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 | * | | |
123 | * 1 | 1 001 100 <- | |
124 | * | | | | |
125 | * 2 | | 2 010 010 | | |
126 | * | | | 3 011 110 <- | | |
127 | * | | | | | | | |
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 | |
df44348d | 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> | |
674f7a69 | 147 | #include <urcu/jhash.h> |
a42cc659 | 148 | #include <urcu/compiler.h> |
abc490a1 | 149 | #include <urcu/rculfhash.h> |
5e28c532 | 150 | #include <stdio.h> |
464a1ec9 | 151 | #include <pthread.h> |
44395fb7 | 152 | |
f9830efd | 153 | #ifdef DEBUG |
f0c29ed7 | 154 | #define dbg_printf(fmt, args...) printf("[debug rculfhash] " fmt, ## args) |
f9830efd | 155 | #else |
e753ff5a | 156 | #define dbg_printf(fmt, args...) |
f9830efd MD |
157 | #endif |
158 | ||
f8994aee MD |
159 | /* |
160 | * Per-CPU split-counters lazily update the global counter each 1024 | |
161 | * addition/removal. It automatically keeps track of resize required. | |
162 | * We use the bucket length as indicator for need to expand for small | |
163 | * tables and machines lacking per-cpu data suppport. | |
164 | */ | |
165 | #define COUNT_COMMIT_ORDER 10 | |
6ea6bc67 MD |
166 | #define CHAIN_LEN_TARGET 1 |
167 | #define CHAIN_LEN_RESIZE_THRESHOLD 3 | |
2ed95849 | 168 | |
abc490a1 MD |
169 | #ifndef max |
170 | #define max(a, b) ((a) > (b) ? (a) : (b)) | |
171 | #endif | |
2ed95849 | 172 | |
d95bd160 MD |
173 | /* |
174 | * The removed flag needs to be updated atomically with the pointer. | |
175 | * The dummy flag does not require to be updated atomically with the | |
176 | * pointer, but it is added as a pointer low bit flag to save space. | |
177 | */ | |
d37166c6 | 178 | #define REMOVED_FLAG (1UL << 0) |
f5596c94 MD |
179 | #define DUMMY_FLAG (1UL << 1) |
180 | #define FLAGS_MASK ((1UL << 2) - 1) | |
d37166c6 | 181 | |
df44348d | 182 | struct ht_items_count { |
3171717f | 183 | unsigned long add, remove; |
df44348d MD |
184 | } __attribute__((aligned(CAA_CACHE_LINE_SIZE))); |
185 | ||
1475579c MD |
186 | struct rcu_level { |
187 | struct rcu_head head; | |
188 | struct _cds_lfht_node nodes[0]; | |
189 | }; | |
190 | ||
395270b6 | 191 | struct rcu_table { |
abc490a1 | 192 | unsigned long size; /* always a power of 2 */ |
f9830efd | 193 | unsigned long resize_target; |
11519af6 | 194 | int resize_initiated; |
abc490a1 | 195 | struct rcu_head head; |
1475579c | 196 | struct rcu_level *tbl[0]; |
395270b6 MD |
197 | }; |
198 | ||
14044b37 | 199 | struct cds_lfht { |
395270b6 | 200 | struct rcu_table *t; /* shared */ |
14044b37 MD |
201 | cds_lfht_hash_fct hash_fct; |
202 | cds_lfht_compare_fct compare_fct; | |
732ad076 | 203 | unsigned long hash_seed; |
464a1ec9 | 204 | pthread_mutex_t resize_mutex; /* resize mutex: add/del mutex */ |
33c7c748 | 205 | unsigned int in_progress_resize, in_progress_destroy; |
14044b37 | 206 | void (*cds_lfht_call_rcu)(struct rcu_head *head, |
abc490a1 | 207 | void (*func)(struct rcu_head *head)); |
1475579c | 208 | void (*cds_lfht_synchronize_rcu)(void); |
df44348d MD |
209 | unsigned long count; /* global approximate item count */ |
210 | struct ht_items_count *percpu_count; /* per-cpu item count */ | |
2ed95849 MD |
211 | }; |
212 | ||
abc490a1 MD |
213 | struct rcu_resize_work { |
214 | struct rcu_head head; | |
14044b37 | 215 | struct cds_lfht *ht; |
abc490a1 | 216 | }; |
2ed95849 | 217 | |
abc490a1 MD |
218 | /* |
219 | * Algorithm to reverse bits in a word by lookup table, extended to | |
220 | * 64-bit words. | |
f9830efd | 221 | * Source: |
abc490a1 | 222 | * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable |
f9830efd | 223 | * Originally from Public Domain. |
abc490a1 MD |
224 | */ |
225 | ||
226 | static const uint8_t BitReverseTable256[256] = | |
2ed95849 | 227 | { |
abc490a1 MD |
228 | #define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64 |
229 | #define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16) | |
230 | #define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 ) | |
231 | R6(0), R6(2), R6(1), R6(3) | |
232 | }; | |
233 | #undef R2 | |
234 | #undef R4 | |
235 | #undef R6 | |
2ed95849 | 236 | |
abc490a1 MD |
237 | static |
238 | uint8_t bit_reverse_u8(uint8_t v) | |
239 | { | |
240 | return BitReverseTable256[v]; | |
241 | } | |
ab7d5fc6 | 242 | |
abc490a1 MD |
243 | static __attribute__((unused)) |
244 | uint32_t bit_reverse_u32(uint32_t v) | |
245 | { | |
246 | return ((uint32_t) bit_reverse_u8(v) << 24) | | |
247 | ((uint32_t) bit_reverse_u8(v >> 8) << 16) | | |
248 | ((uint32_t) bit_reverse_u8(v >> 16) << 8) | | |
249 | ((uint32_t) bit_reverse_u8(v >> 24)); | |
2ed95849 MD |
250 | } |
251 | ||
abc490a1 MD |
252 | static __attribute__((unused)) |
253 | uint64_t bit_reverse_u64(uint64_t v) | |
2ed95849 | 254 | { |
abc490a1 MD |
255 | return ((uint64_t) bit_reverse_u8(v) << 56) | |
256 | ((uint64_t) bit_reverse_u8(v >> 8) << 48) | | |
257 | ((uint64_t) bit_reverse_u8(v >> 16) << 40) | | |
258 | ((uint64_t) bit_reverse_u8(v >> 24) << 32) | | |
259 | ((uint64_t) bit_reverse_u8(v >> 32) << 24) | | |
260 | ((uint64_t) bit_reverse_u8(v >> 40) << 16) | | |
261 | ((uint64_t) bit_reverse_u8(v >> 48) << 8) | | |
262 | ((uint64_t) bit_reverse_u8(v >> 56)); | |
263 | } | |
264 | ||
265 | static | |
266 | unsigned long bit_reverse_ulong(unsigned long v) | |
267 | { | |
268 | #if (CAA_BITS_PER_LONG == 32) | |
269 | return bit_reverse_u32(v); | |
270 | #else | |
271 | return bit_reverse_u64(v); | |
272 | #endif | |
273 | } | |
274 | ||
f9830efd | 275 | /* |
