1288c78d585f983764da0c96eefb6df4f5712cc3
[urcu.git] / rculfhash.c
1 /*
2 * rculfhash.c
3 *
4 * Userspace RCU library - Lock-Free Resizable RCU Hash Table
5 *
6 * Copyright 2010-2011 - Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
7 *
8 * This library is free software; you can redistribute it and/or
9 * modify it under the terms of the GNU Lesser General Public
10 * License as published by the Free Software Foundation; either
11 * version 2.1 of the License, or (at your option) any later version.
12 *
13 * This library is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 * Lesser General Public License for more details.
17 *
18 * You should have received a copy of the GNU Lesser General Public
19 * License along with this library; if not, write to the Free Software
20 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
21 */
22
23 /*
24 * Based on the following articles:
25 * - Ori Shalev and Nir Shavit. Split-ordered lists: Lock-free
26 * extensible hash tables. J. ACM 53, 3 (May 2006), 379-405.
27 * - Michael, M. M. High performance dynamic lock-free hash tables
28 * and list-based sets. In Proceedings of the fourteenth annual ACM
29 * symposium on Parallel algorithms and architectures, ACM Press,
30 * (2002), 73-82.
31 *
32 * Some specificities of this Lock-Free Resizable RCU Hash Table
33 * implementation:
34 *
35 * - RCU read-side critical section allows readers to perform hash
36 * table lookups and use the returned objects safely by delaying
37 * memory reclaim of a grace period.
38 * - Add and remove operations are lock-free, and do not need to
39 * allocate memory. They need to be executed within RCU read-side
40 * critical section to ensure the objects they read are valid and to
41 * deal with the cmpxchg ABA problem.
42 * - add and add_unique operations are supported. add_unique checks if
43 * the node key already exists in the hash table. It ensures no key
44 * duplicata exists.
45 * - The resize operation executes concurrently with add/remove/lookup.
46 * - Hash table nodes are contained within a split-ordered list. This
47 * list is ordered by incrementing reversed-bits-hash value.
48 * - An index of dummy nodes is kept. These dummy nodes are the hash
49 * table "buckets", and they are also chained together in the
50 * split-ordered list, which allows recursive expansion.
51 * - The resize operation 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.
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.
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.
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.
99 *
100 * A bit of ascii art explanation:
101 *
102 * Order index is the off-by-one compare to the actual power of 2 because
103 * we use index 0 to deal with the 0 special-case.
104 *
105 * This shows the nodes for a small table ordered by reversed bits:
106 *
107 * bits reverse
108 * 0 000 000
109 * 4 100 001
110 * 2 010 010
111 * 6 110 011
112 * 1 001 100
113 * 5 101 101
114 * 3 011 110
115 * 7 111 111
116 *
117 * This shows the nodes in order of non-reversed bits, linked by
118 * reversed-bit order.
119 *
120 * order bits reverse
121 * 0 0 000 000
122 * |
123 * 1 | 1 001 100 <- <-
124 * | | | |
125 * 2 | | 2 010 010 | |
126 * | | | 3 011 110 | <- |
127 * | | | | | | |
128 * 3 -> | | | 4 100 001 | |
129 * -> | | 5 101 101 |
130 * -> | 6 110 011
131 * -> 7 111 111
132 */
133
134 #define _LGPL_SOURCE
135 #include <stdlib.h>
136 #include <errno.h>
137 #include <assert.h>
138 #include <stdio.h>
139 #include <stdint.h>
140 #include <string.h>
141
142 #include "config.h"
143 #include <urcu.h>
144 #include <urcu-call-rcu.h>
145 #include <urcu/arch.h>
146 #include <urcu/uatomic.h>
147 #include <urcu/jhash.h>
148 #include <urcu/compiler.h>
149 #include <urcu/rculfhash.h>
150 #include <stdio.h>
151 #include <pthread.h>
152
153 #ifdef DEBUG
154 #define dbg_printf(fmt, args...) printf("[debug rculfhash] " fmt, ## args)
155 #else
156 #define dbg_printf(fmt, args...)
157 #endif
158
159 /*
160 * Per-CPU split-counters lazily update the global counter each 1024
161 * addition/removal. It automatically keeps track of resize required.
162 * We use the bucket length as indicator for need to expand for small
163 * tables and machines lacking per-cpu data suppport.
164 */
165 #define COUNT_COMMIT_ORDER 10
166 #define CHAIN_LEN_TARGET 1
167 #define CHAIN_LEN_RESIZE_THRESHOLD 3
168
169 /*
170 * Define the minimum table size.
171 */
172 #define MIN_TABLE_SIZE 128
173
174 #if (CAA_BITS_PER_LONG == 32)
175 #define MAX_TABLE_ORDER 32
176 #else
177 #define MAX_TABLE_ORDER 64
178 #endif
179
180 #ifndef min
181 #define min(a, b) ((a) < (b) ? (a) : (b))
182 #endif
183
184 #ifndef max
185 #define max(a, b) ((a) > (b) ? (a) : (b))
186 #endif
187
188 /*
189 * The removed flag needs to be updated atomically with the pointer.
190 * The dummy flag does not require to be updated atomically with the
191 * pointer, but it is added as a pointer low bit flag to save space.
192 */
193 #define REMOVED_FLAG (1UL << 0)
194 #define DUMMY_FLAG (1UL << 1)
195 #define FLAGS_MASK ((1UL << 2) - 1)
196
197 struct ht_items_count {
198 unsigned long add, remove;
199 } __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
200
201 struct rcu_level {
202 struct rcu_head head;
203 struct _cds_lfht_node nodes[0];
204 };
205
206 struct rcu_table {
207 unsigned long size; /* always a power of 2, shared (RCU) */
208 unsigned long resize_target;
209 int resize_initiated;
210 struct rcu_level *tbl[MAX_TABLE_ORDER];
211 };
212
213 struct cds_lfht {
214 struct rcu_table t;
215 cds_lfht_hash_fct hash_fct;
216 cds_lfht_compare_fct compare_fct;
217 unsigned long hash_seed;
218 int flags;
219 /*
220 * We need to put the work threads offline (QSBR) when taking this
221 * mutex, because we use synchronize_rcu within this mutex critical
222 * section, which waits on read-side critical sections, and could
223 * therefore cause grace-period deadlock if we hold off RCU G.P.
