Commit | Line | Data |
---|---|---|
5e28c532 | 1 | /* |
abc490a1 MD |
2 | * rculfhash.c |
3 | * | |
4 | * Userspace RCU library - Lock-Free Expandable RCU Hash Table | |
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 | * | |
32 | * Some specificities of this Lock-Free Expandable RCU Hash Table | |
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. | |
51 | * - The resize operation only allows expanding the hash table. | |
52 | * It is triggered either through an API call or automatically by | |
53 | * detecting long chains in the add operation. | |
54 | * - Resize operation initiated by long chain detection is executed by a | |
55 | * call_rcu thread, which keeps lock-freedom of add and remove. | |
56 | * - Resize operations are protected by a mutex. | |
57 | * - The removal operation is split in two parts: first, a "removed" | |
58 | * flag is set in the next pointer within the node to remove. Then, | |
59 | * a "garbage collection" is performed in the bucket containing the | |
60 | * removed node (from the start of the bucket up to the removed node). | |
61 | * All encountered nodes with "removed" flag set in their next | |
62 | * pointers are removed from the linked-list. If the cmpxchg used for | |
63 | * removal fails (due to concurrent garbage-collection or concurrent | |
64 | * add), we retry from the beginning of the bucket. This ensures that | |
65 | * the node with "removed" flag set is removed from the hash table | |
66 | * (not visible to lookups anymore) before the RCU read-side critical | |
67 | * section held across removal ends. Furthermore, this ensures that | |
68 | * the node with "removed" flag set is removed from the linked-list | |
69 | * before its memory is reclaimed. Only the thread which removal | |
70 | * successfully set the "removed" flag (with a cmpxchg) into a node's | |
71 | * next pointer is considered to have succeeded its removal (and thus | |
72 | * owns the node to reclaim). Because we garbage-collect starting from | |
73 | * an invariant node (the start-of-bucket dummy node) up to the | |
74 | * "removed" node (or find a reverse-hash that is higher), we are sure | |
75 | * that a successful traversal of the chain leads to a chain that is | |
76 | * present in the linked-list (the start node is never removed) and | |
77 | * that is does not contain the "removed" node anymore, even if | |
78 | * concurrent delete/add operations are changing the structure of the | |
79 | * list concurrently. | |
80 | * - A RCU "order table" indexed by log2(hash index) is copied and | |
81 | * expanded by the resize operation. This order table allows finding | |
82 | * the "dummy node" tables. | |
83 | * - There is one dummy node table per hash index order. The size of | |
84 | * each dummy node table is half the number of hashes contained in | |
85 | * this order. | |
86 | * - call_rcu is used to garbage-collect the old order table. | |
87 | * - The per-order dummy node tables contain a compact version of the | |
88 | * hash table nodes. These tables are invariant after they are | |
89 | * populated into the hash table. | |
90 | */ | |
91 | ||
2ed95849 MD |
92 | #define _LGPL_SOURCE |
93 | #include <stdlib.h> | |
e0ba718a MD |
94 | #include <errno.h> |
95 | #include <assert.h> | |
96 | #include <stdio.h> | |
abc490a1 | 97 | #include <stdint.h> |
f000907d | 98 | #include <string.h> |
e0ba718a | 99 | |
2ed95849 | 100 | #include <urcu.h> |
abc490a1 | 101 | #include <urcu-call-rcu.h> |
a42cc659 MD |
102 | #include <urcu/arch.h> |
103 | #include <urcu/uatomic.h> | |
674f7a69 | 104 | #include <urcu/jhash.h> |
a42cc659 | 105 | #include <urcu/compiler.h> |
abc490a1 | 106 | #include <urcu/rculfhash.h> |
5e28c532 | 107 | #include <stdio.h> |
464a1ec9 | 108 | #include <pthread.h> |
44395fb7 | 109 | |
f9830efd | 110 | #ifdef DEBUG |
e753ff5a | 111 | #define dbg_printf(fmt, args...) printf(fmt, ## args) |
f9830efd | 112 | #else |
e753ff5a | 113 | #define dbg_printf(fmt, args...) |
f9830efd MD |
114 | #endif |
115 | ||
01370f0b MD |
116 | #define CHAIN_LEN_TARGET 4 |
117 | #define CHAIN_LEN_RESIZE_THRESHOLD 8 | |