24365af7 MD |
276 | * fls: returns the position of the most significant bit. |
277 | * Returns 0 if no bit is set, else returns the position of the most | |
278 | * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit). | |
f9830efd | 279 | */ |
24365af7 MD |
280 | #if defined(__i386) || defined(__x86_64) |
281 | static inline | |
282 | unsigned int fls_u32(uint32_t x) | |
f9830efd | 283 | { |
24365af7 MD |
284 | int r; |
285 | ||
286 | asm("bsrl %1,%0\n\t" | |
287 | "jnz 1f\n\t" | |
288 | "movl $-1,%0\n\t" | |
289 | "1:\n\t" | |
290 | : "=r" (r) : "rm" (x)); | |
291 | return r + 1; | |
292 | } | |
293 | #define HAS_FLS_U32 | |
294 | #endif | |
295 | ||
296 | #if defined(__x86_64) | |
297 | static inline | |
298 | unsigned int fls_u64(uint64_t x) | |
299 | { | |
300 | long r; | |
301 | ||
302 | asm("bsrq %1,%0\n\t" | |
303 | "jnz 1f\n\t" | |
304 | "movq $-1,%0\n\t" | |
305 | "1:\n\t" | |
306 | : "=r" (r) : "rm" (x)); | |
307 | return r + 1; | |
308 | } | |
309 | #define HAS_FLS_U64 | |
310 | #endif | |
311 | ||
312 | #ifndef HAS_FLS_U64 | |
313 | static __attribute__((unused)) | |
314 | unsigned int fls_u64(uint64_t x) | |
315 | { | |
316 | unsigned int r = 64; | |
317 | ||
318 | if (!x) | |
319 | return 0; | |
320 | ||
321 | if (!(x & 0xFFFFFFFF00000000ULL)) { | |
322 | x <<= 32; | |
323 | r -= 32; | |
324 | } | |
325 | if (!(x & 0xFFFF000000000000ULL)) { | |
326 | x <<= 16; | |
327 | r -= 16; | |
328 | } | |
329 | if (!(x & 0xFF00000000000000ULL)) { | |
330 | x <<= 8; | |
331 | r -= 8; | |
332 | } | |
333 | if (!(x & 0xF000000000000000ULL)) { | |
334 | x <<= 4; | |
335 | r -= 4; | |
336 | } | |
337 | if (!(x & 0xC000000000000000ULL)) { | |
338 | x <<= 2; | |
339 | r -= 2; | |
340 | } | |
341 | if (!(x & 0x8000000000000000ULL)) { | |
342 | x <<= 1; | |
343 | r -= 1; | |
344 | } | |
345 | return r; | |
346 | } | |
347 | #endif | |
348 | ||
349 | #ifndef HAS_FLS_U32 | |
350 | static __attribute__((unused)) | |
351 | unsigned int fls_u32(uint32_t x) | |
352 | { | |
353 | unsigned int r = 32; | |
f9830efd | 354 | |
24365af7 MD |
355 | if (!x) |
356 | return 0; | |
357 | if (!(x & 0xFFFF0000U)) { | |
358 | x <<= 16; | |
359 | r -= 16; | |
360 | } | |
361 | if (!(x & 0xFF000000U)) { | |
362 | x <<= 8; | |
363 | r -= 8; | |
364 | } | |
365 | if (!(x & 0xF0000000U)) { | |
366 | x <<= 4; | |
367 | r -= 4; | |
368 | } | |
369 | if (!(x & 0xC0000000U)) { | |
370 | x <<= 2; | |
371 | r -= 2; | |
372 | } | |
373 | if (!(x & 0x80000000U)) { | |
374 | x <<= 1; | |
375 | r -= 1; | |
376 | } | |
377 | return r; | |
378 | } | |
379 | #endif | |
380 | ||
381 | unsigned int fls_ulong(unsigned long x) | |
f9830efd | 382 | { |
24365af7 MD |
383 | #if (CAA_BITS_PER_lONG == 32) |
384 | return fls_u32(x); | |
385 | #else | |
386 | return fls_u64(x); | |
387 | #endif | |
388 | } | |
f9830efd | 389 | |
24365af7 MD |
390 | int get_count_order_u32(uint32_t x) |
391 | { | |
392 | int order; | |
393 | ||
394 | order = fls_u32(x) - 1; | |
395 | if (x & (x - 1)) | |
396 | order++; | |
397 | return order; | |
398 | } | |
399 | ||
400 | int get_count_order_ulong(unsigned long x) | |
401 | { | |
402 | int order; | |
403 | ||
404 | order = fls_ulong(x) - 1; | |
405 | if (x & (x - 1)) | |
406 | order++; | |
407 | return order; | |
f9830efd MD |
408 | } |
409 | ||
410 | static | |
14044b37 | 411 | void cds_lfht_resize_lazy(struct cds_lfht *ht, struct rcu_table *t, int growth); |
f9830efd | 412 | |
df44348d MD |
413 | /* |
414 | * If the sched_getcpu() and sysconf(_SC_NPROCESSORS_CONF) calls are | |
415 | * available, then we support hash table item accounting. | |
416 | * In the unfortunate event the number of CPUs reported would be | |
417 | * inaccurate, we use modulo arithmetic on the number of CPUs we got. | |
418 | */ | |
df44348d MD |
419 | #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) |
420 | ||
f8994aee MD |
421 | static |
422 | void cds_lfht_resize_lazy_count(struct cds_lfht *ht, struct rcu_table *t, | |
423 | unsigned long count); | |
424 | ||
df44348d MD |
425 | static long nr_cpus_mask = -1; |
426 | ||
427 | static | |
428 | struct ht_items_count *alloc_per_cpu_items_count(void) | |
429 | { | |
430 | struct ht_items_count *count; | |
431 | ||
432 | switch (nr_cpus_mask) { | |
433 | case -2: | |
434 | return NULL; | |
435 | case -1: | |
436 | { | |
437 | long maxcpus; | |
438 | ||
439 | maxcpus = sysconf(_SC_NPROCESSORS_CONF); | |
440 | if (maxcpus <= 0) { | |
441 | nr_cpus_mask = -2; | |
442 | return NULL; | |
443 | } | |
444 | /* | |
445 | * round up number of CPUs to next power of two, so we | |
446 | * can use & for modulo. | |
447 | */ | |
448 | maxcpus = 1UL << get_count_order_ulong(maxcpus); | |
449 | nr_cpus_mask = maxcpus - 1; | |
450 | } | |
451 | /* Fall-through */ | |
452 | default: | |
453 | return calloc(nr_cpus_mask + 1, sizeof(*count)); | |
454 | } | |
455 | } | |
456 | ||
457 | static | |
458 | void free_per_cpu_items_count(struct ht_items_count *count) | |
459 | { | |
460 | free(count); | |
461 | } | |
462 | ||
463 | static | |
464 | int ht_get_cpu(void) | |
465 | { | |
466 | int cpu; | |
467 | ||
468 | assert(nr_cpus_mask >= 0); | |
469 | cpu = sched_getcpu(); | |
470 | if (unlikely(cpu < 0)) | |
471 | return cpu; | |
472 | else | |
473 | return cpu & nr_cpus_mask; | |
474 | } | |
475 | ||
476 | static | |
3171717f | 477 | void ht_count_add(struct cds_lfht *ht, struct rcu_table *t) |
df44348d | 478 | { |
3171717f | 479 | unsigned long percpu_count; |
df44348d MD |
480 | int cpu; |
481 | ||
482 | if (unlikely(!ht->percpu_count)) | |