224 * completion.
225 */
226 pthread_mutex_t resize_mutex; /* resize mutex: add/del mutex */
227 unsigned int in_progress_resize, in_progress_destroy;
228 void (*cds_lfht_call_rcu)(struct rcu_head *head,
229 void (*func)(struct rcu_head *head));
230 void (*cds_lfht_synchronize_rcu)(void);
231 void (*cds_lfht_rcu_read_lock)(void);
232 void (*cds_lfht_rcu_read_unlock)(void);
233 void (*cds_lfht_rcu_thread_offline)(void);
234 void (*cds_lfht_rcu_thread_online)(void);
235 unsigned long count; /* global approximate item count */
236 struct ht_items_count *percpu_count; /* per-cpu item count */
237 };
238
239 struct rcu_resize_work {
240 struct rcu_head head;
241 struct cds_lfht *ht;
242 };
243
244 static
245 struct cds_lfht_node *_cds_lfht_add(struct cds_lfht *ht,
246 unsigned long size,
247 struct cds_lfht_node *node,
248 int unique, int dummy);
249
250 /*
251 * Algorithm to reverse bits in a word by lookup table, extended to
252 * 64-bit words.
253 * Source:
254 * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable
255 * Originally from Public Domain.
256 */
257
258 static const uint8_t BitReverseTable256[256] =
259 {
260 #define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64
261 #define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16)
262 #define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 )
263 R6(0), R6(2), R6(1), R6(3)
264 };
265 #undef R2
266 #undef R4
267 #undef R6
268
269 static
270 uint8_t bit_reverse_u8(uint8_t v)
271 {
272 return BitReverseTable256[v];
273 }
274
275 static __attribute__((unused))
276 uint32_t bit_reverse_u32(uint32_t v)
277 {
278 return ((uint32_t) bit_reverse_u8(v) << 24) |
279 ((uint32_t) bit_reverse_u8(v >> 8) << 16) |
280 ((uint32_t) bit_reverse_u8(v >> 16) << 8) |
281 ((uint32_t) bit_reverse_u8(v >> 24));
282 }
283
284 static __attribute__((unused))
285 uint64_t bit_reverse_u64(uint64_t v)
286 {
287 return ((uint64_t) bit_reverse_u8(v) << 56) |
288 ((uint64_t) bit_reverse_u8(v >> 8) << 48) |
289 ((uint64_t) bit_reverse_u8(v >> 16) << 40) |
290 ((uint64_t) bit_reverse_u8(v >> 24) << 32) |
291 ((uint64_t) bit_reverse_u8(v >> 32) << 24) |
292 ((uint64_t) bit_reverse_u8(v >> 40) << 16) |
293 ((uint64_t) bit_reverse_u8(v >> 48) << 8) |
294 ((uint64_t) bit_reverse_u8(v >> 56));
295 }
296
297 static
298 unsigned long bit_reverse_ulong(unsigned long v)
299 {
300 #if (CAA_BITS_PER_LONG == 32)
301 return bit_reverse_u32(v);
302 #else
303 return bit_reverse_u64(v);
304 #endif
305 }
306
307 /*
308 * fls: returns the position of the most significant bit.
309 * Returns 0 if no bit is set, else returns the position of the most
310 * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit).
311 */
312 #if defined(__i386) || defined(__x86_64)
313 static inline
314 unsigned int fls_u32(uint32_t x)
315 {
316 int r;
317
318 asm("bsrl %1,%0\n\t"
319 "jnz 1f\n\t"
320 "movl $-1,%0\n\t"
321 "1:\n\t"
322 : "=r" (r) : "rm" (x));
323 return r + 1;
324 }
325 #define HAS_FLS_U32
326 #endif
327
328 #if defined(__x86_64)
329 static inline
330 unsigned int fls_u64(uint64_t x)
331 {
332 long r;
333
334 asm("bsrq %1,%0\n\t"
335 "jnz 1f\n\t"
336 "movq $-1,%0\n\t"
337 "1:\n\t"
338 : "=r" (r) : "rm" (x));
339 return r + 1;
340 }
341 #define HAS_FLS_U64
342 #endif
343
344 #ifndef HAS_FLS_U64
345 static __attribute__((unused))
346 unsigned int fls_u64(uint64_t x)
347 {
348 unsigned int r = 64;
349
350 if (!x)
351 return 0;
352
353 if (!(x & 0xFFFFFFFF00000000ULL)) {
354 x <<= 32;
355 r -= 32;
356 }
357 if (!(x & 0xFFFF000000000000ULL)) {
358 x <<= 16;
359 r -= 16;
360 }
361 if (!(x & 0xFF00000000000000ULL)) {
362 x <<= 8;
363 r -= 8;
364 }
365 if (!(x & 0xF000000000000000ULL)) {
366 x <<= 4;
367 r -= 4;
368 }
369 if (!(x & 0xC000000000000000ULL)) {
370 x <<= 2;
371 r -= 2;
372 }
373 if (!(x & 0x8000000000000000ULL)) {
374 x <<= 1;
375 r -= 1;
376 }
377 return r;
378 }
379 #endif
380
381 #ifndef HAS_FLS_U32
382 static __attribute__((unused))
383 unsigned int fls_u32(uint32_t x)
384 {
385 unsigned int r = 32;
386
387 if (!x)
388 return 0;
389 if (!(x & 0xFFFF0000U)) {
390 x <<= 16;
391 r -= 16;
392 }
393 if (!(x & 0xFF000000U)) {
394 x <<= 8;
395 r -= 8;
396 }
397 if (!(x & 0xF0000000U)) {
398 x <<= 4;
399 r -= 4;
400 }
401 if (!(x & 0xC0000000U)) {
402 x <<= 2;
403 r -= 2;
404 }
405 if (!(x & 0x80000000U)) {
406 x <<= 1;
407 r -= 1;
408 }
409 return r;
410 }
411 #endif
412
413 unsigned int fls_ulong(unsigned long x)
414 {
415 #if (CAA_BITS_PER_lONG == 32)
416 return fls_u32(x);
417 #else
418 return fls_u64(x);
419 #endif
420 }
421
422 int get_count_order_u32(uint32_t x)
423 {
424 int order;
425
426 order = fls_u32(x) - 1;
427 if (x & (x - 1))
428 order++;
429 return order;
430 }
431
432 int get_count_order_ulong(unsigned long x)
433 {
434 int order;
435
436 order = fls_ulong(x) - 1;
437 if (x & (x - 1))
438 order++;
439 return order;
440 }
441
442 #ifdef POISON_FREE
443 #define poison_free(ptr) \
444 do { \
445 memset(ptr, 0x42, sizeof(*(ptr))); \
446 free(ptr); \
447 } while (0)
448 #else
449 #define poison_free(ptr) free(ptr)
450 #endif
451
452 static
453 void cds_lfht_resize_lazy(struct cds_lfht *ht, unsigned long size, int growth);
454
455 /*
456 * If the sched_getcpu() and sysconf(_SC_NPROCESSORS_CONF) calls are
457 * available, then we support hash table item accounting.
