 4 100 001 | |
110 | * 2 010 010 | |
111 | * 6 110 011 | |
112 | * 1 001 100 | |
113 | * 5 101 101 | |
114 | * 3 011 110 | |
115 | * 7 111 111 | |
116 | * | |
117 | * This shows the nodes in order of non-reversed bits, linked by | |
118 | * reversed-bit order. | |
119 | * | |
120 | * order bits reverse | |
121 | * 0 0 000 000 | |
122 | * | | |
f6fdd688 MD |
123 | * 1 | 1 001 100 <- <- |
124 | * | | | | | |
125 | * 2 | | 2 010 010 | | | |
126 | * | | | 3 011 110 | <- | | |
127 | * | | | | | | | | |
1475579c MD |
128 | * 3 -> | | | 4 100 001 | | |
129 | * -> | | 5 101 101 | | |
130 | * -> | 6 110 011 | |
131 | * -> 7 111 111 | |
e753ff5a MD |
132 | */ |
133 | ||
2ed95849 MD |
134 | #define _LGPL_SOURCE |
135 | #include <stdlib.h> | |
e0ba718a MD |
136 | #include <errno.h> |
137 | #include <assert.h> | |
138 | #include <stdio.h> | |
abc490a1 | 139 | #include <stdint.h> |
f000907d | 140 | #include <string.h> |
e0ba718a | 141 | |
15cfbec7 | 142 | #include "config.h" |
2ed95849 | 143 | #include <urcu.h> |
abc490a1 | 144 | #include <urcu-call-rcu.h> |
a42cc659 MD |
145 | #include <urcu/arch.h> |
146 | #include <urcu/uatomic.h> | |
a42cc659 | 147 | #include <urcu/compiler.h> |
abc490a1 | 148 | #include <urcu/rculfhash.h> |
5e28c532 | 149 | #include <stdio.h> |
464a1ec9 | 150 | #include <pthread.h> |
44395fb7 | 151 | |
f9830efd | 152 | #ifdef DEBUG |
f0c29ed7 | 153 | #define dbg_printf(fmt, args...) printf("[debug rculfhash] " fmt, ## args) |
f9830efd | 154 | #else |
e753ff5a | 155 | #define dbg_printf(fmt, args...) |
f9830efd MD |
156 | #endif |
157 | ||
f8994aee MD |
158 | /* |
159 | * Per-CPU split-counters lazily update the global counter each 1024 | |
160 | * addition/removal. It automatically keeps track of resize required. | |
161 | * We use the bucket length as indicator for need to expand for small | |
162 | * tables and machines lacking per-cpu data suppport. | |
163 | */ | |
164 | #define COUNT_COMMIT_ORDER 10 | |
6ea6bc67 MD |
165 | #define CHAIN_LEN_TARGET 1 |
166 | #define CHAIN_LEN_RESIZE_THRESHOLD 3 | |
2ed95849 | 167 | |
cd95516d | 168 | /* |
76a73da8 | 169 | * Define the minimum table size. |
cd95516d | 170 | */ |
c9edd44a | 171 | #define MIN_TABLE_SIZE 1 |
cd95516d | 172 | |
4105056a MD |
173 | #if (CAA_BITS_PER_LONG == 32) |
174 | #define MAX_TABLE_ORDER 32 | |
175 | #else | |
176 | #define MAX_TABLE_ORDER 64 | |
177 | #endif | |
178 | ||
b7d619b0 MD |
179 | /* |
180 | * Minimum number of dummy nodes to touch per thread to parallelize grow/shrink. | |
181 | */ | |
6083a889 MD |
182 | #define MIN_PARTITION_PER_THREAD_ORDER 12 |
183 | #define MIN_PARTITION_PER_THREAD (1UL << MIN_PARTITION_PER_THREAD_ORDER) | |
b7d619b0 | 184 | |
4105056a MD |
185 | #ifndef min |
186 | #define min(a, b) ((a) < (b) ? (a) : (b)) | |
187 | #endif | |
188 | ||
abc490a1 MD |
189 | #ifndef max |
190 | #define max(a, b) ((a) > (b) ? (a) : (b)) | |
191 | #endif | |
2ed95849 | 192 | |
d95bd160 MD |
193 | /* |
194 | * The removed flag needs to be updated atomically with the pointer. | |
48ed1c18 | 195 | * It indicates that no node must attach to the node scheduled for |
b198f0fd | 196 | * removal, and that node garbage collection must be performed. |
d95bd160 MD |
197 | * The dummy flag does not require to be updated atomically with the |
198 | * pointer, but it is added as a pointer low bit flag to save space. | |
199 | */ | |
d37166c6 | 200 | #define REMOVED_FLAG (1UL << 0) |
b198f0fd MD |
201 | #define DUMMY_FLAG (1UL << 1) |
202 | #define FLAGS_MASK ((1UL << 2) - 1) | |
d37166c6 | 203 | |
bb7b2f26 | 204 | /* Value of the end pointer. Should not interact with flags. */ |
f9c80341 | 205 | #define END_VALUE NULL |
bb7b2f26 | 206 | |
df44348d | 207 | struct ht_items_count { |
860d07e8 | 208 | unsigned long add, del; |
df44348d MD |
209 | } __attribute__((aligned(CAA_CACHE_LINE_SIZE))); |
210 | ||
1475579c | 211 | struct rcu_level { |
0d14ceb2 | 212 | /* Note: manually update allocation length when adding a field */ |
1475579c MD |
213 | struct _cds_lfht_node nodes[0]; |
214 | }; | |
215 | ||
395270b6 | 216 | struct rcu_table { |
4105056a | 217 | unsigned long size; /* always a power of 2, shared (RCU) */ |
f9830efd | 218 | unsigned long resize_target; |
11519af6 | 219 | int resize_initiated; |
4105056a | 220 | struct rcu_level *tbl[MAX_TABLE_ORDER]; |
395270b6 MD |
221 | }; |
222 | ||
14044b37 | 223 | struct cds_lfht { |
4105056a | 224 | struct rcu_table t; |
14044b37 MD |
225 | cds_lfht_hash_fct hash_fct; |
226 | cds_lfht_compare_fct compare_fct; | |
732ad076 | 227 | unsigned long hash_seed; |
b8af5011 | 228 | int flags; |
5f511391 MD |
229 | /* |
230 | * We need to put the work threads offline (QSBR) when taking this | |
231 | * mutex, because we use synchronize_rcu within this mutex critical | |
232 | * section, which waits on read-side critical sections, and could | |
233 | * therefore cause grace-period deadlock if we hold off RCU G.P. | |
234 | * completion. | |
235 | */ | |
464a1ec9 | 236 | pthread_mutex_t resize_mutex; /* resize mutex: add/del mutex */ |
33c7c748 | 237 | unsigned int in_progress_resize, in_progress_destroy; |
14044b37 | 238 | void (*cds_lfht_call_rcu)(struct rcu_head *head, |
abc490a1 | 239 | void (*func)(struct rcu_head *head)); |
1475579c | 240 | void (*cds_lfht_synchronize_rcu)(void); |
01dbfa62 MD |
241 | void (*cds_lfht_rcu_read_lock)(void); |
242 | void (*cds_lfht_rcu_read_unlock)(void); | |
5f511391 MD |
243 | void (*cds_lfht_rcu_thread_offline)(void); |
244 | void (*cds_lfht_rcu_thread_online)(void); | |
b7d619b0 MD |
245 | void (*cds_lfht_rcu_register_thread)(void); |
246 | void (*cds_lfht_rcu_unregister_thread)(void); | |
247 | pthread_attr_t *resize_attr; /* Resize threads attributes */ | |
7de5ccfd | 248 | long count; /* global approximate item count */ |
df44348d | 249 | struct ht_items_count *percpu_count; /* per-cpu item count */ |
2ed95849 MD |
250 | }; |
251 | ||
abc490a1 MD |
252 | struct rcu_resize_work { |
253 | struct rcu_head head; | |
14044b37 | 254 | struct cds_lfht *ht; |
abc490a1 | 255 | }; |
2ed95849 | 256 | |
b7d619b0 | 257 | struct partition_resize_work { |
1af6e26e | 258 | pthread_t thread_id; |
b7d619b0 MD |
259 | struct cds_lfht *ht; |
260 | unsigned long i, start, len; | |
261 | void (*fct)(struct cds_lfht *ht, unsigned long i, | |
262 | unsigned long start, unsigned long len); | |
263 | }; | |
264 | ||
76a73da8 | 265 | static |
83beee94 MD |
266 | void _cds_lfht_add(struct cds_lfht *ht, |
267 | unsigned long size, | |
268 | struct cds_lfht_node *node, | |
269 | struct cds_lfht_iter *unique_ret, | |
270 | int dummy); | |
48ed1c18 | 271 | |
abc490a1 MD |
272 | /* |
273 | * Algorithm to reverse bits in a word by lookup table, extended to | |
274 | * 64-bit words. | |
f9830efd | 275 | * Source: |
abc490a1 | 276 | * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable |
f9830efd | 277 | * Originally from Public Domain. |
abc490a1 MD |
278 | */ |
279 | ||
280 | static const uint8_t BitReverseTable256[256] = | |
2ed95849 | 281 | { |
abc490a1 MD |
282 | #define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64 |
283 | #define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16) | |
284 | #define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 ) | |
285 | R6(0), R6(2), R6(1), R6(3) | |
286 | }; | |
287 | #undef R2 | |
288 | #undef R4 | |
289 | #undef R6 | |
2ed95849 | 290 | |
abc490a1 MD |
291 | static |
292 | uint8_t bit_reverse_u8(uint8_t v) | |
293 | { | |
294 | return BitReverseTable256[v]; | |
295 | } | |
ab7d5fc6 | 296 | |
abc490a1 MD |
297 | static __attribute__((unused)) |
298 | uint32_t bit_reverse_u32(uint32_t v) | |
299 | { | |
300 | return ((uint32_t) bit_reverse_u8(v) << 24) | | |
301 | ((uint32_t) bit_reverse_u8(v >> 8) << 16) | | |
302 | ((uint32_t) bit_reverse_u8(v >> 16) << 8) | | |
303 | ((uint32_t) bit_reverse_u8(v >> 24)); | |
2ed95849 MD |
304 | } |
305 | ||
abc490a1 MD |
306 | static __attribute__((unused)) |
307 | uint64_t bit_reverse_u64(uint64_t v) | |
2ed95849 | 308 | { |
abc490a1 MD |
309 | return ((uint64_t) bit_reverse_u8(v) << 56) | |