3171717f | 483 | return; |
df44348d MD |
484 | cpu = ht_get_cpu(); |
485 | if (unlikely(cpu < 0)) | |
3171717f MD |
486 | return; |
487 | percpu_count = uatomic_add_return(&ht->percpu_count[cpu].add, 1); | |
df44348d MD |
488 | if (unlikely(!(percpu_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) { |
489 | unsigned long count; | |
490 | ||
491 | dbg_printf("add percpu %lu\n", percpu_count); | |
492 | count = uatomic_add_return(&ht->count, | |
493 | 1UL << COUNT_COMMIT_ORDER); | |
494 | /* If power of 2 */ | |
495 | if (!(count & (count - 1))) { | |
f8994aee MD |
496 | if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) |
497 | < t->size) | |
498 | return; | |
499 | dbg_printf("add set global %lu\n", count); | |
500 | cds_lfht_resize_lazy_count(ht, t, | |
6ea6bc67 | 501 | count >> (CHAIN_LEN_TARGET - 1)); |
df44348d MD |
502 | } |
503 | } | |
504 | } | |
505 | ||
506 | static | |
507 | void ht_count_remove(struct cds_lfht *ht, struct rcu_table *t) | |
508 | { | |
509 | unsigned long percpu_count; | |
3171717f | 510 | int cpu; |
df44348d | 511 | |
3171717f MD |
512 | if (unlikely(!ht->percpu_count)) |
513 | return; | |
514 | cpu = ht_get_cpu(); | |
515 | if (unlikely(cpu < 0)) | |
516 | return; | |
517 | percpu_count = uatomic_add_return(&ht->percpu_count[cpu].remove, -1); | |
df44348d MD |
518 | if (unlikely(!(percpu_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) { |
519 | unsigned long count; | |
520 | ||
521 | dbg_printf("remove percpu %lu\n", percpu_count); | |
522 | count = uatomic_add_return(&ht->count, | |
3171717f | 523 | -(1UL << COUNT_COMMIT_ORDER)); |
df44348d MD |
524 | /* If power of 2 */ |
525 | if (!(count & (count - 1))) { | |
f8994aee MD |
526 | if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) |
527 | >= t->size) | |
528 | return; | |
529 | dbg_printf("remove set global %lu\n", count); | |
530 | cds_lfht_resize_lazy_count(ht, t, | |
6ea6bc67 | 531 | count >> (CHAIN_LEN_TARGET - 1)); |
df44348d MD |
532 | } |
533 | } | |
534 | } | |
535 | ||
536 | #else /* #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */ | |
537 | ||
538 | static const long nr_cpus_mask = -1; | |
539 | ||
540 | static | |
541 | struct ht_items_count *alloc_per_cpu_items_count(void) | |
542 | { | |
543 | return NULL; | |
544 | } | |
545 | ||
546 | static | |
547 | void free_per_cpu_items_count(struct ht_items_count *count) | |
548 | { | |
549 | } | |
550 | ||
551 | static | |
f8994aee | 552 | void ht_count_add(struct cds_lfht *ht, struct rcu_table *t) |
df44348d MD |
553 | { |
554 | } | |
555 | ||
556 | static | |
f8994aee | 557 | void ht_count_remove(struct cds_lfht *ht, struct rcu_table *t) |
df44348d MD |
558 | { |
559 | } | |
560 | ||
561 | #endif /* #else #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */ | |
562 | ||
563 | ||
f9830efd | 564 | static |
14044b37 | 565 | void check_resize(struct cds_lfht *ht, struct rcu_table *t, |
f9830efd MD |
566 | uint32_t chain_len) |
567 | { | |
f8994aee MD |
568 | unsigned long count; |
569 | ||
570 | count = uatomic_read(&ht->count); | |
571 | /* | |
572 | * Use bucket-local length for small table expand and for | |
573 | * environments lacking per-cpu data support. | |
574 | */ | |
575 | if (count >= (1UL << COUNT_COMMIT_ORDER)) | |
576 | return; | |
24365af7 | 577 | if (chain_len > 100) |
f0c29ed7 | 578 | dbg_printf("WARNING: large chain length: %u.\n", |
24365af7 | 579 | chain_len); |
3390d470 | 580 | if (chain_len >= CHAIN_LEN_RESIZE_THRESHOLD) |
14044b37 | 581 | cds_lfht_resize_lazy(ht, t, |
01370f0b | 582 | get_count_order_u32(chain_len - (CHAIN_LEN_TARGET - 1))); |
f9830efd MD |
583 | } |
584 | ||
abc490a1 | 585 | static |
14044b37 | 586 | struct cds_lfht_node *clear_flag(struct cds_lfht_node *node) |
abc490a1 | 587 | { |
14044b37 | 588 | return (struct cds_lfht_node *) (((unsigned long) node) & ~FLAGS_MASK); |
abc490a1 MD |
589 | } |
590 | ||
591 | static | |
14044b37 | 592 | int is_removed(struct cds_lfht_node *node) |
abc490a1 | 593 | { |
d37166c6 | 594 | return ((unsigned long) node) & REMOVED_FLAG; |
abc490a1 MD |
595 | } |
596 | ||
597 | static | |
14044b37 | 598 | struct cds_lfht_node *flag_removed(struct cds_lfht_node *node) |
abc490a1 | 599 | { |
14044b37 | 600 | return (struct cds_lfht_node *) (((unsigned long) node) | REMOVED_FLAG); |
abc490a1 MD |
601 | } |
602 | ||
f5596c94 | 603 | static |
14044b37 | 604 | int is_dummy(struct cds_lfht_node *node) |
f5596c94 MD |
605 | { |
606 | return ((unsigned long) node) & DUMMY_FLAG; | |
607 | } | |
608 | ||
609 | static | |
14044b37 | 610 | struct cds_lfht_node *flag_dummy(struct cds_lfht_node *node) |
f5596c94 | 611 | { |
14044b37 | 612 | return (struct cds_lfht_node *) (((unsigned long) node) | DUMMY_FLAG); |
f5596c94 MD |
613 | } |
614 | ||
abc490a1 | 615 | static |
f9830efd | 616 | unsigned long _uatomic_max(unsigned long *ptr, unsigned long v) |
abc490a1 MD |
617 | { |
618 | unsigned long old1, old2; | |
619 | ||
620 | old1 = uatomic_read(ptr); | |
621 | do { | |
622 | old2 = old1; | |
623 | if (old2 >= v) | |
f9830efd | 624 | return old2; |
abc490a1 | 625 | } while ((old1 = uatomic_cmpxchg(ptr, old2, v)) != old2); |
f9830efd | 626 | return v; |
abc490a1 MD |
627 | } |
628 | ||
1475579c MD |
629 | static |
630 | void cds_lfht_free_table_cb(struct rcu_head *head) | |
631 | { | |
632 | struct rcu_table *t = | |
633 | caa_container_of(head, struct rcu_table, head); | |
634 | free(t); | |
635 | } | |
636 | ||
637 | static | |
638 | void cds_lfht_free_level(struct rcu_head *head) | |