458 * In the unfortunate event the number of CPUs reported would be
459 * inaccurate, we use modulo arithmetic on the number of CPUs we got.
460 */
461 #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
462
463 static
464 void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
465 unsigned long count);
466
467 static long nr_cpus_mask = -1;
468
469 static
470 struct ht_items_count *alloc_per_cpu_items_count(void)
471 {
472 struct ht_items_count *count;
473
474 switch (nr_cpus_mask) {
475 case -2:
476 return NULL;
477 case -1:
478 {
479 long maxcpus;
480
481 maxcpus = sysconf(_SC_NPROCESSORS_CONF);
482 if (maxcpus <= 0) {
483 nr_cpus_mask = -2;
484 return NULL;
485 }
486 /*
487 * round up number of CPUs to next power of two, so we
488 * can use & for modulo.
489 */
490 maxcpus = 1UL << get_count_order_ulong(maxcpus);
491 nr_cpus_mask = maxcpus - 1;
492 }
493 /* Fall-through */
494 default:
495 return calloc(nr_cpus_mask + 1, sizeof(*count));
496 }
497 }
498
499 static
500 void free_per_cpu_items_count(struct ht_items_count *count)
501 {
502 poison_free(count);
503 }
504
505 static
506 int ht_get_cpu(void)
507 {
508 int cpu;
509
510 assert(nr_cpus_mask >= 0);
511 cpu = sched_getcpu();
512 if (unlikely(cpu < 0))
513 return cpu;
514 else
515 return cpu & nr_cpus_mask;
516 }
517
518 static
519 void ht_count_add(struct cds_lfht *ht, unsigned long size)
520 {
521 unsigned long percpu_count;
522 int cpu;
523
524 if (unlikely(!ht->percpu_count))
525 return;
526 cpu = ht_get_cpu();
527 if (unlikely(cpu < 0))
528 return;
529 percpu_count = uatomic_add_return(&ht->percpu_count[cpu].add, 1);
530 if (unlikely(!(percpu_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) {
531 unsigned long count;
532
533 dbg_printf("add percpu %lu\n", percpu_count);
534 count = uatomic_add_return(&ht->count,
535 1UL << COUNT_COMMIT_ORDER);
536 /* If power of 2 */
537 if (!(count & (count - 1))) {
538 if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) < size)
539 return;
540 dbg_printf("add set global %lu\n", count);
541 cds_lfht_resize_lazy_count(ht, size,
542 count >> (CHAIN_LEN_TARGET - 1));
543 }
544 }
545 }
546
547 static
548 void ht_count_remove(struct cds_lfht *ht, unsigned long size)
549 {
550 unsigned long percpu_count;
551 int cpu;
552
553 if (unlikely(!ht->percpu_count))
554 return;
555 cpu = ht_get_cpu();
556 if (unlikely(cpu < 0))
557 return;
558 percpu_count = uatomic_add_return(&ht->percpu_count[cpu].remove, -1);
559 if (unlikely(!(percpu_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) {
560 unsigned long count;
561
562 dbg_printf("remove percpu %lu\n", percpu_count);
563 count = uatomic_add_return(&ht->count,
564 -(1UL << COUNT_COMMIT_ORDER));
565 /* If power of 2 */
566 if (!(count & (count - 1))) {
567 if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) >= size)
568 return;
569 dbg_printf("remove set global %lu\n", count);
570 cds_lfht_resize_lazy_count(ht, size,
571 count >> (CHAIN_LEN_TARGET - 1));
572 }
573 }
574 }
575
576 #else /* #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
577
578 static const long nr_cpus_mask = -1;
579
580 static
581 struct ht_items_count *alloc_per_cpu_items_count(void)
582 {
583 return NULL;
584 }
585
586 static
587 void free_per_cpu_items_count(struct ht_items_count *count)
588 {
589 }
590
591 static
592 void ht_count_add(struct cds_lfht *ht, unsigned long size)
593 {
594 }
595
596 static
597 void ht_count_remove(struct cds_lfht *ht, unsigned long size)
598 {
599 }
600
601 #endif /* #else #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
602
603
604 static
605 void check_resize(struct cds_lfht *ht, unsigned long size, uint32_t chain_len)
606 {
607 unsigned long count;
608
609 if (!(ht->flags & CDS_LFHT_AUTO_RESIZE))
610 return;
611 count = uatomic_read(&ht->count);
612 /*
613 * Use bucket-local length for small table expand and for
614 * environments lacking per-cpu data support.
615 */
616 if (count >= (1UL << COUNT_COMMIT_ORDER))
617 return;
618 if (chain_len > 100)
619 dbg_printf("WARNING: large chain length: %u.\n",
620 chain_len);
621 if (chain_len >= CHAIN_LEN_RESIZE_THRESHOLD)
622 cds_lfht_resize_lazy(ht, size,
623 get_count_order_u32(chain_len - (CHAIN_LEN_TARGET - 1)));
624 }
625
626 static
627 struct cds_lfht_node *clear_flag(struct cds_lfht_node *node)
628 {
629 return (struct cds_lfht_node *) (((unsigned long) node) & ~FLAGS_MASK);
630 }
631
632 static
633 int is_removed(struct cds_lfht_node *node)
634 {
635 return ((unsigned long) node) & REMOVED_FLAG;
636 }
637
638 static
639 struct cds_lfht_node *flag_removed(struct cds_lfht_node *node)
640 {
641 return (struct cds_lfht_node *) (((unsigned long) node) | REMOVED_FLAG);
642 }
643
644 static
645 int is_dummy(struct cds_lfht_node *node)
646 {
647 return ((unsigned long) node) & DUMMY_FLAG;
648 }
649
650 static
651 struct cds_lfht_node *flag_dummy(struct cds_lfht_node *node)
652 {
653 return (struct cds_lfht_node *) (((unsigned long) node) | DUMMY_FLAG);
654 }
655
656 static
657 unsigned long _uatomic_max(unsigned long *ptr, unsigned long v)
658 {
659 unsigned long old1, old2;
660
661 old1 = uatomic_read(ptr);
662 do {
663 old2 = old1;
664 if (old2 >= v)
665 return old2;
666 } while ((old1 = uatomic_cmpxchg(ptr, old2, v)) != old2);
667 return v;
668 }
669
670 static
671 void cds_lfht_free_level(struct rcu_head *head)
672 {
673 struct rcu_level *l =
674 caa_container_of(head, struct rcu_level, head);
675 poison_free(l);
676 }
677
678 /*
679 * Remove all logically deleted nodes from a bucket up to a certain node key.
