2bb26e55034d69f7e7a694802a31856a5d7cfd87
[lttng-tools.git] / hashtable / 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 <urcu/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/compiler.h>
148 #include "rculfhash.h"
149 #include <stdio.h>
150 #include <pthread.h>
151
152 #ifdef DEBUG
153 #define dbg_printf(fmt, args...) printf("[debug rculfhash] " fmt, ## args)
154 #else
155 #define dbg_printf(fmt, args...)
156 #endif
157
158 /*
159 * Per-CPU split-counters lazily update the global counter each 1024
160 * addition/removal. It automatically keeps track of resize required.
161 * We use the bucket length as indicator for need to expand for small
162 * tables and machines lacking per-cpu data suppport.
163 */
164 #define COUNT_COMMIT_ORDER 10
165 #define CHAIN_LEN_TARGET 1
166 #define CHAIN_LEN_RESIZE_THRESHOLD 3
167
168 /*
169 * Define the minimum table size.
170 */
171 #define MIN_TABLE_SIZE 1
172
173 #if (CAA_BITS_PER_LONG == 32)
174 #define MAX_TABLE_ORDER 32
175 #else
176 #define MAX_TABLE_ORDER 64
177 #endif
178
179 /*
180 * Minimum number of dummy nodes to touch per thread to parallelize grow/shrink.
181 */
182 #define MIN_PARTITION_PER_THREAD_ORDER 12
183 #define MIN_PARTITION_PER_THREAD (1UL << MIN_PARTITION_PER_THREAD_ORDER)
184
185 #ifndef min
186 #define min(a, b) ((a) < (b) ? (a) : (b))
187 #endif
188
189 #ifndef max
190 #define max(a, b) ((a) > (b) ? (a) : (b))
191 #endif
192
193 /*
194 * The removed flag needs to be updated atomically with the pointer.
195 * It indicates that no node must attach to the node scheduled for
196 * removal, and that node garbage collection must be performed.
197 * The dummy flag does not require to be updated atomically with the
198 * pointer, but it is added as a pointer low bit flag to save space.
199 */
200 #define REMOVED_FLAG (1UL << 0)
201 #define DUMMY_FLAG (1UL << 1)
202 #define FLAGS_MASK ((1UL << 2) - 1)
203
204 /* Value of the end pointer. Should not interact with flags. */
205 #define END_VALUE NULL
206
207 struct ht_items_count {
208 unsigned long add, del;
209 } __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
210
211 struct rcu_level {
212 struct rcu_head head;
213 struct _cds_lfht_node nodes[0];
214 };
215
216 struct rcu_table {
217 unsigned long size; /* always a power of 2, shared (RCU) */
218 unsigned long resize_target;
219 int resize_initiated;
220 struct rcu_level *tbl[MAX_TABLE_ORDER];
221 };
222
223 struct cds_lfht {
224 struct rcu_table t;
225 cds_lfht_hash_fct hash_fct;
226 cds_lfht_compare_fct compare_fct;
227 unsigned long hash_seed;
228 int flags;
229 /*
230 * We need to put the work threads offline (QSBR) when taking this
231 * mutex, because we use synchronize_rcu within this mutex critical
232 * section, which waits on read-side critical sections, and could
233 * therefore cause grace-period deadlock if we hold off RCU G.P.
234 * completion.
235 */
236 pthread_mutex_t resize_mutex; /* resize mutex: add/del mutex */
237 unsigned int in_progress_resize, in_progress_destroy;
238 void (*cds_lfht_call_rcu)(struct rcu_head *head,
239 void (*func)(struct rcu_head *head));
240 void (*cds_lfht_synchronize_rcu)(void);
241 void (*cds_lfht_rcu_read_lock)(void);
242 void (*cds_lfht_rcu_read_unlock)(void);
243 void (*cds_lfht_rcu_thread_offline)(void);
244 void (*cds_lfht_rcu_thread_online)(void);
245 void (*cds_lfht_rcu_register_thread)(void);
246 void (*cds_lfht_rcu_unregister_thread)(void);
247 pthread_attr_t *resize_attr; /* Resize threads attributes */
248 long count; /* global approximate item count */
249 struct ht_items_count *percpu_count; /* per-cpu item count */
250 };
251
252 struct rcu_resize_work {
253 struct rcu_head head;
254 struct cds_lfht *ht;
255 };
256
257 struct partition_resize_work {
258 struct rcu_head head;
259 struct cds_lfht *ht;
260 unsigned long i, start, len;
261 void (*fct)(struct cds_lfht *ht, unsigned long i,
262 unsigned long start, unsigned long len);
263 };
264
265 enum add_mode {
266 ADD_DEFAULT = 0,
267 ADD_UNIQUE = 1,
268 ADD_REPLACE = 2,
269 };
270
271 static
272 struct cds_lfht_node *_cds_lfht_add(struct cds_lfht *ht,
273 unsigned long size,
274 struct cds_lfht_node *node,
275 enum add_mode mode, int dummy);
276
277 /*
278 * Algorithm to reverse bits in a word by lookup table, extended to
279 * 64-bit words.