639 | { | |
640 | struct rcu_level *l = | |
641 | caa_container_of(head, struct rcu_level, head); | |
642 | free(l); | |
643 | } | |
644 | ||
273399de MD |
645 | /* |
646 | * Remove all logically deleted nodes from a bucket up to a certain node key. | |
647 | */ | |
648 | static | |
14044b37 | 649 | void _cds_lfht_gc_bucket(struct cds_lfht_node *dummy, struct cds_lfht_node *node) |
273399de | 650 | { |
14044b37 | 651 | struct cds_lfht_node *iter_prev, *iter, *next, *new_next; |
273399de MD |
652 | |
653 | for (;;) { | |
654 | iter_prev = dummy; | |
655 | /* We can always skip the dummy node initially */ | |
cc4fcb10 MD |
656 | iter = rcu_dereference(iter_prev->p.next); |
657 | assert(iter_prev->p.reverse_hash <= node->p.reverse_hash); | |
273399de | 658 | for (;;) { |
a2974903 | 659 | if (unlikely(!clear_flag(iter))) |
479c8a32 | 660 | return; |
76412f24 | 661 | if (likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash)) |
273399de | 662 | return; |
cc4fcb10 | 663 | next = rcu_dereference(clear_flag(iter)->p.next); |
76412f24 | 664 | if (likely(is_removed(next))) |
273399de | 665 | break; |
b453eae1 | 666 | iter_prev = clear_flag(iter); |
273399de MD |
667 | iter = next; |
668 | } | |
669 | assert(!is_removed(iter)); | |
f5596c94 MD |
670 | if (is_dummy(iter)) |
671 | new_next = flag_dummy(clear_flag(next)); | |
672 | else | |
673 | new_next = clear_flag(next); | |
674 | (void) uatomic_cmpxchg(&iter_prev->p.next, iter, new_next); | |
273399de MD |
675 | } |
676 | } | |
677 | ||
abc490a1 | 678 | static |
14044b37 MD |
679 | struct cds_lfht_node *_cds_lfht_add(struct cds_lfht *ht, struct rcu_table *t, |
680 | struct cds_lfht_node *node, int unique, int dummy) | |
abc490a1 | 681 | { |
14044b37 | 682 | struct cds_lfht_node *iter_prev, *iter, *next, *new_node, *new_next, |
f5596c94 | 683 | *dummy_node; |
14044b37 | 684 | struct _cds_lfht_node *lookup; |
24365af7 | 685 | unsigned long hash, index, order; |
abc490a1 | 686 | |
18117871 | 687 | if (!t->size) { |
f5596c94 MD |
688 | assert(dummy); |
689 | node->p.next = flag_dummy(NULL); | |
18117871 MD |
690 | return node; /* Initial first add (head) */ |
691 | } | |
cc4fcb10 | 692 | hash = bit_reverse_ulong(node->p.reverse_hash); |
abc490a1 | 693 | for (;;) { |
f9830efd | 694 | uint32_t chain_len = 0; |
abc490a1 | 695 | |
11519af6 MD |
696 | /* |
697 | * iter_prev points to the non-removed node prior to the | |
698 | * insert location. | |
11519af6 | 699 | */ |
24365af7 MD |
700 | index = hash & (t->size - 1); |
701 | order = get_count_order_ulong(index + 1); | |
1475579c | 702 | lookup = &t->tbl[order]->nodes[index & ((1UL << (order - 1)) - 1)]; |
14044b37 | 703 | iter_prev = (struct cds_lfht_node *) lookup; |
11519af6 | 704 | /* We can always skip the dummy node initially */ |
cc4fcb10 MD |
705 | iter = rcu_dereference(iter_prev->p.next); |
706 | assert(iter_prev->p.reverse_hash <= node->p.reverse_hash); | |
abc490a1 | 707 | for (;;) { |
a2974903 | 708 | if (unlikely(!clear_flag(iter))) |
273399de | 709 | goto insert; |
76412f24 | 710 | if (likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash)) |
273399de | 711 | goto insert; |
cc4fcb10 | 712 | next = rcu_dereference(clear_flag(iter)->p.next); |
76412f24 | 713 | if (unlikely(is_removed(next))) |
9dba85be | 714 | goto gc_node; |
e43f23f8 | 715 | if (unique |
1b81fe1a | 716 | && !is_dummy(next) |
e43f23f8 MD |
717 | && !ht->compare_fct(node->key, node->key_len, |
718 | clear_flag(iter)->key, | |
719 | clear_flag(iter)->key_len)) | |
18117871 | 720 | return clear_flag(iter); |
11519af6 | 721 | /* Only account for identical reverse hash once */ |
24365af7 MD |
722 | if (iter_prev->p.reverse_hash != clear_flag(iter)->p.reverse_hash |
723 | && !is_dummy(next)) | |
11519af6 MD |
724 | check_resize(ht, t, ++chain_len); |
725 | iter_prev = clear_flag(iter); | |
273399de | 726 | iter = next; |
abc490a1 | 727 | } |
273399de | 728 | insert: |
7ec59d3b | 729 | assert(node != clear_flag(iter)); |
11519af6 | 730 | assert(!is_removed(iter_prev)); |
f000907d | 731 | assert(iter_prev != node); |
f5596c94 | 732 | if (!dummy) |
1b81fe1a | 733 | node->p.next = clear_flag(iter); |
f5596c94 | 734 | else |
1b81fe1a | 735 | node->p.next = flag_dummy(clear_flag(iter)); |
f5596c94 MD |
736 | if (is_dummy(iter)) |
737 | new_node = flag_dummy(node); | |
738 | else | |
739 | new_node = node; | |
cc4fcb10 | 740 | if (uatomic_cmpxchg(&iter_prev->p.next, iter, |
f5596c94 | 741 | new_node) != iter) |
273399de | 742 | continue; /* retry */ |
11519af6 | 743 | else |
273399de | 744 | goto gc_end; |
9dba85be MD |
745 | gc_node: |
746 | assert(!is_removed(iter)); | |
f5596c94 MD |
747 | if (is_dummy(iter)) |
748 | new_next = flag_dummy(clear_flag(next)); | |
749 | else | |
750 | new_next = clear_flag(next); | |
751 | (void) uatomic_cmpxchg(&iter_prev->p.next, iter, new_next); | |
273399de | 752 | /* retry */ |
464a1ec9 | 753 | } |
273399de MD |
754 | gc_end: |
755 | /* Garbage collect logically removed nodes in the bucket */ | |
24365af7 MD |
756 | index = hash & (t->size - 1); |
757 | order = get_count_order_ulong(index + 1); | |
1475579c | 758 | lookup = &t->tbl[order]->nodes[index & ((1UL << (order - 1)) - 1)]; |
14044b37 MD |
759 | dummy_node = (struct cds_lfht_node *) lookup; |
760 | _cds_lfht_gc_bucket(dummy_node, node); | |
18117871 | 761 | return node; |
abc490a1 | 762 | } |
464a1ec9 | 763 | |
abc490a1 | 764 | static |
14044b37 | 765 | int _cds_lfht_remove(struct cds_lfht *ht, struct rcu_table *t, |
1475579c | 766 | struct cds_lfht_node *node, int dummy_removal) |
