034c7f9053cfbd881496b592cbb3bcddf92a1d8f
[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 1
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 /*
181 * Minimum number of dummy nodes to touch per thread to parallelize grow/shrink.
182 */
183 #define MIN_PARTITION_PER_THREAD_ORDER 12
184 #define MIN_PARTITION_PER_THREAD (1UL << MIN_PARTITION_PER_THREAD_ORDER)
185
186 #ifndef min
187 #define min(a, b) ((a) < (b) ? (a) : (b))
188 #endif
189
190 #ifndef max
191 #define max(a, b) ((a) > (b) ? (a) : (b))
192 #endif
193
194 /*
195 * The removed flag needs to be updated atomically with the pointer.
196 * It indicates that no node must attach to the node scheduled for
197 * removal. The gc flag also needs to be updated atomically with the
198 * pointer. It indicates that node garbage collection must be performed.
199 * The dummy flag does not require to be updated atomically with the
200 * pointer, but it is added as a pointer low bit flag to save space.
201 */
202 #define REMOVED_FLAG (1UL << 0)
203 #define GC_FLAG (1UL << 1)
204 #define DUMMY_FLAG (1UL << 2)
205 #define FLAGS_MASK ((1UL << 3) - 1)
206
207 /* Value of the end pointer. Should not interact with flags. */
208 #define END_VALUE NULL
209
210 struct ht_items_count {
211 unsigned long add, del;
212 } __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
213
214 struct rcu_level {
215 struct rcu_head head;
216 struct _cds_lfht_node nodes[0];
217 };
218
219 struct rcu_table {
220 unsigned long size; /* always a power of 2, shared (RCU) */
221 unsigned long resize_target;
222 int resize_initiated;
223 struct rcu_level *tbl[MAX_TABLE_ORDER];
224 };
225
226 struct cds_lfht {
227 struct rcu_table t;
228 cds_lfht_hash_fct hash_fct;
229 cds_lfht_compare_fct compare_fct;
230 unsigned long hash_seed;
231 int flags;
232 /*
233 * We need to put the work threads offline (QSBR) when taking this
234 * mutex, because we use synchronize_rcu within this mutex critical
235 * section, which waits on read-side critical sections, and could
236 * therefore cause grace-period deadlock if we hold off RCU G.P.
237 * completion.
238 */
239 pthread_mutex_t resize_mutex; /* resize mutex: add/del mutex */
240 unsigned int in_progress_resize, in_progress_destroy;
241 void (*cds_lfht_call_rcu)(struct rcu_head *head,
242 void (*func)(struct rcu_head *head));
243 void (*cds_lfht_synchronize_rcu)(void);
244 void (*cds_lfht_rcu_read_lock)(void);
245 void (*cds_lfht_rcu_read_unlock)(void);
246 void (*cds_lfht_rcu_thread_offline)(void);
247 void (*cds_lfht_rcu_thread_online)(void);
248 void (*cds_lfht_rcu_register_thread)(void);
249 void (*cds_lfht_rcu_unregister_thread)(void);
250 pthread_attr_t *resize_attr; /* Resize threads attributes */
251 unsigned long count; /* global approximate item count */
252 struct ht_items_count *percpu_count; /* per-cpu item count */
253 };
254
255 struct rcu_resize_work {
256 struct rcu_head head;
257 struct cds_lfht *ht;
258 };
259
260 struct partition_resize_work {
261 struct rcu_head head;
262 struct cds_lfht *ht;
263 unsigned long i, start, len;
264 void (*fct)(struct cds_lfht *ht, unsigned long i,
265 unsigned long start, unsigned long len);
266 };
267
268 enum add_mode {
269 ADD_DEFAULT = 0,
270 ADD_UNIQUE = 1,
271 ADD_REPLACE = 2,
272 };
273
274 static
275 struct cds_lfht_node *_cds_lfht_add(struct cds_lfht *ht,
276 unsigned long size,
277 struct cds_lfht_node *node,
278 enum add_mode mode, int dummy);
279
280 static
281 int _cds_lfht_del(struct cds_lfht *ht, unsigned long size,
282 struct cds_lfht_node *node,
283 int dummy_removal, int do_gc);
284
285 /*
286 * Algorithm to reverse bits in a word by lookup table, extended to
287 * 64-bit words.
288 * Source:
289 * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable
290 * Originally from Public Domain.
291 */
292
293 static const uint8_t BitReverseTable256[256] =
294 {
295 #define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64
296 #define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16)
297 #define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 )
298 R6(0), R6(2), R6(1), R6(3)
299 };
300 #undef R2
301 #undef R4
302 #undef R6
303
304 static
305 uint8_t bit_reverse_u8(uint8_t v)
306 {
307 return BitReverseTable256[v];
308 }
309
310 static __attribute__((unused))
311 uint32_t bit_reverse_u32(uint32_t v)
312 {
313 return ((uint32_t) bit_reverse_u8(v) << 24) |
314 ((uint32_t) bit_reverse_u8(v >> 8) << 16) |
315 ((uint32_t) bit_reverse_u8(v >> 16) << 8) |
316 ((uint32_t) bit_reverse_u8(v >> 24));
317 }
318
319 static __attribute__((unused))
320 uint64_t bit_reverse_u64(uint64_t v)
321 {
322 return ((uint64_t) bit_reverse_u8(v) << 56) |
323 ((uint64_t) bit_reverse_u8(v >> 8) << 48) |
324 ((uint64_t) bit_reverse_u8(v >> 16) << 40) |
325 ((uint64_t) bit_reverse_u8(v >> 24) << 32) |
326 ((uint64_t) bit_reverse_u8(v >> 32) << 24) |
327 ((uint64_t) bit_reverse_u8(v >> 40) << 16) |
328 ((uint64_t) bit_reverse_u8(v >> 48) << 8) |
329 ((uint64_t) bit_reverse_u8(v >> 56));
330 }
331
332 static
333 unsigned long bit_reverse_ulong(unsigned long v)
334 {
335 #if (CAA_BITS_PER_LONG == 32)
336 return bit_reverse_u32(v);
337 #else
338 return bit_reverse_u64(v);
339 #endif
340 }
341
342 /*
343 * fls: returns the position of the most significant bit.
344 * Returns 0 if no bit is set, else returns the position of the most
345 * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit).