680 */
681 static
682 int _cds_lfht_gc_bucket(struct cds_lfht_node *dummy, struct cds_lfht_node *node)
683 {
684 struct cds_lfht_node *iter_prev, *iter, *next, *new_next;
685
686 assert(!is_dummy(dummy));
687 assert(!is_removed(dummy));
688 assert(!is_dummy(node));
689 assert(!is_removed(node));
690 for (;;) {
691 iter_prev = dummy;
692 /* We can always skip the dummy node initially */
693 iter = rcu_dereference(iter_prev->p.next);
694 if (unlikely(iter == NULL)) {
695 /*
696 * We are executing concurrently with a hash table
697 * expand, so we see a dummy node with NULL next value.
698 * Help expand by linking this node into the list and
699 * retry.
700 */
701 return 1;
702 }
703 assert(iter_prev->p.reverse_hash <= node->p.reverse_hash);
704 /*
705 * We should never be called with dummy (start of chain)
706 * and logically removed node (end of path compression
707 * marker) being the actual same node. This would be a
708 * bug in the algorithm implementation.
709 */
710 assert(dummy != node);
711 for (;;) {
712 if (unlikely(!clear_flag(iter)))
713 return 0;
714 if (likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash))
715 return 0;
716 next = rcu_dereference(clear_flag(iter)->p.next);
717 if (likely(is_removed(next)))
718 break;
719 iter_prev = clear_flag(iter);
720 iter = next;
721 }
722 assert(!is_removed(iter));
723 if (is_dummy(iter))
724 new_next = flag_dummy(clear_flag(next));
725 else
726 new_next = clear_flag(next);
727 (void) uatomic_cmpxchg(&iter_prev->p.next, iter, new_next);
728 }
729 return 0;
730 }
731
732 static
733 struct cds_lfht_node *_cds_lfht_add(struct cds_lfht *ht,
734 unsigned long size,
735 struct cds_lfht_node *node,
736 int unique, int dummy)
737 {
738 struct cds_lfht_node *iter_prev, *iter, *next, *new_node, *new_next,
739 *dummy_node;
740 struct _cds_lfht_node *lookup;
741 unsigned long hash, index, order;
742
743 assert(!is_dummy(node));
744 assert(!is_removed(node));
745 if (!size) {
746 assert(dummy);
747 node->p.next = flag_dummy(NULL);
748 return node; /* Initial first add (head) */
749 }
750 hash = bit_reverse_ulong(node->p.reverse_hash);
751 for (;;) {
752 uint32_t chain_len = 0;
753
754 /*
755 * iter_prev points to the non-removed node prior to the
756 * insert location.
757 */
758 index = hash & (size - 1);
759 order = get_count_order_ulong(index + 1);
760 lookup = &ht->t.tbl[order]->nodes[index & ((!order ? 0 : (1UL << (order - 1))) - 1)];
761 iter_prev = (struct cds_lfht_node *) lookup;
762 /* We can always skip the dummy node initially */
763 iter = rcu_dereference(iter_prev->p.next);
764 if (unlikely(iter == NULL)) {
765 /*
766 * We are executing concurrently with a hash table
767 * expand, so we see a dummy node with NULL next value.
768 * Help expand by linking this node into the list and
769 * retry.
770 */
771 (void) _cds_lfht_add(ht, size >> 1, iter_prev, 0, 1);
772 continue; /* retry */
773 }
774 assert(iter_prev->p.reverse_hash <= node->p.reverse_hash);
775 for (;;) {
776 /*
777 * When adding a dummy node, we allow concurrent
778 * add/removal to help. If we find the dummy node in
779 * place, skip its insertion.