abc490a1 | 767 | { |
14044b37 MD |
768 | struct cds_lfht_node *dummy, *next, *old; |
769 | struct _cds_lfht_node *lookup; | |
abc490a1 | 770 | int flagged = 0; |
24365af7 | 771 | unsigned long hash, index, order; |
5e28c532 | 772 | |
7ec59d3b | 773 | /* logically delete the node */ |
cc4fcb10 | 774 | old = rcu_dereference(node->p.next); |
7ec59d3b MD |
775 | do { |
776 | next = old; | |
76412f24 | 777 | if (unlikely(is_removed(next))) |
7ec59d3b | 778 | goto end; |
1475579c MD |
779 | if (dummy_removal) |
780 | assert(is_dummy(next)); | |
781 | else | |
782 | assert(!is_dummy(next)); | |
cc4fcb10 | 783 | old = uatomic_cmpxchg(&node->p.next, next, |
7ec59d3b MD |
784 | flag_removed(next)); |
785 | } while (old != next); | |
786 | ||
787 | /* We performed the (logical) deletion. */ | |
788 | flagged = 1; | |
789 | ||
1475579c MD |
790 | if (dummy_removal) |
791 | node = clear_flag(node); | |
792 | ||
7ec59d3b MD |
793 | /* |
794 | * Ensure that the node is not visible to readers anymore: lookup for | |
273399de MD |
795 | * the node, and remove it (along with any other logically removed node) |
796 | * if found. | |
11519af6 | 797 | */ |
cc4fcb10 | 798 | hash = bit_reverse_ulong(node->p.reverse_hash); |
1475579c MD |
799 | /* |
800 | * When removing a dummy node, we need to consider the lower | |
801 | * order table, so we don't end up looking up the dummy nodes we | |
802 | * are currently removing. | |
803 | */ | |
804 | ||
805 | if (dummy_removal) | |
dd51e0ad | 806 | index = hash & ((t->size == 1) ? 0 : (t->size >> 1) - 1); |
1475579c MD |
807 | else |
808 | index = hash & (t->size - 1); | |
24365af7 | 809 | order = get_count_order_ulong(index + 1); |
1475579c | 810 | lookup = &t->tbl[order]->nodes[index & ((1UL << (order - 1)) - 1)]; |
14044b37 MD |
811 | dummy = (struct cds_lfht_node *) lookup; |
812 | _cds_lfht_gc_bucket(dummy, node); | |
2ed95849 | 813 | end: |
11519af6 MD |
814 | /* |
815 | * Only the flagging action indicated that we (and no other) | |
816 | * removed the node from the hash. | |
817 | */ | |
7ec59d3b | 818 | if (flagged) { |
cc4fcb10 | 819 | assert(is_removed(rcu_dereference(node->p.next))); |
11519af6 | 820 | return 0; |
7ec59d3b | 821 | } else |
11519af6 | 822 | return -ENOENT; |
abc490a1 | 823 | } |
2ed95849 | 824 | |
abc490a1 | 825 | static |
14044b37 | 826 | void init_table(struct cds_lfht *ht, struct rcu_table *t, |
24365af7 MD |
827 | unsigned long first_order, unsigned long len_order) |
828 | { | |
829 | unsigned long i, end_order; | |
830 | ||
f0c29ed7 | 831 | dbg_printf("init table: first_order %lu end_order %lu\n", |
24365af7 MD |
832 | first_order, first_order + len_order); |
833 | end_order = first_order + len_order; | |
834 | t->size = !first_order ? 0 : (1UL << (first_order - 1)); | |
835 | for (i = first_order; i < end_order; i++) { | |
836 | unsigned long j, len; | |
837 | ||
838 | len = !i ? 1 : 1UL << (i - 1); | |
f0c29ed7 | 839 | dbg_printf("init order %lu len: %lu\n", i, len); |
1475579c MD |
840 | t->tbl[i] = calloc(1, sizeof(struct rcu_level) |
841 | + (len * sizeof(struct _cds_lfht_node))); | |
24365af7 | 842 | for (j = 0; j < len; j++) { |
1475579c MD |
843 | struct cds_lfht_node *new_node = |
844 | (struct cds_lfht_node *) &t->tbl[i]->nodes[j]; | |
845 | ||
f0c29ed7 | 846 | dbg_printf("init entry: i %lu j %lu hash %lu\n", |
24365af7 | 847 | i, j, !i ? 0 : (1UL << (i - 1)) + j); |
24365af7 MD |
848 | new_node->p.reverse_hash = |
849 | bit_reverse_ulong(!i ? 0 : (1UL << (i - 1)) + j); | |
14044b37 | 850 | (void) _cds_lfht_add(ht, t, new_node, 0, 1); |
33c7c748 MD |
851 | if (CMM_LOAD_SHARED(ht->in_progress_destroy)) |
852 | break; | |
24365af7 MD |
853 | } |
854 | /* Update table size */ | |
855 | t->size = !i ? 1 : (1UL << i); | |
f0c29ed7 | 856 | dbg_printf("init new size: %lu\n", t->size); |
33c7c748 MD |
857 | if (CMM_LOAD_SHARED(ht->in_progress_destroy)) |
858 | break; | |
abc490a1 | 859 | } |
24365af7 | 860 | t->resize_target = t->size; |
11519af6 | 861 | t->resize_initiated = 0; |
2ed95849 MD |
862 | } |
863 | ||
1475579c MD |
864 | static |
865 | void fini_table(struct cds_lfht *ht, struct rcu_table *t, | |
866 | unsigned long first_order, unsigned long len_order) | |
867 | { | |
868 | long i, end_order; | |
869 | ||
870 | dbg_printf("fini table: first_order %lu end_order %lu\n", | |
871 | first_order, first_order + len_order); | |
872 | end_order = first_order + len_order; | |
873 | assert(first_order > 0); | |
874 | assert(t->size == (1UL << (end_order - 1))); | |
875 | for (i = end_order - 1; i >= first_order; i--) { | |
876 | unsigned long j, len; | |
877 | ||
878 | len = !i ? 1 : 1UL << (i - 1); | |
879 | dbg_printf("fini order %lu len: %lu\n", i, len); | |
880 | /* Unlink */ | |
881 | for (j = 0; j < len; j++) { | |
882 | struct cds_lfht_node *new_node = | |
883 | (struct cds_lfht_node *) &t->tbl[i]->nodes[j]; | |
884 | ||
885 | dbg_printf("fini entry: i %lu j %lu hash %lu\n", | |
886 | i, j, !i ? 0 : (1UL << (i - 1)) + j); | |
887 | new_node->p.reverse_hash = | |
888 | bit_reverse_ulong(!i ? 0 : (1UL << (i - 1)) + j); | |
889 | (void) _cds_lfht_remove(ht, t, new_node, 1); | |
890 | if (CMM_LOAD_SHARED(ht->in_progress_destroy)) | |
891 | break; | |
892 | } | |
893 | ht->cds_lfht_call_rcu(&t->tbl[i]->head, cds_lfht_free_level); | |
894 | /* Update table size */ | |
895 | t->size = (i == 1) ? 0 : 1UL << (i - 2); | |
896 | dbg_printf("fini new size: %lu\n", t->size); | |
897 | if (CMM_LOAD_SHARED(ht->in_progress_destroy)) | |
898 | break; | |