346 */
347 #if defined(__i386) || defined(__x86_64)
348 static inline
349 unsigned int fls_u32(uint32_t x)
350 {
351 int r;
352
353 asm("bsrl %1,%0\n\t"
354 "jnz 1f\n\t"
355 "movl $-1,%0\n\t"
356 "1:\n\t"
357 : "=r" (r) : "rm" (x));
358 return r + 1;
359 }
360 #define HAS_FLS_U32
361 #endif
362
363 #if defined(__x86_64)
364 static inline
365 unsigned int fls_u64(uint64_t x)
366 {
367 long r;
368
369 asm("bsrq %1,%0\n\t"
370 "jnz 1f\n\t"
371 "movq $-1,%0\n\t"
372 "1:\n\t"
373 : "=r" (r) : "rm" (x));
374 return r + 1;
375 }
376 #define HAS_FLS_U64
377 #endif
378
379 #ifndef HAS_FLS_U64
380 static __attribute__((unused))
381 unsigned int fls_u64(uint64_t x)
382 {
383 unsigned int r = 64;
384
385 if (!x)
386 return 0;
387
388 if (!(x & 0xFFFFFFFF00000000ULL)) {
389 x <<= 32;
390 r -= 32;
391 }
392 if (!(x & 0xFFFF000000000000ULL)) {
393 x <<= 16;
394 r -= 16;
395 }
396 if (!(x & 0xFF00000000000000ULL)) {
397 x <<= 8;
398 r -= 8;
399 }
400 if (!(x & 0xF000000000000000ULL)) {
401 x <<= 4;
402 r -= 4;
403 }
404 if (!(x & 0xC000000000000000ULL)) {
405 x <<= 2;
406 r -= 2;
407 }
408 if (!(x & 0x8000000000000000ULL)) {
409 x <<= 1;
410 r -= 1;
411 }
412 return r;
413 }
414 #endif
415
416 #ifndef HAS_FLS_U32
417 static __attribute__((unused))
418 unsigned int fls_u32(uint32_t x)
419 {
420 unsigned int r = 32;
421
422 if (!x)
423 return 0;
424 if (!(x & 0xFFFF0000U)) {
425 x <<= 16;
426 r -= 16;
427 }
428 if (!(x & 0xFF000000U)) {
429 x <<= 8;
430 r -= 8;
431 }
432 if (!(x & 0xF0000000U)) {
433 x <<= 4;
434 r -= 4;
435 }
436 if (!(x & 0xC0000000U)) {
437 x <<= 2;
438 r -= 2;
439 }
440 if (!(x & 0x80000000U)) {
441 x <<= 1;
442 r -= 1;
443 }
444 return r;
445 }
446 #endif
447
448 unsigned int fls_ulong(unsigned long x)
449 {
450 #if (CAA_BITS_PER_lONG == 32)
451 return fls_u32(x);
452 #else
453 return fls_u64(x);
454 #endif
455 }
456
457 int get_count_order_u32(uint32_t x)
458 {
459 int order;
460
461 order = fls_u32(x) - 1;
462 if (x & (x - 1))
463 order++;
464 return order;
465 }
466
467 int get_count_order_ulong(unsigned long x)
468 {
469 int order;
470
471 order = fls_ulong(x) - 1;
472 if (x & (x - 1))
473 order++;
474 return order;
475 }
476
477 #ifdef POISON_FREE
478 #define poison_free(ptr) \
479 do { \
480 memset(ptr, 0x42, sizeof(*(ptr))); \
481 free(ptr); \
482 } while (0)
483 #else
484 #define poison_free(ptr) free(ptr)
485 #endif
486
487 static
488 void cds_lfht_resize_lazy(struct cds_lfht *ht, unsigned long size, int growth);
489
490 /*
491 * If the sched_getcpu() and sysconf(_SC_NPROCESSORS_CONF) calls are
492 * available, then we support hash table item accounting.
493 * In the unfortunate event the number of CPUs reported would be
494 * inaccurate, we use modulo arithmetic on the number of CPUs we got.
495 */
496 #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
497
498 static
499 void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
500 unsigned long count);
501
502 static long nr_cpus_mask = -1;
503
504 static
505 struct ht_items_count *alloc_per_cpu_items_count(void)
506 {
507 struct ht_items_count *count;
508
509 switch (nr_cpus_mask) {
510 case -2:
511 return NULL;
512 case -1:
513 {
514 long maxcpus;
515
516 maxcpus = sysconf(_SC_NPROCESSORS_CONF);
517 if (maxcpus <= 0) {
518 nr_cpus_mask = -2;
519 return NULL;
520 }
521 /*
522 * round up number of CPUs to next power of two, so we
523 * can use & for modulo.
524 */
525 maxcpus = 1UL << get_count_order_ulong(maxcpus);
526 nr_cpus_mask = maxcpus - 1;
527 }
528 /* Fall-through */
529 default:
530 return calloc(nr_cpus_mask + 1, sizeof(*count));
531 }
532 }
533
534 static
535 void free_per_cpu_items_count(struct ht_items_count *count)
536 {
537 poison_free(count);
538 }
539
540 static
541 int ht_get_cpu(void)
542 {
543 int cpu;
544
545 assert(nr_cpus_mask >= 0);
546 cpu = sched_getcpu();
547 if (unlikely(cpu < 0))
548 return cpu;
549 else
550 return cpu & nr_cpus_mask;
551 }
552
553 static
554 void ht_count_add(struct cds_lfht *ht, unsigned long size)
555 {
556 unsigned long percpu_count;
557 int cpu;
558
559 if (unlikely(!ht->percpu_count))
560 return;
561 cpu = ht_get_cpu();
562 if (unlikely(cpu < 0))
563 return;
564 percpu_count = uatomic_add_return(&ht->percpu_count[cpu].add, 1);
565 if (unlikely(!(percpu_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) {
566 unsigned long count;
567
568 dbg_printf("add percpu %lu\n", percpu_count);
569 count = uatomic_add_return(&ht->count,
570 1UL << COUNT_COMMIT_ORDER);
571 /* If power of 2 */
572 if (!(count & (count - 1))) {
573 if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) < size)
574 return;
575 dbg_printf("add set global %lu\n", count);
576 cds_lfht_resize_lazy_count(ht, size,
577 count >> (CHAIN_LEN_TARGET - 1));
578 }
579 }
580 }
581
582 static
583 void ht_count_del(struct cds_lfht *ht, unsigned long size)
584 {
585 unsigned long percpu_count;
586 int cpu;
587
588 if (unlikely(!ht->percpu_count))
589 return;
590 cpu = ht_get_cpu();
591 if (unlikely(cpu < 0))
592 return;
593 percpu_count = uatomic_add_return(&ht->percpu_count[cpu].del, -1);
594 if (unlikely(!(percpu_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) {
595 unsigned long count;
596
597 dbg_printf("del percpu %lu\n", percpu_count);
598 count = uatomic_add_return(&ht->count,
599 -(1UL << COUNT_COMMIT_ORDER));
600 /* If power of 2 */
601 if (!(count & (count - 1))) {
602 if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) >= size)
603 return;
604 dbg_printf("del set global %lu\n", count);
605 cds_lfht_resize_lazy_count(ht, size,
606 count >> (CHAIN_LEN_TARGET - 1));
607 }
608 }
609 }
610
611 #else /* #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
612
613 static const long nr_cpus_mask = -1;
614
615 static
616 struct ht_items_count *alloc_per_cpu_items_count(void)
617 {
618 return NULL;
619 }
620
621 static
622 void free_per_cpu_items_count(struct ht_items_count *count)
623 {
624 }
625
626 static
627 void ht_count_add(struct cds_lfht *ht, unsigned long size)
628 {
629 }
630
631 static
632 void ht_count_del(struct cds_lfht *ht, unsigned long size)
633 {
634 }
635
636 #endif /* #else #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
637
638
639 static
640 void check_resize(struct cds_lfht *ht, unsigned long size, uint32_t chain_len)
641 {
642 unsigned long count;
643
644 if (!(ht->flags & CDS_LFHT_AUTO_RESIZE))
645 return;
646 count = uatomic_read(&ht->count);
647 /*
648 * Use bucket-local length for small table expand and for
649 * environments lacking per-cpu data support.