310 | ((uint64_t) bit_reverse_u8(v >> 8) << 48) | | |
311 | ((uint64_t) bit_reverse_u8(v >> 16) << 40) | | |
312 | ((uint64_t) bit_reverse_u8(v >> 24) << 32) | | |
313 | ((uint64_t) bit_reverse_u8(v >> 32) << 24) | | |
314 | ((uint64_t) bit_reverse_u8(v >> 40) << 16) | | |
315 | ((uint64_t) bit_reverse_u8(v >> 48) << 8) | | |
316 | ((uint64_t) bit_reverse_u8(v >> 56)); | |
317 | } | |
318 | ||
319 | static | |
320 | unsigned long bit_reverse_ulong(unsigned long v) | |
321 | { | |
322 | #if (CAA_BITS_PER_LONG == 32) | |
323 | return bit_reverse_u32(v); | |
324 | #else | |
325 | return bit_reverse_u64(v); | |
326 | #endif | |
327 | } | |
328 | ||
f9830efd | 329 | /* |
24365af7 MD |
330 | * fls: returns the position of the most significant bit. |
331 | * Returns 0 if no bit is set, else returns the position of the most | |
332 | * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit). | |
f9830efd | 333 | */ |
24365af7 MD |
334 | #if defined(__i386) || defined(__x86_64) |
335 | static inline | |
336 | unsigned int fls_u32(uint32_t x) | |
f9830efd | 337 | { |
24365af7 MD |
338 | int r; |
339 | ||
340 | asm("bsrl %1,%0\n\t" | |
341 | "jnz 1f\n\t" | |
342 | "movl $-1,%0\n\t" | |
343 | "1:\n\t" | |
344 | : "=r" (r) : "rm" (x)); | |
345 | return r + 1; | |
346 | } | |
347 | #define HAS_FLS_U32 | |
348 | #endif | |
349 | ||
350 | #if defined(__x86_64) | |
351 | static inline | |
352 | unsigned int fls_u64(uint64_t x) | |
353 | { | |
354 | long r; | |
355 | ||
356 | asm("bsrq %1,%0\n\t" | |
357 | "jnz 1f\n\t" | |
358 | "movq $-1,%0\n\t" | |
359 | "1:\n\t" | |
360 | : "=r" (r) : "rm" (x)); | |
361 | return r + 1; | |
362 | } | |
363 | #define HAS_FLS_U64 | |
364 | #endif | |
365 | ||
366 | #ifndef HAS_FLS_U64 | |
367 | static __attribute__((unused)) | |
368 | unsigned int fls_u64(uint64_t x) | |
369 | { | |
370 | unsigned int r = 64; | |
371 | ||
372 | if (!x) | |
373 | return 0; | |
374 | ||
375 | if (!(x & 0xFFFFFFFF00000000ULL)) { | |
376 | x <<= 32; | |
377 | r -= 32; | |
378 | } | |
379 | if (!(x & 0xFFFF000000000000ULL)) { | |
380 | x <<= 16; | |
381 | r -= 16; | |
382 | } | |
383 | if (!(x & 0xFF00000000000000ULL)) { | |
384 | x <<= 8; | |
385 | r -= 8; | |
386 | } | |
387 | if (!(x & 0xF000000000000000ULL)) { | |
388 | x <<= 4; | |
389 | r -= 4; | |
390 | } | |
391 | if (!(x & 0xC000000000000000ULL)) { | |
392 | x <<= 2; | |
393 | r -= 2; | |
394 | } | |
395 | if (!(x & 0x8000000000000000ULL)) { | |
396 | x <<= 1; | |
397 | r -= 1; | |
398 | } | |
399 | return r; | |
400 | } | |
401 | #endif | |
402 | ||
403 | #ifndef HAS_FLS_U32 | |
404 | static __attribute__((unused)) | |
405 | unsigned int fls_u32(uint32_t x) | |
406 | { | |
407 | unsigned int r = 32; | |
f9830efd | 408 | |
24365af7 MD |
409 | if (!x) |
410 | return 0; | |
411 | if (!(x & 0xFFFF0000U)) { | |
412 | x <<= 16; | |
413 | r -= 16; | |
414 | } | |
415 | if (!(x & 0xFF000000U)) { | |
416 | x <<= 8; | |
417 | r -= 8; | |
418 | } | |
419 | if (!(x & 0xF0000000U)) { | |
420 | x <<= 4; | |
421 | r -= 4; | |
422 | } | |
423 | if (!(x & 0xC0000000U)) { | |
424 | x <<= 2; | |
425 | r -= 2; | |
426 | } | |
427 | if (!(x & 0x80000000U)) { | |
428 | x <<= 1; | |
429 | r -= 1; | |
430 | } | |
431 | return r; | |
432 | } | |
433 | #endif | |
434 | ||
435 | unsigned int fls_ulong(unsigned long x) | |
f9830efd | 436 | { |
24365af7 MD |
437 | #if (CAA_BITS_PER_lONG == 32) |
438 | return fls_u32(x); | |
439 | #else | |
440 | return fls_u64(x); | |
441 | #endif | |
442 | } | |
f9830efd | 443 | |
920f8ef6 LJ |
444 | /* |
445 | * Return the minimum order for which x <= (1UL << order). | |
446 | * Return -1 if x is 0. | |
447 | */ | |
24365af7 MD |
448 | int get_count_order_u32(uint32_t x) |
449 | { | |
920f8ef6 LJ |
450 | if (!x) |
451 | return -1; | |
24365af7 | 452 | |
920f8ef6 | 453 | return fls_u32(x - 1); |
24365af7 MD |
454 | } |
455 | ||
920f8ef6 LJ |
456 | /* |
457 | * Return the minimum order for which x <= (1UL << order). | |
458 | * Return -1 if x is 0. | |
459 | */ | |
24365af7 MD |
460 | int get_count_order_ulong(unsigned long x) |
461 | { | |
920f8ef6 LJ |
462 | if (!x) |
463 | return -1; | |
24365af7 | 464 | |
920f8ef6 | 465 | return fls_ulong(x - 1); |
f9830efd MD |
466 | } |
467 | ||
98808fb1 MD |
468 | #ifdef POISON_FREE |
469 | #define poison_free(ptr) \ | |
470 | do { \ | |
471 | memset(ptr, 0x42, sizeof(*(ptr))); \ | |
472 | free(ptr); \ | |
473 | } while (0) | |
474 | #else | |
475 | #define poison_free(ptr) free(ptr) | |
476 | #endif | |
477 | ||
f9830efd | 478 | static |
4105056a | 479 | void cds_lfht_resize_lazy(struct cds_lfht *ht, unsigned long size, int growth); |
f9830efd | 480 | |
df44348d MD |
481 | /* |
482 | * If the sched_getcpu() and sysconf(_SC_NPROCESSORS_CONF) calls are | |
483 | * available, then we support hash table item accounting. | |
484 | * In the unfortunate event the number of CPUs reported would be | |
485 | * inaccurate, we use modulo arithmetic on the number of CPUs we got. | |
486 | */ | |
df44348d MD |
487 | #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) |
488 | ||
f8994aee | 489 | static |
4105056a | 490 | void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size, |
f8994aee MD |
491 | unsigned long count); |
492 | ||
df44348d MD |
493 | static long nr_cpus_mask = -1; |
494 | ||
495 | static | |
496 | struct ht_items_count *alloc_per_cpu_items_count(void) | |
497 | { | |
498 | struct ht_items_count *count; | |
499 | ||
500 | switch (nr_cpus_mask) { | |
501 | case -2: | |
502 | return NULL; | |
503 | case -1: | |
504 | { | |
505 | long maxcpus; | |
506 | ||
507 | maxcpus = sysconf(_SC_NPROCESSORS_CONF); | |
508 | if (maxcpus <= 0) { | |
509 | nr_cpus_mask = -2; | |
510 | return NULL; | |
511 | } | |
512 | /* | |
513 | * round up number of CPUs to next power of two, so we | |
514 | * can use & for modulo. | |
515 | */ | |
516 | maxcpus = 1UL << get_count_order_ulong(maxcpus); | |
517 | nr_cpus_mask = maxcpus - 1; | |
518 | } | |
519 | /* Fall-through */ | |
520 | default: | |
521 | return calloc(nr_cpus_mask + 1, sizeof(*count)); | |
522 | } | |
523 | } | |
524 | ||
525 | static | |
526 | void free_per_cpu_items_count(struct ht_items_count *count) | |
527 | { | |
98808fb1 | 528 | poison_free(count); |
df44348d MD |
529 | } |
530 | ||
531 | static | |
532 | int ht_get_cpu(void) | |
533 | { | |
534 | int cpu; | |
535 | ||
536 | assert(nr_cpus_mask >= 0); | |
537 | cpu = sched_getcpu(); | |
538 | if (unlikely(cpu < 0)) | |
539 | return cpu; | |
540 | else | |
541 | return cpu & nr_cpus_mask; | |
542 | } | |
543 | ||
544 | static | |
4105056a | 545 | void ht_count_add(struct cds_lfht *ht, unsigned long size) |
df44348d | 546 | { |
3171717f | 547 | unsigned long percpu_count; |
df44348d MD |
548 | int cpu; |
549 | ||
550 | if (unlikely(!ht->percpu_count)) | |
3171717f | 551 | return; |
df44348d MD |
552 | cpu = ht_get_cpu(); |
553 | if (unlikely(cpu < 0)) | |
3171717f MD |
554 | return; |
555 | percpu_count = uatomic_add_return(&ht->percpu_count[cpu].add, 1); | |
df44348d | 556 | if (unlikely(!(percpu_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) { |
e3ecefd6 | 557 | long count; |
df44348d MD |
558 | |
559 | dbg_printf("add percpu %lu\n", percpu_count); | |
560 | count = uatomic_add_return(&ht->count, | |
561 | 1UL << COUNT_COMMIT_ORDER); | |
562 | /* If power of 2 */ | |
563 | if (!(count & (count - 1))) { | |
4105056a | 564 | if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) < size) |
f8994aee | 565 | return; |
e3ecefd6 | 566 | dbg_printf("add set global %ld\n", count); |
4105056a | 567 | cds_lfht_resize_lazy_count(ht, size, |
6ea6bc67 | 568 | count >> (CHAIN_LEN_TARGET - 1)); |
df44348d MD |
569 | } |
570 | } | |
571 | } | |
572 | ||
573 | static | |
860d07e8 | 574 | void ht_count_del(struct cds_lfht *ht, unsigned long size) |
df44348d MD |
575 | { |
576 | unsigned long percpu_count; | |
3171717f | 577 | int cpu; |
df44348d | 578 | |
3171717f MD |
579 | if (unlikely(!ht->percpu_count)) |
580 | return; | |
581 | cpu = ht_get_cpu(); | |
582 | if (unlikely(cpu < 0)) | |
583 | return; | |
