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 | |
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 | ||
98808fb1 MD |
159 | /* For testing */ |
160 | #define POISON_FREE | |
161 | ||
f8994aee MD |
162 | /* |
163 | * Per-CPU split-counters lazily update the global counter each 1024 | |
164 | * addition/removal. It automatically keeps track of resize required. | |
165 | * We use the bucket length as indicator for need to expand for small | |
166 | * tables and machines lacking per-cpu data suppport. | |
167 | */ | |
168 | #define COUNT_COMMIT_ORDER 10 | |
6ea6bc67 MD |
169 | #define CHAIN_LEN_TARGET 1 |
170 | #define CHAIN_LEN_RESIZE_THRESHOLD 3 | |
2ed95849 | 171 | |
cd95516d MD |
172 | /* |
173 | * Define the minimum table size. Protects against hash table resize overload | |
174 | * when too many entries are added quickly before the resize can complete. | |
175 | * This is especially the case if the table could be shrinked to a size of 1. | |
176 | * TODO: we might want to make the add/remove operations help the resize to | |
177 | * add or remove dummy nodes when a resize is ongoing to ensure upper-bound on | |
178 | * chain length. | |
179 | */ | |
180 | #define MIN_TABLE_SIZE 128 | |
181 | ||
abc490a1 MD |
182 | #ifndef max |
183 | #define max(a, b) ((a) > (b) ? (a) : (b)) | |
184 | #endif | |
2ed95849 | 185 | |
d95bd160 MD |
186 | /* |
187 | * The removed flag needs to be updated atomically with the pointer. | |
188 | * The dummy flag does not require to be updated atomically with the | |
189 | * pointer, but it is added as a pointer low bit flag to save space. | |
190 | */ | |
d37166c6 | 191 | #define REMOVED_FLAG (1UL << 0) |
f5596c94 MD |
192 | #define DUMMY_FLAG (1UL << 1) |
193 | #define FLAGS_MASK ((1UL << 2) - 1) | |
d37166c6 | 194 | |
df44348d | 195 | struct ht_items_count { |
3171717f | 196 | unsigned long add, remove; |
df44348d MD |
197 | } __attribute__((aligned(CAA_CACHE_LINE_SIZE))); |
198 | ||
1475579c MD |
199 | struct rcu_level { |
200 | struct rcu_head head; | |
201 | struct _cds_lfht_node nodes[0]; | |
202 | }; | |
203 | ||
395270b6 | 204 | struct rcu_table { |
abc490a1 | 205 | unsigned long size; /* always a power of 2 */ |
f9830efd | 206 | unsigned long resize_target; |
11519af6 | 207 | int resize_initiated; |
abc490a1 | 208 | struct rcu_head head; |
1475579c | 209 | struct rcu_level *tbl[0]; |
395270b6 MD |
210 | }; |
211 | ||
14044b37 | 212 | struct cds_lfht { |
395270b6 | 213 | struct rcu_table *t; /* shared */ |
14044b37 MD |
214 | cds_lfht_hash_fct hash_fct; |
215 | cds_lfht_compare_fct compare_fct; | |
732ad076 | 216 | unsigned long hash_seed; |
b8af5011 | 217 | int flags; |
464a1ec9 | 218 | pthread_mutex_t resize_mutex; /* resize mutex: add/del mutex */ |
33c7c748 | 219 | unsigned int in_progress_resize, in_progress_destroy; |
14044b37 | 220 | void (*cds_lfht_call_rcu)(struct rcu_head *head, |
abc490a1 | 221 | void (*func)(struct rcu_head *head)); |
1475579c | 222 | void (*cds_lfht_synchronize_rcu)(void); |
01dbfa62 MD |
223 | void (*cds_lfht_rcu_read_lock)(void); |
224 | void (*cds_lfht_rcu_read_unlock)(void); | |
df44348d MD |
225 | unsigned long count; /* global approximate item count */ |
226 | struct ht_items_count *percpu_count; /* per-cpu item count */ | |
2ed95849 MD |
227 | }; |
228 | ||
abc490a1 MD |
229 | struct rcu_resize_work { |
230 | struct rcu_head head; | |
14044b37 | 231 | struct cds_lfht *ht; |
abc490a1 | 232 | }; |
2ed95849 | 233 | |
abc490a1 MD |
234 | /* |
235 | * Algorithm to reverse bits in a word by lookup table, extended to | |
236 | * 64-bit words. | |
f9830efd | 237 | * Source: |
abc490a1 | 238 | * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable |
f9830efd | 239 | * Originally from Public Domain. |
abc490a1 MD |
240 | */ |
241 | ||
242 | static const uint8_t BitReverseTable256[256] = | |
2ed95849 | 243 | { |
abc490a1 MD |
244 | #define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64 |
245 | #define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16) | |
246 | #define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 ) | |
247 | R6(0), R6(2), R6(1), R6(3) | |
248 | }; | |
249 | #undef R2 | |
250 | #undef R4 | |
251 | #undef R6 | |
2ed95849 | 252 | |
abc490a1 MD |
253 | static |
254 | uint8_t bit_reverse_u8(uint8_t v) | |
255 | { | |
256 | return BitReverseTable256[v]; | |
257 | } | |
ab7d5fc6 | 258 | |
abc490a1 MD |
259 | static __attribute__((unused)) |
260 | uint32_t bit_reverse_u32(uint32_t v) | |
261 | { | |
262 | return ((uint32_t) bit_reverse_u8(v) << 24) | | |
263 | ((uint32_t) bit_reverse_u8(v >> 8) << 16) | | |
264 | ((uint32_t) bit_reverse_u8(v >> 16) << 8) | | |
265 | ((uint32_t) bit_reverse_u8(v >> 24)); | |
2ed95849 MD |
266 | } |
267 | ||
abc490a1 MD |
268 | static __attribute__((unused)) |
269 | uint64_t bit_reverse_u64(uint64_t v) | |
2ed95849 | 270 | { |
abc490a1 MD |
271 | return ((uint64_t) bit_reverse_u8(v) << 56) | |
272 | ((uint64_t) bit_reverse_u8(v >> 8) << 48) | | |
273 | ((uint64_t) bit_reverse_u8(v >> 16) << 40) | | |
274 | ((uint64_t) bit_reverse_u8(v >> 24) << 32) | | |
275 | ((uint64_t) bit_reverse_u8(v >> 32) << 24) | | |
276 | ((uint64_t) bit_reverse_u8(v >> 40) << 16) | | |
277 | ((uint64_t) bit_reverse_u8(v >> 48) << 8) | | |
278 | ((uint64_t) bit_reverse_u8(v >> 56)); | |
279 | } | |
280 | ||
281 | static | |
282 | unsigned long bit_reverse_ulong(unsigned long v) | |
283 | { | |
284 | #if (CAA_BITS_PER_LONG == 32) | |
285 | return bit_reverse_u32(v); | |
286 | #else | |
287 | return bit_reverse_u64(v); | |
288 | #endif | |
289 | } | |
290 | ||
f9830efd | 291 | /* |
24365af7 MD |
292 | * fls: returns the position of the most significant bit. |
293 | * Returns 0 if no bit is set, else returns the position of the most | |
294 | * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit). | |
f9830efd | 295 | */ |
24365af7 MD |
296 | #if defined(__i386) || defined(__x86_64) |