2ed95849 | 118 | |
abc490a1 MD |
119 | #ifndef max |
120 | #define max(a, b) ((a) > (b) ? (a) : (b)) | |
121 | #endif | |
2ed95849 | 122 | |
d95bd160 MD |
123 | /* |
124 | * The removed flag needs to be updated atomically with the pointer. | |
125 | * The dummy flag does not require to be updated atomically with the | |
126 | * pointer, but it is added as a pointer low bit flag to save space. | |
127 | */ | |
d37166c6 | 128 | #define REMOVED_FLAG (1UL << 0) |
f5596c94 MD |
129 | #define DUMMY_FLAG (1UL << 1) |
130 | #define FLAGS_MASK ((1UL << 2) - 1) | |
d37166c6 | 131 | |
395270b6 | 132 | struct rcu_table { |
abc490a1 | 133 | unsigned long size; /* always a power of 2 */ |
f9830efd | 134 | unsigned long resize_target; |
11519af6 | 135 | int resize_initiated; |
abc490a1 | 136 | struct rcu_head head; |
24365af7 | 137 | struct _rcu_ht_node *tbl[0]; |
395270b6 MD |
138 | }; |
139 | ||
2ed95849 | 140 | struct rcu_ht { |
395270b6 | 141 | struct rcu_table *t; /* shared */ |
2ed95849 | 142 | ht_hash_fct hash_fct; |
732ad076 MD |
143 | ht_compare_fct compare_fct; |
144 | unsigned long hash_seed; | |
464a1ec9 | 145 | pthread_mutex_t resize_mutex; /* resize mutex: add/del mutex */ |
33c7c748 | 146 | unsigned int in_progress_resize, in_progress_destroy; |
abc490a1 MD |
147 | void (*ht_call_rcu)(struct rcu_head *head, |
148 | void (*func)(struct rcu_head *head)); | |
2ed95849 MD |
149 | }; |
150 | ||
abc490a1 MD |
151 | struct rcu_resize_work { |
152 | struct rcu_head head; | |
2ed95849 | 153 | struct rcu_ht *ht; |
abc490a1 | 154 | }; |
2ed95849 | 155 | |
abc490a1 MD |
156 | /* |
157 | * Algorithm to reverse bits in a word by lookup table, extended to | |
158 | * 64-bit words. | |
f9830efd | 159 | * Source: |
abc490a1 | 160 | * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable |
f9830efd | 161 | * Originally from Public Domain. |
abc490a1 MD |
162 | */ |
163 | ||
164 | static const uint8_t BitReverseTable256[256] = | |
2ed95849 | 165 | { |
abc490a1 MD |
166 | #define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64 |
167 | #define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16) | |
168 | #define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 ) | |
169 | R6(0), R6(2), R6(1), R6(3) | |
170 | }; | |
171 | #undef R2 | |
172 | #undef R4 | |
173 | #undef R6 | |
2ed95849 | 174 | |
abc490a1 MD |
175 | static |
176 | uint8_t bit_reverse_u8(uint8_t v) | |
177 | { | |
178 | return BitReverseTable256[v]; | |
179 | } | |
ab7d5fc6 | 180 | |
abc490a1 MD |
181 | static __attribute__((unused)) |
182 | uint32_t bit_reverse_u32(uint32_t v) | |
183 | { | |
184 | return ((uint32_t) bit_reverse_u8(v) << 24) | | |
185 | ((uint32_t) bit_reverse_u8(v >> 8) << 16) | | |
186 | ((uint32_t) bit_reverse_u8(v >> 16) << 8) | | |
187 | ((uint32_t) bit_reverse_u8(v >> 24)); | |
2ed95849 MD |
188 | } |
189 | ||
abc490a1 MD |
190 | static __attribute__((unused)) |
191 | uint64_t bit_reverse_u64(uint64_t v) | |
2ed95849 | 192 | { |
abc490a1 MD |
193 | return ((uint64_t) bit_reverse_u8(v) << 56) | |
194 | ((uint64_t) bit_reverse_u8(v >> 8) << 48) | | |
195 | ((uint64_t) bit_reverse_u8(v >> 16) << 40) | | |
196 | ((uint64_t) bit_reverse_u8(v >> 24) << 32) | | |
197 | ((uint64_t) bit_reverse_u8(v >> 32) << 24) | | |
198 | ((uint64_t) bit_reverse_u8(v >> 40) << 16) | | |
199 | ((uint64_t) bit_reverse_u8(v >> 48) << 8) | | |
200 | ((uint64_t) bit_reverse_u8(v >> 56)); | |
201 | } | |
202 | ||
203 | static | |
204 | unsigned long bit_reverse_ulong(unsigned long v) | |
205 | { | |
206 | #if (CAA_BITS_PER_LONG == 32) | |
207 | return bit_reverse_u32(v); | |
208 | #else | |
209 | return bit_reverse_u64(v); | |
210 | #endif | |
211 | } | |
212 | ||
f9830efd | 213 | /* |
24365af7 MD |
214 | * fls: returns the position of the most significant bit. |
215 | * Returns 0 if no bit is set, else returns the position of the most | |
216 | * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit). | |
f9830efd | 217 | */ |
24365af7 MD |
218 | #if defined(__i386) || defined(__x86_64) |
219 | static inline | |