80d90c06 | 584 | percpu_count = uatomic_add_return(&ht->percpu_count[cpu].del, 1); |
df44348d | 585 | if (unlikely(!(percpu_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) { |
e3ecefd6 | 586 | long count; |
df44348d | 587 | |
860d07e8 | 588 | dbg_printf("del percpu %lu\n", percpu_count); |
df44348d | 589 | count = uatomic_add_return(&ht->count, |
3171717f | 590 | -(1UL << COUNT_COMMIT_ORDER)); |
df44348d MD |
591 | /* If power of 2 */ |
592 | if (!(count & (count - 1))) { | |
4105056a | 593 | if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) >= size) |
f8994aee | 594 | return; |
e3ecefd6 MD |
595 | dbg_printf("del set global %ld\n", count); |
596 | /* | |
c941bb9e | 597 | * Don't shrink table if the number of nodes is below a |
e3ecefd6 MD |
598 | * certain threshold. |
599 | */ | |
c941bb9e | 600 | if (count < (1UL << COUNT_COMMIT_ORDER) * (nr_cpus_mask + 1)) |
e3ecefd6 | 601 | return; |
4105056a | 602 | cds_lfht_resize_lazy_count(ht, size, |
6ea6bc67 | 603 | count >> (CHAIN_LEN_TARGET - 1)); |
df44348d MD |
604 | } |
605 | } | |
606 | } | |
607 | ||
608 | #else /* #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */ | |
609 | ||
91452a6a | 610 | static const long nr_cpus_mask = -2; |
df44348d MD |
611 | |
612 | static | |
613 | struct ht_items_count *alloc_per_cpu_items_count(void) | |
614 | { | |
615 | return NULL; | |
616 | } | |
617 | ||
618 | static | |
619 | void free_per_cpu_items_count(struct ht_items_count *count) | |
620 | { | |
621 | } | |
622 | ||
623 | static | |
4105056a | 624 | void ht_count_add(struct cds_lfht *ht, unsigned long size) |
df44348d MD |
625 | { |
626 | } | |
627 | ||
628 | static | |
860d07e8 | 629 | void ht_count_del(struct cds_lfht *ht, unsigned long size) |
df44348d MD |
630 | { |
631 | } | |
632 | ||
633 | #endif /* #else #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */ | |
634 | ||
635 | ||
f9830efd | 636 | static |
4105056a | 637 | void check_resize(struct cds_lfht *ht, unsigned long size, uint32_t chain_len) |
f9830efd | 638 | { |
f8994aee MD |
639 | unsigned long count; |
640 | ||
b8af5011 MD |
641 | if (!(ht->flags & CDS_LFHT_AUTO_RESIZE)) |
642 | return; | |
f8994aee MD |
643 | count = uatomic_read(&ht->count); |
644 | /* | |
645 | * Use bucket-local length for small table expand and for | |
646 | * environments lacking per-cpu data support. | |
647 | */ | |
648 | if (count >= (1UL << COUNT_COMMIT_ORDER)) | |
649 | return; | |
24365af7 | 650 | if (chain_len > 100) |
f0c29ed7 | 651 | dbg_printf("WARNING: large chain length: %u.\n", |
24365af7 | 652 | chain_len); |
3390d470 | 653 | if (chain_len >= CHAIN_LEN_RESIZE_THRESHOLD) |
4105056a | 654 | cds_lfht_resize_lazy(ht, size, |
01370f0b | 655 | get_count_order_u32(chain_len - (CHAIN_LEN_TARGET - 1))); |
f9830efd MD |
656 | } |
657 | ||
abc490a1 | 658 | static |
14044b37 | 659 | struct cds_lfht_node *clear_flag(struct cds_lfht_node *node) |
abc490a1 | 660 | { |
14044b37 | 661 | return (struct cds_lfht_node *) (((unsigned long) node) & ~FLAGS_MASK); |
abc490a1 MD |
662 | } |
663 | ||
664 | static | |
14044b37 | 665 | int is_removed(struct cds_lfht_node *node) |
abc490a1 | 666 | { |
d37166c6 | 667 | return ((unsigned long) node) & REMOVED_FLAG; |
abc490a1 MD |
668 | } |
669 | ||
670 | static | |
14044b37 | 671 | struct cds_lfht_node *flag_removed(struct cds_lfht_node *node) |
abc490a1 | 672 | { |
14044b37 | 673 | return (struct cds_lfht_node *) (((unsigned long) node) | REMOVED_FLAG); |
abc490a1 MD |
674 | } |
675 | ||
f5596c94 | 676 | static |
14044b37 | 677 | int is_dummy(struct cds_lfht_node *node) |
f5596c94 MD |
678 | { |
679 | return ((unsigned long) node) & DUMMY_FLAG; | |
680 | } | |
681 | ||
682 | static | |
14044b37 | 683 | struct cds_lfht_node *flag_dummy(struct cds_lfht_node *node) |
f5596c94 | 684 | { |
14044b37 | 685 | return (struct cds_lfht_node *) (((unsigned long) node) | DUMMY_FLAG); |
f5596c94 | 686 | } |
bb7b2f26 MD |
687 | |
688 | static | |
689 | struct cds_lfht_node *get_end(void) | |
690 | { | |
691 | return (struct cds_lfht_node *) END_VALUE; | |
692 | } | |
693 | ||
694 | static | |
695 | int is_end(struct cds_lfht_node *node) | |
696 | { | |
697 | return clear_flag(node) == (struct cds_lfht_node *) END_VALUE; | |
698 | } | |
699 | ||
abc490a1 | 700 | static |
f9830efd | 701 | unsigned long _uatomic_max(unsigned long *ptr, unsigned long v) |
abc490a1 MD |
702 | { |
703 | unsigned long old1, old2; | |
704 | ||
705 | old1 = uatomic_read(ptr); | |
706 | do { | |
707 | old2 = old1; | |
708 | if (old2 >= v) | |
f9830efd | 709 | return old2; |
abc490a1 | 710 | } while ((old1 = uatomic_cmpxchg(ptr, old2, v)) != old2); |
f9830efd | 711 | return v; |
abc490a1 MD |
712 | } |
713 | ||
f4a9cc0b LJ |
714 | static |
715 | struct _cds_lfht_node *lookup_bucket(struct cds_lfht *ht, unsigned long size, | |
716 | unsigned long hash) | |
717 | { | |
718 | unsigned long index, order; | |
719 | ||
720 | assert(size > 0); | |
721 | index = hash & (size - 1); | |
a4ea2223 LJ |
722 | /* |
723 | * equivalent to get_count_order_ulong(index + 1), but optimizes | |
724 | * away the non-existing 0 special-case for | |
725 | * get_count_order_ulong. | |
726 | */ | |
727 | order = fls_ulong(index); | |
f4a9cc0b LJ |
728 | |
729 | dbg_printf("lookup hash %lu index %lu order %lu aridx %lu\n", | |
730 | hash, index, order, index & (!order ? 0 : ((1UL << (order - 1)) - 1))); | |
731 | ||
732 | return &ht->t.tbl[order]->nodes[index & (!order ? 0 : ((1UL << (order - 1)) - 1))]; | |
733 | } | |
734 | ||
273399de MD |
735 | /* |
736 | * Remove all logically deleted nodes from a bucket up to a certain node key. | |
737 | */ | |
738 | static | |
f9c80341 | 739 | void _cds_lfht_gc_bucket(struct cds_lfht_node *dummy, struct cds_lfht_node *node) |
273399de | 740 | { |
14044b37 | 741 | struct cds_lfht_node *iter_prev, *iter, *next, *new_next; |
273399de | 742 | |
c90201ac MD |
743 | assert(!is_dummy(dummy)); |
744 | assert(!is_removed(dummy)); | |
745 | assert(!is_dummy(node)); | |
746 | assert(!is_removed(node)); | |
273399de MD |
747 | for (;;) { |
748 | iter_prev = dummy; | |
749 | /* We can always skip the dummy node initially */ | |
cc4fcb10 | 750 | iter = rcu_dereference(iter_prev->p.next); |
b4cb483f | 751 | assert(!is_removed(iter)); |
cc4fcb10 | 752 | assert(iter_prev->p.reverse_hash <= node->p.reverse_hash); |
bd4db153 MD |
753 | /* |
754 | * We should never be called with dummy (start of chain) | |
755 | * and logically removed node (end of path compression | |
756 | * marker) being the actual same node. This would be a | |
757 | * bug in the algorithm implementation. | |
758 | */ | |
759 | assert(dummy != node); | |
273399de | 760 | for (;;) { |
bb7b2f26 | 761 | if (unlikely(is_end(iter))) |
f9c80341 | 762 | return; |
76412f24 | 763 | if (likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash)) |
f9c80341 | 764 | return; |
cc4fcb10 | 765 | next = rcu_dereference(clear_flag(iter)->p.next); |
b198f0fd | 766 | if (likely(is_removed(next))) |
273399de | 767 | break; |
b453eae1 | 768 | iter_prev = clear_flag(iter); |
273399de MD |
769 | iter = next; |
770 | } | |
b198f0fd | 771 | assert(!is_removed(iter)); |
f5596c94 MD |
772 | if (is_dummy(iter)) |
773 | new_next = flag_dummy(clear_flag(next)); | |
774 | else | |
775 | new_next = clear_flag(next); | |
776 | (void) uatomic_cmpxchg(&iter_prev->p.next, iter, new_next); | |
273399de | 777 | } |
f9c80341 | 778 | return; |
273399de MD |
779 | } |
780 | ||
9357c415 MD |
781 | static |
782 | int _cds_lfht_replace(struct cds_lfht *ht, unsigned long size, | |
783 | struct cds_lfht_node *old_node, | |
3fb86f26 | 784 | struct cds_lfht_node *old_next, |
9357c415 MD |
785 | struct cds_lfht_node *new_node) |
786 | { | |
3fb86f26 | 787 | struct cds_lfht_node *dummy, *ret_next; |
9357c415 | 788 | struct _cds_lfht_node *lookup; |
9357c415 MD |
789 | |
790 | if (!old_node) /* Return -ENOENT if asked to replace NULL node */ | |
7801dadd | 791 | return -ENOENT; |
9357c415 MD |
792 | |
793 | assert(!is_removed(old_node)); | |
794 | assert(!is_dummy(old_node)); | |
795 | assert(!is_removed(new_node)); | |
796 | assert(!is_dummy(new_node)); | |