297 | static inline | |
298 | unsigned int fls_u32(uint32_t x) | |
f9830efd | 299 | { |
24365af7 MD |
300 | int r; |
301 | ||
302 | asm("bsrl %1,%0\n\t" | |
303 | "jnz 1f\n\t" | |
304 | "movl $-1,%0\n\t" | |
305 | "1:\n\t" | |
306 | : "=r" (r) : "rm" (x)); | |
307 | return r + 1; | |
308 | } | |
309 | #define HAS_FLS_U32 | |
310 | #endif | |
311 | ||
312 | #if defined(__x86_64) | |
313 | static inline | |
314 | unsigned int fls_u64(uint64_t x) | |
315 | { | |
316 | long r; | |
317 | ||
318 | asm("bsrq %1,%0\n\t" | |
319 | "jnz 1f\n\t" | |
320 | "movq $-1,%0\n\t" | |
321 | "1:\n\t" | |
322 | : "=r" (r) : "rm" (x)); | |
323 | return r + 1; | |
324 | } | |
325 | #define HAS_FLS_U64 | |
326 | #endif | |
327 | ||
328 | #ifndef HAS_FLS_U64 | |
329 | static __attribute__((unused)) | |
330 | unsigned int fls_u64(uint64_t x) | |
331 | { | |
332 | unsigned int r = 64; | |
333 | ||
334 | if (!x) | |
335 | return 0; | |
336 | ||
337 | if (!(x & 0xFFFFFFFF00000000ULL)) { | |
338 | x <<= 32; | |
339 | r -= 32; | |
340 | } | |
341 | if (!(x & 0xFFFF000000000000ULL)) { | |
342 | x <<= 16; | |
343 | r -= 16; | |
344 | } | |
345 | if (!(x & 0xFF00000000000000ULL)) { | |
346 | x <<= 8; | |
347 | r -= 8; | |
348 | } | |
349 | if (!(x & 0xF000000000000000ULL)) { | |
350 | x <<= 4; | |
351 | r -= 4; | |
352 | } | |
353 | if (!(x & 0xC000000000000000ULL)) { | |
354 | x <<= 2; | |
355 | r -= 2; | |
356 | } | |
357 | if (!(x & 0x8000000000000000ULL)) { | |
358 | x <<= 1; | |
359 | r -= 1; | |
360 | } | |
361 | return r; | |
362 | } | |
363 | #endif | |
364 | ||
365 | #ifndef HAS_FLS_U32 | |
366 | static __attribute__((unused)) | |
367 | unsigned int fls_u32(uint32_t x) | |
368 | { | |
369 | unsigned int r = 32; | |
f9830efd | 370 | |
24365af7 MD |
371 | if (!x) |
372 | return 0; | |
373 | if (!(x & 0xFFFF0000U)) { | |
374 | x <<= 16; | |
375 | r -= 16; | |
376 | } | |
377 | if (!(x & 0xFF000000U)) { | |
378 | x <<= 8; | |
379 | r -= 8; | |
380 | } | |
381 | if (!(x & 0xF0000000U)) { | |
382 | x <<= 4; | |
383 | r -= 4; | |
384 | } | |
385 | if (!(x & 0xC0000000U)) { | |
386 | x <<= 2; | |
387 | r -= 2; | |
388 | } | |
389 | if (!(x & 0x80000000U)) { | |
390 | x <<= 1; | |
391 | r -= 1; | |
392 | } | |
393 | return r; | |
394 | } | |
395 | #endif | |
396 | ||
397 | unsigned int fls_ulong(unsigned long x) | |
f9830efd | 398 | { |
24365af7 MD |
399 | #if (CAA_BITS_PER_lONG == 32) |
400 | return fls_u32(x); | |
401 | #else | |
402 | return fls_u64(x); | |
403 | #endif | |
404 | } | |
f9830efd | 405 | |
24365af7 MD |
406 | int get_count_order_u32(uint32_t x) |
407 | { | |
408 | int order; | |
409 | ||
410 | order = fls_u32(x) - 1; | |
411 | if (x & (x - 1)) | |
412 | order++; | |
413 | return order; | |
414 | } | |
415 | ||
416 | int get_count_order_ulong(unsigned long x) | |
417 | { | |
418 | int order; | |
419 | ||
420 | order = fls_ulong(x) - 1; | |
421 | if (x & (x - 1)) | |
422 | order++; | |
423 | return order; | |
f9830efd MD |
424 | } |
425 | ||
98808fb1 MD |
426 | #ifdef POISON_FREE |
427 | #define poison_free(ptr) \ | |
428 | do { \ | |
429 | memset(ptr, 0x42, sizeof(*(ptr))); \ | |
430 | free(ptr); \ | |
431 | } while (0) | |
432 | #else | |
433 | #define poison_free(ptr) free(ptr) | |
434 | #endif | |
435 | ||
f9830efd | 436 | static |
14044b37 | 437 | void cds_lfht_resize_lazy(struct cds_lfht *ht, struct rcu_table *t, int growth); |
f9830efd | 438 | |
df44348d MD |
439 | /* |
440 | * If the sched_getcpu() and sysconf(_SC_NPROCESSORS_CONF) calls are | |
441 | * available, then we support hash table item accounting. | |
442 | * In the unfortunate event the number of CPUs reported would be | |
443 | * inaccurate, we use modulo arithmetic on the number of CPUs we got. | |
444 | */ | |
df44348d MD |
445 | #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) |
446 | ||
f8994aee MD |
447 | static |
448 | void cds_lfht_resize_lazy_count(struct cds_lfht *ht, struct rcu_table *t, | |
449 | unsigned long count); | |
450 | ||
df44348d MD |
451 | static long nr_cpus_mask = -1; |
452 | ||
453 | static | |
454 | struct ht_items_count *alloc_per_cpu_items_count(void) | |
455 | { | |
456 | struct ht_items_count *count; | |
457 | ||
458 | switch (nr_cpus_mask) { | |
459 | case -2: | |
460 | return NULL; | |
461 | case -1: | |
462 | { | |
463 | long maxcpus; | |
464 | ||
465 | maxcpus = sysconf(_SC_NPROCESSORS_CONF); | |
466 | if (maxcpus <= 0) { | |
467 | nr_cpus_mask = -2; | |
468 | return NULL; | |
469 | } | |
470 | /* | |
471 | * round up number of CPUs to next power of two, so we | |
472 | * can use & for modulo. | |
473 | */ | |
474 | maxcpus = 1UL << get_count_order_ulong(maxcpus); | |
475 | nr_cpus_mask = maxcpus - 1; | |
476 | } | |
477 | /* Fall-through */ | |
478 | default: | |
479 | return calloc(nr_cpus_mask + 1, sizeof(*count)); | |
480 | } | |
481 | } | |
482 | ||
483 | static | |
484 | void free_per_cpu_items_count(struct ht_items_count *count) | |
485 | { | |
98808fb1 | 486 | poison_free(count); |
df44348d MD |
487 | } |
488 | ||
489 | static | |
490 | int ht_get_cpu(void) | |
491 | { | |
492 | int cpu; | |
493 | ||
494 | assert(nr_cpus_mask >= 0); | |
495 | cpu = sched_getcpu(); | |
496 | if (unlikely(cpu < 0)) | |
497 | return cpu; | |
498 | else | |
499 | return cpu & nr_cpus_mask; | |
500 | } | |
501 | ||
502 | static | |
3171717f | 503 | void ht_count_add(struct cds_lfht *ht, struct rcu_table *t) |
df44348d | 504 | { |
3171717f | 505 | unsigned long percpu_count; |
df44348d MD |
506 | int cpu; |
507 | ||
508 | if (unlikely(!ht->percpu_count)) | |
3171717f | 509 | return; |
df44348d MD |
510 | cpu = ht_get_cpu(); |
511 | if (unlikely(cpu < 0)) | |
3171717f MD |
512 | return; |
513 | percpu_count = uatomic_add_return(&ht->percpu_count[cpu].add, 1); | |
df44348d MD |
514 | if (unlikely(!(percpu_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) { |
515 | unsigned long count; | |
516 | ||