220 | unsigned int fls_u32(uint32_t x) | |
f9830efd | 221 | { |
24365af7 MD |
222 | int r; |
223 | ||
224 | asm("bsrl %1,%0\n\t" | |
225 | "jnz 1f\n\t" | |
226 | "movl $-1,%0\n\t" | |
227 | "1:\n\t" | |
228 | : "=r" (r) : "rm" (x)); | |
229 | return r + 1; | |
230 | } | |
231 | #define HAS_FLS_U32 | |
232 | #endif | |
233 | ||
234 | #if defined(__x86_64) | |
235 | static inline | |
236 | unsigned int fls_u64(uint64_t x) | |
237 | { | |
238 | long r; | |
239 | ||
240 | asm("bsrq %1,%0\n\t" | |
241 | "jnz 1f\n\t" | |
242 | "movq $-1,%0\n\t" | |
243 | "1:\n\t" | |
244 | : "=r" (r) : "rm" (x)); | |
245 | return r + 1; | |
246 | } | |
247 | #define HAS_FLS_U64 | |
248 | #endif | |
249 | ||
250 | #ifndef HAS_FLS_U64 | |
251 | static __attribute__((unused)) | |
252 | unsigned int fls_u64(uint64_t x) | |
253 | { | |
254 | unsigned int r = 64; | |
255 | ||
256 | if (!x) | |
257 | return 0; | |
258 | ||
259 | if (!(x & 0xFFFFFFFF00000000ULL)) { | |
260 | x <<= 32; | |
261 | r -= 32; | |
262 | } | |
263 | if (!(x & 0xFFFF000000000000ULL)) { | |
264 | x <<= 16; | |
265 | r -= 16; | |
266 | } | |
267 | if (!(x & 0xFF00000000000000ULL)) { | |
268 | x <<= 8; | |
269 | r -= 8; | |
270 | } | |
271 | if (!(x & 0xF000000000000000ULL)) { | |
272 | x <<= 4; | |
273 | r -= 4; | |
274 | } | |
275 | if (!(x & 0xC000000000000000ULL)) { | |
276 | x <<= 2; | |
277 | r -= 2; | |
278 | } | |
279 | if (!(x & 0x8000000000000000ULL)) { | |
280 | x <<= 1; | |
281 | r -= 1; | |
282 | } | |
283 | return r; | |
284 | } | |
285 | #endif | |
286 | ||
287 | #ifndef HAS_FLS_U32 | |
288 | static __attribute__((unused)) | |
289 | unsigned int fls_u32(uint32_t x) | |
290 | { | |
291 | unsigned int r = 32; | |
f9830efd | 292 | |
24365af7 MD |
293 | if (!x) |
294 | return 0; | |
295 | if (!(x & 0xFFFF0000U)) { | |
296 | x <<= 16; | |
297 | r -= 16; | |
298 | } | |
299 | if (!(x & 0xFF000000U)) { | |
300 | x <<= 8; | |
301 | r -= 8; | |
302 | } | |
303 | if (!(x & 0xF0000000U)) { | |
304 | x <<= 4; | |
305 | r -= 4; | |
306 | } | |
307 | if (!(x & 0xC0000000U)) { | |
308 | x <<= 2; | |
309 | r -= 2; | |
310 | } | |
311 | if (!(x & 0x80000000U)) { | |
312 | x <<= 1; | |
313 | r -= 1; | |
314 | } | |
315 | return r; | |
316 | } | |
317 | #endif | |
318 | ||
319 | unsigned int fls_ulong(unsigned long x) | |
f9830efd | 320 | { |
24365af7 MD |
321 | #if (CAA_BITS_PER_lONG == 32) |
322 | return fls_u32(x); | |
323 | #else | |
324 | return fls_u64(x); | |
325 | #endif | |
326 | } | |
f9830efd | 327 | |
24365af7 MD |
328 | int get_count_order_u32(uint32_t x) |
329 | { | |
330 | int order; | |
331 | ||
332 | order = fls_u32(x) - 1; | |
333 | if (x & (x - 1)) | |
334 | order++; | |
335 | return order; | |
336 | } | |
337 | ||
338 | int get_count_order_ulong(unsigned long x) | |
339 | { | |
340 | int order; | |
341 | ||
342 | order = fls_ulong(x) - 1; | |
343 | if (x & (x - 1)) | |
344 | order++; | |
345 | return order; | |
f9830efd MD |
346 | } |
347 | ||
348 | static | |
349 | void ht_resize_lazy(struct rcu_ht *ht, struct rcu_table *t, int growth); | |
350 | ||
351 | static | |
352 | void check_resize(struct rcu_ht *ht, struct rcu_table *t, | |
353 | uint32_t chain_len) | |
354 | { | |
24365af7 MD |
355 | if (chain_len > 100) |
356 | dbg_printf("rculfhash: WARNING: large chain length: %u.\n", | |
357 | chain_len); | |
3390d470 MD |
358 | if (chain_len >= CHAIN_LEN_RESIZE_THRESHOLD) |
359 | ht_resize_lazy(ht, t, | |
01370f0b | 360 | get_count_order_u32(chain_len - (CHAIN_LEN_TARGET - 1))); |
f9830efd MD |
361 | } |
362 | ||
abc490a1 MD |
363 | static |
364 | struct rcu_ht_node *clear_flag(struct rcu_ht_node *node) | |
365 | { | |
d37166c6 | 366 | return (struct rcu_ht_node *) (((unsigned long) node) & ~FLAGS_MASK); |
abc490a1 MD |
367 | } |
368 | ||
369 | static | |
370 | int is_removed(struct rcu_ht_node *node) | |
371 | { | |
d37166c6 | 372 | return ((unsigned long) node) & REMOVED_FLAG; |
abc490a1 MD |
373 | } |
374 | ||
375 | static | |
376 | struct rcu_ht_node *flag_removed(struct rcu_ht_node *node) | |
377 | { | |
d37166c6 | 378 | return (struct rcu_ht_node *) (((unsigned long) node) | REMOVED_FLAG); |