899 | } | |
900 | t->resize_target = t->size; | |
901 | t->resize_initiated = 0; | |
902 | } | |
903 | ||
14044b37 MD |
904 | struct cds_lfht *cds_lfht_new(cds_lfht_hash_fct hash_fct, |
905 | cds_lfht_compare_fct compare_fct, | |
906 | unsigned long hash_seed, | |
907 | unsigned long init_size, | |
908 | void (*cds_lfht_call_rcu)(struct rcu_head *head, | |
1475579c MD |
909 | void (*func)(struct rcu_head *head)), |
910 | void (*cds_lfht_synchronize_rcu)(void)) | |
abc490a1 | 911 | { |
14044b37 | 912 | struct cds_lfht *ht; |
24365af7 | 913 | unsigned long order; |
abc490a1 | 914 | |
8129be4e | 915 | /* init_size must be power of two */ |
49619ea0 | 916 | if (init_size && (init_size & (init_size - 1))) |
8129be4e | 917 | return NULL; |
14044b37 | 918 | ht = calloc(1, sizeof(struct cds_lfht)); |
abc490a1 | 919 | ht->hash_fct = hash_fct; |
732ad076 MD |
920 | ht->compare_fct = compare_fct; |
921 | ht->hash_seed = hash_seed; | |
14044b37 | 922 | ht->cds_lfht_call_rcu = cds_lfht_call_rcu; |
1475579c | 923 | ht->cds_lfht_synchronize_rcu = cds_lfht_synchronize_rcu; |
848d4088 | 924 | ht->in_progress_resize = 0; |
df44348d | 925 | ht->percpu_count = alloc_per_cpu_items_count(); |
abc490a1 MD |
926 | /* this mutex should not nest in read-side C.S. */ |
927 | pthread_mutex_init(&ht->resize_mutex, NULL); | |
24365af7 | 928 | order = get_count_order_ulong(max(init_size, 1)) + 1; |
14044b37 | 929 | ht->t = calloc(1, sizeof(struct cds_lfht) |
1475579c | 930 | + (order * sizeof(struct rcu_level *))); |
abc490a1 | 931 | ht->t->size = 0; |
f000907d | 932 | pthread_mutex_lock(&ht->resize_mutex); |
24365af7 | 933 | init_table(ht, ht->t, 0, order); |
f000907d | 934 | pthread_mutex_unlock(&ht->resize_mutex); |
abc490a1 MD |
935 | return ht; |
936 | } | |
937 | ||
14044b37 | 938 | struct cds_lfht_node *cds_lfht_lookup(struct cds_lfht *ht, void *key, size_t key_len) |
2ed95849 | 939 | { |
395270b6 | 940 | struct rcu_table *t; |
14044b37 MD |
941 | struct cds_lfht_node *node, *next; |
942 | struct _cds_lfht_node *lookup; | |
24365af7 | 943 | unsigned long hash, reverse_hash, index, order; |
2ed95849 | 944 | |
732ad076 | 945 | hash = ht->hash_fct(key, key_len, ht->hash_seed); |
abc490a1 | 946 | reverse_hash = bit_reverse_ulong(hash); |
464a1ec9 | 947 | |
395270b6 | 948 | t = rcu_dereference(ht->t); |
24365af7 MD |
949 | index = hash & (t->size - 1); |
950 | order = get_count_order_ulong(index + 1); | |
1475579c | 951 | lookup = &t->tbl[order]->nodes[index & ((1UL << (order - 1)) - 1)]; |
f0c29ed7 | 952 | dbg_printf("lookup hash %lu index %lu order %lu aridx %lu\n", |
24365af7 | 953 | hash, index, order, index & ((1UL << (order - 1)) - 1)); |
14044b37 | 954 | node = (struct cds_lfht_node *) lookup; |
2ed95849 | 955 | for (;;) { |
abc490a1 MD |
956 | if (unlikely(!node)) |
957 | break; | |
cc4fcb10 | 958 | if (unlikely(node->p.reverse_hash > reverse_hash)) { |
abc490a1 MD |
959 | node = NULL; |
960 | break; | |
2ed95849 | 961 | } |
1b81fe1a MD |
962 | next = rcu_dereference(node->p.next); |
963 | if (likely(!is_removed(next)) | |
964 | && !is_dummy(next) | |
49c2e2d6 | 965 | && likely(!ht->compare_fct(node->key, node->key_len, key, key_len))) { |
273399de | 966 | break; |
2ed95849 | 967 | } |
1b81fe1a | 968 | node = clear_flag(next); |
2ed95849 | 969 | } |
1b81fe1a | 970 | assert(!node || !is_dummy(rcu_dereference(node->p.next))); |
abc490a1 MD |
971 | return node; |
972 | } | |
e0ba718a | 973 | |
a481e5ff MD |
974 | struct cds_lfht_node *cds_lfht_next(struct cds_lfht *ht, |
975 | struct cds_lfht_node *node) | |
976 | { | |
977 | struct cds_lfht_node *next; | |
978 | unsigned long reverse_hash; | |
979 | void *key; | |
980 | size_t key_len; | |
981 | ||
982 | reverse_hash = node->p.reverse_hash; | |
983 | key = node->key; | |
984 | key_len = node->key_len; | |
985 | next = rcu_dereference(node->p.next); | |
986 | node = clear_flag(next); | |
987 | ||
988 | for (;;) { | |
989 | if (unlikely(!node)) | |
990 | break; | |
991 | if (unlikely(node->p.reverse_hash > reverse_hash)) { | |
992 | node = NULL; | |
993 | break; | |
994 | } | |
995 | next = rcu_dereference(node->p.next); | |
996 | if (likely(!is_removed(next)) | |
997 | && !is_dummy(next) | |
998 | && likely(!ht->compare_fct(node->key, node->key_len, key, key_len))) { | |
999 | break; | |
1000 | } | |
1001 | node = clear_flag(next); | |
1002 | } | |
1003 | assert(!node || !is_dummy(rcu_dereference(node->p.next))); | |
1004 | return node; | |
1005 | } | |
1006 | ||
14044b37 | 1007 | void cds_lfht_add(struct cds_lfht *ht, struct cds_lfht_node *node) |
abc490a1 MD |
1008 | { |
1009 | struct rcu_table *t; | |
49c2e2d6 | 1010 | unsigned long hash; |
ab7d5fc6 | 1011 | |
49c2e2d6 | 1012 | hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed); |
cc4fcb10 | 1013 | node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash); |
2ed95849 | 1014 | |
abc490a1 | 1015 | t = rcu_dereference(ht->t); |
14044b37 | 1016 | (void) _cds_lfht_add(ht, t, node, 0, 0); |
df44348d | 1017 | ht_count_add(ht, t); |
3eca1b8c MD |
1018 | } |
1019 | ||
14044b37 MD |
1020 | struct cds_lfht_node *cds_lfht_add_unique(struct cds_lfht *ht, |
1021 | struct cds_lfht_node *node) | |
3eca1b8c MD |
1022 | { |
1023 | struct rcu_table *t; | |
49c2e2d6 | 1024 | unsigned long hash; |
df44348d | 1025 | struct cds_lfht_node *ret; |
3eca1b8c | 1026 | |
49c2e2d6 | 1027 | hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed); |
cc4fcb10 | 1028 | node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash); |
3eca1b8c MD |
1029 | |