517 | dbg_printf("add percpu %lu\n", percpu_count); | |
518 | count = uatomic_add_return(&ht->count, | |
519 | 1UL << COUNT_COMMIT_ORDER); | |
520 | /* If power of 2 */ | |
521 | if (!(count & (count - 1))) { | |
f8994aee MD |
522 | if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) |
523 | < t->size) | |
524 | return; | |
525 | dbg_printf("add set global %lu\n", count); | |
526 | cds_lfht_resize_lazy_count(ht, t, | |
6ea6bc67 | 527 | count >> (CHAIN_LEN_TARGET - 1)); |
df44348d MD |
528 | } |
529 | } | |
530 | } | |
531 | ||
532 | static | |
533 | void ht_count_remove(struct cds_lfht *ht, struct rcu_table *t) | |
534 | { | |
535 | unsigned long percpu_count; | |
3171717f | 536 | int cpu; |
df44348d | 537 | |
3171717f MD |
538 | if (unlikely(!ht->percpu_count)) |
539 | return; | |
540 | cpu = ht_get_cpu(); | |
541 | if (unlikely(cpu < 0)) | |
542 | return; | |
543 | percpu_count = uatomic_add_return(&ht->percpu_count[cpu].remove, -1); | |
df44348d MD |
544 | if (unlikely(!(percpu_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) { |
545 | unsigned long count; | |
546 | ||
547 | dbg_printf("remove percpu %lu\n", percpu_count); | |
548 | count = uatomic_add_return(&ht->count, | |
3171717f | 549 | -(1UL << COUNT_COMMIT_ORDER)); |
df44348d MD |
550 | /* If power of 2 */ |
551 | if (!(count & (count - 1))) { | |
f8994aee MD |
552 | if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) |
553 | >= t->size) | |
554 | return; | |
555 | dbg_printf("remove set global %lu\n", count); | |
556 | cds_lfht_resize_lazy_count(ht, t, | |
6ea6bc67 | 557 | count >> (CHAIN_LEN_TARGET - 1)); |
df44348d MD |
558 | } |
559 | } | |
560 | } | |
561 | ||
562 | #else /* #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */ | |
563 | ||
564 | static const long nr_cpus_mask = -1; | |
565 | ||
566 | static | |
567 | struct ht_items_count *alloc_per_cpu_items_count(void) | |
568 | { | |
569 | return NULL; | |
570 | } | |
571 | ||
572 | static | |
573 | void free_per_cpu_items_count(struct ht_items_count *count) | |
574 | { | |
575 | } | |
576 | ||
577 | static | |
f8994aee | 578 | void ht_count_add(struct cds_lfht *ht, struct rcu_table *t) |
df44348d MD |
579 | { |
580 | } | |
581 | ||
582 | static | |
f8994aee | 583 | void ht_count_remove(struct cds_lfht *ht, struct rcu_table *t) |
df44348d MD |
584 | { |
585 | } | |
586 | ||
587 | #endif /* #else #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */ | |
588 | ||
589 | ||
f9830efd | 590 | static |
14044b37 | 591 | void check_resize(struct cds_lfht *ht, struct rcu_table *t, |
f9830efd MD |
592 | uint32_t chain_len) |
593 | { | |
f8994aee MD |
594 | unsigned long count; |
595 | ||
b8af5011 MD |
596 | if (!(ht->flags & CDS_LFHT_AUTO_RESIZE)) |
597 | return; | |
f8994aee MD |
598 | count = uatomic_read(&ht->count); |
599 | /* | |
600 | * Use bucket-local length for small table expand and for | |
601 | * environments lacking per-cpu data support. | |
602 | */ | |
603 | if (count >= (1UL << COUNT_COMMIT_ORDER)) | |
604 | return; | |
24365af7 | 605 | if (chain_len > 100) |
f0c29ed7 | 606 | dbg_printf("WARNING: large chain length: %u.\n", |
24365af7 | 607 | chain_len); |
3390d470 | 608 | if (chain_len >= CHAIN_LEN_RESIZE_THRESHOLD) |
14044b37 | 609 | cds_lfht_resize_lazy(ht, t, |
01370f0b | 610 | get_count_order_u32(chain_len - (CHAIN_LEN_TARGET - 1))); |
f9830efd MD |
611 | } |
612 | ||
abc490a1 | 613 | static |
14044b37 | 614 | struct cds_lfht_node *clear_flag(struct cds_lfht_node *node) |
abc490a1 | 615 | { |
14044b37 | 616 | return (struct cds_lfht_node *) (((unsigned long) node) & ~FLAGS_MASK); |
abc490a1 MD |
617 | } |
618 | ||
619 | static | |
14044b37 | 620 | int is_removed(struct cds_lfht_node *node) |
abc490a1 | 621 | { |
d37166c6 | 622 | return ((unsigned long) node) & REMOVED_FLAG; |
abc490a1 MD |
623 | } |
624 | ||
625 | static | |
14044b37 | 626 | struct cds_lfht_node *flag_removed(struct cds_lfht_node *node) |
abc490a1 | 627 | { |
14044b37 | 628 | return (struct cds_lfht_node *) (((unsigned long) node) | REMOVED_FLAG); |
abc490a1 MD |
629 | } |
630 | ||
f5596c94 | 631 | static |
14044b37 | 632 | int is_dummy(struct cds_lfht_node *node) |
f5596c94 MD |
633 | { |
634 | return ((unsigned long) node) & DUMMY_FLAG; | |
635 | } | |
636 | ||
637 | static | |
14044b37 | 638 | struct cds_lfht_node *flag_dummy(struct cds_lfht_node *node) |
f5596c94 | 639 | { |
14044b37 | 640 | return (struct cds_lfht_node *) (((unsigned long) node) | DUMMY_FLAG); |
f5596c94 MD |
641 | } |
642 | ||
abc490a1 | 643 | static |
f9830efd | 644 | unsigned long _uatomic_max(unsigned long *ptr, unsigned long v) |
abc490a1 MD |
645 | { |
646 | unsigned long old1, old2; | |
647 | ||
648 | old1 = uatomic_read(ptr); | |
649 | do { | |
650 | old2 = old1; | |
651 | if (old2 >= v) | |
f9830efd | 652 | return old2; |
abc490a1 | 653 | } while ((old1 = uatomic_cmpxchg(ptr, old2, v)) != old2); |
f9830efd | 654 | return v; |
abc490a1 MD |
655 | } |
656 | ||
1475579c MD |
657 | static |
658 | void cds_lfht_free_table_cb(struct rcu_head *head) | |
659 | { | |
660 | struct rcu_table *t = | |
661 | caa_container_of(head, struct rcu_table, head); | |
98808fb1 | 662 | poison_free(t); |
1475579c MD |
663 | } |
664 | ||
665 | static | |
666 | void cds_lfht_free_level(struct rcu_head *head) | |
667 | { | |
668 | struct rcu_level *l = | |
669 | caa_container_of(head, struct rcu_level, head); | |
98808fb1 | 670 | poison_free(l); |
1475579c MD |
671 | } |
672 | ||
273399de MD |
673 | /* |
674 | * Remove all logically deleted nodes from a bucket up to a certain node key. | |
675 | */ | |
676 | static | |
14044b37 | 677 | void _cds_lfht_gc_bucket(struct cds_lfht_node *dummy, struct cds_lfht_node *node) |
273399de | 678 | { |
14044b37 | 679 | struct cds_lfht_node *iter_prev, *iter, *next, *new_next; |
53960c9f MD |
680 | struct cds_lfht_node *iter_trace[64]; |
681 | unsigned long trace_idx = 0; | |
273399de | 682 | |