abc490a1 MD |
379 | } |
380 | ||
f5596c94 MD |
381 | static |
382 | int is_dummy(struct rcu_ht_node *node) | |
383 | { | |
384 | return ((unsigned long) node) & DUMMY_FLAG; | |
385 | } | |
386 | ||
387 | static | |
388 | struct rcu_ht_node *flag_dummy(struct rcu_ht_node *node) | |
389 | { | |
390 | return (struct rcu_ht_node *) (((unsigned long) node) | DUMMY_FLAG); | |
391 | } | |
392 | ||
abc490a1 | 393 | static |
f9830efd | 394 | unsigned long _uatomic_max(unsigned long *ptr, unsigned long v) |
abc490a1 MD |
395 | { |
396 | unsigned long old1, old2; | |
397 | ||
398 | old1 = uatomic_read(ptr); | |
399 | do { | |
400 | old2 = old1; | |
401 | if (old2 >= v) | |
f9830efd | 402 | return old2; |
abc490a1 | 403 | } while ((old1 = uatomic_cmpxchg(ptr, old2, v)) != old2); |
f9830efd | 404 | return v; |
abc490a1 MD |
405 | } |
406 | ||
273399de MD |
407 | /* |
408 | * Remove all logically deleted nodes from a bucket up to a certain node key. | |
409 | */ | |
410 | static | |
411 | void _ht_gc_bucket(struct rcu_ht_node *dummy, struct rcu_ht_node *node) | |
412 | { | |
f5596c94 | 413 | struct rcu_ht_node *iter_prev, *iter, *next, *new_next; |
273399de MD |
414 | |
415 | for (;;) { | |
416 | iter_prev = dummy; | |
417 | /* We can always skip the dummy node initially */ | |
cc4fcb10 MD |
418 | iter = rcu_dereference(iter_prev->p.next); |
419 | assert(iter_prev->p.reverse_hash <= node->p.reverse_hash); | |
273399de | 420 | for (;;) { |
a2974903 | 421 | if (unlikely(!clear_flag(iter))) |
479c8a32 | 422 | return; |
cc4fcb10 | 423 | if (clear_flag(iter)->p.reverse_hash > node->p.reverse_hash) |
273399de | 424 | return; |
cc4fcb10 | 425 | next = rcu_dereference(clear_flag(iter)->p.next); |
273399de MD |
426 | if (is_removed(next)) |
427 | break; | |
b453eae1 | 428 | iter_prev = clear_flag(iter); |
273399de MD |
429 | iter = next; |
430 | } | |
431 | assert(!is_removed(iter)); | |
f5596c94 MD |
432 | if (is_dummy(iter)) |
433 | new_next = flag_dummy(clear_flag(next)); | |
434 | else | |
435 | new_next = clear_flag(next); | |
436 | (void) uatomic_cmpxchg(&iter_prev->p.next, iter, new_next); | |
273399de MD |
437 | } |
438 | } | |
439 | ||
abc490a1 | 440 | static |
18117871 | 441 | struct rcu_ht_node *_ht_add(struct rcu_ht *ht, struct rcu_table *t, |
f5596c94 | 442 | struct rcu_ht_node *node, int unique, int dummy) |
abc490a1 | 443 | { |
f5596c94 MD |
444 | struct rcu_ht_node *iter_prev, *iter, *next, *new_node, *new_next, |
445 | *dummy_node; | |
24365af7 MD |
446 | struct _rcu_ht_node *lookup; |
447 | unsigned long hash, index, order; | |
abc490a1 | 448 | |
18117871 | 449 | if (!t->size) { |
f5596c94 MD |
450 | assert(dummy); |
451 | node->p.next = flag_dummy(NULL); | |
18117871 MD |
452 | return node; /* Initial first add (head) */ |
453 | } | |
cc4fcb10 | 454 | hash = bit_reverse_ulong(node->p.reverse_hash); |
abc490a1 | 455 | for (;;) { |
f9830efd | 456 | uint32_t chain_len = 0; |
abc490a1 | 457 | |
11519af6 MD |
458 | /* |
459 | * iter_prev points to the non-removed node prior to the | |
460 | * insert location. | |
11519af6 | 461 | */ |
24365af7 MD |
462 | index = hash & (t->size - 1); |
463 | order = get_count_order_ulong(index + 1); | |
464 | lookup = &t->tbl[order][index & ((1UL << (order - 1)) - 1)]; | |
465 | iter_prev = (struct rcu_ht_node *) lookup; | |
11519af6 | 466 | /* We can always skip the dummy node initially */ |
cc4fcb10 MD |
467 | iter = rcu_dereference(iter_prev->p.next); |
468 | assert(iter_prev->p.reverse_hash <= node->p.reverse_hash); | |
abc490a1 | 469 | for (;;) { |
a2974903 | 470 | if (unlikely(!clear_flag(iter))) |
273399de | 471 | goto insert; |
cc4fcb10 | 472 | if (clear_flag(iter)->p.reverse_hash > node->p.reverse_hash) |
273399de | 473 | goto insert; |
cc4fcb10 | 474 | next = rcu_dereference(clear_flag(iter)->p.next); |
273399de | 475 | if (is_removed(next)) |
9dba85be | 476 | goto gc_node; |
e43f23f8 | 477 | if (unique |
1b81fe1a | 478 | && !is_dummy(next) |
e43f23f8 MD |
479 | && !ht->compare_fct(node->key, node->key_len, |
480 | clear_flag(iter)->key, | |