650 */
651 if (count >= (1UL << COUNT_COMMIT_ORDER))
652 return;
653 if (chain_len > 100)
654 dbg_printf("WARNING: large chain length: %u.\n",
655 chain_len);
656 if (chain_len >= CHAIN_LEN_RESIZE_THRESHOLD)
657 cds_lfht_resize_lazy(ht, size,
658 get_count_order_u32(chain_len - (CHAIN_LEN_TARGET - 1)));
659 }
660
661 static
662 struct cds_lfht_node *clear_flag(struct cds_lfht_node *node)
663 {
664 return (struct cds_lfht_node *) (((unsigned long) node) & ~FLAGS_MASK);
665 }
666
667 static
668 int is_removed(struct cds_lfht_node *node)
669 {
670 return ((unsigned long) node) & REMOVED_FLAG;
671 }
672
673 static
674 struct cds_lfht_node *flag_removed(struct cds_lfht_node *node)
675 {
676 return (struct cds_lfht_node *) (((unsigned long) node) | REMOVED_FLAG);
677 }
678
679 static
680 int is_gc(struct cds_lfht_node *node)
681 {
682 return ((unsigned long) node) & GC_FLAG;
683 }
684
685 static
686 struct cds_lfht_node *flag_gc(struct cds_lfht_node *node)
687 {
688 return (struct cds_lfht_node *) (((unsigned long) node) | GC_FLAG);
689 }
690
691 static
692 int is_dummy(struct cds_lfht_node *node)
693 {
694 return ((unsigned long) node) & DUMMY_FLAG;
695 }
696
697 static
698 struct cds_lfht_node *flag_dummy(struct cds_lfht_node *node)
699 {
700 return (struct cds_lfht_node *) (((unsigned long) node) | DUMMY_FLAG);
701 }
702
703 static
704 struct cds_lfht_node *get_end(void)
705 {
706 return (struct cds_lfht_node *) END_VALUE;
707 }
708
709 static
710 int is_end(struct cds_lfht_node *node)
711 {
712 return clear_flag(node) == (struct cds_lfht_node *) END_VALUE;
713 }
714
715 static
716 unsigned long _uatomic_max(unsigned long *ptr, unsigned long v)
717 {
718 unsigned long old1, old2;
719
720 old1 = uatomic_read(ptr);
721 do {
722 old2 = old1;
723 if (old2 >= v)
724 return old2;
725 } while ((old1 = uatomic_cmpxchg(ptr, old2, v)) != old2);
726 return v;
727 }
728
729 static
730 void cds_lfht_free_level(struct rcu_head *head)
731 {
732 struct rcu_level *l =
733 caa_container_of(head, struct rcu_level, head);
734 poison_free(l);
735 }
736
737 /*
738 * Remove all logically deleted nodes from a bucket up to a certain node key.
739 */
740 static
741 void _cds_lfht_gc_bucket(struct cds_lfht_node *dummy, struct cds_lfht_node *node)
742 {
743 struct cds_lfht_node *iter_prev, *iter, *next, *new_next;
744
745 assert(!is_dummy(dummy));
746 assert(!is_gc(dummy));
747 assert(!is_removed(dummy));
748 assert(!is_dummy(node));
749 assert(!is_gc(node));
750 assert(!is_removed(node));
751 for (;;) {
752 iter_prev = dummy;
753 /* We can always skip the dummy node initially */
754 iter = rcu_dereference(iter_prev->p.next);
755 assert(iter_prev->p.reverse_hash <= node->p.reverse_hash);
756 /*
757 * We should never be called with dummy (start of chain)
758 * and logically removed node (end of path compression
759 * marker) being the actual same node. This would be a
760 * bug in the algorithm implementation.
761 */
762 assert(dummy != node);
763 for (;;) {
764 if (unlikely(is_end(iter)))
765 return;
766 if (likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash))
767 return;
768 next = rcu_dereference(clear_flag(iter)->p.next);
769 if (likely(is_gc(next)))
770 break;
771 iter_prev = clear_flag(iter);
772 iter = next;
773 }
774 assert(!is_gc(iter));
775 if (is_dummy(iter))
776 new_next = flag_dummy(clear_flag(next));
777 else
778 new_next = clear_flag(next);
779 if (is_removed(iter))
780 new_next = flag_removed(new_next);
781 (void) uatomic_cmpxchg(&iter_prev->p.next, iter, new_next);
782 }
783 return;
784 }
785
786 static
787 struct cds_lfht_node *_cds_lfht_add(struct cds_lfht *ht,
788 unsigned long size,
789 struct cds_lfht_node *node,
790 enum add_mode mode, int dummy)
791 {
792 struct cds_lfht_node *iter_prev, *iter, *next, *new_node, *new_next,
793 *dummy_node, *return_node;
794 struct _cds_lfht_node *lookup;
795 unsigned long hash, index, order;
796
797 assert(!is_dummy(node));
798 assert(!is_gc(node));
799 assert(!is_removed(node));
800 if (!size) {
801 assert(dummy);
802 node->p.next = flag_dummy(get_end());
803 return node; /* Initial first add (head) */
804 }
805 hash = bit_reverse_ulong(node->p.reverse_hash);
806 for (;;) {
807 uint32_t chain_len = 0;
808
809 /*
810 * iter_prev points to the non-removed node prior to the
811 * insert location.