1030 | t = rcu_dereference(ht->t); | |
df44348d MD |
1031 | ret = _cds_lfht_add(ht, t, node, 1, 0); |
1032 | if (ret != node) | |
1033 | ht_count_add(ht, t); | |
1034 | return ret; | |
2ed95849 MD |
1035 | } |
1036 | ||
14044b37 | 1037 | int cds_lfht_remove(struct cds_lfht *ht, struct cds_lfht_node *node) |
2ed95849 | 1038 | { |
abc490a1 | 1039 | struct rcu_table *t; |
df44348d | 1040 | int ret; |
abc490a1 MD |
1041 | |
1042 | t = rcu_dereference(ht->t); | |
1475579c | 1043 | ret = _cds_lfht_remove(ht, t, node, 0); |
df44348d MD |
1044 | if (!ret) |
1045 | ht_count_remove(ht, t); | |
1046 | return ret; | |
2ed95849 | 1047 | } |
ab7d5fc6 | 1048 | |
abc490a1 | 1049 | static |
14044b37 | 1050 | int cds_lfht_delete_dummy(struct cds_lfht *ht) |
674f7a69 | 1051 | { |
395270b6 | 1052 | struct rcu_table *t; |
14044b37 MD |
1053 | struct cds_lfht_node *node; |
1054 | struct _cds_lfht_node *lookup; | |
24365af7 | 1055 | unsigned long order, i; |
674f7a69 | 1056 | |
abc490a1 MD |
1057 | t = ht->t; |
1058 | /* Check that the table is empty */ | |
1475579c | 1059 | lookup = &t->tbl[0]->nodes[0]; |
14044b37 | 1060 | node = (struct cds_lfht_node *) lookup; |
abc490a1 | 1061 | do { |
1b81fe1a MD |
1062 | node = clear_flag(node)->p.next; |
1063 | if (!is_dummy(node)) | |
abc490a1 | 1064 | return -EPERM; |
273399de | 1065 | assert(!is_removed(node)); |
a2974903 | 1066 | } while (clear_flag(node)); |
abc490a1 | 1067 | /* Internal sanity check: all nodes left should be dummy */ |
24365af7 MD |
1068 | for (order = 0; order < get_count_order_ulong(t->size) + 1; order++) { |
1069 | unsigned long len; | |
1070 | ||
1071 | len = !order ? 1 : 1UL << (order - 1); | |
1072 | for (i = 0; i < len; i++) { | |
f0c29ed7 | 1073 | dbg_printf("delete order %lu i %lu hash %lu\n", |
24365af7 | 1074 | order, i, |
1475579c MD |
1075 | bit_reverse_ulong(t->tbl[order]->nodes[i].reverse_hash)); |
1076 | assert(is_dummy(t->tbl[order]->nodes[i].next)); | |
24365af7 MD |
1077 | } |
1078 | free(t->tbl[order]); | |
674f7a69 | 1079 | } |
abc490a1 | 1080 | return 0; |
674f7a69 MD |
1081 | } |
1082 | ||
1083 | /* | |
1084 | * Should only be called when no more concurrent readers nor writers can | |
1085 | * possibly access the table. | |
1086 | */ | |
14044b37 | 1087 | int cds_lfht_destroy(struct cds_lfht *ht) |
674f7a69 | 1088 | { |
5e28c532 MD |
1089 | int ret; |
1090 | ||
848d4088 | 1091 | /* Wait for in-flight resize operations to complete */ |
33c7c748 | 1092 | CMM_STORE_SHARED(ht->in_progress_destroy, 1); |
848d4088 MD |
1093 | while (uatomic_read(&ht->in_progress_resize)) |
1094 | poll(NULL, 0, 100); /* wait for 100ms */ | |
14044b37 | 1095 | ret = cds_lfht_delete_dummy(ht); |
abc490a1 MD |
1096 | if (ret) |
1097 | return ret; | |
395270b6 | 1098 | free(ht->t); |
df44348d | 1099 | free_per_cpu_items_count(ht->percpu_count); |
674f7a69 | 1100 | free(ht); |
5e28c532 | 1101 | return ret; |
674f7a69 MD |
1102 | } |
1103 | ||
14044b37 | 1104 | void cds_lfht_count_nodes(struct cds_lfht *ht, |
273399de MD |
1105 | unsigned long *count, |
1106 | unsigned long *removed) | |
1107 | { | |
1108 | struct rcu_table *t; | |
14044b37 MD |
1109 | struct cds_lfht_node *node, *next; |
1110 | struct _cds_lfht_node *lookup; | |
24365af7 | 1111 | unsigned long nr_dummy = 0; |
273399de MD |
1112 | |
1113 | *count = 0; | |
1114 | *removed = 0; | |
1115 | ||
1116 | t = rcu_dereference(ht->t); | |
24365af7 | 1117 | /* Count non-dummy nodes in the table */ |
1475579c | 1118 | lookup = &t->tbl[0]->nodes[0]; |
14044b37 | 1119 | node = (struct cds_lfht_node *) lookup; |
273399de | 1120 | do { |
cc4fcb10 | 1121 | next = rcu_dereference(node->p.next); |
273399de | 1122 | if (is_removed(next)) { |
1b81fe1a | 1123 | assert(!is_dummy(next)); |
273399de | 1124 | (*removed)++; |
1b81fe1a | 1125 | } else if (!is_dummy(next)) |
273399de | 1126 | (*count)++; |
24365af7 MD |
1127 | else |
1128 | (nr_dummy)++; | |
273399de MD |
1129 | node = clear_flag(next); |
1130 | } while (node); | |
f0c29ed7 | 1131 | dbg_printf("number of dummy nodes: %lu\n", nr_dummy); |
273399de MD |
1132 | } |
1133 | ||
1475579c | 1134 | /* called with resize mutex held */ |
abc490a1 | 1135 | static |
1475579c MD |
1136 | void _do_cds_lfht_grow(struct cds_lfht *ht, struct rcu_table *old_t, |
1137 | unsigned long old_size, unsigned long new_size) | |
abc490a1 | 1138 | { |
1475579c MD |
1139 | unsigned long old_order, new_order; |
1140 | struct rcu_table *new_t; | |
1141 | ||
1142 | old_order = get_count_order_ulong(old_size) + 1; | |
1143 | new_order = get_count_order_ulong(new_size) + 1; | |
1144 | printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n", | |
1145 | old_size, old_order, new_size, new_order); | |
1146 | new_t = malloc(sizeof(struct cds_lfht) | |
1147 | + (new_order * sizeof(struct rcu_level *))); | |
1148 | assert(new_size > old_size); | |
1149 | memcpy(&new_t->tbl, &old_t->tbl, | |
1150 | old_order * sizeof(struct rcu_level *)); | |
1151 | init_table(ht, new_t, old_order, new_order - old_order); | |
1152 | /* Changing table and size atomically wrt lookups */ | |
1153 | rcu_assign_pointer(ht->t, new_t); | |
1154 | ht->cds_lfht_call_rcu(&old_t->head, cds_lfht_free_table_cb); | |
abc490a1 MD |
1155 | } |
1156 | ||
1157 | /* called with resize mutex held */ | |
1158 | static | |
1475579c MD |
1159 | void _do_cds_lfht_shrink(struct cds_lfht *ht, struct rcu_table *old_t, |
1160 | unsigned long old_size, unsigned long new_size) | |
464a1ec9 | 1161 | { |
1475579c MD |
1162 | unsigned long old_order, new_order; |