53960c9f | 683 | memset(iter_trace, 0, sizeof(iter_trace)); |
c90201ac MD |
684 | assert(!is_dummy(dummy)); |
685 | assert(!is_removed(dummy)); | |
686 | assert(!is_dummy(node)); | |
687 | assert(!is_removed(node)); | |
273399de | 688 | for (;;) { |
53960c9f | 689 | iter_trace[trace_idx++ & (64 - 1)] = (void *) 0x1; |
273399de MD |
690 | iter_prev = dummy; |
691 | /* We can always skip the dummy node initially */ | |
cc4fcb10 | 692 | iter = rcu_dereference(iter_prev->p.next); |
53960c9f | 693 | iter_trace[trace_idx++ & (64 - 1)] = iter; |
cc4fcb10 | 694 | assert(iter_prev->p.reverse_hash <= node->p.reverse_hash); |
bd4db153 MD |
695 | /* |
696 | * We should never be called with dummy (start of chain) | |
697 | * and logically removed node (end of path compression | |
698 | * marker) being the actual same node. This would be a | |
699 | * bug in the algorithm implementation. | |
700 | */ | |
701 | assert(dummy != node); | |
273399de | 702 | for (;;) { |
a2974903 | 703 | if (unlikely(!clear_flag(iter))) |
479c8a32 | 704 | return; |
76412f24 | 705 | if (likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash)) |
273399de | 706 | return; |
cc4fcb10 | 707 | next = rcu_dereference(clear_flag(iter)->p.next); |
76412f24 | 708 | if (likely(is_removed(next))) |
273399de | 709 | break; |
b453eae1 | 710 | iter_prev = clear_flag(iter); |
273399de | 711 | iter = next; |
53960c9f | 712 | iter_trace[trace_idx++ & (64 - 1)] = iter; |
273399de MD |
713 | } |
714 | assert(!is_removed(iter)); | |
f5596c94 MD |
715 | if (is_dummy(iter)) |
716 | new_next = flag_dummy(clear_flag(next)); | |
717 | else | |
718 | new_next = clear_flag(next); | |
719 | (void) uatomic_cmpxchg(&iter_prev->p.next, iter, new_next); | |
53960c9f | 720 | iter_trace[trace_idx++ & (64 - 1)] = (void *) 0x2; |
273399de MD |
721 | } |
722 | } | |
723 | ||
abc490a1 | 724 | static |
14044b37 MD |
725 | struct cds_lfht_node *_cds_lfht_add(struct cds_lfht *ht, struct rcu_table *t, |
726 | struct cds_lfht_node *node, int unique, int dummy) | |
abc490a1 | 727 | { |
14044b37 | 728 | struct cds_lfht_node *iter_prev, *iter, *next, *new_node, *new_next, |
f5596c94 | 729 | *dummy_node; |
14044b37 | 730 | struct _cds_lfht_node *lookup; |
24365af7 | 731 | unsigned long hash, index, order; |
abc490a1 | 732 | |
c90201ac MD |
733 | assert(!is_dummy(node)); |
734 | assert(!is_removed(node)); | |
18117871 | 735 | if (!t->size) { |
f5596c94 MD |
736 | assert(dummy); |
737 | node->p.next = flag_dummy(NULL); | |
18117871 MD |
738 | return node; /* Initial first add (head) */ |
739 | } | |
cc4fcb10 | 740 | hash = bit_reverse_ulong(node->p.reverse_hash); |
abc490a1 | 741 | for (;;) { |
f9830efd | 742 | uint32_t chain_len = 0; |
abc490a1 | 743 | |
11519af6 MD |
744 | /* |
745 | * iter_prev points to the non-removed node prior to the | |
746 | * insert location. | |
11519af6 | 747 | */ |
24365af7 MD |
748 | index = hash & (t->size - 1); |
749 | order = get_count_order_ulong(index + 1); | |
554c284e | 750 | lookup = &t->tbl[order]->nodes[index & ((!order ? 0 : (1UL << (order - 1))) - 1)]; |
14044b37 | 751 | iter_prev = (struct cds_lfht_node *) lookup; |
11519af6 | 752 | /* We can always skip the dummy node initially */ |
cc4fcb10 MD |
753 | iter = rcu_dereference(iter_prev->p.next); |
754 | assert(iter_prev->p.reverse_hash <= node->p.reverse_hash); | |
abc490a1 | 755 | for (;;) { |
c90201ac | 756 | /* TODO: check if removed */ |
a2974903 | 757 | if (unlikely(!clear_flag(iter))) |
273399de | 758 | goto insert; |
c90201ac | 759 | /* TODO: check if removed */ |
76412f24 | 760 | if (likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash)) |
273399de | 761 | goto insert; |
cc4fcb10 | 762 | next = rcu_dereference(clear_flag(iter)->p.next); |
76412f24 | 763 | if (unlikely(is_removed(next))) |
9dba85be | 764 | goto gc_node; |
e43f23f8 | 765 | if (unique |
1b81fe1a | 766 | && !is_dummy(next) |
e43f23f8 MD |
767 | && !ht->compare_fct(node->key, node->key_len, |
768 | clear_flag(iter)->key, | |
769 | clear_flag(iter)->key_len)) | |
18117871 | 770 | return clear_flag(iter); |
11519af6 | 771 | /* Only account for identical reverse hash once */ |
24365af7 MD |
772 | if (iter_prev->p.reverse_hash != clear_flag(iter)->p.reverse_hash |
773 | && !is_dummy(next)) | |
11519af6 MD |
774 | check_resize(ht, t, ++chain_len); |
775 | iter_prev = clear_flag(iter); | |
273399de | 776 | iter = next; |
abc490a1 | 777 | } |
273399de | 778 | insert: |
7ec59d3b | 779 | assert(node != clear_flag(iter)); |
11519af6 | 780 | assert(!is_removed(iter_prev)); |
c90201ac | 781 | assert(!is_removed(iter)); |
f000907d | 782 | assert(iter_prev != node); |
f5596c94 | 783 | if (!dummy) |
1b81fe1a | 784 | node->p.next = clear_flag(iter); |
f5596c94 | 785 | else |
1b81fe1a | 786 | node->p.next = flag_dummy(clear_flag(iter)); |
f5596c94 MD |
787 | if (is_dummy(iter)) |
788 | new_node = flag_dummy(node); | |
789 | else | |
790 | new_node = node; | |
cc4fcb10 | 791 | if (uatomic_cmpxchg(&iter_prev->p.next, iter, |
f5596c94 | 792 | new_node) != iter) |
273399de | 793 | continue; /* retry */ |
11519af6 | 794 | else |
273399de | 795 | goto gc_end; |
9dba85be MD |
796 | gc_node: |
797 | assert(!is_removed(iter)); | |
f5596c94 MD |
798 | if (is_dummy(iter)) |
799 | new_next = flag_dummy(clear_flag(next)); | |
800 | else | |
801 | new_next = clear_flag(next); | |
802 | (void) uatomic_cmpxchg(&iter_prev->p.next, iter, new_next); | |
273399de | 803 | /* retry */ |
464a1ec9 | 804 | } |
273399de MD |
805 | gc_end: |
806 | /* Garbage collect logically removed nodes in the bucket */ | |
24365af7 MD |
807 | index = hash & (t->size - 1); |
808 | order = get_count_order_ulong(index + 1); | |
554c284e | 809 | lookup = &t->tbl[order]->nodes[index & (!order ? 0 : ((1UL << (order - 1)) - 1))]; |
14044b37 MD |
810 | dummy_node = (struct cds_lfht_node *) lookup; |
811 | _cds_lfht_gc_bucket(dummy_node, node); | |
18117871 | 812 | return node; |