812 */
813 index = hash & (size - 1);
814 order = get_count_order_ulong(index + 1);
815 lookup = &ht->t.tbl[order]->nodes[index & ((!order ? 0 : (1UL << (order - 1))) - 1)];
816 iter_prev = (struct cds_lfht_node *) lookup;
817 /* We can always skip the dummy node initially */
818 iter = rcu_dereference(iter_prev->p.next);
819 assert(iter_prev->p.reverse_hash <= node->p.reverse_hash);
820 for (;;) {
821 if (unlikely(is_end(iter)))
822 goto insert;
823 if (likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash))
824 goto insert;
825 next = rcu_dereference(clear_flag(iter)->p.next);
826 if (unlikely(is_gc(next)))
827 goto gc_node;
828 assert(!is_removed(next));
829 if ((mode == ADD_UNIQUE || mode == ADD_REPLACE)
830 && !is_dummy(next)
831 && !ht->compare_fct(node->key, node->key_len,
832 clear_flag(iter)->key,
833 clear_flag(iter)->key_len)) {
834 if (mode == ADD_UNIQUE)
835 return clear_flag(iter);
836 else /* mode == ADD_REPLACE */
837 goto replace;
838 }
839 /* Only account for identical reverse hash once */
840 if (iter_prev->p.reverse_hash != clear_flag(iter)->p.reverse_hash
841 && !is_dummy(next))
842 check_resize(ht, size, ++chain_len);
843 iter_prev = clear_flag(iter);
844 iter = next;
845 }
846
847 insert:
848 assert(node != clear_flag(iter));
849 assert(!is_removed(iter_prev));
850 assert(!is_removed(iter));
851 assert(!is_gc(iter_prev));
852 assert(!is_gc(iter));
853 assert(iter_prev != node);
854 if (!dummy)
855 node->p.next = clear_flag(iter);
856 else
857 node->p.next = flag_dummy(clear_flag(iter));
858 if (is_dummy(iter))
859 new_node = flag_dummy(node);
860 else
861 new_node = node;
862 if (uatomic_cmpxchg(&iter_prev->p.next, iter,
863 new_node) != iter) {
864 continue; /* retry */
865 } else {
866 if (mode == ADD_REPLACE)
867 return_node = NULL;
868 else /* ADD_DEFAULT and ADD_UNIQUE */
869 return_node = node;
870 goto gc_end;
871 }
872
873 replace:
874 /* Insert after node to be replaced */
875 iter_prev = clear_flag(iter);
876 iter = next;
877 assert(node != clear_flag(iter));
878 assert(!is_removed(iter_prev));
879 assert(!is_removed(iter));
880 assert(!is_gc(iter_prev));
881 assert(!is_gc(iter));
882 assert(iter_prev != node);
883 assert(!dummy);
884 node->p.next = clear_flag(iter);
885 if (is_dummy(iter))
886 new_node = flag_dummy(node);
887 else
888 new_node = node;
889 /*
890 * Here is the whole trick for lock-free replace: we add
891 * the replacement node _after_ the node we want to
892 * replace by atomically setting its next pointer at the
893 * same time we set its removal and gc flags. Given that
894 * the lookups/get next use an iterator aware of the
895 * next pointer, they will either skip the old node due
896 * to the removal/gc flag and see the new node, or use
897 * the old new, but will not see the new one.
898 */
899 new_node = flag_removed(new_node);
900 new_node = flag_gc(new_node);
901 if (uatomic_cmpxchg(&iter_prev->p.next,
902 iter, new_node) != iter) {
903 continue; /* retry */
904 } else {
905 return_node = iter_prev;
906 goto gc_end;
907 }
908
909 gc_node:
910 assert(!is_removed(iter));
911 assert(!is_gc(iter));
912 if (is_dummy(iter))
913 new_next = flag_dummy(clear_flag(next));
914 else
915 new_next = clear_flag(next);
916 (void) uatomic_cmpxchg(&iter_prev->p.next, iter, new_next);
917 /* retry */
918 }
919 gc_end:
920 /* Garbage collect logically removed nodes in the bucket */
921 index = hash & (size - 1);
922 order = get_count_order_ulong(index + 1);
923 lookup = &ht->t.tbl[order]->nodes[index & (!order ? 0 : ((1UL << (order - 1)) - 1))];
924 dummy_node = (struct cds_lfht_node *) lookup;
925 _cds_lfht_gc_bucket(dummy_node, node);
926 return return_node;
927 }
928
929 static
930 int _cds_lfht_del(struct cds_lfht *ht, unsigned long size,
931 struct cds_lfht_node *node,
932 int dummy_removal, int do_gc)
933 {
934 struct cds_lfht_node *dummy, *next, *old;
935 struct _cds_lfht_node *lookup;
936 int flagged = 0;
937 unsigned long hash, index, order;
938
939 /* logically delete the node */
940 assert(!is_dummy(node));
941 assert(!is_gc(node));
942 assert(!is_removed(node));
943 old = rcu_dereference(node->p.next);
944 do {
945 struct cds_lfht_node *new_next;
946
947 next = old;
948 if (unlikely(is_removed(next)))
949 goto end;
950 if (dummy_removal)
951 assert(is_dummy(next));
952 else
953 assert(!is_dummy(next));
954 new_next = flag_removed(next);
955 if (do_gc)
956 new_next = flag_gc(new_next);
957 old = uatomic_cmpxchg(&node->p.next, next, new_next);
958 } while (old != next);
959
960 /* We performed the (logical) deletion. */
961 flagged = 1;
962
963 if (!do_gc)
964 goto end;
965
966 /*
967 * Ensure that the node is not visible to readers anymore: lookup for
968 * the node, and remove it (along with any other logically removed node)
969 * if found.
970 */
971 hash = bit_reverse_ulong(node->p.reverse_hash);
972 assert(size > 0);
973 index = hash & (size - 1);
974 order = get_count_order_ulong(index + 1);
975 lookup = &ht->t.tbl[order]->nodes[index & (!order ? 0 : ((1UL << (order - 1)) - 1))];
976 dummy = (struct cds_lfht_node *) lookup;
977 _cds_lfht_gc_bucket(dummy, node);
978 end:
979 /*
980 * Only the flagging action indicated that we (and no other)
981 * removed the node from the hash.
