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