abc490a1 | 813 | } |
464a1ec9 | 814 | |
abc490a1 | 815 | static |
14044b37 | 816 | int _cds_lfht_remove(struct cds_lfht *ht, struct rcu_table *t, |
1475579c | 817 | struct cds_lfht_node *node, int dummy_removal) |
abc490a1 | 818 | { |
14044b37 MD |
819 | struct cds_lfht_node *dummy, *next, *old; |
820 | struct _cds_lfht_node *lookup; | |
abc490a1 | 821 | int flagged = 0; |
24365af7 | 822 | unsigned long hash, index, order; |
5e28c532 | 823 | |
7ec59d3b | 824 | /* logically delete the node */ |
c90201ac MD |
825 | assert(!is_dummy(node)); |
826 | assert(!is_removed(node)); | |
cc4fcb10 | 827 | old = rcu_dereference(node->p.next); |
7ec59d3b MD |
828 | do { |
829 | next = old; | |
76412f24 | 830 | if (unlikely(is_removed(next))) |
7ec59d3b | 831 | goto end; |
1475579c MD |
832 | if (dummy_removal) |
833 | assert(is_dummy(next)); | |
834 | else | |
835 | assert(!is_dummy(next)); | |
cc4fcb10 | 836 | old = uatomic_cmpxchg(&node->p.next, next, |
7ec59d3b MD |
837 | flag_removed(next)); |
838 | } while (old != next); | |
839 | ||
840 | /* We performed the (logical) deletion. */ | |
841 | flagged = 1; | |
842 | ||
843 | /* | |
844 | * Ensure that the node is not visible to readers anymore: lookup for | |
273399de MD |
845 | * the node, and remove it (along with any other logically removed node) |
846 | * if found. | |
11519af6 | 847 | */ |
cc4fcb10 | 848 | hash = bit_reverse_ulong(node->p.reverse_hash); |
4c3a0329 MD |
849 | assert(t->size > 0); |
850 | index = hash & (t->size - 1); | |
24365af7 | 851 | order = get_count_order_ulong(index + 1); |
554c284e | 852 | lookup = &t->tbl[order]->nodes[index & (!order ? 0 : ((1UL << (order - 1)) - 1))]; |
14044b37 MD |
853 | dummy = (struct cds_lfht_node *) lookup; |
854 | _cds_lfht_gc_bucket(dummy, node); | |
2ed95849 | 855 | end: |
11519af6 MD |
856 | /* |
857 | * Only the flagging action indicated that we (and no other) | |
858 | * removed the node from the hash. | |
859 | */ | |
7ec59d3b | 860 | if (flagged) { |
cc4fcb10 | 861 | assert(is_removed(rcu_dereference(node->p.next))); |
11519af6 | 862 | return 0; |
7ec59d3b | 863 | } else |
11519af6 | 864 | return -ENOENT; |
abc490a1 | 865 | } |
2ed95849 | 866 | |
01dbfa62 MD |
867 | /* |
868 | * Holding RCU read lock to protect _cds_lfht_add against memory | |
869 | * reclaim that could be performed by other call_rcu worker threads (ABA | |
870 | * problem). | |
871 | */ | |
abc490a1 | 872 | static |
14044b37 | 873 | void init_table(struct cds_lfht *ht, struct rcu_table *t, |
24365af7 MD |
874 | unsigned long first_order, unsigned long len_order) |
875 | { | |
876 | unsigned long i, end_order; | |
877 | ||
f0c29ed7 | 878 | dbg_printf("init table: first_order %lu end_order %lu\n", |
24365af7 MD |
879 | first_order, first_order + len_order); |
880 | end_order = first_order + len_order; | |
881 | t->size = !first_order ? 0 : (1UL << (first_order - 1)); | |
882 | for (i = first_order; i < end_order; i++) { | |
883 | unsigned long j, len; | |
884 | ||
885 | len = !i ? 1 : 1UL << (i - 1); | |
f0c29ed7 | 886 | dbg_printf("init order %lu len: %lu\n", i, len); |
1475579c MD |
887 | t->tbl[i] = calloc(1, sizeof(struct rcu_level) |
888 | + (len * sizeof(struct _cds_lfht_node))); | |
01dbfa62 | 889 | ht->cds_lfht_rcu_read_lock(); |
24365af7 | 890 | for (j = 0; j < len; j++) { |
1475579c MD |
891 | struct cds_lfht_node *new_node = |
892 | (struct cds_lfht_node *) &t->tbl[i]->nodes[j]; | |
893 | ||
f0c29ed7 | 894 | dbg_printf("init entry: i %lu j %lu hash %lu\n", |
24365af7 | 895 | i, j, !i ? 0 : (1UL << (i - 1)) + j); |
24365af7 MD |
896 | new_node->p.reverse_hash = |
897 | bit_reverse_ulong(!i ? 0 : (1UL << (i - 1)) + j); | |
14044b37 | 898 | (void) _cds_lfht_add(ht, t, new_node, 0, 1); |
33c7c748 MD |
899 | if (CMM_LOAD_SHARED(ht->in_progress_destroy)) |
900 | break; | |
24365af7 | 901 | } |
01dbfa62 | 902 | ht->cds_lfht_rcu_read_unlock(); |
24365af7 MD |
903 | /* Update table size */ |
904 | t->size = !i ? 1 : (1UL << i); | |
f0c29ed7 | 905 | dbg_printf("init new size: %lu\n", t->size); |
33c7c748 MD |
906 | if (CMM_LOAD_SHARED(ht->in_progress_destroy)) |
907 | break; | |
abc490a1 | 908 | } |
24365af7 | 909 | t->resize_target = t->size; |
11519af6 | 910 | t->resize_initiated = 0; |
2ed95849 MD |
911 | } |
912 | ||
01dbfa62 MD |
913 | /* |
914 | * Holding RCU read lock to protect _cds_lfht_remove against memory | |
915 | * reclaim that could be performed by other call_rcu worker threads (ABA | |
916 | * problem). | |
917 | */ | |
1475579c MD |
918 | static |
919 | void fini_table(struct cds_lfht *ht, struct rcu_table *t, | |
920 | unsigned long first_order, unsigned long len_order) | |
921 | { | |
922 | long i, end_order; | |
923 | ||
924 | dbg_printf("fini table: first_order %lu end_order %lu\n", | |
925 | first_order, first_order + len_order); | |
926 | end_order = first_order + len_order; | |
927 | assert(first_order > 0); | |
928 | assert(t->size == (1UL << (end_order - 1))); | |
929 | for (i = end_order - 1; i >= first_order; i--) { | |
930 | unsigned long j, len; | |
931 | ||
932 | len = !i ? 1 : 1UL << (i - 1); | |
933 | dbg_printf("fini order %lu len: %lu\n", i, len); | |
21263e21 MD |
934 | /* |
935 | * Update table size. Need to shrink this table prior to | |
936 | * removal so gc lookups use non-logically-removed dummy | |
937 | * nodes. | |
938 | */ | |
e1a68f1d | 939 | t->size = 1UL << (i - 1); |
1475579c | 940 | /* Unlink */ |
01dbfa62 | 941 | ht->cds_lfht_rcu_read_lock(); |
1475579c | 942 | for (j = 0; j < len; j++) { |
c90201ac | 943 | struct cds_lfht_node *fini_node = |
1475579c MD |
944 | (struct cds_lfht_node *) &t->tbl[i]->nodes[j]; |
945 | ||
946 | dbg_printf("fini entry: i %lu j %lu hash %lu\n", | |
947 | i, j, !i ? 0 : (1UL << (i - 1)) + j); | |
c90201ac | 948 | fini_node->p.reverse_hash = |
1475579c | 949 | bit_reverse_ulong(!i ? 0 : (1UL << (i - 1)) + j); |
c90201ac | 950 | (void) _cds_lfht_remove(ht, t, fini_node, 1); |