481 | clear_flag(iter)->key_len)) | |
18117871 | 482 | return clear_flag(iter); |
11519af6 | 483 | /* Only account for identical reverse hash once */ |
24365af7 MD |
484 | if (iter_prev->p.reverse_hash != clear_flag(iter)->p.reverse_hash |
485 | && !is_dummy(next)) | |
11519af6 MD |
486 | check_resize(ht, t, ++chain_len); |
487 | iter_prev = clear_flag(iter); | |
273399de | 488 | iter = next; |
abc490a1 | 489 | } |
273399de | 490 | insert: |
7ec59d3b | 491 | assert(node != clear_flag(iter)); |
11519af6 | 492 | assert(!is_removed(iter_prev)); |
f000907d | 493 | assert(iter_prev != node); |
f5596c94 | 494 | if (!dummy) |
1b81fe1a | 495 | node->p.next = clear_flag(iter); |
f5596c94 | 496 | else |
1b81fe1a | 497 | node->p.next = flag_dummy(clear_flag(iter)); |
f5596c94 MD |
498 | if (is_dummy(iter)) |
499 | new_node = flag_dummy(node); | |
500 | else | |
501 | new_node = node; | |
cc4fcb10 | 502 | if (uatomic_cmpxchg(&iter_prev->p.next, iter, |
f5596c94 | 503 | new_node) != iter) |
273399de | 504 | continue; /* retry */ |
11519af6 | 505 | else |
273399de | 506 | goto gc_end; |
9dba85be MD |
507 | gc_node: |
508 | assert(!is_removed(iter)); | |
f5596c94 MD |
509 | if (is_dummy(iter)) |
510 | new_next = flag_dummy(clear_flag(next)); | |
511 | else | |
512 | new_next = clear_flag(next); | |
513 | (void) uatomic_cmpxchg(&iter_prev->p.next, iter, new_next); | |
273399de | 514 | /* retry */ |
464a1ec9 | 515 | } |
273399de MD |
516 | gc_end: |
517 | /* Garbage collect logically removed nodes in the bucket */ | |
24365af7 MD |
518 | index = hash & (t->size - 1); |
519 | order = get_count_order_ulong(index + 1); | |
520 | lookup = &t->tbl[order][index & ((1UL << (order - 1)) - 1)]; | |
521 | dummy_node = (struct rcu_ht_node *) lookup; | |
f5596c94 | 522 | _ht_gc_bucket(dummy_node, node); |
18117871 | 523 | return node; |
abc490a1 | 524 | } |
464a1ec9 | 525 | |
abc490a1 MD |
526 | static |
527 | int _ht_remove(struct rcu_ht *ht, struct rcu_table *t, struct rcu_ht_node *node) | |
528 | { | |
273399de | 529 | struct rcu_ht_node *dummy, *next, *old; |
24365af7 | 530 | struct _rcu_ht_node *lookup; |
abc490a1 | 531 | int flagged = 0; |
24365af7 | 532 | unsigned long hash, index, order; |
5e28c532 | 533 | |
7ec59d3b | 534 | /* logically delete the node */ |
cc4fcb10 | 535 | old = rcu_dereference(node->p.next); |
7ec59d3b MD |
536 | do { |
537 | next = old; | |
538 | if (is_removed(next)) | |
539 | goto end; | |
1b81fe1a | 540 | assert(!is_dummy(next)); |
cc4fcb10 | 541 | old = uatomic_cmpxchg(&node->p.next, next, |
7ec59d3b MD |
542 | flag_removed(next)); |
543 | } while (old != next); | |
544 | ||
545 | /* We performed the (logical) deletion. */ | |
546 | flagged = 1; | |
547 | ||
548 | /* | |
549 | * Ensure that the node is not visible to readers anymore: lookup for | |
273399de MD |
550 | * the node, and remove it (along with any other logically removed node) |
551 | * if found. | |
11519af6 | 552 | */ |
cc4fcb10 | 553 | hash = bit_reverse_ulong(node->p.reverse_hash); |
24365af7 MD |
554 | index = hash & (t->size - 1); |
555 | order = get_count_order_ulong(index + 1); | |
556 | lookup = &t->tbl[order][index & ((1UL << (order - 1)) - 1)]; | |
557 | dummy = (struct rcu_ht_node *) lookup; | |
273399de | 558 | _ht_gc_bucket(dummy, node); |
2ed95849 | 559 | end: |
11519af6 MD |
560 | /* |
561 | * Only the flagging action indicated that we (and no other) | |
562 | * removed the node from the hash. | |
563 | */ | |
7ec59d3b | 564 | if (flagged) { |
cc4fcb10 | 565 | assert(is_removed(rcu_dereference(node->p.next))); |
11519af6 | 566 | return 0; |
7ec59d3b | 567 | } else |
11519af6 | 568 | return -ENOENT; |
abc490a1 | 569 | } |
2ed95849 | 570 | |
abc490a1 MD |
571 | static |
572 | void init_table(struct rcu_ht *ht, struct rcu_table *t, | |
24365af7 MD |
573 | unsigned long first_order, unsigned long len_order) |
574 | { | |
575 | unsigned long i, end_order; | |
576 | ||
577 | dbg_printf("rculfhash: init table: first_order %lu end_order %lu\n", | |
578 | first_order, first_order + len_order); | |
579 | end_order = first_order + len_order; | |