982 */
983 if (flagged) {
984 assert(is_removed(rcu_dereference(node->p.next)));
985 return 0;
986 } else
987 return -ENOENT;
988 }
989
990 static
991 void *partition_resize_thread(void *arg)
992 {
993 struct partition_resize_work *work = arg;
994
995 work->ht->cds_lfht_rcu_register_thread();
996 work->fct(work->ht, work->i, work->start, work->len);
997 work->ht->cds_lfht_rcu_unregister_thread();
998 return NULL;
999 }
1000
1001 static
1002 void partition_resize_helper(struct cds_lfht *ht, unsigned long i,
1003 unsigned long len,
1004 void (*fct)(struct cds_lfht *ht, unsigned long i,
1005 unsigned long start, unsigned long len))
1006 {
1007 unsigned long partition_len;
1008 struct partition_resize_work *work;
1009 int thread, ret;
1010 unsigned long nr_threads;
1011 pthread_t *thread_id;
1012
1013 /*
1014 * Note: nr_cpus_mask + 1 is always power of 2.
1015 * We spawn just the number of threads we need to satisfy the minimum
1016 * partition size, up to the number of CPUs in the system.
1017 */
1018 nr_threads = min(nr_cpus_mask + 1,
1019 len >> MIN_PARTITION_PER_THREAD_ORDER);
1020 partition_len = len >> get_count_order_ulong(nr_threads);
1021 work = calloc(nr_threads, sizeof(*work));
1022 thread_id = calloc(nr_threads, sizeof(*thread_id));
1023 assert(work);
1024 for (thread = 0; thread < nr_threads; thread++) {
1025 work[thread].ht = ht;
1026 work[thread].i = i;
1027 work[thread].len = partition_len;
1028 work[thread].start = thread * partition_len;
1029 work[thread].fct = fct;
1030 ret = pthread_create(&thread_id[thread], ht->resize_attr,
1031 partition_resize_thread, &work[thread]);
1032 assert(!ret);
1033 }
1034 for (thread = 0; thread < nr_threads; thread++) {
1035 ret = pthread_join(thread_id[thread], NULL);
1036 assert(!ret);
1037 }
1038 free(work);
1039 free(thread_id);
1040 }
1041
1042 /*
1043 * Holding RCU read lock to protect _cds_lfht_add against memory
1044 * reclaim that could be performed by other call_rcu worker threads (ABA
1045 * problem).
1046 *
1047 * When we reach a certain length, we can split this population phase over
1048 * many worker threads, based on the number of CPUs available in the system.
1049 * This should therefore take care of not having the expand lagging behind too
1050 * many concurrent insertion threads by using the scheduler's ability to
1051 * schedule dummy node population fairly with insertions.
1052 */
1053 static
1054 void init_table_populate_partition(struct cds_lfht *ht, unsigned long i,
1055 unsigned long start, unsigned long len)
1056 {
1057 unsigned long j;
1058
1059 ht->cds_lfht_rcu_read_lock();
1060 for (j = start; j < start + len; j++) {
1061 struct cds_lfht_node *new_node =
1062 (struct cds_lfht_node *) &ht->t.tbl[i]->nodes[j];
1063
1064 dbg_printf("init populate: i %lu j %lu hash %lu\n",
1065 i, j, !i ? 0 : (1UL << (i - 1)) + j);
1066 new_node->p.reverse_hash =
1067 bit_reverse_ulong(!i ? 0 : (1UL << (i - 1)) + j);
1068 (void) _cds_lfht_add(ht, !i ? 0 : (1UL << (i - 1)),
1069 new_node, ADD_DEFAULT, 1);
1070 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1071 break;
1072 }
1073 ht->cds_lfht_rcu_read_unlock();
1074 }
1075
1076 static
1077 void init_table_populate(struct cds_lfht *ht, unsigned long i,
1078 unsigned long len)
1079 {
1080 assert(nr_cpus_mask != -1);
1081 if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD) {
1082 ht->cds_lfht_rcu_thread_online();
1083 init_table_populate_partition(ht, i, 0, len);
1084 ht->cds_lfht_rcu_thread_offline();
1085 return;
1086 }
1087 partition_resize_helper(ht, i, len, init_table_populate_partition);
1088 }
1089
1090 static
1091 void init_table(struct cds_lfht *ht,
1092 unsigned long first_order, unsigned long len_order)
1093 {
1094 unsigned long i, end_order;
1095
1096 dbg_printf("init table: first_order %lu end_order %lu\n",
1097 first_order, first_order + len_order);
1098 end_order = first_order + len_order;
1099 for (i = first_order; i < end_order; i++) {
1100 unsigned long len;
1101
1102 len = !i ? 1 : 1UL << (i - 1);
1103 dbg_printf("init order %lu len: %lu\n", i, len);
1104
1105 /* Stop expand if the resize target changes under us */
1106 if (CMM_LOAD_SHARED(ht->t.resize_target) < (!i ? 1 : (1UL << i)))
1107 break;
1108
1109 ht->t.tbl[i] = calloc(1, sizeof(struct rcu_level)
1110 + (len * sizeof(struct _cds_lfht_node)));
1111 assert(ht->t.tbl[i]);
1112
1113 /*
1114 * Set all dummy nodes reverse hash values for a level and
1115 * link all dummy nodes into the table.
1116 */
1117 init_table_populate(ht, i, len);
1118
1119 /*
1120 * Update table size.
1121 */
1122 cmm_smp_wmb(); /* populate data before RCU size */
1123 CMM_STORE_SHARED(ht->t.size, !i ? 1 : (1UL << i));
1124
1125 dbg_printf("init new size: %lu\n", !i ? 1 : (1UL << i));
1126 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1127 break;
1128 }
1129 }
1130
1131 /*
1132 * Holding RCU read lock to protect _cds_lfht_remove against memory
1133 * reclaim that could be performed by other call_rcu worker threads (ABA
1134 * problem).
1135 * For a single level, we logically remove and garbage collect each node.
1136 *
1137 * As a design choice, we perform logical removal and garbage collection on a
1138 * node-per-node basis to simplify this algorithm. We also assume keeping good
1139 * cache locality of the operation would overweight possible performance gain
1140 * that could be achieved by batching garbage collection for multiple levels.
1141 * However, this would have to be justified by benchmarks.
1142 *
1143 * Concurrent removal and add operations are helping us perform garbage
1144 * collection of logically removed nodes. We guarantee that all logically
1145 * removed nodes have been garbage-collected (unlinked) before call_rcu is
1146 * invoked to free a hole level of dummy nodes (after a grace period).
1147 *
1148 * Logical removal and garbage collection can therefore be done in batch or on a
1149 * node-per-node basis, as long as the guarantee above holds.
1150 *
1151 * When we reach a certain length, we can split this removal over many worker
1152 * threads, based on the number of CPUs available in the system. This should
1153 * take care of not letting resize process lag behind too many concurrent
1154 * updater threads actively inserting into the hash table.