780 */
781 if (unlikely(dummy && clear_flag(iter) == node))
782 return node;
783 if (unlikely(!clear_flag(iter)))
784 goto insert;
785 if (likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash))
786 goto insert;
787 next = rcu_dereference(clear_flag(iter)->p.next);
788 if (unlikely(is_removed(next)))
789 goto gc_node;
790 if (unique
791 && !is_dummy(next)
792 && !ht->compare_fct(node->key, node->key_len,
793 clear_flag(iter)->key,
794 clear_flag(iter)->key_len))
795 return clear_flag(iter);
796 /* Only account for identical reverse hash once */
797 if (iter_prev->p.reverse_hash != clear_flag(iter)->p.reverse_hash
798 && !is_dummy(next))
799 check_resize(ht, size, ++chain_len);
800 iter_prev = clear_flag(iter);
801 iter = next;
802 }
803 insert:
804 assert(node != clear_flag(iter));
805 assert(!is_removed(iter_prev));
806 assert(!is_removed(iter));
807 assert(iter_prev != node);
808 if (!dummy)
809 node->p.next = clear_flag(iter);
810 else
811 node->p.next = flag_dummy(clear_flag(iter));
812 if (is_dummy(iter))
813 new_node = flag_dummy(node);
814 else
815 new_node = node;
816 if (uatomic_cmpxchg(&iter_prev->p.next, iter,
817 new_node) != iter)
818 continue; /* retry */
819 else
820 goto gc_end;
821 gc_node:
822 assert(!is_removed(iter));
823 if (is_dummy(iter))
824 new_next = flag_dummy(clear_flag(next));
825 else
826 new_next = clear_flag(next);
827 (void) uatomic_cmpxchg(&iter_prev->p.next, iter, new_next);
828 /* retry */
829 }
830 gc_end:
831 /* Garbage collect logically removed nodes in the bucket */
832 index = hash & (size - 1);
833 order = get_count_order_ulong(index + 1);
834 lookup = &ht->t.tbl[order]->nodes[index & (!order ? 0 : ((1UL << (order - 1)) - 1))];
835 dummy_node = (struct cds_lfht_node *) lookup;
836 if (_cds_lfht_gc_bucket(dummy_node, node)) {
837 /* Help expand */
838 (void) _cds_lfht_add(ht, size >> 1, dummy_node, 0, 1);
839 goto gc_end; /* retry */
840 }
841 return node;
842 }
843
844 static
845 int _cds_lfht_remove(struct cds_lfht *ht, unsigned long size,
846 struct cds_lfht_node *node,
847 int dummy_removal)
848 {
849 struct cds_lfht_node *dummy, *next, *old;
850 struct _cds_lfht_node *lookup;
851 int flagged = 0;
852 unsigned long hash, index, order;
853
854 /* logically delete the node */
855 assert(!is_dummy(node));
856 assert(!is_removed(node));
857 old = rcu_dereference(node->p.next);
858 do {
859 next = old;
860 if (unlikely(is_removed(next)))
861 goto end;
862 if (dummy_removal)
863 assert(is_dummy(next));
864 else
865 assert(!is_dummy(next));
866 old = uatomic_cmpxchg(&node->p.next, next,
867 flag_removed(next));
868 } while (old != next);
869
870 /* We performed the (logical) deletion. */
871 flagged = 1;
872
873 /*
874 * Ensure that the node is not visible to readers anymore: lookup for
875 * the node, and remove it (along with any other logically removed node)
876 * if found.
877 */
878 gc_retry:
879 hash = bit_reverse_ulong(node->p.reverse_hash);
880 assert(size > 0);
881 index = hash & (size - 1);
882 order = get_count_order_ulong(index + 1);
883 lookup = &ht->t.tbl[order]->nodes[index & (!order ? 0 : ((1UL << (order - 1)) - 1))];
884 dummy = (struct cds_lfht_node *) lookup;
885 if (_cds_lfht_gc_bucket(dummy, node)) {
886 /* Help expand */
887 (void) _cds_lfht_add(ht, size >> 1, dummy, 0, 1);
888 goto gc_retry; /* retry */
889 }
890 end:
891 /*
892 * Only the flagging action indicated that we (and no other)
893 * removed the node from the hash.
894 */
895 if (flagged) {
896 assert(is_removed(rcu_dereference(node->p.next)));
897 return 0;
898 } else
899 return -ENOENT;
900 }
901
902 static
903 void init_table_hash(struct cds_lfht *ht, unsigned long i,
904 unsigned long len)
905 {
906 unsigned long j;
907
908 for (j = 0; j < len; j++) {
909 struct cds_lfht_node *new_node =
910 (struct cds_lfht_node *) &ht->t.tbl[i]->nodes[j];
911
912 dbg_printf("init hash entry: i %lu j %lu hash %lu\n",
913 i, j, !i ? 0 : (1UL << (i - 1)) + j);
914 new_node->p.reverse_hash =
915 bit_reverse_ulong(!i ? 0 : (1UL << (i - 1)) + j);
916 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
917 break;
918 }
919 }
920
921 /*
922 * Holding RCU read lock to protect _cds_lfht_add against memory
923 * reclaim that could be performed by other call_rcu worker threads (ABA
924 * problem).
925 */
926 static
927 void init_table_link(struct cds_lfht *ht, unsigned long i, unsigned long len)
928 {
929 unsigned long j;
930
931 ht->cds_lfht_rcu_thread_online();
932 ht->cds_lfht_rcu_read_lock();
933 for (j = 0; j < len; j++) {
934 struct cds_lfht_node *new_node =
935 (struct cds_lfht_node *) &ht->t.tbl[i]->nodes[j];
936
937 dbg_printf("init link: i %lu j %lu hash %lu\n",
938 i, j, !i ? 0 : (1UL << (i - 1)) + j);
939 (void) _cds_lfht_add(ht, !i ? 0 : (1UL << (i - 1)),
940 new_node, 0, 1);
941 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
942 break;
943 }
944 ht->cds_lfht_rcu_read_unlock();
945 ht->cds_lfht_rcu_thread_offline();
946 }
947
948 static
949 void init_table(struct cds_lfht *ht,
950 unsigned long first_order, unsigned long len_order)
951 {
952 unsigned long i, end_order;
953
954 dbg_printf("init table: first_order %lu end_order %lu\n",
955 first_order, first_order + len_order);
956 end_order = first_order + len_order;
957 for (i = first_order; i < end_order; i++) {
958 unsigned long len;
959
960 len = !i ? 1 : 1UL << (i - 1);
961 dbg_printf("init order %lu len: %lu\n", i, len);
962
963 /* Stop expand if the resize target changes under us */
964 if (CMM_LOAD_SHARED(ht->t.resize_target) < (!i ? 1 : (1UL << i)))
965 break;
966
967 ht->t.tbl[i] = calloc(1, sizeof(struct rcu_level)
968 + (len * sizeof(struct _cds_lfht_node)));
969
970 /* Set all dummy nodes reverse hash values for a level */
971 init_table_hash(ht, i, len);
972
973 /*
974 * Update table size. At this point, concurrent add/remove see
975 * dummy nodes with correctly initialized reverse hash value,
976 * but with NULL next pointers. If they do, they can help us
977 * link the dummy nodes into the list and retry.
978 */
979 cmm_smp_wmb(); /* populate data before RCU size */
980 CMM_STORE_SHARED(ht->t.size, !i ? 1 : (1UL << i));
981
982 /*
983 * Link all dummy nodes into the table. Concurrent
984 * add/remove are helping us.
985 */
986 init_table_link(ht, i, len);
987
988 dbg_printf("init new size: %lu\n", !i ? 1 : (1UL << i));
989 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
990 break;
991 }
992 }
993
994 /*
995 * Holding RCU read lock to protect _cds_lfht_remove against memory
996 * reclaim that could be performed by other call_rcu worker threads (ABA
997 * problem).
998 * For a single level, we logically remove and garbage collect each node.