797 | assert(new_node != old_node); | |
3fb86f26 | 798 | for (;;) { |
9357c415 | 799 | /* Insert after node to be replaced */ |
9357c415 MD |
800 | if (is_removed(old_next)) { |
801 | /* | |
802 | * Too late, the old node has been removed under us | |
803 | * between lookup and replace. Fail. | |
804 | */ | |
7801dadd | 805 | return -ENOENT; |
9357c415 MD |
806 | } |
807 | assert(!is_dummy(old_next)); | |
808 | assert(new_node != clear_flag(old_next)); | |
809 | new_node->p.next = clear_flag(old_next); | |
810 | /* | |
811 | * Here is the whole trick for lock-free replace: we add | |
812 | * the replacement node _after_ the node we want to | |
813 | * replace by atomically setting its next pointer at the | |
814 | * same time we set its removal flag. Given that | |
815 | * the lookups/get next use an iterator aware of the | |
816 | * next pointer, they will either skip the old node due | |
817 | * to the removal flag and see the new node, or use | |
818 | * the old node, but will not see the new one. | |
819 | */ | |
820 | ret_next = uatomic_cmpxchg(&old_node->p.next, | |
821 | old_next, flag_removed(new_node)); | |
3fb86f26 | 822 | if (ret_next == old_next) |
7801dadd | 823 | break; /* We performed the replacement. */ |
3fb86f26 LJ |
824 | old_next = ret_next; |
825 | } | |
9357c415 | 826 | |
9357c415 MD |
827 | /* |
828 | * Ensure that the old node is not visible to readers anymore: | |
829 | * lookup for the node, and remove it (along with any other | |
830 | * logically removed node) if found. | |
831 | */ | |
f4a9cc0b | 832 | lookup = lookup_bucket(ht, size, bit_reverse_ulong(old_node->p.reverse_hash)); |
9357c415 MD |
833 | dummy = (struct cds_lfht_node *) lookup; |
834 | _cds_lfht_gc_bucket(dummy, new_node); | |
7801dadd LJ |
835 | |
836 | assert(is_removed(rcu_dereference(old_node->p.next))); | |
837 | return 0; | |
9357c415 MD |
838 | } |
839 | ||
83beee94 MD |
840 | /* |
841 | * A non-NULL unique_ret pointer uses the "add unique" (or uniquify) add | |
842 | * mode. A NULL unique_ret allows creation of duplicate keys. | |
843 | */ | |
abc490a1 | 844 | static |
83beee94 MD |
845 | void _cds_lfht_add(struct cds_lfht *ht, |
846 | unsigned long size, | |
847 | struct cds_lfht_node *node, | |
848 | struct cds_lfht_iter *unique_ret, | |
849 | int dummy) | |
abc490a1 | 850 | { |
14044b37 | 851 | struct cds_lfht_node *iter_prev, *iter, *next, *new_node, *new_next, |
960c9e4f | 852 | *return_node; |
14044b37 | 853 | struct _cds_lfht_node *lookup; |
abc490a1 | 854 | |
c90201ac MD |
855 | assert(!is_dummy(node)); |
856 | assert(!is_removed(node)); | |
4105056a | 857 | if (!size) { |
f5596c94 | 858 | assert(dummy); |
83beee94 | 859 | assert(!unique_ret); |
bb7b2f26 | 860 | node->p.next = flag_dummy(get_end()); |
83beee94 | 861 | return; /* Initial first add (head) */ |
18117871 | 862 | } |
f4a9cc0b | 863 | lookup = lookup_bucket(ht, size, bit_reverse_ulong(node->p.reverse_hash)); |
abc490a1 | 864 | for (;;) { |
adc0de68 | 865 | uint32_t chain_len = 0; |
abc490a1 | 866 | |
11519af6 MD |
867 | /* |
868 | * iter_prev points to the non-removed node prior to the | |
869 | * insert location. | |
11519af6 | 870 | */ |
14044b37 | 871 | iter_prev = (struct cds_lfht_node *) lookup; |
11519af6 | 872 | /* We can always skip the dummy node initially */ |
cc4fcb10 MD |
873 | iter = rcu_dereference(iter_prev->p.next); |
874 | assert(iter_prev->p.reverse_hash <= node->p.reverse_hash); | |
abc490a1 | 875 | for (;;) { |
bb7b2f26 | 876 | if (unlikely(is_end(iter))) |
273399de | 877 | goto insert; |
76412f24 | 878 | if (likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash)) |
273399de | 879 | goto insert; |
194fdbd1 LJ |
880 | /* dummy node is the first node of the identical-hash-value chain */ |
881 | if (dummy && clear_flag(iter)->p.reverse_hash == node->p.reverse_hash) | |
882 | goto insert; | |
cc4fcb10 | 883 | next = rcu_dereference(clear_flag(iter)->p.next); |
b198f0fd | 884 | if (unlikely(is_removed(next))) |
9dba85be | 885 | goto gc_node; |
83beee94 | 886 | if (unique_ret |
1b81fe1a | 887 | && !is_dummy(next) |
ae450da3 | 888 | && clear_flag(iter)->p.reverse_hash == node->p.reverse_hash |
e43f23f8 MD |
889 | && !ht->compare_fct(node->key, node->key_len, |
890 | clear_flag(iter)->key, | |
48ed1c18 | 891 | clear_flag(iter)->key_len)) { |
83beee94 MD |
892 | unique_ret->node = clear_flag(iter); |
893 | unique_ret->next = next; | |
894 | return; | |
48ed1c18 | 895 | } |
11519af6 | 896 | /* Only account for identical reverse hash once */ |
24365af7 MD |
897 | if (iter_prev->p.reverse_hash != clear_flag(iter)->p.reverse_hash |
898 | && !is_dummy(next)) | |
4105056a | 899 | check_resize(ht, size, ++chain_len); |
11519af6 | 900 | iter_prev = clear_flag(iter); |
273399de | 901 | iter = next; |
abc490a1 | 902 | } |
48ed1c18 | 903 | |
273399de | 904 | insert: |
7ec59d3b | 905 | assert(node != clear_flag(iter)); |
11519af6 | 906 | assert(!is_removed(iter_prev)); |
c90201ac | 907 | assert(!is_removed(iter)); |
f000907d | 908 | assert(iter_prev != node); |
f9c80341 | 909 | if (!dummy) |
1b81fe1a | 910 | node->p.next = clear_flag(iter); |
f9c80341 MD |
911 | else |
912 | node->p.next = flag_dummy(clear_flag(iter)); | |
f5596c94 MD |
913 | if (is_dummy(iter)) |
914 | new_node = flag_dummy(node); | |
915 | else | |
916 | new_node = node; | |
cc4fcb10 | 917 | if (uatomic_cmpxchg(&iter_prev->p.next, iter, |
48ed1c18 | 918 | new_node) != iter) { |
273399de | 919 | continue; /* retry */ |
48ed1c18 | 920 | } else { |
83beee94 | 921 | return_node = node; |
960c9e4f | 922 | goto end; |
48ed1c18 MD |
923 | } |
924 | ||
9dba85be MD |
925 | gc_node: |
926 | assert(!is_removed(iter)); | |
f5596c94 MD |
927 | if (is_dummy(iter)) |
928 | new_next = flag_dummy(clear_flag(next)); | |
929 | else | |
930 | new_next = clear_flag(next); | |
931 | (void) uatomic_cmpxchg(&iter_prev->p.next, iter, new_next); | |
273399de | 932 | /* retry */ |
464a1ec9 | 933 | } |
9357c415 | 934 | end: |
83beee94 MD |
935 | if (unique_ret) { |
936 | unique_ret->node = return_node; | |
937 | /* unique_ret->next left unset, never used. */ | |
938 | } | |
abc490a1 | 939 | } |
464a1ec9 | 940 | |
abc490a1 | 941 | static |
860d07e8 | 942 | int _cds_lfht_del(struct cds_lfht *ht, unsigned long size, |
4105056a | 943 | struct cds_lfht_node *node, |
b198f0fd | 944 | int dummy_removal) |
abc490a1 | 945 | { |
14044b37 MD |
946 | struct cds_lfht_node *dummy, *next, *old; |
947 | struct _cds_lfht_node *lookup; | |
5e28c532 | 948 | |
9357c415 | 949 | if (!node) /* Return -ENOENT if asked to delete NULL node */ |
743f9143 | 950 | return -ENOENT; |
9357c415 | 951 | |
7ec59d3b | 952 | /* logically delete the node */ |
c90201ac MD |
953 | assert(!is_dummy(node)); |
954 | assert(!is_removed(node)); | |
cc4fcb10 | 955 | old = rcu_dereference(node->p.next); |
7ec59d3b | 956 | do { |
48ed1c18 MD |
957 | struct cds_lfht_node *new_next; |
958 | ||
7ec59d3b | 959 | next = old; |
76412f24 | 960 | if (unlikely(is_removed(next))) |
743f9143 | 961 | return -ENOENT; |
1475579c MD |
962 | if (dummy_removal) |
963 | assert(is_dummy(next)); | |
964 | else | |
965 | assert(!is_dummy(next)); | |
48ed1c18 | 966 | new_next = flag_removed(next); |
48ed1c18 | 967 | old = uatomic_cmpxchg(&node->p.next, next, new_next); |
7ec59d3b | 968 | } while (old != next); |
7ec59d3b | 969 | /* We performed the (logical) deletion. */ |
7ec59d3b MD |
970 | |
971 | /* | |
972 | * Ensure that the node is not visible to readers anymore: lookup for | |
273399de MD |
973 | * the node, and remove it (along with any other logically removed node) |
974 | * if found. | |
11519af6 | 975 | */ |
f4a9cc0b | 976 | lookup = lookup_bucket(ht, size, bit_reverse_ulong(node->p.reverse_hash)); |
14044b37 | 977 | dummy = (struct cds_lfht_node *) lookup; |
f9c80341 | 978 | _cds_lfht_gc_bucket(dummy, node); |
743f9143 LJ |
979 | |
980 | assert(is_removed(rcu_dereference(node->p.next))); | |
981 | return 0; | |
abc490a1 | 982 | } |
2ed95849 | 983 | |
b7d619b0 MD |
984 | static |
985 | void *partition_resize_thread(void *arg) | |
986 | { | |