1163 | struct rcu_table *new_t; | |
464a1ec9 | 1164 | |
1475579c | 1165 | new_size = max(new_size, 1); |
24365af7 | 1166 | old_order = get_count_order_ulong(old_size) + 1; |
24365af7 | 1167 | new_order = get_count_order_ulong(new_size) + 1; |
df44348d | 1168 | printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n", |
df03fab8 | 1169 | old_size, old_order, new_size, new_order); |
14044b37 | 1170 | new_t = malloc(sizeof(struct cds_lfht) |
1475579c MD |
1171 | + (new_order * sizeof(struct rcu_level *))); |
1172 | assert(new_size < old_size); | |
f000907d | 1173 | memcpy(&new_t->tbl, &old_t->tbl, |
1475579c MD |
1174 | new_order * sizeof(struct rcu_level *)); |
1175 | new_t->size = !new_order ? 1 : (1UL << (new_order - 1)); | |
1176 | new_t->resize_target = new_t->size; | |
1177 | new_t->resize_initiated = 0; | |
1178 | ||
f000907d MD |
1179 | /* Changing table and size atomically wrt lookups */ |
1180 | rcu_assign_pointer(ht->t, new_t); | |
1475579c MD |
1181 | |
1182 | /* | |
1183 | * We need to wait for all reader threads to reach Q.S. (and | |
1184 | * thus use the new table for lookups) before we can start | |
1185 | * releasing the old dummy nodes. | |
1186 | */ | |
1187 | ht->cds_lfht_synchronize_rcu(); | |
1188 | ||
1189 | /* Unlink and remove all now-unused dummy node pointers. */ | |
1190 | fini_table(ht, old_t, new_order, old_order - new_order); | |
14044b37 | 1191 | ht->cds_lfht_call_rcu(&old_t->head, cds_lfht_free_table_cb); |
464a1ec9 MD |
1192 | } |
1193 | ||
1475579c MD |
1194 | |
1195 | /* called with resize mutex held */ | |
1196 | static | |
1197 | void _do_cds_lfht_resize(struct cds_lfht *ht) | |
1198 | { | |
1199 | unsigned long new_size, old_size; | |
1200 | struct rcu_table *old_t; | |
1201 | ||
1202 | old_t = ht->t; | |
1203 | old_size = old_t->size; | |
1204 | new_size = CMM_LOAD_SHARED(old_t->resize_target); | |
1205 | if (old_size < new_size) | |
1206 | _do_cds_lfht_grow(ht, old_t, old_size, new_size); | |
1207 | else if (old_size > new_size) | |
1208 | _do_cds_lfht_shrink(ht, old_t, old_size, new_size); | |
1209 | else | |
1210 | CMM_STORE_SHARED(old_t->resize_initiated, 0); | |
1211 | } | |
1212 | ||
abc490a1 | 1213 | static |
f9830efd MD |
1214 | unsigned long resize_target_update(struct rcu_table *t, |
1215 | int growth_order) | |
464a1ec9 | 1216 | { |
f9830efd MD |
1217 | return _uatomic_max(&t->resize_target, |
1218 | t->size << growth_order); | |
464a1ec9 MD |
1219 | } |
1220 | ||
1475579c MD |
1221 | static |
1222 | unsigned long resize_target_update_count(struct rcu_table *t, | |
1223 | unsigned long count) | |
1224 | { | |
1225 | count = max(count, 1); | |
1226 | return uatomic_set(&t->resize_target, count); | |
1227 | } | |
1228 | ||
1229 | void cds_lfht_resize(struct cds_lfht *ht, unsigned long new_size) | |
464a1ec9 | 1230 | { |
f9830efd MD |
1231 | struct rcu_table *t = rcu_dereference(ht->t); |
1232 | unsigned long target_size; | |
1233 | ||
1475579c MD |
1234 | target_size = resize_target_update_count(t, new_size); |
1235 | CMM_STORE_SHARED(t->resize_initiated, 1); | |
1236 | pthread_mutex_lock(&ht->resize_mutex); | |
1237 | _do_cds_lfht_resize(ht); | |
1238 | pthread_mutex_unlock(&ht->resize_mutex); | |
abc490a1 | 1239 | } |
464a1ec9 | 1240 | |
abc490a1 MD |
1241 | static |
1242 | void do_resize_cb(struct rcu_head *head) | |
1243 | { | |
1244 | struct rcu_resize_work *work = | |
1245 | caa_container_of(head, struct rcu_resize_work, head); | |
14044b37 | 1246 | struct cds_lfht *ht = work->ht; |
abc490a1 MD |
1247 | |
1248 | pthread_mutex_lock(&ht->resize_mutex); | |
14044b37 | 1249 | _do_cds_lfht_resize(ht); |
abc490a1 MD |
1250 | pthread_mutex_unlock(&ht->resize_mutex); |
1251 | free(work); | |
848d4088 MD |
1252 | cmm_smp_mb(); /* finish resize before decrement */ |
1253 | uatomic_dec(&ht->in_progress_resize); | |
464a1ec9 MD |
1254 | } |
1255 | ||
abc490a1 | 1256 | static |
14044b37 | 1257 | void cds_lfht_resize_lazy(struct cds_lfht *ht, struct rcu_table *t, int growth) |
ab7d5fc6 | 1258 | { |
abc490a1 | 1259 | struct rcu_resize_work *work; |
f9830efd | 1260 | unsigned long target_size; |
abc490a1 | 1261 | |
f9830efd | 1262 | target_size = resize_target_update(t, growth); |
11519af6 | 1263 | if (!CMM_LOAD_SHARED(t->resize_initiated) && t->size < target_size) { |
848d4088 MD |
1264 | uatomic_inc(&ht->in_progress_resize); |
1265 | cmm_smp_mb(); /* increment resize count before calling it */ | |
f9830efd MD |
1266 | work = malloc(sizeof(*work)); |
1267 | work->ht = ht; | |
14044b37 | 1268 | ht->cds_lfht_call_rcu(&work->head, do_resize_cb); |
11519af6 | 1269 | CMM_STORE_SHARED(t->resize_initiated, 1); |
f9830efd | 1270 | } |
ab7d5fc6 | 1271 | } |
3171717f | 1272 | |
f8994aee MD |
1273 | #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) |
1274 | ||
3171717f MD |
1275 | static |
1276 | void cds_lfht_resize_lazy_count(struct cds_lfht *ht, struct rcu_table *t, | |
1277 | unsigned long count) | |
1278 | { | |
1279 | struct rcu_resize_work *work; | |
1280 | unsigned long target_size; | |
1281 | ||
1282 | target_size = resize_target_update_count(t, count); | |
1475579c | 1283 | if (!CMM_LOAD_SHARED(t->resize_initiated)) { |
3171717f MD |
1284 | uatomic_inc(&ht->in_progress_resize); |
1285 | cmm_smp_mb(); /* increment resize count before calling it */ | |
1286 | work = malloc(sizeof(*work)); | |
1287 | work->ht = ht; | |
1288 | ht->cds_lfht_call_rcu(&work->head, do_resize_cb); | |
1289 | CMM_STORE_SHARED(t->resize_initiated, 1); | |
1290 | } | |
1291 | } | |
f8994aee MD |
1292 | |
1293 | #endif |