1155 */
1156 static
1157 void remove_table_partition(struct cds_lfht *ht, unsigned long i,
1158 unsigned long start, unsigned long len)
1159 {
1160 unsigned long j;
1161
1162 ht->cds_lfht_rcu_read_lock();
1163 for (j = start; j < start + len; j++) {
1164 struct cds_lfht_node *fini_node =
1165 (struct cds_lfht_node *) &ht->t.tbl[i]->nodes[j];
1166
1167 dbg_printf("remove entry: i %lu j %lu hash %lu\n",
1168 i, j, !i ? 0 : (1UL << (i - 1)) + j);
1169 fini_node->p.reverse_hash =
1170 bit_reverse_ulong(!i ? 0 : (1UL << (i - 1)) + j);
1171 (void) _cds_lfht_del(ht, !i ? 0 : (1UL << (i - 1)),
1172 fini_node, 1, 1);
1173 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1174 break;
1175 }
1176 ht->cds_lfht_rcu_read_unlock();
1177 }
1178
1179 static
1180 void remove_table(struct cds_lfht *ht, unsigned long i, unsigned long len)
1181 {
1182
1183 assert(nr_cpus_mask != -1);
1184 if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD) {
1185 ht->cds_lfht_rcu_thread_online();
1186 remove_table_partition(ht, i, 0, len);
1187 ht->cds_lfht_rcu_thread_offline();
1188 return;
1189 }
1190 partition_resize_helper(ht, i, len, remove_table_partition);
1191 }
1192
1193 static
1194 void fini_table(struct cds_lfht *ht,
1195 unsigned long first_order, unsigned long len_order)
1196 {
1197 long i, end_order;
1198
1199 dbg_printf("fini table: first_order %lu end_order %lu\n",
1200 first_order, first_order + len_order);
1201 end_order = first_order + len_order;
1202 assert(first_order > 0);
1203 for (i = end_order - 1; i >= first_order; i--) {
1204 unsigned long len;
1205
1206 len = !i ? 1 : 1UL << (i - 1);
1207 dbg_printf("fini order %lu len: %lu\n", i, len);
1208
1209 /* Stop shrink if the resize target changes under us */
1210 if (CMM_LOAD_SHARED(ht->t.resize_target) > (1UL << (i - 1)))
1211 break;
1212
1213 cmm_smp_wmb(); /* populate data before RCU size */
1214 CMM_STORE_SHARED(ht->t.size, 1UL << (i - 1));
1215
1216 /*
1217 * We need to wait for all add operations to reach Q.S. (and
1218 * thus use the new table for lookups) before we can start
1219 * releasing the old dummy nodes. Otherwise their lookup will
1220 * return a logically removed node as insert position.
1221 */
1222 ht->cds_lfht_synchronize_rcu();
1223
1224 /*
1225 * Set "removed" flag in dummy nodes about to be removed.
1226 * Unlink all now-logically-removed dummy node pointers.
1227 * Concurrent add/remove operation are helping us doing
1228 * the gc.
1229 */
1230 remove_table(ht, i, len);
1231
1232 ht->cds_lfht_call_rcu(&ht->t.tbl[i]->head, cds_lfht_free_level);
1233
1234 dbg_printf("fini new size: %lu\n", 1UL << i);
1235 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1236 break;
1237 }
1238 }
1239
1240 struct cds_lfht *_cds_lfht_new(cds_lfht_hash_fct hash_fct,
1241 cds_lfht_compare_fct compare_fct,
1242 unsigned long hash_seed,
1243 unsigned long init_size,
1244 int flags,
1245 void (*cds_lfht_call_rcu)(struct rcu_head *head,
1246 void (*func)(struct rcu_head *head)),
1247 void (*cds_lfht_synchronize_rcu)(void),
1248 void (*cds_lfht_rcu_read_lock)(void),
1249 void (*cds_lfht_rcu_read_unlock)(void),
1250 void (*cds_lfht_rcu_thread_offline)(void),
1251 void (*cds_lfht_rcu_thread_online)(void),
1252 void (*cds_lfht_rcu_register_thread)(void),
1253 void (*cds_lfht_rcu_unregister_thread)(void),
1254 pthread_attr_t *attr)
1255 {
1256 struct cds_lfht *ht;
1257 unsigned long order;
1258
1259 /* init_size must be power of two */
1260 if (init_size && (init_size & (init_size - 1)))
1261 return NULL;
1262 ht = calloc(1, sizeof(struct cds_lfht));
1263 assert(ht);
1264 ht->hash_fct = hash_fct;
1265 ht->compare_fct = compare_fct;
1266 ht->hash_seed = hash_seed;
1267 ht->cds_lfht_call_rcu = cds_lfht_call_rcu;
1268 ht->cds_lfht_synchronize_rcu = cds_lfht_synchronize_rcu;
1269 ht->cds_lfht_rcu_read_lock = cds_lfht_rcu_read_lock;
1270 ht->cds_lfht_rcu_read_unlock = cds_lfht_rcu_read_unlock;
1271 ht->cds_lfht_rcu_thread_offline = cds_lfht_rcu_thread_offline;
1272 ht->cds_lfht_rcu_thread_online = cds_lfht_rcu_thread_online;
1273 ht->cds_lfht_rcu_register_thread = cds_lfht_rcu_register_thread;
1274 ht->cds_lfht_rcu_unregister_thread = cds_lfht_rcu_unregister_thread;
1275 ht->resize_attr = attr;
1276 ht->percpu_count = alloc_per_cpu_items_count();
1277 /* this mutex should not nest in read-side C.S. */
1278 pthread_mutex_init(&ht->resize_mutex, NULL);
1279 order = get_count_order_ulong(max(init_size, MIN_TABLE_SIZE)) + 1;
1280 ht->flags = flags;
1281 ht->cds_lfht_rcu_thread_offline();
1282 pthread_mutex_lock(&ht->resize_mutex);
1283 ht->t.resize_target = 1UL << (order - 1);
1284 init_table(ht, 0, order);
1285 pthread_mutex_unlock(&ht->resize_mutex);
1286 ht->cds_lfht_rcu_thread_online();
1287 return ht;
1288 }
1289
1290 void cds_lfht_lookup(struct cds_lfht *ht, void *key, size_t key_len,
1291 struct cds_lfht_iter *iter)
1292 {
1293 struct cds_lfht_node *node, *next, *dummy_node;
1294 struct _cds_lfht_node *lookup;
1295 unsigned long hash, reverse_hash, index, order, size;
1296