987 | struct partition_resize_work *work = arg; | |
988 | ||
989 | work->ht->cds_lfht_rcu_register_thread(); | |
990 | work->fct(work->ht, work->i, work->start, work->len); | |
991 | work->ht->cds_lfht_rcu_unregister_thread(); | |
992 | return NULL; | |
993 | } | |
994 | ||
995 | static | |
996 | void partition_resize_helper(struct cds_lfht *ht, unsigned long i, | |
997 | unsigned long len, | |
998 | void (*fct)(struct cds_lfht *ht, unsigned long i, | |
999 | unsigned long start, unsigned long len)) | |
1000 | { | |
1001 | unsigned long partition_len; | |
1002 | struct partition_resize_work *work; | |
6083a889 MD |
1003 | int thread, ret; |
1004 | unsigned long nr_threads; | |
b7d619b0 | 1005 | |
6083a889 MD |
1006 | /* |
1007 | * Note: nr_cpus_mask + 1 is always power of 2. | |
1008 | * We spawn just the number of threads we need to satisfy the minimum | |
1009 | * partition size, up to the number of CPUs in the system. | |
1010 | */ | |
91452a6a MD |
1011 | if (nr_cpus_mask > 0) { |
1012 | nr_threads = min(nr_cpus_mask + 1, | |
1013 | len >> MIN_PARTITION_PER_THREAD_ORDER); | |
1014 | } else { | |
1015 | nr_threads = 1; | |
1016 | } | |
6083a889 MD |
1017 | partition_len = len >> get_count_order_ulong(nr_threads); |
1018 | work = calloc(nr_threads, sizeof(*work)); | |
b7d619b0 | 1019 | assert(work); |
6083a889 MD |
1020 | for (thread = 0; thread < nr_threads; thread++) { |
1021 | work[thread].ht = ht; | |
1022 | work[thread].i = i; | |
1023 | work[thread].len = partition_len; | |
1024 | work[thread].start = thread * partition_len; | |
1025 | work[thread].fct = fct; | |
1af6e26e | 1026 | ret = pthread_create(&(work[thread].thread_id), ht->resize_attr, |
6083a889 | 1027 | partition_resize_thread, &work[thread]); |
b7d619b0 MD |
1028 | assert(!ret); |
1029 | } | |
6083a889 | 1030 | for (thread = 0; thread < nr_threads; thread++) { |
1af6e26e | 1031 | ret = pthread_join(work[thread].thread_id, NULL); |
b7d619b0 MD |
1032 | assert(!ret); |
1033 | } | |
1034 | free(work); | |
b7d619b0 MD |
1035 | } |
1036 | ||
e8de508e MD |
1037 | /* |
1038 | * Holding RCU read lock to protect _cds_lfht_add against memory | |
1039 | * reclaim that could be performed by other call_rcu worker threads (ABA | |
1040 | * problem). | |
9ee0fc9a | 1041 | * |
b7d619b0 | 1042 | * When we reach a certain length, we can split this population phase over |
9ee0fc9a MD |
1043 | * many worker threads, based on the number of CPUs available in the system. |
1044 | * This should therefore take care of not having the expand lagging behind too | |
1045 | * many concurrent insertion threads by using the scheduler's ability to | |
1046 | * schedule dummy node population fairly with insertions. | |
e8de508e | 1047 | */ |
4105056a | 1048 | static |
b7d619b0 MD |
1049 | void init_table_populate_partition(struct cds_lfht *ht, unsigned long i, |
1050 | unsigned long start, unsigned long len) | |
4105056a MD |
1051 | { |
1052 | unsigned long j; | |
1053 | ||
1054 | ht->cds_lfht_rcu_read_lock(); | |
b7d619b0 | 1055 | for (j = start; j < start + len; j++) { |
4105056a MD |
1056 | struct cds_lfht_node *new_node = |
1057 | (struct cds_lfht_node *) &ht->t.tbl[i]->nodes[j]; | |
1058 | ||
dc1da8f6 | 1059 | dbg_printf("init populate: i %lu j %lu hash %lu\n", |
4105056a | 1060 | i, j, !i ? 0 : (1UL << (i - 1)) + j); |
dc1da8f6 MD |
1061 | new_node->p.reverse_hash = |
1062 | bit_reverse_ulong(!i ? 0 : (1UL << (i - 1)) + j); | |
83beee94 MD |
1063 | _cds_lfht_add(ht, !i ? 0 : (1UL << (i - 1)), |
1064 | new_node, NULL, 1); | |
4105056a MD |
1065 | } |
1066 | ht->cds_lfht_rcu_read_unlock(); | |
b7d619b0 MD |
1067 | } |
1068 | ||
1069 | static | |
1070 | void init_table_populate(struct cds_lfht *ht, unsigned long i, | |
1071 | unsigned long len) | |
1072 | { | |
1073 | assert(nr_cpus_mask != -1); | |
6083a889 | 1074 | if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD) { |
b7d619b0 MD |
1075 | ht->cds_lfht_rcu_thread_online(); |
1076 | init_table_populate_partition(ht, i, 0, len); | |
1077 | ht->cds_lfht_rcu_thread_offline(); | |
1078 | return; | |
1079 | } | |
1080 | partition_resize_helper(ht, i, len, init_table_populate_partition); | |
4105056a MD |
1081 | } |
1082 | ||
abc490a1 | 1083 | static |
4105056a | 1084 | void init_table(struct cds_lfht *ht, |
24365af7 MD |
1085 | unsigned long first_order, unsigned long len_order) |
1086 | { | |
1087 | unsigned long i, end_order; | |
1088 | ||
f0c29ed7 | 1089 | dbg_printf("init table: first_order %lu end_order %lu\n", |
24365af7 MD |
1090 | first_order, first_order + len_order); |
1091 | end_order = first_order + len_order; | |
24365af7 | 1092 | for (i = first_order; i < end_order; i++) { |
4105056a | 1093 | unsigned long len; |
24365af7 MD |
1094 | |
1095 | len = !i ? 1 : 1UL << (i - 1); | |
f0c29ed7 | 1096 | dbg_printf("init order %lu len: %lu\n", i, len); |
4d676753 MD |
1097 | |
1098 | /* Stop expand if the resize target changes under us */ | |
1099 | if (CMM_LOAD_SHARED(ht->t.resize_target) < (!i ? 1 : (1UL << i))) | |
1100 | break; | |
1101 | ||
0d14ceb2 | 1102 | ht->t.tbl[i] = calloc(1, len * sizeof(struct _cds_lfht_node)); |
b7d619b0 | 1103 | assert(ht->t.tbl[i]); |
4105056a | 1104 | |
4105056a | 1105 | /* |
dc1da8f6 MD |
1106 | * Set all dummy nodes reverse hash values for a level and |
1107 | * link all dummy nodes into the table. | |
4105056a | 1108 | */ |
dc1da8f6 | 1109 | init_table_populate(ht, i, len); |
4105056a | 1110 | |
f9c80341 MD |
1111 | /* |
1112 | * Update table size. | |
1113 | */ | |
1114 | cmm_smp_wmb(); /* populate data before RCU size */ | |
1115 | CMM_STORE_SHARED(ht->t.size, !i ? 1 : (1UL << i)); | |
1116 | ||
4105056a MD |
1117 | dbg_printf("init new size: %lu\n", !i ? 1 : (1UL << i)); |
1118 | if (CMM_LOAD_SHARED(ht->in_progress_destroy)) | |
1119 | break; | |
1120 | } | |
1121 | } | |
1122 | ||
e8de508e MD |
1123 | /* |
1124 | * Holding RCU read lock to protect _cds_lfht_remove against memory | |
1125 | * reclaim that could be performed by other call_rcu worker threads (ABA | |
1126 | * problem). | |
1127 | * For a single level, we logically remove and garbage collect each node. | |
1128 | * | |
1129 | * As a design choice, we perform logical removal and garbage collection on a | |
1130 | * node-per-node basis to simplify this algorithm. We also assume keeping good | |
1131 | * cache locality of the operation would overweight possible performance gain | |
1132 | * that could be achieved by batching garbage collection for multiple levels. | |
1133 | * However, this would have to be justified by benchmarks. | |
1134 | * | |
1135 | * Concurrent removal and add operations are helping us perform garbage | |
1136 | * collection of logically removed nodes. We guarantee that all logically | |
1137 | * removed nodes have been garbage-collected (unlinked) before call_rcu is | |
1138 | * invoked to free a hole level of dummy nodes (after a grace period). | |
1139 | * | |
1140 | * Logical removal and garbage collection can therefore be done in batch or on a | |
1141 | * node-per-node basis, as long as the guarantee above holds. | |
9ee0fc9a | 1142 | * |
b7d619b0 MD |
1143 | * When we reach a certain length, we can split this removal over many worker |
1144 | * threads, based on the number of CPUs available in the system. This should | |
1145 | * take care of not letting resize process lag behind too many concurrent | |
9ee0fc9a | 1146 | * updater threads actively inserting into the hash table. |
e8de508e | 1147 | */ |
4105056a | 1148 | static |
b7d619b0 MD |
1149 | void remove_table_partition(struct cds_lfht *ht, unsigned long i, |
1150 | unsigned long start, unsigned long len) | |
4105056a MD |
1151 | { |
1152 | unsigned long j; | |
1153 | ||
1154 | ht->cds_lfht_rcu_read_lock(); | |
b7d619b0 | 1155 | for (j = start; j < start + len; j++) { |
4105056a MD |
1156 | struct cds_lfht_node *fini_node = |
1157 | (struct cds_lfht_node *) &ht->t.tbl[i]->nodes[j]; | |
1158 | ||
1159 | dbg_printf("remove entry: i %lu j %lu hash %lu\n", | |
1160 | i, j, !i ? 0 : (1UL << (i - 1)) + j); | |
1161 | fini_node->p.reverse_hash = | |