580 | t->size = !first_order ? 0 : (1UL << (first_order - 1)); | |
581 | for (i = first_order; i < end_order; i++) { | |
582 | unsigned long j, len; | |
583 | ||
584 | len = !i ? 1 : 1UL << (i - 1); | |
585 | dbg_printf("rculfhash: init order %lu len: %lu\n", i, len); | |
586 | t->tbl[i] = calloc(len, sizeof(struct _rcu_ht_node)); | |
587 | for (j = 0; j < len; j++) { | |
588 | dbg_printf("rculfhash: init entry: i %lu j %lu hash %lu\n", | |
589 | i, j, !i ? 0 : (1UL << (i - 1)) + j); | |
590 | struct rcu_ht_node *new_node = | |
591 | (struct rcu_ht_node *) &t->tbl[i][j]; | |
592 | new_node->p.reverse_hash = | |
593 | bit_reverse_ulong(!i ? 0 : (1UL << (i - 1)) + j); | |
594 | (void) _ht_add(ht, t, new_node, 0, 1); | |
33c7c748 MD |
595 | if (CMM_LOAD_SHARED(ht->in_progress_destroy)) |
596 | break; | |
24365af7 MD |
597 | } |
598 | /* Update table size */ | |
599 | t->size = !i ? 1 : (1UL << i); | |
600 | dbg_printf("rculfhash: init new size: %lu\n", t->size); | |
33c7c748 MD |
601 | if (CMM_LOAD_SHARED(ht->in_progress_destroy)) |
602 | break; | |
abc490a1 | 603 | } |
24365af7 | 604 | t->resize_target = t->size; |
11519af6 | 605 | t->resize_initiated = 0; |
2ed95849 MD |
606 | } |
607 | ||
abc490a1 | 608 | struct rcu_ht *ht_new(ht_hash_fct hash_fct, |
732ad076 MD |
609 | ht_compare_fct compare_fct, |
610 | unsigned long hash_seed, | |
abc490a1 MD |
611 | unsigned long init_size, |
612 | void (*ht_call_rcu)(struct rcu_head *head, | |
613 | void (*func)(struct rcu_head *head))) | |
614 | { | |
615 | struct rcu_ht *ht; | |
24365af7 | 616 | unsigned long order; |
abc490a1 MD |
617 | |
618 | ht = calloc(1, sizeof(struct rcu_ht)); | |
619 | ht->hash_fct = hash_fct; | |
732ad076 MD |
620 | ht->compare_fct = compare_fct; |
621 | ht->hash_seed = hash_seed; | |
f000907d | 622 | ht->ht_call_rcu = ht_call_rcu; |
848d4088 | 623 | ht->in_progress_resize = 0; |
abc490a1 MD |
624 | /* this mutex should not nest in read-side C.S. */ |
625 | pthread_mutex_init(&ht->resize_mutex, NULL); | |
24365af7 | 626 | order = get_count_order_ulong(max(init_size, 1)) + 1; |
abc490a1 | 627 | ht->t = calloc(1, sizeof(struct rcu_table) |
24365af7 | 628 | + (order * sizeof(struct _rcu_ht_node *))); |
abc490a1 | 629 | ht->t->size = 0; |
f000907d | 630 | pthread_mutex_lock(&ht->resize_mutex); |
24365af7 | 631 | init_table(ht, ht->t, 0, order); |
f000907d | 632 | pthread_mutex_unlock(&ht->resize_mutex); |
abc490a1 MD |
633 | return ht; |
634 | } | |
635 | ||
732ad076 | 636 | struct rcu_ht_node *ht_lookup(struct rcu_ht *ht, void *key, size_t key_len) |
2ed95849 | 637 | { |
395270b6 | 638 | struct rcu_table *t; |
1b81fe1a | 639 | struct rcu_ht_node *node, *next; |
24365af7 MD |
640 | struct _rcu_ht_node *lookup; |
641 | unsigned long hash, reverse_hash, index, order; | |
2ed95849 | 642 | |
732ad076 | 643 | hash = ht->hash_fct(key, key_len, ht->hash_seed); |
abc490a1 | 644 | reverse_hash = bit_reverse_ulong(hash); |
464a1ec9 | 645 | |
395270b6 | 646 | t = rcu_dereference(ht->t); |
24365af7 MD |
647 | index = hash & (t->size - 1); |
648 | order = get_count_order_ulong(index + 1); | |
649 | lookup = &t->tbl[order][index & ((1UL << (order - 1)) - 1)]; | |
650 | dbg_printf("rculfhash: lookup hash %lu index %lu order %lu aridx %lu\n", | |
651 | hash, index, order, index & ((1UL << (order - 1)) - 1)); | |
652 | node = (struct rcu_ht_node *) lookup; | |
2ed95849 | 653 | for (;;) { |
abc490a1 MD |
654 | if (unlikely(!node)) |
655 | break; | |
cc4fcb10 | 656 | if (unlikely(node->p.reverse_hash > reverse_hash)) { |
abc490a1 MD |
657 | node = NULL; |
658 | break; | |
2ed95849 | 659 | } |
1b81fe1a MD |
660 | next = rcu_dereference(node->p.next); |
661 | if (likely(!is_removed(next)) | |
662 | && !is_dummy(next) | |
49c2e2d6 | 663 | && likely(!ht->compare_fct(node->key, node->key_len, key, key_len))) { |
273399de | 664 | break; |
2ed95849 | 665 | } |
1b81fe1a | 666 | node = clear_flag(next); |
2ed95849 | 667 | } |
1b81fe1a | 668 | assert(!node || !is_dummy(rcu_dereference(node->p.next))); |
abc490a1 MD |
669 | return node; |
670 | } | |
e0ba718a | 671 | |