1297 hash = ht->hash_fct(key, key_len, ht->hash_seed);
1298 reverse_hash = bit_reverse_ulong(hash);
1299
1300 size = rcu_dereference(ht->t.size);
1301 index = hash & (size - 1);
1302 order = get_count_order_ulong(index + 1);
1303 lookup = &ht->t.tbl[order]->nodes[index & (!order ? 0 : ((1UL << (order - 1))) - 1)];
1304 dbg_printf("lookup hash %lu index %lu order %lu aridx %lu\n",
1305 hash, index, order, index & (!order ? 0 : ((1UL << (order - 1)) - 1)));
1306 dummy_node = (struct cds_lfht_node *) lookup;
1307 /* We can always skip the dummy node initially */
1308 node = rcu_dereference(dummy_node->p.next);
1309 node = clear_flag(node);
1310 for (;;) {
1311 if (unlikely(is_end(node))) {
1312 node = NULL;
1313 break;
1314 }
1315 if (unlikely(node->p.reverse_hash > reverse_hash)) {
1316 node = NULL;
1317 break;
1318 }
1319 next = rcu_dereference(node->p.next);
1320 if (likely(!is_removed(next))
1321 && !is_dummy(next)
1322 && likely(!ht->compare_fct(node->key, node->key_len, key, key_len))) {
1323 break;
1324 }
1325 node = clear_flag(next);
1326 }
1327 assert(!node || !is_dummy(rcu_dereference(node->p.next)));
1328 iter->node = node;
1329 iter->next = next;
1330 }
1331
1332 void cds_lfht_next(struct cds_lfht *ht, struct cds_lfht_iter *iter)
1333 {
1334 struct cds_lfht_node *node, *next;
1335 unsigned long reverse_hash;
1336 void *key;
1337 size_t key_len;
1338
1339 node = iter->node;
1340 reverse_hash = node->p.reverse_hash;
1341 key = node->key;
1342 key_len = node->key_len;
1343 next = iter->next;
1344 node = clear_flag(next);
1345
1346 for (;;) {
1347 if (unlikely(is_end(node))) {
1348 node = NULL;
1349 break;
1350 }
1351 if (unlikely(node->p.reverse_hash > reverse_hash)) {
1352 node = NULL;
1353 break;
1354 }
1355 next = rcu_dereference(node->p.next);
1356 if (likely(!is_removed(next))
1357 && !is_dummy(next)
1358 && likely(!ht->compare_fct(node->key, node->key_len, key, key_len))) {
1359 break;
1360 }
1361 node = clear_flag(next);
1362 }
1363 assert(!node || !is_dummy(rcu_dereference(node->p.next)));
1364 iter->node = node;
1365 iter->next = next;
1366 }
1367
1368 void cds_lfht_add(struct cds_lfht *ht, struct cds_lfht_node *node)
1369 {
1370 unsigned long hash, size;
1371
1372 hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed);
1373 node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash);
1374
1375 size = rcu_dereference(ht->t.size);
1376 (void) _cds_lfht_add(ht, size, node, ADD_DEFAULT, 0);
1377 ht_count_add(ht, size);
1378 }
1379
1380 struct cds_lfht_node *cds_lfht_add_unique(struct cds_lfht *ht,
1381 struct cds_lfht_node *node)
1382 {
1383 unsigned long hash, size;
1384 struct cds_lfht_node *ret;
1385
1386 hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed);
1387 node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash);
1388
1389 size = rcu_dereference(ht->t.size);
1390 ret = _cds_lfht_add(ht, size, node, ADD_UNIQUE, 0);
1391 if (ret == node)
1392 ht_count_add(ht, size);
1393 return ret;
1394 }
1395
1396 struct cds_lfht_node *cds_lfht_replace(struct cds_lfht *ht,
1397 struct cds_lfht_node *node)
1398 {
1399 unsigned long hash, size;
1400 struct cds_lfht_node *ret;
1401
1402 hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed);
1403 node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash);
1404
1405 size = rcu_dereference(ht->t.size);
1406 ret = _cds_lfht_add(ht, size, node, ADD_REPLACE, 0);
1407 if (ret == NULL)
1408 ht_count_add(ht, size);
1409 return ret;
1410 }
1411
1412 int cds_lfht_del(struct cds_lfht *ht, struct cds_lfht_node *node)
1413 {
1414 unsigned long size;
1415 int ret;
1416
1417 size = rcu_dereference(ht->t.size);
1418 ret = _cds_lfht_del(ht, size, node, 0, 1);
1419 if (!ret)
1420 ht_count_del(ht, size);
1421 return ret;
1422 }
1423
1424 static
1425 int cds_lfht_delete_dummy(struct cds_lfht *ht)
1426 {
1427 struct cds_lfht_node *node;
1428 struct _cds_lfht_node *lookup;
1429 unsigned long order, i, size;
1430
1431 /* Check that the table is empty */
1432 lookup = &ht->t.tbl[0]->nodes[0];
1433 node = (struct cds_lfht_node *) lookup;
1434 do {
1435 node = clear_flag(node)->p.next;
1436 if (!is_dummy(node))
1437 return -EPERM;
1438 assert(!is_removed(node));
1439 assert(!is_gc(node));
1440 } while (!is_end(node));
1441 /*
1442 * size accessed without rcu_dereference because hash table is
1443 * being destroyed.
1444 */
1445 size = ht->t.size;
1446 /* Internal sanity check: all nodes left should be dummy */
1447 for (order = 0; order < get_count_order_ulong(size) + 1; order++) {
1448 unsigned long len;
1449
1450 len = !order ? 1 : 1UL << (order - 1);
1451 for (i = 0; i < len; i++) {
1452 dbg_printf("delete order %lu i %lu hash %lu\n",
1453 order, i,
1454 bit_reverse_ulong(ht->t.tbl[order]->nodes[i].reverse_hash));
1455 assert(is_dummy(ht->t.tbl[order]->nodes[i].next));
1456 }
1457 poison_free(ht->t.tbl[order]);
1458 }
1459 return 0;
1460 }
1461
1462 /*
1463 * Should only be called when no more concurrent readers nor writers can
1464 * possibly access the table.