1475579c MD |
951 | if (CMM_LOAD_SHARED(ht->in_progress_destroy)) |
952 | break; | |
953 | } | |
01dbfa62 | 954 | ht->cds_lfht_rcu_read_unlock(); |
1475579c | 955 | ht->cds_lfht_call_rcu(&t->tbl[i]->head, cds_lfht_free_level); |
1475579c MD |
956 | dbg_printf("fini new size: %lu\n", t->size); |
957 | if (CMM_LOAD_SHARED(ht->in_progress_destroy)) | |
958 | break; | |
959 | } | |
960 | t->resize_target = t->size; | |
961 | t->resize_initiated = 0; | |
962 | } | |
963 | ||
14044b37 MD |
964 | struct cds_lfht *cds_lfht_new(cds_lfht_hash_fct hash_fct, |
965 | cds_lfht_compare_fct compare_fct, | |
966 | unsigned long hash_seed, | |
967 | unsigned long init_size, | |
b8af5011 | 968 | int flags, |
14044b37 | 969 | void (*cds_lfht_call_rcu)(struct rcu_head *head, |
1475579c | 970 | void (*func)(struct rcu_head *head)), |
01dbfa62 MD |
971 | void (*cds_lfht_synchronize_rcu)(void), |
972 | void (*cds_lfht_rcu_read_lock)(void), | |
973 | void (*cds_lfht_rcu_read_unlock)(void)) | |
abc490a1 | 974 | { |
14044b37 | 975 | struct cds_lfht *ht; |
24365af7 | 976 | unsigned long order; |
abc490a1 | 977 | |
8129be4e | 978 | /* init_size must be power of two */ |
49619ea0 | 979 | if (init_size && (init_size & (init_size - 1))) |
8129be4e | 980 | return NULL; |
14044b37 | 981 | ht = calloc(1, sizeof(struct cds_lfht)); |
abc490a1 | 982 | ht->hash_fct = hash_fct; |
732ad076 MD |
983 | ht->compare_fct = compare_fct; |
984 | ht->hash_seed = hash_seed; | |
14044b37 | 985 | ht->cds_lfht_call_rcu = cds_lfht_call_rcu; |
1475579c | 986 | ht->cds_lfht_synchronize_rcu = cds_lfht_synchronize_rcu; |
01dbfa62 MD |
987 | ht->cds_lfht_rcu_read_lock = cds_lfht_rcu_read_lock; |
988 | ht->cds_lfht_rcu_read_unlock = cds_lfht_rcu_read_unlock; | |
848d4088 | 989 | ht->in_progress_resize = 0; |
df44348d | 990 | ht->percpu_count = alloc_per_cpu_items_count(); |
abc490a1 MD |
991 | /* this mutex should not nest in read-side C.S. */ |
992 | pthread_mutex_init(&ht->resize_mutex, NULL); | |
cd95516d | 993 | order = get_count_order_ulong(max(init_size, MIN_TABLE_SIZE)) + 1; |
14044b37 | 994 | ht->t = calloc(1, sizeof(struct cds_lfht) |
1475579c | 995 | + (order * sizeof(struct rcu_level *))); |
abc490a1 | 996 | ht->t->size = 0; |
b8af5011 | 997 | ht->flags = flags; |
f000907d | 998 | pthread_mutex_lock(&ht->resize_mutex); |
24365af7 | 999 | init_table(ht, ht->t, 0, order); |
f000907d | 1000 | pthread_mutex_unlock(&ht->resize_mutex); |
abc490a1 MD |
1001 | return ht; |
1002 | } | |
1003 | ||
14044b37 | 1004 | struct cds_lfht_node *cds_lfht_lookup(struct cds_lfht *ht, void *key, size_t key_len) |
2ed95849 | 1005 | { |
395270b6 | 1006 | struct rcu_table *t; |
14044b37 MD |
1007 | struct cds_lfht_node *node, *next; |
1008 | struct _cds_lfht_node *lookup; | |
24365af7 | 1009 | unsigned long hash, reverse_hash, index, order; |
2ed95849 | 1010 | |
732ad076 | 1011 | hash = ht->hash_fct(key, key_len, ht->hash_seed); |
abc490a1 | 1012 | reverse_hash = bit_reverse_ulong(hash); |
464a1ec9 | 1013 | |
395270b6 | 1014 | t = rcu_dereference(ht->t); |
24365af7 MD |
1015 | index = hash & (t->size - 1); |
1016 | order = get_count_order_ulong(index + 1); | |
554c284e | 1017 | lookup = &t->tbl[order]->nodes[index & (!order ? 0 : ((1UL << (order - 1))) - 1)]; |
f0c29ed7 | 1018 | dbg_printf("lookup hash %lu index %lu order %lu aridx %lu\n", |
554c284e | 1019 | hash, index, order, index & (!order ? 0 : ((1UL << (order - 1)) - 1))); |
14044b37 | 1020 | node = (struct cds_lfht_node *) lookup; |
2ed95849 | 1021 | for (;;) { |
abc490a1 MD |
1022 | if (unlikely(!node)) |
1023 | break; | |
cc4fcb10 | 1024 | if (unlikely(node->p.reverse_hash > reverse_hash)) { |
abc490a1 MD |
1025 | node = NULL; |
1026 | break; | |
2ed95849 | 1027 | } |
1b81fe1a MD |
1028 | next = rcu_dereference(node->p.next); |
1029 | if (likely(!is_removed(next)) | |
1030 | && !is_dummy(next) | |
49c2e2d6 | 1031 | && likely(!ht->compare_fct(node->key, node->key_len, key, key_len))) { |
273399de | 1032 | break; |
2ed95849 | 1033 | } |
1b81fe1a | 1034 | node = clear_flag(next); |
2ed95849 | 1035 | } |
1b81fe1a | 1036 | assert(!node || !is_dummy(rcu_dereference(node->p.next))); |
abc490a1 MD |
1037 | return node; |
1038 | } | |
e0ba718a | 1039 | |
a481e5ff MD |
1040 | struct cds_lfht_node *cds_lfht_next(struct cds_lfht *ht, |
1041 | struct cds_lfht_node *node) | |
1042 | { | |
1043 | struct cds_lfht_node *next; | |
1044 | unsigned long reverse_hash; | |
1045 | void *key; | |
1046 | size_t key_len; | |
1047 | ||
1048 | reverse_hash = node->p.reverse_hash; | |
1049 | key = node->key; | |
1050 | key_len = node->key_len; | |
1051 | next = rcu_dereference(node->p.next); | |
1052 | node = clear_flag(next); | |
1053 | ||
1054 | for (;;) { | |
1055 | if (unlikely(!node)) | |
1056 | break; | |
1057 | if (unlikely(node->p.reverse_hash > reverse_hash)) { | |
1058 | node = NULL; | |
1059 | break; | |
1060 | } | |
1061 | next = rcu_dereference(node->p.next); | |
1062 | if (likely(!is_removed(next)) | |
1063 | && !is_dummy(next) | |
1064 | && likely(!ht->compare_fct(node->key, node->key_len, key, key_len))) { | |
1065 | break; | |
1066 | } | |
1067 | node = clear_flag(next); | |
1068 | } | |
1069 | assert(!node || !is_dummy(rcu_dereference(node->p.next))); | |
1070 | return node; | |
1071 | } | |
1072 | ||
14044b37 | 1073 | void cds_lfht_add(struct cds_lfht *ht, struct cds_lfht_node *node) |
abc490a1 MD |
1074 | { |
1075 | struct rcu_table *t; | |
49c2e2d6 | 1076 | unsigned long hash; |
ab7d5fc6 | 1077 | |
49c2e2d6 | 1078 | hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed); |
cc4fcb10 | 1079 | node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash); |
2ed95849 | 1080 | |
abc490a1 | 1081 | t = rcu_dereference(ht->t); |
14044b37 | 1082 | (void) _cds_lfht_add(ht, t, node, 0, 0); |
df44348d | 1083 | ht_count_add(ht, t); |
3eca1b8c MD |
1084 | } |
1085 | ||
14044b37 MD |