f000907d | 672 | void ht_add(struct rcu_ht *ht, struct rcu_ht_node *node) |
abc490a1 MD |
673 | { |
674 | struct rcu_table *t; | |
49c2e2d6 | 675 | unsigned long hash; |
ab7d5fc6 | 676 | |
49c2e2d6 | 677 | hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed); |
cc4fcb10 | 678 | node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash); |
2ed95849 | 679 | |
abc490a1 | 680 | t = rcu_dereference(ht->t); |
f5596c94 | 681 | (void) _ht_add(ht, t, node, 0, 0); |
3eca1b8c MD |
682 | } |
683 | ||
18117871 | 684 | struct rcu_ht_node *ht_add_unique(struct rcu_ht *ht, struct rcu_ht_node *node) |
3eca1b8c MD |
685 | { |
686 | struct rcu_table *t; | |
49c2e2d6 | 687 | unsigned long hash; |
3eca1b8c | 688 | |
49c2e2d6 | 689 | hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed); |
cc4fcb10 | 690 | node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash); |
3eca1b8c MD |
691 | |
692 | t = rcu_dereference(ht->t); | |
f5596c94 | 693 | return _ht_add(ht, t, node, 1, 0); |
2ed95849 MD |
694 | } |
695 | ||
abc490a1 | 696 | int ht_remove(struct rcu_ht *ht, struct rcu_ht_node *node) |
2ed95849 | 697 | { |
abc490a1 MD |
698 | struct rcu_table *t; |
699 | ||
700 | t = rcu_dereference(ht->t); | |
abc490a1 | 701 | return _ht_remove(ht, t, node); |
2ed95849 | 702 | } |
ab7d5fc6 | 703 | |
abc490a1 MD |
704 | static |
705 | int ht_delete_dummy(struct rcu_ht *ht) | |
674f7a69 | 706 | { |
395270b6 | 707 | struct rcu_table *t; |
abc490a1 | 708 | struct rcu_ht_node *node; |
24365af7 MD |
709 | struct _rcu_ht_node *lookup; |
710 | unsigned long order, i; | |
674f7a69 | 711 | |
abc490a1 MD |
712 | t = ht->t; |
713 | /* Check that the table is empty */ | |
24365af7 MD |
714 | lookup = &t->tbl[0][0]; |
715 | node = (struct rcu_ht_node *) lookup; | |
abc490a1 | 716 | do { |
1b81fe1a MD |
717 | node = clear_flag(node)->p.next; |
718 | if (!is_dummy(node)) | |
abc490a1 | 719 | return -EPERM; |
273399de | 720 | assert(!is_removed(node)); |
a2974903 | 721 | } while (clear_flag(node)); |
abc490a1 | 722 | /* Internal sanity check: all nodes left should be dummy */ |
24365af7 MD |
723 | for (order = 0; order < get_count_order_ulong(t->size) + 1; order++) { |
724 | unsigned long len; | |
725 | ||
726 | len = !order ? 1 : 1UL << (order - 1); | |
727 | for (i = 0; i < len; i++) { | |
728 | dbg_printf("rculfhash: delete order %lu i %lu hash %lu\n", | |
729 | order, i, | |
730 | bit_reverse_ulong(t->tbl[order][i].reverse_hash)); | |
731 | assert(is_dummy(t->tbl[order][i].next)); | |
732 | } | |
733 | free(t->tbl[order]); | |
674f7a69 | 734 | } |
abc490a1 | 735 | return 0; |
674f7a69 MD |
736 | } |
737 | ||
738 | /* | |
739 | * Should only be called when no more concurrent readers nor writers can | |
740 | * possibly access the table. | |
741 | */ | |
5e28c532 | 742 | int ht_destroy(struct rcu_ht *ht) |
674f7a69 | 743 | { |
5e28c532 MD |
744 | int ret; |
745 | ||
848d4088 | 746 | /* Wait for in-flight resize operations to complete */ |
33c7c748 | 747 | CMM_STORE_SHARED(ht->in_progress_destroy, 1); |
848d4088 MD |
748 | while (uatomic_read(&ht->in_progress_resize)) |
749 | poll(NULL, 0, 100); /* wait for 100ms */ | |
abc490a1 MD |
750 | ret = ht_delete_dummy(ht); |
751 | if (ret) | |
752 | return ret; | |
395270b6 | 753 | free(ht->t); |
674f7a69 | 754 | free(ht); |
5e28c532 | 755 | return ret; |
674f7a69 MD |
756 | } |
757 | ||
273399de MD |
758 | void ht_count_nodes(struct rcu_ht *ht, |
759 | unsigned long *count, | |
760 | unsigned long *removed) | |
761 | { | |
762 | struct rcu_table *t; | |
763 | struct rcu_ht_node *node, *next; | |
24365af7 MD |
764 | struct _rcu_ht_node *lookup; |
765 | unsigned long nr_dummy = 0; | |
273399de MD |
766 | |
767 | *count = 0; | |
768 | *removed = 0; | |
769 | ||
770 | t = rcu_dereference(ht->t); | |
24365af7 MD |
771 | /* Count non-dummy nodes in the table */ |
772 | lookup = &t->tbl[0][0]; | |
773 | node = (struct rcu_ht_node *) lookup; | |
273399de | 774 | do { |
cc4fcb10 | 775 | next = rcu_dereference(node->p.next); |
273399de | 776 | if (is_removed(next)) { |
1b81fe1a | 777 | assert(!is_dummy(next)); |