1162 | bit_reverse_ulong(!i ? 0 : (1UL << (i - 1)) + j); | |
860d07e8 | 1163 | (void) _cds_lfht_del(ht, !i ? 0 : (1UL << (i - 1)), |
b198f0fd | 1164 | fini_node, 1); |
abc490a1 | 1165 | } |
4105056a | 1166 | ht->cds_lfht_rcu_read_unlock(); |
b7d619b0 MD |
1167 | } |
1168 | ||
1169 | static | |
1170 | void remove_table(struct cds_lfht *ht, unsigned long i, unsigned long len) | |
1171 | { | |
1172 | ||
1173 | assert(nr_cpus_mask != -1); | |
6083a889 | 1174 | if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD) { |
b7d619b0 MD |
1175 | ht->cds_lfht_rcu_thread_online(); |
1176 | remove_table_partition(ht, i, 0, len); | |
1177 | ht->cds_lfht_rcu_thread_offline(); | |
1178 | return; | |
1179 | } | |
1180 | partition_resize_helper(ht, i, len, remove_table_partition); | |
2ed95849 MD |
1181 | } |
1182 | ||
1475579c | 1183 | static |
4105056a | 1184 | void fini_table(struct cds_lfht *ht, |
1475579c MD |
1185 | unsigned long first_order, unsigned long len_order) |
1186 | { | |
1187 | long i, end_order; | |
0d14ceb2 | 1188 | void *free_by_rcu = NULL; |
1475579c MD |
1189 | |
1190 | dbg_printf("fini table: first_order %lu end_order %lu\n", | |
1191 | first_order, first_order + len_order); | |
1192 | end_order = first_order + len_order; | |
1193 | assert(first_order > 0); | |
1475579c | 1194 | for (i = end_order - 1; i >= first_order; i--) { |
4105056a | 1195 | unsigned long len; |
1475579c MD |
1196 | |
1197 | len = !i ? 1 : 1UL << (i - 1); | |
1198 | dbg_printf("fini order %lu len: %lu\n", i, len); | |
4105056a | 1199 | |
4d676753 MD |
1200 | /* Stop shrink if the resize target changes under us */ |
1201 | if (CMM_LOAD_SHARED(ht->t.resize_target) > (1UL << (i - 1))) | |
1202 | break; | |
1203 | ||
1204 | cmm_smp_wmb(); /* populate data before RCU size */ | |
1205 | CMM_STORE_SHARED(ht->t.size, 1UL << (i - 1)); | |
1206 | ||
1207 | /* | |
1208 | * We need to wait for all add operations to reach Q.S. (and | |
1209 | * thus use the new table for lookups) before we can start | |
1210 | * releasing the old dummy nodes. Otherwise their lookup will | |
1211 | * return a logically removed node as insert position. | |
1212 | */ | |
1213 | ht->cds_lfht_synchronize_rcu(); | |
0d14ceb2 LJ |
1214 | if (free_by_rcu) |
1215 | free(free_by_rcu); | |
4d676753 | 1216 | |
21263e21 | 1217 | /* |
4105056a MD |
1218 | * Set "removed" flag in dummy nodes about to be removed. |
1219 | * Unlink all now-logically-removed dummy node pointers. | |
1220 | * Concurrent add/remove operation are helping us doing | |
1221 | * the gc. | |
21263e21 | 1222 | */ |
4105056a MD |
1223 | remove_table(ht, i, len); |
1224 | ||
0d14ceb2 | 1225 | free_by_rcu = ht->t.tbl[i]; |
4105056a MD |
1226 | |
1227 | dbg_printf("fini new size: %lu\n", 1UL << i); | |
1475579c MD |
1228 | if (CMM_LOAD_SHARED(ht->in_progress_destroy)) |
1229 | break; | |
1230 | } | |
0d14ceb2 LJ |
1231 | |
1232 | if (free_by_rcu) { | |
1233 | ht->cds_lfht_synchronize_rcu(); | |
1234 | free(free_by_rcu); | |
1235 | } | |
1475579c MD |
1236 | } |
1237 | ||
7a9dcf9b | 1238 | struct cds_lfht *_cds_lfht_new(cds_lfht_hash_fct hash_fct, |
14044b37 MD |
1239 | cds_lfht_compare_fct compare_fct, |
1240 | unsigned long hash_seed, | |
1241 | unsigned long init_size, | |
b8af5011 | 1242 | int flags, |
14044b37 | 1243 | void (*cds_lfht_call_rcu)(struct rcu_head *head, |
1475579c | 1244 | void (*func)(struct rcu_head *head)), |
01dbfa62 MD |
1245 | void (*cds_lfht_synchronize_rcu)(void), |
1246 | void (*cds_lfht_rcu_read_lock)(void), | |
5f511391 MD |
1247 | void (*cds_lfht_rcu_read_unlock)(void), |
1248 | void (*cds_lfht_rcu_thread_offline)(void), | |
b7d619b0 MD |
1249 | void (*cds_lfht_rcu_thread_online)(void), |
1250 | void (*cds_lfht_rcu_register_thread)(void), | |
1251 | void (*cds_lfht_rcu_unregister_thread)(void), | |
1252 | pthread_attr_t *attr) | |
abc490a1 | 1253 | { |
14044b37 | 1254 | struct cds_lfht *ht; |
24365af7 | 1255 | unsigned long order; |
abc490a1 | 1256 | |
8129be4e | 1257 | /* init_size must be power of two */ |
49619ea0 | 1258 | if (init_size && (init_size & (init_size - 1))) |
8129be4e | 1259 | return NULL; |
14044b37 | 1260 | ht = calloc(1, sizeof(struct cds_lfht)); |
b7d619b0 | 1261 | assert(ht); |
abc490a1 | 1262 | ht->hash_fct = hash_fct; |
732ad076 MD |
1263 | ht->compare_fct = compare_fct; |
1264 | ht->hash_seed = hash_seed; | |
14044b37 | 1265 | ht->cds_lfht_call_rcu = cds_lfht_call_rcu; |
1475579c | 1266 | ht->cds_lfht_synchronize_rcu = cds_lfht_synchronize_rcu; |
01dbfa62 MD |
1267 | ht->cds_lfht_rcu_read_lock = cds_lfht_rcu_read_lock; |
1268 | ht->cds_lfht_rcu_read_unlock = cds_lfht_rcu_read_unlock; | |
5f511391 MD |
1269 | ht->cds_lfht_rcu_thread_offline = cds_lfht_rcu_thread_offline; |
1270 | ht->cds_lfht_rcu_thread_online = cds_lfht_rcu_thread_online; | |
b7d619b0 MD |
1271 | ht->cds_lfht_rcu_register_thread = cds_lfht_rcu_register_thread; |
1272 | ht->cds_lfht_rcu_unregister_thread = cds_lfht_rcu_unregister_thread; | |
1273 | ht->resize_attr = attr; | |
df44348d | 1274 | ht->percpu_count = alloc_per_cpu_items_count(); |
abc490a1 MD |
1275 | /* this mutex should not nest in read-side C.S. */ |
1276 | pthread_mutex_init(&ht->resize_mutex, NULL); | |
cd95516d | 1277 | order = get_count_order_ulong(max(init_size, MIN_TABLE_SIZE)) + 1; |
b8af5011 | 1278 | ht->flags = flags; |
5f511391 | 1279 | ht->cds_lfht_rcu_thread_offline(); |
f000907d | 1280 | pthread_mutex_lock(&ht->resize_mutex); |
4d676753 | 1281 | ht->t.resize_target = 1UL << (order - 1); |
4105056a | 1282 | init_table(ht, 0, order); |
f000907d | 1283 | pthread_mutex_unlock(&ht->resize_mutex); |
5f511391 | 1284 | ht->cds_lfht_rcu_thread_online(); |
abc490a1 MD |
1285 | return ht; |
1286 | } | |
1287 | ||
adc0de68 MD |
1288 | void cds_lfht_lookup(struct cds_lfht *ht, void *key, size_t key_len, |
1289 | struct cds_lfht_iter *iter) | |
2ed95849 | 1290 | { |
bb7b2f26 | 1291 | struct cds_lfht_node *node, *next, *dummy_node; |
14044b37 | 1292 | struct _cds_lfht_node *lookup; |
f4a9cc0b | 1293 | unsigned long hash, reverse_hash, size; |
2ed95849 | 1294 | |
732ad076 | 1295 | hash = ht->hash_fct(key, key_len, ht->hash_seed); |
abc490a1 | 1296 | reverse_hash = bit_reverse_ulong(hash); |
464a1ec9 | 1297 | |
4105056a | 1298 | size = rcu_dereference(ht->t.size); |
f4a9cc0b | 1299 | lookup = lookup_bucket(ht, size, hash); |
bb7b2f26 MD |
1300 | dummy_node = (struct cds_lfht_node *) lookup; |
1301 | /* We can always skip the dummy node initially */ | |
1302 | node = rcu_dereference(dummy_node->p.next); | |
bb7b2f26 | 1303 | node = clear_flag(node); |
2ed95849 | 1304 | for (;;) { |
bb7b2f26 | 1305 | if (unlikely(is_end(node))) { |
96ad1112 | 1306 | node = next = NULL; |
abc490a1 | 1307 | break; |
bb7b2f26 | 1308 | } |
cc4fcb10 | 1309 | if (unlikely(node->p.reverse_hash > reverse_hash)) { |
96ad1112 | 1310 | node = next = NULL; |
abc490a1 | 1311 | break; |
2ed95849 | 1312 | } |
1b81fe1a | 1313 | next = rcu_dereference(node->p.next); |
adc0de68 | 1314 | if (likely(!is_removed(next)) |
1b81fe1a | 1315 | && !is_dummy(next) |
ae450da3 | 1316 | && clear_flag(node)->p.reverse_hash == reverse_hash |
49c2e2d6 | 1317 | && likely(!ht->compare_fct(node->key, node->key_len, key, key_len))) { |
273399de | 1318 | break; |
2ed95849 | 1319 | } |
1b81fe1a | 1320 | node = clear_flag(next); |
2ed95849 | 1321 | } |
1b81fe1a | 1322 | assert(!node || !is_dummy(rcu_dereference(node->p.next))); |
adc0de68 MD |
1323 | iter->node = node; |
1324 | iter->next = next; | |
abc490a1 | 1325 | } |
e0ba718a | 1326 | |
3883c0e5 | 1327 | void cds_lfht_next_duplicate(struct cds_lfht *ht, struct cds_lfht_iter *iter) |
a481e5ff | 1328 | { |
adc0de68 | 1329 | struct cds_lfht_node *node, *next; |
a481e5ff MD |
1330 | unsigned long reverse_hash; |
1331 | void *key; | |
1332 | size_t key_len; | |
1333 | ||
adc0de68 | 1334 | node = iter->node; |
a481e5ff MD |
1335 | reverse_hash = node->p.reverse_hash; |
1336 | key = node->key; | |
1337 | key_len = node->key_len; | |
adc0de68 | 1338 | next = iter->next; |
a481e5ff MD |
1339 | node = clear_flag(next); |
1340 | ||
1341 | for (;;) { | |