999 *
1000 * As a design choice, we perform logical removal and garbage collection on a
1001 * node-per-node basis to simplify this algorithm. We also assume keeping good
1002 * cache locality of the operation would overweight possible performance gain
1003 * that could be achieved by batching garbage collection for multiple levels.
1004 * However, this would have to be justified by benchmarks.
1005 *
1006 * Concurrent removal and add operations are helping us perform garbage
1007 * collection of logically removed nodes. We guarantee that all logically
1008 * removed nodes have been garbage-collected (unlinked) before call_rcu is
1009 * invoked to free a hole level of dummy nodes (after a grace period).
1010 *
1011 * Logical removal and garbage collection can therefore be done in batch or on a
1012 * node-per-node basis, as long as the guarantee above holds.
1013 */
1014 static
1015 void remove_table(struct cds_lfht *ht, unsigned long i, unsigned long len)
1016 {
1017 unsigned long j;
1018
1019 ht->cds_lfht_rcu_thread_online();
1020 ht->cds_lfht_rcu_read_lock();
1021 for (j = 0; j < len; j++) {
1022 struct cds_lfht_node *fini_node =
1023 (struct cds_lfht_node *) &ht->t.tbl[i]->nodes[j];
1024
1025 dbg_printf("remove entry: i %lu j %lu hash %lu\n",
1026 i, j, !i ? 0 : (1UL << (i - 1)) + j);
1027 fini_node->p.reverse_hash =
1028 bit_reverse_ulong(!i ? 0 : (1UL << (i - 1)) + j);
1029 (void) _cds_lfht_remove(ht, !i ? 0 : (1UL << (i - 1)),
1030 fini_node, 1);
1031 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1032 break;
1033 }
1034 ht->cds_lfht_rcu_read_unlock();
1035 ht->cds_lfht_rcu_thread_offline();
1036 }
1037
1038 static
1039 void fini_table(struct cds_lfht *ht,
1040 unsigned long first_order, unsigned long len_order)
1041 {
1042 long i, end_order;
1043
1044 dbg_printf("fini table: first_order %lu end_order %lu\n",
1045 first_order, first_order + len_order);
1046 end_order = first_order + len_order;
1047 assert(first_order > 0);
1048 for (i = end_order - 1; i >= first_order; i--) {
1049 unsigned long len;
1050
1051 len = !i ? 1 : 1UL << (i - 1);
1052 dbg_printf("fini order %lu len: %lu\n", i, len);
1053
1054 /* Stop shrink if the resize target changes under us */
1055 if (CMM_LOAD_SHARED(ht->t.resize_target) > (1UL << (i - 1)))
1056 break;
1057
1058 cmm_smp_wmb(); /* populate data before RCU size */
1059 CMM_STORE_SHARED(ht->t.size, 1UL << (i - 1));
1060
1061 /*
1062 * We need to wait for all add operations to reach Q.S. (and
1063 * thus use the new table for lookups) before we can start
1064 * releasing the old dummy nodes. Otherwise their lookup will
1065 * return a logically removed node as insert position.
1066 */
1067 ht->cds_lfht_synchronize_rcu();
1068
1069 /*
1070 * Set "removed" flag in dummy nodes about to be removed.
1071 * Unlink all now-logically-removed dummy node pointers.
1072 * Concurrent add/remove operation are helping us doing
1073 * the gc.
1074 */
1075 remove_table(ht, i, len);
1076
1077 ht->cds_lfht_call_rcu(&ht->t.tbl[i]->head, cds_lfht_free_level);
1078
1079 dbg_printf("fini new size: %lu\n", 1UL << i);
1080 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1081 break;
1082 }
1083 }
1084
1085 struct cds_lfht *cds_lfht_new(cds_lfht_hash_fct hash_fct,
1086 cds_lfht_compare_fct compare_fct,
1087 unsigned long hash_seed,
1088 unsigned long init_size,
1089 int flags,
1090 void (*cds_lfht_call_rcu)(struct rcu_head *head,
1091 void (*func)(struct rcu_head *head)),
1092 void (*cds_lfht_synchronize_rcu)(void),
1093 void (*cds_lfht_rcu_read_lock)(void),
1094 void (*cds_lfht_rcu_read_unlock)(void),
1095 void (*cds_lfht_rcu_thread_offline)(void),
1096 void (*cds_lfht_rcu_thread_online)(void))
1097 {
1098 struct cds_lfht *ht;
1099 unsigned long order;
1100
1101 /* init_size must be power of two */
1102 if (init_size && (init_size & (init_size - 1)))
1103 return NULL;
1104 ht = calloc(1, sizeof(struct cds_lfht));
1105 ht->hash_fct = hash_fct;
1106 ht->compare_fct = compare_fct;
1107 ht->hash_seed = hash_seed;
1108 ht->cds_lfht_call_rcu = cds_lfht_call_rcu;
1109 ht->cds_lfht_synchronize_rcu = cds_lfht_synchronize_rcu;
1110 ht->cds_lfht_rcu_read_lock = cds_lfht_rcu_read_lock;
1111 ht->cds_lfht_rcu_read_unlock = cds_lfht_rcu_read_unlock;
1112 ht->cds_lfht_rcu_thread_offline = cds_lfht_rcu_thread_offline;
1113 ht->cds_lfht_rcu_thread_online = cds_lfht_rcu_thread_online;
1114 ht->percpu_count = alloc_per_cpu_items_count();
1115 /* this mutex should not nest in read-side C.S. */
1116 pthread_mutex_init(&ht->resize_mutex, NULL);
1117 order = get_count_order_ulong(max(init_size, MIN_TABLE_SIZE)) + 1;
1118 ht->flags = flags;
1119 ht->cds_lfht_rcu_thread_offline();