1086 | struct cds_lfht_node *cds_lfht_add_unique(struct cds_lfht *ht, |
1087 | struct cds_lfht_node *node) | |
3eca1b8c MD |
1088 | { |
1089 | struct rcu_table *t; | |
49c2e2d6 | 1090 | unsigned long hash; |
df44348d | 1091 | struct cds_lfht_node *ret; |
3eca1b8c | 1092 | |
49c2e2d6 | 1093 | hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed); |
cc4fcb10 | 1094 | node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash); |
3eca1b8c MD |
1095 | |
1096 | t = rcu_dereference(ht->t); | |
df44348d MD |
1097 | ret = _cds_lfht_add(ht, t, node, 1, 0); |
1098 | if (ret != node) | |
1099 | ht_count_add(ht, t); | |
1100 | return ret; | |
2ed95849 MD |
1101 | } |
1102 | ||
14044b37 | 1103 | int cds_lfht_remove(struct cds_lfht *ht, struct cds_lfht_node *node) |
2ed95849 | 1104 | { |
abc490a1 | 1105 | struct rcu_table *t; |
df44348d | 1106 | int ret; |
abc490a1 MD |
1107 | |
1108 | t = rcu_dereference(ht->t); | |
1475579c | 1109 | ret = _cds_lfht_remove(ht, t, node, 0); |
df44348d MD |
1110 | if (!ret) |
1111 | ht_count_remove(ht, t); | |
1112 | return ret; | |
2ed95849 | 1113 | } |
ab7d5fc6 | 1114 | |
abc490a1 | 1115 | static |
14044b37 | 1116 | int cds_lfht_delete_dummy(struct cds_lfht *ht) |
674f7a69 | 1117 | { |
395270b6 | 1118 | struct rcu_table *t; |
14044b37 MD |
1119 | struct cds_lfht_node *node; |
1120 | struct _cds_lfht_node *lookup; | |
24365af7 | 1121 | unsigned long order, i; |
674f7a69 | 1122 | |
abc490a1 MD |
1123 | t = ht->t; |
1124 | /* Check that the table is empty */ | |
1475579c | 1125 | lookup = &t->tbl[0]->nodes[0]; |
14044b37 | 1126 | node = (struct cds_lfht_node *) lookup; |
abc490a1 | 1127 | do { |
1b81fe1a MD |
1128 | node = clear_flag(node)->p.next; |
1129 | if (!is_dummy(node)) | |
abc490a1 | 1130 | return -EPERM; |
273399de | 1131 | assert(!is_removed(node)); |
a2974903 | 1132 | } while (clear_flag(node)); |
abc490a1 | 1133 | /* Internal sanity check: all nodes left should be dummy */ |
24365af7 MD |
1134 | for (order = 0; order < get_count_order_ulong(t->size) + 1; order++) { |
1135 | unsigned long len; | |
1136 | ||
1137 | len = !order ? 1 : 1UL << (order - 1); | |
1138 | for (i = 0; i < len; i++) { | |
f0c29ed7 | 1139 | dbg_printf("delete order %lu i %lu hash %lu\n", |
24365af7 | 1140 | order, i, |
1475579c MD |
1141 | bit_reverse_ulong(t->tbl[order]->nodes[i].reverse_hash)); |
1142 | assert(is_dummy(t->tbl[order]->nodes[i].next)); | |
24365af7 | 1143 | } |
98808fb1 | 1144 | poison_free(t->tbl[order]); |
674f7a69 | 1145 | } |
abc490a1 | 1146 | return 0; |
674f7a69 MD |
1147 | } |
1148 | ||
1149 | /* | |
1150 | * Should only be called when no more concurrent readers nor writers can | |
1151 | * possibly access the table. | |
1152 | */ | |
14044b37 | 1153 | int cds_lfht_destroy(struct cds_lfht *ht) |
674f7a69 | 1154 | { |
5e28c532 MD |
1155 | int ret; |
1156 | ||
848d4088 | 1157 | /* Wait for in-flight resize operations to complete */ |
33c7c748 | 1158 | CMM_STORE_SHARED(ht->in_progress_destroy, 1); |
848d4088 MD |
1159 | while (uatomic_read(&ht->in_progress_resize)) |
1160 | poll(NULL, 0, 100); /* wait for 100ms */ | |
14044b37 | 1161 | ret = cds_lfht_delete_dummy(ht); |
abc490a1 MD |
1162 | if (ret) |
1163 | return ret; | |
98808fb1 | 1164 | poison_free(ht->t); |
df44348d | 1165 | free_per_cpu_items_count(ht->percpu_count); |
98808fb1 | 1166 | poison_free(ht); |
5e28c532 | 1167 | return ret; |
674f7a69 MD |
1168 | } |
1169 | ||
14044b37 | 1170 | void cds_lfht_count_nodes(struct cds_lfht *ht, |
273399de MD |
1171 | unsigned long *count, |
1172 | unsigned long *removed) | |
1173 | { | |
1174 | struct rcu_table *t; | |
14044b37 MD |
1175 | struct cds_lfht_node *node, *next; |
1176 | struct _cds_lfht_node *lookup; | |
24365af7 | 1177 | unsigned long nr_dummy = 0; |
273399de MD |
1178 | |
1179 | *count = 0; | |
1180 | *removed = 0; | |
1181 | ||
1182 | t = rcu_dereference(ht->t); | |
24365af7 | 1183 | /* Count non-dummy nodes in the table */ |
1475579c | 1184 | lookup = &t->tbl[0]->nodes[0]; |
14044b37 | 1185 | node = (struct cds_lfht_node *) lookup; |
273399de | 1186 | do { |
cc4fcb10 | 1187 | next = rcu_dereference(node->p.next); |
273399de | 1188 | if (is_removed(next)) { |
1b81fe1a | 1189 | assert(!is_dummy(next)); |
273399de | 1190 | (*removed)++; |
1b81fe1a | 1191 | } else if (!is_dummy(next)) |
273399de | 1192 | (*count)++; |
24365af7 MD |
1193 | else |
1194 | (nr_dummy)++; | |
273399de MD |
1195 | node = clear_flag(next); |
1196 | } while (node); | |
f0c29ed7 | 1197 | dbg_printf("number of dummy nodes: %lu\n", nr_dummy); |
273399de MD |
1198 | } |
1199 | ||
1475579c | 1200 | /* called with resize mutex held */ |
abc490a1 | 1201 | static |
1475579c MD |
1202 | void _do_cds_lfht_grow(struct cds_lfht *ht, struct rcu_table *old_t, |
1203 | unsigned long old_size, unsigned long new_size) | |
abc490a1 | 1204 | { |
1475579c MD |
1205 | unsigned long old_order, new_order; |
1206 | struct rcu_table *new_t; | |
1207 | ||
1208 | old_order = get_count_order_ulong(old_size) + 1; | |
1209 | new_order = get_count_order_ulong(new_size) + 1; | |
1210 | printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n", | |
1211 | old_size, old_order, new_size, new_order); | |
1212 | new_t = malloc(sizeof(struct cds_lfht) | |
1213 | + (new_order * sizeof(struct rcu_level *))); | |
1214 | assert(new_size > old_size); | |
1215 | memcpy(&new_t->tbl, &old_t->tbl, | |
1216 | old_order * sizeof(struct rcu_level *)); | |
1217 | init_table(ht, new_t, old_order, new_order - old_order); | |
1218 | /* Changing table and size atomically wrt lookups */ | |
1219 | rcu_assign_pointer(ht->t, new_t); | |
1220 | ht->cds_lfht_call_rcu(&old_t->head, cds_lfht_free_table_cb); | |
abc490a1 MD |
1221 | } |
1222 | ||
1223 | /* called with resize mutex held */ | |
1224 | static | |
1475579c MD |
1225 | void _do_cds_lfht_shrink(struct cds_lfht *ht, struct rcu_table *old_t, |
1226 | unsigned long old_size, unsigned long new_size) | |