bb7b2f26 | 1342 | if (unlikely(is_end(node))) { |
96ad1112 | 1343 | node = next = NULL; |
a481e5ff | 1344 | break; |
bb7b2f26 | 1345 | } |
a481e5ff | 1346 | if (unlikely(node->p.reverse_hash > reverse_hash)) { |
96ad1112 | 1347 | node = next = NULL; |
a481e5ff MD |
1348 | break; |
1349 | } | |
1350 | next = rcu_dereference(node->p.next); | |
adc0de68 | 1351 | if (likely(!is_removed(next)) |
a481e5ff MD |
1352 | && !is_dummy(next) |
1353 | && likely(!ht->compare_fct(node->key, node->key_len, key, key_len))) { | |
1354 | break; | |
1355 | } | |
1356 | node = clear_flag(next); | |
1357 | } | |
1358 | assert(!node || !is_dummy(rcu_dereference(node->p.next))); | |
adc0de68 MD |
1359 | iter->node = node; |
1360 | iter->next = next; | |
a481e5ff MD |
1361 | } |
1362 | ||
4e9b9fbf MD |
1363 | void cds_lfht_next(struct cds_lfht *ht, struct cds_lfht_iter *iter) |
1364 | { | |
1365 | struct cds_lfht_node *node, *next; | |
1366 | ||
853395e1 | 1367 | node = clear_flag(iter->next); |
4e9b9fbf MD |
1368 | for (;;) { |
1369 | if (unlikely(is_end(node))) { | |
1370 | node = next = NULL; | |
1371 | break; | |
1372 | } | |
1373 | next = rcu_dereference(node->p.next); | |
1374 | if (likely(!is_removed(next)) | |
1375 | && !is_dummy(next)) { | |
1376 | break; | |
1377 | } | |
1378 | node = clear_flag(next); | |
1379 | } | |
1380 | assert(!node || !is_dummy(rcu_dereference(node->p.next))); | |
1381 | iter->node = node; | |
1382 | iter->next = next; | |
1383 | } | |
1384 | ||
1385 | void cds_lfht_first(struct cds_lfht *ht, struct cds_lfht_iter *iter) | |
1386 | { | |
1387 | struct _cds_lfht_node *lookup; | |
1388 | ||
1389 | /* | |
1390 | * Get next after first dummy node. The first dummy node is the | |
1391 | * first node of the linked list. | |
1392 | */ | |
1393 | lookup = &ht->t.tbl[0]->nodes[0]; | |
853395e1 | 1394 | iter->next = lookup->next; |
4e9b9fbf MD |
1395 | cds_lfht_next(ht, iter); |
1396 | } | |
1397 | ||
14044b37 | 1398 | void cds_lfht_add(struct cds_lfht *ht, struct cds_lfht_node *node) |
abc490a1 | 1399 | { |
4105056a | 1400 | unsigned long hash, size; |
ab7d5fc6 | 1401 | |
49c2e2d6 | 1402 | hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed); |
cc4fcb10 | 1403 | node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash); |
2ed95849 | 1404 | |
4105056a | 1405 | size = rcu_dereference(ht->t.size); |
83beee94 | 1406 | _cds_lfht_add(ht, size, node, NULL, 0); |
4105056a | 1407 | ht_count_add(ht, size); |
3eca1b8c MD |
1408 | } |
1409 | ||
14044b37 | 1410 | struct cds_lfht_node *cds_lfht_add_unique(struct cds_lfht *ht, |
48ed1c18 | 1411 | struct cds_lfht_node *node) |
3eca1b8c | 1412 | { |
4105056a | 1413 | unsigned long hash, size; |
83beee94 | 1414 | struct cds_lfht_iter iter; |
3eca1b8c | 1415 | |
49c2e2d6 | 1416 | hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed); |
cc4fcb10 | 1417 | node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash); |
3eca1b8c | 1418 | |
4105056a | 1419 | size = rcu_dereference(ht->t.size); |
83beee94 MD |
1420 | _cds_lfht_add(ht, size, node, &iter, 0); |
1421 | if (iter.node == node) | |
4105056a | 1422 | ht_count_add(ht, size); |
83beee94 | 1423 | return iter.node; |
2ed95849 MD |
1424 | } |
1425 | ||
9357c415 | 1426 | struct cds_lfht_node *cds_lfht_add_replace(struct cds_lfht *ht, |
48ed1c18 MD |
1427 | struct cds_lfht_node *node) |
1428 | { | |
1429 | unsigned long hash, size; | |
83beee94 | 1430 | struct cds_lfht_iter iter; |
48ed1c18 MD |
1431 | |
1432 | hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed); | |
1433 | node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash); | |
1434 | ||
1435 | size = rcu_dereference(ht->t.size); | |
83beee94 MD |
1436 | for (;;) { |
1437 | _cds_lfht_add(ht, size, node, &iter, 0); | |
1438 | if (iter.node == node) { | |
1439 | ht_count_add(ht, size); | |
1440 | return NULL; | |
1441 | } | |
1442 | ||
1443 | if (!_cds_lfht_replace(ht, size, iter.node, iter.next, node)) | |
1444 | return iter.node; | |
1445 | } | |
48ed1c18 MD |
1446 | } |
1447 | ||
9357c415 MD |
1448 | int cds_lfht_replace(struct cds_lfht *ht, struct cds_lfht_iter *old_iter, |
1449 | struct cds_lfht_node *new_node) | |
1450 | { | |
1451 | unsigned long size; | |
1452 | ||
1453 | size = rcu_dereference(ht->t.size); | |
1454 | return _cds_lfht_replace(ht, size, old_iter->node, old_iter->next, | |
1455 | new_node); | |
1456 | } | |
1457 | ||
1458 | int cds_lfht_del(struct cds_lfht *ht, struct cds_lfht_iter *iter) | |
2ed95849 | 1459 | { |
4105056a | 1460 | unsigned long size; |
df44348d | 1461 | int ret; |
abc490a1 | 1462 | |
4105056a | 1463 | size = rcu_dereference(ht->t.size); |
9357c415 | 1464 | ret = _cds_lfht_del(ht, size, iter->node, 0); |
df44348d | 1465 | if (!ret) |
860d07e8 | 1466 | ht_count_del(ht, size); |
df44348d | 1467 | return ret; |
2ed95849 | 1468 | } |
ab7d5fc6 | 1469 | |
abc490a1 | 1470 | static |
14044b37 | 1471 | int cds_lfht_delete_dummy(struct cds_lfht *ht) |
674f7a69 | 1472 | { |
14044b37 MD |
1473 | struct cds_lfht_node *node; |
1474 | struct _cds_lfht_node *lookup; | |
4105056a | 1475 | unsigned long order, i, size; |
674f7a69 | 1476 | |
abc490a1 | 1477 | /* Check that the table is empty */ |
4105056a | 1478 | lookup = &ht->t.tbl[0]->nodes[0]; |
14044b37 | 1479 | node = (struct cds_lfht_node *) lookup; |
abc490a1 | 1480 | do { |
1b81fe1a MD |
1481 | node = clear_flag(node)->p.next; |
1482 | if (!is_dummy(node)) | |
abc490a1 | 1483 | return -EPERM; |
273399de | 1484 | assert(!is_removed(node)); |
bb7b2f26 | 1485 | } while (!is_end(node)); |
4105056a MD |
1486 | /* |
1487 | * size accessed without rcu_dereference because hash table is | |
1488 | * being destroyed. | |
1489 | */ | |
1490 | size = ht->t.size; | |
abc490a1 | 1491 | /* Internal sanity check: all nodes left should be dummy */ |
4105056a | 1492 | for (order = 0; order < get_count_order_ulong(size) + 1; order++) { |
24365af7 MD |
1493 | unsigned long len; |
1494 | ||
1495 | len = !order ? 1 : 1UL << (order - 1); | |
1496 | for (i = 0; i < len; i++) { | |
f0c29ed7 | 1497 | dbg_printf("delete order %lu i %lu hash %lu\n", |
24365af7 | 1498 | order, i, |
4105056a MD |
1499 | bit_reverse_ulong(ht->t.tbl[order]->nodes[i].reverse_hash)); |
1500 | assert(is_dummy(ht->t.tbl[order]->nodes[i].next)); | |
24365af7 | 1501 | } |
4105056a | 1502 | poison_free(ht->t.tbl[order]); |
674f7a69 | 1503 | } |
abc490a1 | 1504 | return 0; |
674f7a69 MD |
1505 | } |
1506 | ||
1507 | /* | |
1508 | * Should only be called when no more concurrent readers nor writers can | |
1509 | * possibly access the table. | |
1510 | */ | |
b7d619b0 | 1511 | int cds_lfht_destroy(struct cds_lfht *ht, pthread_attr_t **attr) |
674f7a69 | 1512 | { |
5e28c532 MD |
1513 | int ret; |
1514 | ||
848d4088 | 1515 | /* Wait for in-flight resize operations to complete */ |
24953e08 MD |
1516 | _CMM_STORE_SHARED(ht->in_progress_destroy, 1); |
1517 | cmm_smp_mb(); /* Store destroy before load resize */ | |
848d4088 MD |
1518 | while (uatomic_read(&ht->in_progress_resize)) |
1519 | poll(NULL, 0, 100); /* wait for 100ms */ | |
14044b37 | 1520 | ret = cds_lfht_delete_dummy(ht); |
abc490a1 MD |
1521 | if (ret) |
1522 | return ret; | |
df44348d | 1523 | free_per_cpu_items_count(ht->percpu_count); |
b7d619b0 MD |
1524 | if (attr) |
1525 | *attr = ht->resize_attr; | |
98808fb1 | 1526 | poison_free(ht); |
5e28c532 | 1527 | return ret; |
674f7a69 MD |
1528 | } |
1529 | ||
14044b37 | 1530 | void cds_lfht_count_nodes(struct cds_lfht *ht, |
d933dd0e | 1531 | long *approx_before, |
273399de | 1532 | unsigned long *count, |
973e5e1b | 1533 | unsigned long *removed, |
d933dd0e | 1534 | long *approx_after) |
273399de | 1535 | { |
14044b37 MD |
1536 | struct cds_lfht_node *node, *next; |
1537 | struct _cds_lfht_node *lookup; | |
24365af7 | 1538 | unsigned long nr_dummy = 0; |
273399de | 1539 | |
7ed7682f | 1540 | *approx_before = 0; |
973e5e1b MD |
1541 | if (nr_cpus_mask >= 0) { |
1542 | int i; | |
1543 | ||
1544 | for (i = 0; i < nr_cpus_mask + 1; i++) { | |
1545 | *approx_before += uatomic_read(&ht->percpu_count[i].add); | |
1546 | *approx_before -= uatomic_read(&ht->percpu_count[i].del); | |