1120 pthread_mutex_lock(&ht->resize_mutex);
1121 ht->t.resize_target = 1UL << (order - 1);
1122 init_table(ht, 0, order);
1123 pthread_mutex_unlock(&ht->resize_mutex);
1124 ht->cds_lfht_rcu_thread_online();
1125 return ht;
1126 }
1127
1128 struct cds_lfht_node *cds_lfht_lookup(struct cds_lfht *ht, void *key, size_t key_len)
1129 {
1130 struct cds_lfht_node *node, *next;
1131 struct _cds_lfht_node *lookup;
1132 unsigned long hash, reverse_hash, index, order, size;
1133
1134 hash = ht->hash_fct(key, key_len, ht->hash_seed);
1135 reverse_hash = bit_reverse_ulong(hash);
1136
1137 size = rcu_dereference(ht->t.size);
1138 index = hash & (size - 1);
1139 order = get_count_order_ulong(index + 1);
1140 lookup = &ht->t.tbl[order]->nodes[index & (!order ? 0 : ((1UL << (order - 1))) - 1)];
1141 dbg_printf("lookup hash %lu index %lu order %lu aridx %lu\n",
1142 hash, index, order, index & (!order ? 0 : ((1UL << (order - 1)) - 1)));
1143 node = (struct cds_lfht_node *) lookup;
1144 for (;;) {
1145 if (unlikely(!node))
1146 break;
1147 if (unlikely(node->p.reverse_hash > reverse_hash)) {
1148 node = NULL;
1149 break;
1150 }
1151 next = rcu_dereference(node->p.next);
1152 if (likely(!is_removed(next))
1153 && !is_dummy(next)
1154 && likely(!ht->compare_fct(node->key, node->key_len, key, key_len))) {
1155 break;
1156 }
1157 node = clear_flag(next);
1158 }
1159 assert(!node || !is_dummy(rcu_dereference(node->p.next)));
1160 return node;
1161 }
1162
1163 struct cds_lfht_node *cds_lfht_next(struct cds_lfht *ht,
1164 struct cds_lfht_node *node)
1165 {
1166 struct cds_lfht_node *next;
1167 unsigned long reverse_hash;
1168 void *key;
1169 size_t key_len;
1170
1171 reverse_hash = node->p.reverse_hash;
1172 key = node->key;
1173 key_len = node->key_len;
1174 next = rcu_dereference(node->p.next);
1175 node = clear_flag(next);
1176
1177 for (;;) {
1178 if (unlikely(!node))
1179 break;
1180 if (unlikely(node->p.reverse_hash > reverse_hash)) {
1181 node = NULL;
1182 break;
1183 }
1184 next = rcu_dereference(node->p.next);
1185 if (likely(!is_removed(next))
1186 && !is_dummy(next)
1187 && likely(!ht->compare_fct(node->key, node->key_len, key, key_len))) {
1188 break;
1189 }
1190 node = clear_flag(next);
1191 }
1192 assert(!node || !is_dummy(rcu_dereference(node->p.next)));
1193 return node;
1194 }
1195
1196 void cds_lfht_add(struct cds_lfht *ht, struct cds_lfht_node *node)
1197 {
1198 unsigned long hash, size;
1199
1200 hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed);
1201 node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash);
1202
1203 size = rcu_dereference(ht->t.size);
1204 (void) _cds_lfht_add(ht, size, node, 0, 0);
1205 ht_count_add(ht, size);
1206 }
1207
1208 struct cds_lfht_node *cds_lfht_add_unique(struct cds_lfht *ht,
1209 struct cds_lfht_node *node)
1210 {
1211 unsigned long hash, size;
1212 struct cds_lfht_node *ret;
1213
1214 hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed);
1215 node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash);
1216
1217 size = rcu_dereference(ht->t.size);
1218 ret = _cds_lfht_add(ht, size, node, 1, 0);
1219 if (ret != node)
1220 ht_count_add(ht, size);
1221 return ret;
1222 }
1223
1224 int cds_lfht_remove(struct cds_lfht *ht, struct cds_lfht_node *node)
1225 {
1226 unsigned long size;
1227 int ret;
1228
1229 size = rcu_dereference(ht->t.size);
1230 ret = _cds_lfht_remove(ht, size, node, 0);
1231 if (!ret)
1232 ht_count_remove(ht, size);
1233 return ret;
1234 }
1235
1236 static
1237 int cds_lfht_delete_dummy(struct cds_lfht *ht)
1238 {
1239 struct cds_lfht_node *node;
1240 struct _cds_lfht_node *lookup;
1241 unsigned long order, i, size;
1242
1243 /* Check that the table is empty */
1244 lookup = &ht->t.tbl[0]->nodes[0];
1245 node = (struct cds_lfht_node *) lookup;
1246 do {
1247 node = clear_flag(node)->p.next;
1248 if (!is_dummy(node))
1249 return -EPERM;
1250 assert(!is_removed(node));
1251 } while (clear_flag(node));
1252 /*
1253 * size accessed without rcu_dereference because hash table is
1254 * being destroyed.
1255 */
1256 size = ht->t.size;
1257 /* Internal sanity check: all nodes left should be dummy */
1258 for (order = 0; order < get_count_order_ulong(size) + 1; order++) {
1259 unsigned long len;
1260
1261 len = !order ? 1 : 1UL << (order - 1);
1262 for (i = 0; i < len; i++) {
1263 dbg_printf("delete order %lu i %lu hash %lu\n",
1264 order, i,
1265 bit_reverse_ulong(ht->t.tbl[order]->nodes[i].reverse_hash));
1266 assert(is_dummy(ht->t.tbl[order]->nodes[i].next));
1267 }
1268 poison_free(ht->t.tbl[order]);
1269 }
1270 return 0;
1271 }
1272
1273 /*
1274 * Should only be called when no more concurrent readers nor writers can
1275 * possibly access the table.