1465 */
1466 int cds_lfht_destroy(struct cds_lfht *ht, pthread_attr_t **attr)
1467 {
1468 int ret;
1469
1470 /* Wait for in-flight resize operations to complete */
1471 CMM_STORE_SHARED(ht->in_progress_destroy, 1);
1472 while (uatomic_read(&ht->in_progress_resize))
1473 poll(NULL, 0, 100); /* wait for 100ms */
1474 ret = cds_lfht_delete_dummy(ht);
1475 if (ret)
1476 return ret;
1477 free_per_cpu_items_count(ht->percpu_count);
1478 if (attr)
1479 *attr = ht->resize_attr;
1480 poison_free(ht);
1481 return ret;
1482 }
1483
1484 void cds_lfht_count_nodes(struct cds_lfht *ht,
1485 unsigned long *count,
1486 unsigned long *removed)
1487 {
1488 struct cds_lfht_node *node, *next;
1489 struct _cds_lfht_node *lookup;
1490 unsigned long nr_dummy = 0;
1491
1492 *count = 0;
1493 *removed = 0;
1494
1495 /* Count non-dummy nodes in the table */
1496 lookup = &ht->t.tbl[0]->nodes[0];
1497 node = (struct cds_lfht_node *) lookup;
1498 do {
1499 next = rcu_dereference(node->p.next);
1500 if (is_removed(next) || is_gc(next)) {
1501 assert(!is_dummy(next));
1502 (*removed)++;
1503 } else if (!is_dummy(next))
1504 (*count)++;
1505 else
1506 (nr_dummy)++;
1507 node = clear_flag(next);
1508 } while (!is_end(node));
1509 dbg_printf("number of dummy nodes: %lu\n", nr_dummy);
1510 }
1511
1512 /* called with resize mutex held */
1513 static
1514 void _do_cds_lfht_grow(struct cds_lfht *ht,
1515 unsigned long old_size, unsigned long new_size)
1516 {
1517 unsigned long old_order, new_order;
1518
1519 old_order = get_count_order_ulong(old_size) + 1;
1520 new_order = get_count_order_ulong(new_size) + 1;
1521 printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1522 old_size, old_order, new_size, new_order);
1523 assert(new_size > old_size);
1524 init_table(ht, old_order, new_order - old_order);
1525 }
1526
1527 /* called with resize mutex held */
1528 static
1529 void _do_cds_lfht_shrink(struct cds_lfht *ht,
1530 unsigned long old_size, unsigned long new_size)
1531 {
1532 unsigned long old_order, new_order;
1533
1534 new_size = max(new_size, MIN_TABLE_SIZE);
1535 old_order = get_count_order_ulong(old_size) + 1;
1536 new_order = get_count_order_ulong(new_size) + 1;
1537 printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1538 old_size, old_order, new_size, new_order);
1539 assert(new_size < old_size);
1540
1541 /* Remove and unlink all dummy nodes to remove. */
1542 fini_table(ht, new_order, old_order - new_order);
1543 }
1544
1545
1546 /* called with resize mutex held */
1547 static
1548 void _do_cds_lfht_resize(struct cds_lfht *ht)
1549 {
1550 unsigned long new_size, old_size;
1551
1552 /*
1553 * Resize table, re-do if the target size has changed under us.
1554 */
1555 do {
1556 ht->t.resize_initiated = 1;
1557 old_size = ht->t.size;
1558 new_size = CMM_LOAD_SHARED(ht->t.resize_target);
1559 if (old_size < new_size)
1560 _do_cds_lfht_grow(ht, old_size, new_size);
1561 else if (old_size > new_size)
1562 _do_cds_lfht_shrink(ht, old_size, new_size);
1563 ht->t.resize_initiated = 0;
1564 /* write resize_initiated before read resize_target */
1565 cmm_smp_mb();
1566 } while (ht->t.size != CMM_LOAD_SHARED(ht->t.resize_target));
1567 }
1568
1569 static
1570 unsigned long resize_target_update(struct cds_lfht *ht, unsigned long size,
1571 int growth_order)
1572 {
1573 return _uatomic_max(&ht->t.resize_target,
1574 size << growth_order);
1575 }
1576
1577 static
1578 void resize_target_update_count(struct cds_lfht *ht,
1579 unsigned long count)
1580 {
1581 count = max(count, MIN_TABLE_SIZE);
1582 uatomic_set(&ht->t.resize_target, count);
1583 }
1584
1585 void cds_lfht_resize(struct cds_lfht *ht, unsigned long new_size)
1586 {
1587 resize_target_update_count(ht, new_size);
1588 CMM_STORE_SHARED(ht->t.resize_initiated, 1);
1589 ht->cds_lfht_rcu_thread_offline();
1590 pthread_mutex_lock(&ht->resize_mutex);
1591 _do_cds_lfht_resize(ht);
1592 pthread_mutex_unlock(&ht->resize_mutex);
1593 ht->cds_lfht_rcu_thread_online();
1594 }
1595
1596 static
1597 void do_resize_cb(struct rcu_head *head)
1598 {
1599 struct rcu_resize_work *work =
1600 caa_container_of(head, struct rcu_resize_work, head);
1601 struct cds_lfht *ht = work->ht;
1602
1603 ht->cds_lfht_rcu_thread_offline();
1604 pthread_mutex_lock(&ht->resize_mutex);
1605 _do_cds_lfht_resize(ht);
1606 pthread_mutex_unlock(&ht->resize_mutex);
1607 ht->cds_lfht_rcu_thread_online();
1608 poison_free(work);
1609 cmm_smp_mb(); /* finish resize before decrement */
1610 uatomic_dec(&ht->in_progress_resize);
1611 }
1612
1613 static
1614 void cds_lfht_resize_lazy(struct cds_lfht *ht, unsigned long size, int growth)
1615 {
1616 struct rcu_resize_work *work;
1617 unsigned long target_size;
1618
1619 target_size = resize_target_update(ht, size, growth);
1620 /* Store resize_target before read resize_initiated */
1621 cmm_smp_mb();
1622 if (!CMM_LOAD_SHARED(ht->t.resize_initiated) && size < target_size) {
1623 uatomic_inc(&ht->in_progress_resize);
1624 cmm_smp_mb(); /* increment resize count before calling it */
1625 work = malloc(sizeof(*work));
1626 work->ht = ht;
1627 ht->cds_lfht_call_rcu(&work->head, do_resize_cb);
1628 CMM_STORE_SHARED(ht->t.resize_initiated, 1);
1629 }
1630 }
1631
1632 #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
1633
1634 static
1635 void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
1636 unsigned long count)
1637 {
1638 struct rcu_resize_work *work;
1639
1640 if (!(ht->flags & CDS_LFHT_AUTO_RESIZE))
1641 return;
1642 resize_target_update_count(ht, count);
1643 /* Store resize_target before read resize_initiated */
1644 cmm_smp_mb();
1645 if (!CMM_LOAD_SHARED(ht->t.resize_initiated)) {
1646 uatomic_inc(&ht->in_progress_resize);
1647 cmm_smp_mb(); /* increment resize count before calling it */
1648 work = malloc(sizeof(*work));
1649 work->ht = ht;
1650 ht->cds_lfht_call_rcu(&work->head, do_resize_cb);
1651 CMM_STORE_SHARED(ht->t.resize_initiated, 1);
1652 }
1653 }
1654
1655 #endif
This page took 0.101855 seconds and 3 git commands to generate.