464a1ec9 | 1227 | { |
1475579c MD |
1228 | unsigned long old_order, new_order; |
1229 | struct rcu_table *new_t; | |
464a1ec9 | 1230 | |
cd95516d | 1231 | new_size = max(new_size, MIN_TABLE_SIZE); |
24365af7 | 1232 | old_order = get_count_order_ulong(old_size) + 1; |
24365af7 | 1233 | new_order = get_count_order_ulong(new_size) + 1; |
df44348d | 1234 | printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n", |
df03fab8 | 1235 | old_size, old_order, new_size, new_order); |
14044b37 | 1236 | new_t = malloc(sizeof(struct cds_lfht) |
1475579c MD |
1237 | + (new_order * sizeof(struct rcu_level *))); |
1238 | assert(new_size < old_size); | |
f000907d | 1239 | memcpy(&new_t->tbl, &old_t->tbl, |
1475579c MD |
1240 | new_order * sizeof(struct rcu_level *)); |
1241 | new_t->size = !new_order ? 1 : (1UL << (new_order - 1)); | |
c90201ac | 1242 | assert(new_t->size == new_size); |
1475579c MD |
1243 | new_t->resize_target = new_t->size; |
1244 | new_t->resize_initiated = 0; | |
1245 | ||
f000907d MD |
1246 | /* Changing table and size atomically wrt lookups */ |
1247 | rcu_assign_pointer(ht->t, new_t); | |
1475579c MD |
1248 | |
1249 | /* | |
c90201ac | 1250 | * We need to wait for all add operations to reach Q.S. (and |
1475579c | 1251 | * thus use the new table for lookups) before we can start |
c90201ac MD |
1252 | * releasing the old dummy nodes. Otherwise their lookup will |
1253 | * return a logically removed node as insert position. | |
1475579c MD |
1254 | */ |
1255 | ht->cds_lfht_synchronize_rcu(); | |
1256 | ||
1257 | /* Unlink and remove all now-unused dummy node pointers. */ | |
1258 | fini_table(ht, old_t, new_order, old_order - new_order); | |
14044b37 | 1259 | ht->cds_lfht_call_rcu(&old_t->head, cds_lfht_free_table_cb); |
464a1ec9 MD |
1260 | } |
1261 | ||
1475579c MD |
1262 | |
1263 | /* called with resize mutex held */ | |
1264 | static | |
1265 | void _do_cds_lfht_resize(struct cds_lfht *ht) | |
1266 | { | |
1267 | unsigned long new_size, old_size; | |
1268 | struct rcu_table *old_t; | |
1269 | ||
1270 | old_t = ht->t; | |
1271 | old_size = old_t->size; | |
1272 | new_size = CMM_LOAD_SHARED(old_t->resize_target); | |
1273 | if (old_size < new_size) | |
1274 | _do_cds_lfht_grow(ht, old_t, old_size, new_size); | |
1275 | else if (old_size > new_size) | |
1276 | _do_cds_lfht_shrink(ht, old_t, old_size, new_size); | |
1277 | else | |
1278 | CMM_STORE_SHARED(old_t->resize_initiated, 0); | |
1279 | } | |
1280 | ||
abc490a1 | 1281 | static |
f9830efd MD |
1282 | unsigned long resize_target_update(struct rcu_table *t, |
1283 | int growth_order) | |
464a1ec9 | 1284 | { |
f9830efd MD |
1285 | return _uatomic_max(&t->resize_target, |
1286 | t->size << growth_order); | |
464a1ec9 MD |
1287 | } |
1288 | ||
1475579c | 1289 | static |
b8af5011 MD |
1290 | void resize_target_update_count(struct rcu_table *t, |
1291 | unsigned long count) | |
1475579c | 1292 | { |
cd95516d | 1293 | count = max(count, MIN_TABLE_SIZE); |
b8af5011 | 1294 | uatomic_set(&t->resize_target, count); |
1475579c MD |
1295 | } |
1296 | ||
1297 | void cds_lfht_resize(struct cds_lfht *ht, unsigned long new_size) | |
464a1ec9 | 1298 | { |
f9830efd | 1299 | struct rcu_table *t = rcu_dereference(ht->t); |
f9830efd | 1300 | |
b8af5011 | 1301 | resize_target_update_count(t, new_size); |
1475579c MD |
1302 | CMM_STORE_SHARED(t->resize_initiated, 1); |
1303 | pthread_mutex_lock(&ht->resize_mutex); | |
1304 | _do_cds_lfht_resize(ht); | |
1305 | pthread_mutex_unlock(&ht->resize_mutex); | |
abc490a1 | 1306 | } |
464a1ec9 | 1307 | |
abc490a1 MD |
1308 | static |
1309 | void do_resize_cb(struct rcu_head *head) | |
1310 | { | |
1311 | struct rcu_resize_work *work = | |
1312 | caa_container_of(head, struct rcu_resize_work, head); | |
14044b37 | 1313 | struct cds_lfht *ht = work->ht; |
abc490a1 MD |
1314 | |
1315 | pthread_mutex_lock(&ht->resize_mutex); | |
14044b37 | 1316 | _do_cds_lfht_resize(ht); |
abc490a1 | 1317 | pthread_mutex_unlock(&ht->resize_mutex); |
98808fb1 | 1318 | poison_free(work); |
848d4088 MD |
1319 | cmm_smp_mb(); /* finish resize before decrement */ |
1320 | uatomic_dec(&ht->in_progress_resize); | |
464a1ec9 MD |
1321 | } |
1322 | ||
abc490a1 | 1323 | static |
14044b37 | 1324 | void cds_lfht_resize_lazy(struct cds_lfht *ht, struct rcu_table *t, int growth) |
ab7d5fc6 | 1325 | { |
abc490a1 | 1326 | struct rcu_resize_work *work; |
f9830efd | 1327 | unsigned long target_size; |
abc490a1 | 1328 | |
f9830efd | 1329 | target_size = resize_target_update(t, growth); |
11519af6 | 1330 | if (!CMM_LOAD_SHARED(t->resize_initiated) && t->size < target_size) { |
848d4088 MD |
1331 | uatomic_inc(&ht->in_progress_resize); |
1332 | cmm_smp_mb(); /* increment resize count before calling it */ | |
f9830efd MD |
1333 | work = malloc(sizeof(*work)); |
1334 | work->ht = ht; | |
14044b37 | 1335 | ht->cds_lfht_call_rcu(&work->head, do_resize_cb); |
11519af6 | 1336 | CMM_STORE_SHARED(t->resize_initiated, 1); |
f9830efd | 1337 | } |
ab7d5fc6 | 1338 | } |
3171717f | 1339 | |
f8994aee MD |
1340 | #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) |
1341 | ||
3171717f MD |
1342 | static |
1343 | void cds_lfht_resize_lazy_count(struct cds_lfht *ht, struct rcu_table *t, | |
1344 | unsigned long count) | |
1345 | { | |
1346 | struct rcu_resize_work *work; | |
3171717f | 1347 | |
b8af5011 MD |
1348 | if (!(ht->flags & CDS_LFHT_AUTO_RESIZE)) |
1349 | return; | |
1350 | resize_target_update_count(t, count); | |
1475579c | 1351 | if (!CMM_LOAD_SHARED(t->resize_initiated)) { |
3171717f MD |
1352 | uatomic_inc(&ht->in_progress_resize); |
1353 | cmm_smp_mb(); /* increment resize count before calling it */ | |
1354 | work = malloc(sizeof(*work)); | |
1355 | work->ht = ht; | |
1356 | ht->cds_lfht_call_rcu(&work->head, do_resize_cb); | |
1357 | CMM_STORE_SHARED(t->resize_initiated, 1); | |
1358 | } | |
1359 | } | |
f8994aee MD |
1360 | |
1361 | #endif |