1276 */
1277 int cds_lfht_destroy(struct cds_lfht *ht)
1278 {
1279 int ret;
1280
1281 /* Wait for in-flight resize operations to complete */
1282 CMM_STORE_SHARED(ht->in_progress_destroy, 1);
1283 while (uatomic_read(&ht->in_progress_resize))
1284 poll(NULL, 0, 100); /* wait for 100ms */
1285 ret = cds_lfht_delete_dummy(ht);
1286 if (ret)
1287 return ret;
1288 free_per_cpu_items_count(ht->percpu_count);
1289 poison_free(ht);
1290 return ret;
1291 }
1292
1293 void cds_lfht_count_nodes(struct cds_lfht *ht,
1294 unsigned long *count,
1295 unsigned long *removed)
1296 {
1297 struct cds_lfht_node *node, *next;
1298 struct _cds_lfht_node *lookup;
1299 unsigned long nr_dummy = 0;
1300
1301 *count = 0;
1302 *removed = 0;
1303
1304 /* Count non-dummy nodes in the table */
1305 lookup = &ht->t.tbl[0]->nodes[0];
1306 node = (struct cds_lfht_node *) lookup;
1307 do {
1308 next = rcu_dereference(node->p.next);
1309 if (is_removed(next)) {
1310 assert(!is_dummy(next));
1311 (*removed)++;
1312 } else if (!is_dummy(next))
1313 (*count)++;
1314 else
1315 (nr_dummy)++;
1316 node = clear_flag(next);
1317 } while (node);
1318 dbg_printf("number of dummy nodes: %lu\n", nr_dummy);
1319 }
1320
1321 /* called with resize mutex held */
1322 static
1323 void _do_cds_lfht_grow(struct cds_lfht *ht,
1324 unsigned long old_size, unsigned long new_size)
1325 {
1326 unsigned long old_order, new_order;
1327
1328 old_order = get_count_order_ulong(old_size) + 1;
1329 new_order = get_count_order_ulong(new_size) + 1;
1330 printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1331 old_size, old_order, new_size, new_order);
1332 assert(new_size > old_size);
1333 init_table(ht, old_order, new_order - old_order);
1334 }
1335
1336 /* called with resize mutex held */
1337 static
1338 void _do_cds_lfht_shrink(struct cds_lfht *ht,
1339 unsigned long old_size, unsigned long new_size)
1340 {
1341 unsigned long old_order, new_order;
1342
1343 new_size = max(new_size, MIN_TABLE_SIZE);
1344 old_order = get_count_order_ulong(old_size) + 1;
1345 new_order = get_count_order_ulong(new_size) + 1;
1346 printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1347 old_size, old_order, new_size, new_order);
1348 assert(new_size < old_size);
1349
1350 /* Remove and unlink all dummy nodes to remove. */
1351 fini_table(ht, new_order, old_order - new_order);
1352 }
1353
1354
1355 /* called with resize mutex held */
1356 static
1357 void _do_cds_lfht_resize(struct cds_lfht *ht)
1358 {
1359 unsigned long new_size, old_size;
1360
1361 /*
1362 * Resize table, re-do if the target size has changed under us.
1363 */
1364 do {
1365 ht->t.resize_initiated = 1;
1366 old_size = ht->t.size;
1367 new_size = CMM_LOAD_SHARED(ht->t.resize_target);
1368 if (old_size < new_size)
1369 _do_cds_lfht_grow(ht, old_size, new_size);
1370 else if (old_size > new_size)
1371 _do_cds_lfht_shrink(ht, old_size, new_size);
1372 ht->t.resize_initiated = 0;
1373 /* write resize_initiated before read resize_target */
1374 cmm_smp_mb();
1375 } while (ht->t.size != CMM_LOAD_SHARED(ht->t.resize_target));
1376 }
1377
1378 static
1379 unsigned long resize_target_update(struct cds_lfht *ht, unsigned long size,
1380 int growth_order)
1381 {
1382 return _uatomic_max(&ht->t.resize_target,
1383 size << growth_order);
1384 }
1385
1386 static
1387 void resize_target_update_count(struct cds_lfht *ht,
1388 unsigned long count)
1389 {
1390 count = max(count, MIN_TABLE_SIZE);
1391 uatomic_set(&ht->t.resize_target, count);
1392 }
1393
1394 void cds_lfht_resize(struct cds_lfht *ht, unsigned long new_size)
1395 {
1396 resize_target_update_count(ht, new_size);
1397 CMM_STORE_SHARED(ht->t.resize_initiated, 1);
1398 ht->cds_lfht_rcu_thread_offline();
1399 pthread_mutex_lock(&ht->resize_mutex);
1400 _do_cds_lfht_resize(ht);
1401 pthread_mutex_unlock(&ht->resize_mutex);
1402 ht->cds_lfht_rcu_thread_online();
1403 }
1404
1405 static
1406 void do_resize_cb(struct rcu_head *head)
1407 {
1408 struct rcu_resize_work *work =
1409 caa_container_of(head, struct rcu_resize_work, head);
1410 struct cds_lfht *ht = work->ht;
1411
1412 ht->cds_lfht_rcu_thread_offline();
1413 pthread_mutex_lock(&ht->resize_mutex);
1414 _do_cds_lfht_resize(ht);
1415 pthread_mutex_unlock(&ht->resize_mutex);
1416 ht->cds_lfht_rcu_thread_online();
1417 poison_free(work);
1418 cmm_smp_mb(); /* finish resize before decrement */
1419 uatomic_dec(&ht->in_progress_resize);
1420 }
1421
1422 static
1423 void cds_lfht_resize_lazy(struct cds_lfht *ht, unsigned long size, int growth)
1424 {
1425 struct rcu_resize_work *work;
1426 unsigned long target_size;
1427
1428 target_size = resize_target_update(ht, size, growth);
1429 /* Store resize_target before read resize_initiated */
1430 cmm_smp_mb();
1431 if (!CMM_LOAD_SHARED(ht->t.resize_initiated) && size < target_size) {
1432 uatomic_inc(&ht->in_progress_resize);
1433 cmm_smp_mb(); /* increment resize count before calling it */
1434 work = malloc(sizeof(*work));
1435 work->ht = ht;
1436 ht->cds_lfht_call_rcu(&work->head, do_resize_cb);
1437 CMM_STORE_SHARED(ht->t.resize_initiated, 1);
1438 }
1439 }
1440
1441 #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
1442
1443 static
1444 void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
1445 unsigned long count)
1446 {
1447 struct rcu_resize_work *work;
1448
1449 if (!(ht->flags & CDS_LFHT_AUTO_RESIZE))
1450 return;
1451 resize_target_update_count(ht, count);
1452 /* Store resize_target before read resize_initiated */
1453 cmm_smp_mb();
1454 if (!CMM_LOAD_SHARED(ht->t.resize_initiated)) {
1455 uatomic_inc(&ht->in_progress_resize);
1456 cmm_smp_mb(); /* increment resize count before calling it */
1457 work = malloc(sizeof(*work));
1458 work->ht = ht;
1459 ht->cds_lfht_call_rcu(&work->head, do_resize_cb);
1460 CMM_STORE_SHARED(ht->t.resize_initiated, 1);
1461 }
1462 }
1463
1464 #endif
This page took 0.095876 seconds and 3 git commands to generate.