1547 | } | |
1548 | } | |
1549 | ||
273399de MD |
1550 | *count = 0; |
1551 | *removed = 0; | |
1552 | ||
24365af7 | 1553 | /* Count non-dummy nodes in the table */ |
4105056a | 1554 | lookup = &ht->t.tbl[0]->nodes[0]; |
14044b37 | 1555 | node = (struct cds_lfht_node *) lookup; |
273399de | 1556 | do { |
cc4fcb10 | 1557 | next = rcu_dereference(node->p.next); |
b198f0fd | 1558 | if (is_removed(next)) { |
973e5e1b MD |
1559 | if (!is_dummy(next)) |
1560 | (*removed)++; | |
1561 | else | |
1562 | (nr_dummy)++; | |
1b81fe1a | 1563 | } else if (!is_dummy(next)) |
273399de | 1564 | (*count)++; |
24365af7 MD |
1565 | else |
1566 | (nr_dummy)++; | |
273399de | 1567 | node = clear_flag(next); |
bb7b2f26 | 1568 | } while (!is_end(node)); |
f0c29ed7 | 1569 | dbg_printf("number of dummy nodes: %lu\n", nr_dummy); |
7ed7682f | 1570 | *approx_after = 0; |
973e5e1b MD |
1571 | if (nr_cpus_mask >= 0) { |
1572 | int i; | |
1573 | ||
1574 | for (i = 0; i < nr_cpus_mask + 1; i++) { | |
1575 | *approx_after += uatomic_read(&ht->percpu_count[i].add); | |
1576 | *approx_after -= uatomic_read(&ht->percpu_count[i].del); | |
1577 | } | |
1578 | } | |
273399de MD |
1579 | } |
1580 | ||
1475579c | 1581 | /* called with resize mutex held */ |
abc490a1 | 1582 | static |
4105056a | 1583 | void _do_cds_lfht_grow(struct cds_lfht *ht, |
1475579c | 1584 | unsigned long old_size, unsigned long new_size) |
abc490a1 | 1585 | { |
1475579c | 1586 | unsigned long old_order, new_order; |
1475579c MD |
1587 | |
1588 | old_order = get_count_order_ulong(old_size) + 1; | |
1589 | new_order = get_count_order_ulong(new_size) + 1; | |
1a401918 LJ |
1590 | dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n", |
1591 | old_size, old_order, new_size, new_order); | |
1475579c | 1592 | assert(new_size > old_size); |
4105056a | 1593 | init_table(ht, old_order, new_order - old_order); |
abc490a1 MD |
1594 | } |
1595 | ||
1596 | /* called with resize mutex held */ | |
1597 | static | |
4105056a | 1598 | void _do_cds_lfht_shrink(struct cds_lfht *ht, |
1475579c | 1599 | unsigned long old_size, unsigned long new_size) |
464a1ec9 | 1600 | { |
1475579c | 1601 | unsigned long old_order, new_order; |
464a1ec9 | 1602 | |
cd95516d | 1603 | new_size = max(new_size, MIN_TABLE_SIZE); |
24365af7 | 1604 | old_order = get_count_order_ulong(old_size) + 1; |
24365af7 | 1605 | new_order = get_count_order_ulong(new_size) + 1; |
1a401918 LJ |
1606 | dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n", |
1607 | old_size, old_order, new_size, new_order); | |
1475579c | 1608 | assert(new_size < old_size); |
1475579c | 1609 | |
4105056a MD |
1610 | /* Remove and unlink all dummy nodes to remove. */ |
1611 | fini_table(ht, new_order, old_order - new_order); | |
464a1ec9 MD |
1612 | } |
1613 | ||
1475579c MD |
1614 | |
1615 | /* called with resize mutex held */ | |
1616 | static | |
1617 | void _do_cds_lfht_resize(struct cds_lfht *ht) | |
1618 | { | |
1619 | unsigned long new_size, old_size; | |
4105056a MD |
1620 | |
1621 | /* | |
1622 | * Resize table, re-do if the target size has changed under us. | |
1623 | */ | |
1624 | do { | |
d2be3620 MD |
1625 | assert(uatomic_read(&ht->in_progress_resize)); |
1626 | if (CMM_LOAD_SHARED(ht->in_progress_destroy)) | |
1627 | break; | |
4105056a MD |
1628 | ht->t.resize_initiated = 1; |
1629 | old_size = ht->t.size; | |
1630 | new_size = CMM_LOAD_SHARED(ht->t.resize_target); | |
1631 | if (old_size < new_size) | |
1632 | _do_cds_lfht_grow(ht, old_size, new_size); | |
1633 | else if (old_size > new_size) | |
1634 | _do_cds_lfht_shrink(ht, old_size, new_size); | |
1635 | ht->t.resize_initiated = 0; | |
1636 | /* write resize_initiated before read resize_target */ | |
1637 | cmm_smp_mb(); | |
4d676753 | 1638 | } while (ht->t.size != CMM_LOAD_SHARED(ht->t.resize_target)); |
1475579c MD |
1639 | } |
1640 | ||
abc490a1 | 1641 | static |
4105056a | 1642 | unsigned long resize_target_update(struct cds_lfht *ht, unsigned long size, |
f9830efd | 1643 | int growth_order) |
464a1ec9 | 1644 | { |
4105056a MD |
1645 | return _uatomic_max(&ht->t.resize_target, |
1646 | size << growth_order); | |
464a1ec9 MD |
1647 | } |
1648 | ||
1475579c | 1649 | static |
4105056a | 1650 | void resize_target_update_count(struct cds_lfht *ht, |
b8af5011 | 1651 | unsigned long count) |
1475579c | 1652 | { |
cd95516d | 1653 | count = max(count, MIN_TABLE_SIZE); |
4105056a | 1654 | uatomic_set(&ht->t.resize_target, count); |
1475579c MD |
1655 | } |
1656 | ||
1657 | void cds_lfht_resize(struct cds_lfht *ht, unsigned long new_size) | |
464a1ec9 | 1658 | { |
4105056a MD |
1659 | resize_target_update_count(ht, new_size); |
1660 | CMM_STORE_SHARED(ht->t.resize_initiated, 1); | |
5f511391 | 1661 | ht->cds_lfht_rcu_thread_offline(); |
1475579c MD |
1662 | pthread_mutex_lock(&ht->resize_mutex); |
1663 | _do_cds_lfht_resize(ht); | |
1664 | pthread_mutex_unlock(&ht->resize_mutex); | |
5f511391 | 1665 | ht->cds_lfht_rcu_thread_online(); |
abc490a1 | 1666 | } |
464a1ec9 | 1667 | |
abc490a1 MD |
1668 | static |
1669 | void do_resize_cb(struct rcu_head *head) | |
1670 | { | |
1671 | struct rcu_resize_work *work = | |
1672 | caa_container_of(head, struct rcu_resize_work, head); | |
14044b37 | 1673 | struct cds_lfht *ht = work->ht; |
abc490a1 | 1674 | |
5f511391 | 1675 | ht->cds_lfht_rcu_thread_offline(); |
abc490a1 | 1676 | pthread_mutex_lock(&ht->resize_mutex); |
14044b37 | 1677 | _do_cds_lfht_resize(ht); |
abc490a1 | 1678 | pthread_mutex_unlock(&ht->resize_mutex); |
5f511391 | 1679 | ht->cds_lfht_rcu_thread_online(); |
98808fb1 | 1680 | poison_free(work); |
848d4088 MD |
1681 | cmm_smp_mb(); /* finish resize before decrement */ |
1682 | uatomic_dec(&ht->in_progress_resize); | |
464a1ec9 MD |
1683 | } |
1684 | ||
abc490a1 | 1685 | static |
4105056a | 1686 | void cds_lfht_resize_lazy(struct cds_lfht *ht, unsigned long size, int growth) |
ab7d5fc6 | 1687 | { |
abc490a1 | 1688 | struct rcu_resize_work *work; |
f9830efd | 1689 | unsigned long target_size; |
abc490a1 | 1690 | |
4105056a MD |
1691 | target_size = resize_target_update(ht, size, growth); |
1692 | /* Store resize_target before read resize_initiated */ | |
1693 | cmm_smp_mb(); | |
1694 | if (!CMM_LOAD_SHARED(ht->t.resize_initiated) && size < target_size) { | |
848d4088 | 1695 | uatomic_inc(&ht->in_progress_resize); |
59290e9d | 1696 | cmm_smp_mb(); /* increment resize count before load destroy */ |
ed35e6d8 MD |
1697 | if (CMM_LOAD_SHARED(ht->in_progress_destroy)) { |
1698 | uatomic_dec(&ht->in_progress_resize); | |
59290e9d | 1699 | return; |
ed35e6d8 | 1700 | } |
f9830efd MD |
1701 | work = malloc(sizeof(*work)); |
1702 | work->ht = ht; | |
14044b37 | 1703 | ht->cds_lfht_call_rcu(&work->head, do_resize_cb); |
4105056a | 1704 | CMM_STORE_SHARED(ht->t.resize_initiated, 1); |
f9830efd | 1705 | } |
ab7d5fc6 | 1706 | } |
3171717f | 1707 | |
f8994aee MD |
1708 | #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) |
1709 | ||
3171717f | 1710 | static |
4105056a | 1711 | void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size, |
3171717f MD |
1712 | unsigned long count) |
1713 | { | |
1714 | struct rcu_resize_work *work; | |
3171717f | 1715 | |
b8af5011 MD |
1716 | if (!(ht->flags & CDS_LFHT_AUTO_RESIZE)) |
1717 | return; | |
4105056a MD |
1718 | resize_target_update_count(ht, count); |
1719 | /* Store resize_target before read resize_initiated */ | |
1720 | cmm_smp_mb(); | |
1721 | if (!CMM_LOAD_SHARED(ht->t.resize_initiated)) { | |
3171717f | 1722 | uatomic_inc(&ht->in_progress_resize); |
59290e9d | 1723 | cmm_smp_mb(); /* increment resize count before load destroy */ |
ed35e6d8 MD |
1724 | if (CMM_LOAD_SHARED(ht->in_progress_destroy)) { |
1725 | uatomic_dec(&ht->in_progress_resize); | |
59290e9d | 1726 | return; |
ed35e6d8 | 1727 | } |
3171717f MD |
1728 | work = malloc(sizeof(*work)); |
1729 | work->ht = ht; | |
1730 | ht->cds_lfht_call_rcu(&work->head, do_resize_cb); | |
4105056a | 1731 | CMM_STORE_SHARED(ht->t.resize_initiated, 1); |
3171717f MD |
1732 | } |
1733 | } | |
f8994aee MD |
1734 | |
1735 | #endif |