273399de | 778 | (*removed)++; |
1b81fe1a | 779 | } else if (!is_dummy(next)) |
273399de | 780 | (*count)++; |
24365af7 MD |
781 | else |
782 | (nr_dummy)++; | |
273399de MD |
783 | node = clear_flag(next); |
784 | } while (node); | |
24365af7 | 785 | dbg_printf("rculfhash: number of dummy nodes: %lu\n", nr_dummy); |
273399de MD |
786 | } |
787 | ||
abc490a1 MD |
788 | static |
789 | void ht_free_table_cb(struct rcu_head *head) | |
790 | { | |
791 | struct rcu_table *t = | |
792 | caa_container_of(head, struct rcu_table, head); | |
793 | free(t); | |
794 | } | |
795 | ||
796 | /* called with resize mutex held */ | |
797 | static | |
798 | void _do_ht_resize(struct rcu_ht *ht) | |
464a1ec9 | 799 | { |
24365af7 | 800 | unsigned long new_size, old_size, old_order, new_order; |
395270b6 | 801 | struct rcu_table *new_t, *old_t; |
464a1ec9 | 802 | |
395270b6 MD |
803 | old_t = ht->t; |
804 | old_size = old_t->size; | |
24365af7 | 805 | old_order = get_count_order_ulong(old_size) + 1; |
464a1ec9 | 806 | |
f9830efd | 807 | new_size = CMM_LOAD_SHARED(old_t->resize_target); |
abc490a1 | 808 | if (old_size == new_size) |
464a1ec9 | 809 | return; |
24365af7 | 810 | new_order = get_count_order_ulong(new_size) + 1; |
df03fab8 MD |
811 | printf("rculfhash: resize from %lu (order %lu) to %lu (order %lu) buckets\n", |
812 | old_size, old_order, new_size, new_order); | |
f000907d | 813 | new_t = malloc(sizeof(struct rcu_table) |
24365af7 | 814 | + (new_order * sizeof(struct _rcu_ht_node *))); |
f000907d MD |
815 | assert(new_size > old_size); |
816 | memcpy(&new_t->tbl, &old_t->tbl, | |
24365af7 MD |
817 | old_order * sizeof(struct _rcu_ht_node *)); |
818 | init_table(ht, new_t, old_order, new_order - old_order); | |
f000907d MD |
819 | /* Changing table and size atomically wrt lookups */ |
820 | rcu_assign_pointer(ht->t, new_t); | |
821 | ht->ht_call_rcu(&old_t->head, ht_free_table_cb); | |
464a1ec9 MD |
822 | } |
823 | ||
abc490a1 | 824 | static |
f9830efd MD |
825 | unsigned long resize_target_update(struct rcu_table *t, |
826 | int growth_order) | |
464a1ec9 | 827 | { |
f9830efd MD |
828 | return _uatomic_max(&t->resize_target, |
829 | t->size << growth_order); | |
464a1ec9 MD |
830 | } |
831 | ||
464a1ec9 MD |
832 | void ht_resize(struct rcu_ht *ht, int growth) |
833 | { | |
f9830efd MD |
834 | struct rcu_table *t = rcu_dereference(ht->t); |
835 | unsigned long target_size; | |
836 | ||
837 | target_size = resize_target_update(t, growth); | |
838 | if (t->size < target_size) { | |
11519af6 | 839 | CMM_STORE_SHARED(t->resize_initiated, 1); |
f9830efd MD |
840 | pthread_mutex_lock(&ht->resize_mutex); |
841 | _do_ht_resize(ht); | |
842 | pthread_mutex_unlock(&ht->resize_mutex); | |
843 | } | |
abc490a1 | 844 | } |
464a1ec9 | 845 | |
abc490a1 MD |
846 | static |
847 | void do_resize_cb(struct rcu_head *head) | |
848 | { | |
849 | struct rcu_resize_work *work = | |
850 | caa_container_of(head, struct rcu_resize_work, head); | |
851 | struct rcu_ht *ht = work->ht; | |
852 | ||
853 | pthread_mutex_lock(&ht->resize_mutex); | |
854 | _do_ht_resize(ht); | |
855 | pthread_mutex_unlock(&ht->resize_mutex); | |
856 | free(work); | |
848d4088 MD |
857 | cmm_smp_mb(); /* finish resize before decrement */ |
858 | uatomic_dec(&ht->in_progress_resize); | |
464a1ec9 MD |
859 | } |
860 | ||
abc490a1 | 861 | static |
f000907d | 862 | void ht_resize_lazy(struct rcu_ht *ht, struct rcu_table *t, int growth) |
ab7d5fc6 | 863 | { |
abc490a1 | 864 | struct rcu_resize_work *work; |
f9830efd | 865 | unsigned long target_size; |
abc490a1 | 866 | |
f9830efd | 867 | target_size = resize_target_update(t, growth); |
11519af6 | 868 | if (!CMM_LOAD_SHARED(t->resize_initiated) && t->size < target_size) { |
848d4088 MD |
869 | uatomic_inc(&ht->in_progress_resize); |
870 | cmm_smp_mb(); /* increment resize count before calling it */ | |
f9830efd MD |
871 | work = malloc(sizeof(*work)); |
872 | work->ht = ht; | |
873 | ht->ht_call_rcu(&work->head, do_resize_cb); | |
11519af6 | 874 | CMM_STORE_SHARED(t->resize_initiated, 1); |
f9830efd | 875 | } |
ab7d5fc6 | 876 | } |