280 * Source:
281 * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable
282 * Originally from Public Domain.
283 */
284
285 static const uint8_t BitReverseTable256[256] =
286 {
287 #define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64
288 #define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16)
289 #define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 )
290 R6(0), R6(2), R6(1), R6(3)
291 };
292 #undef R2
293 #undef R4
294 #undef R6
295
296 static
297 uint8_t bit_reverse_u8(uint8_t v)
298 {
299 return BitReverseTable256[v];
300 }
301
302 static __attribute__((unused))
303 uint32_t bit_reverse_u32(uint32_t v)
304 {
305 return ((uint32_t) bit_reverse_u8(v) << 24) |
306 ((uint32_t) bit_reverse_u8(v >> 8) << 16) |
307 ((uint32_t) bit_reverse_u8(v >> 16) << 8) |
308 ((uint32_t) bit_reverse_u8(v >> 24));
309 }
310
311 static __attribute__((unused))
312 uint64_t bit_reverse_u64(uint64_t v)
313 {
314 return ((uint64_t) bit_reverse_u8(v) << 56) |
315 ((uint64_t) bit_reverse_u8(v >> 8) << 48) |
316 ((uint64_t) bit_reverse_u8(v >> 16) << 40) |
317 ((uint64_t) bit_reverse_u8(v >> 24) << 32) |
318 ((uint64_t) bit_reverse_u8(v >> 32) << 24) |
319 ((uint64_t) bit_reverse_u8(v >> 40) << 16) |
320 ((uint64_t) bit_reverse_u8(v >> 48) << 8) |
321 ((uint64_t) bit_reverse_u8(v >> 56));
322 }
323
324 static
325 unsigned long bit_reverse_ulong(unsigned long v)
326 {
327 #if (CAA_BITS_PER_LONG == 32)
328 return bit_reverse_u32(v);
329 #else
330 return bit_reverse_u64(v);
331 #endif
332 }
333
334 /*
335 * fls: returns the position of the most significant bit.
336 * Returns 0 if no bit is set, else returns the position of the most
337 * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit).
338 */
339 #if defined(__i386) || defined(__x86_64)
340 static inline
341 unsigned int fls_u32(uint32_t x)
342 {
343 int r;
344
345 asm("bsrl %1,%0\n\t"
346 "jnz 1f\n\t"
347 "movl $-1,%0\n\t"
348 "1:\n\t"
349 : "=r" (r) : "rm" (x));
350 return r + 1;
351 }
352 #define HAS_FLS_U32
353 #endif
354
355 #if defined(__x86_64)
356 static inline
357 unsigned int fls_u64(uint64_t x)
358 {
359 long r;
360
361 asm("bsrq %1,%0\n\t"
362 "jnz 1f\n\t"
363 "movq $-1,%0\n\t"
364 "1:\n\t"
365 : "=r" (r) : "rm" (x));
366 return r + 1;
367 }
368 #define HAS_FLS_U64
369 #endif
370
371 #ifndef HAS_FLS_U64
372 static __attribute__((unused))
373 unsigned int fls_u64(uint64_t x)
374 {
375 unsigned int r = 64;
376
377 if (!x)
378 return 0;
379
380 if (!(x & 0xFFFFFFFF00000000ULL)) {
381 x <<= 32;
382 r -= 32;
383 }
384 if (!(x & 0xFFFF000000000000ULL)) {
385 x <<= 16;
386 r -= 16;
387 }
388 if (!(x & 0xFF00000000000000ULL)) {
389 x <<= 8;
390 r -= 8;
391 }
392 if (!(x & 0xF000000000000000ULL)) {
393 x <<= 4;
394 r -= 4;
395 }
396 if (!(x & 0xC000000000000000ULL)) {
397 x <<= 2;
398 r -= 2;
399 }
400 if (!(x & 0x8000000000000000ULL)) {
401 x <<= 1;
402 r -= 1;
403 }
404 return r;
405 }
406 #endif
407
408 #ifndef HAS_FLS_U32
409 static __attribute__((unused))
410 unsigned int fls_u32(uint32_t x)
411 {
412 unsigned int r = 32;
413
414 if (!x)
415 return 0;
416 if (!(x & 0xFFFF0000U)) {
417 x <<= 16;
418 r -= 16;
419 }
420 if (!(x & 0xFF000000U)) {
421 x <<= 8;
422 r -= 8;
423 }
424 if (!(x & 0xF0000000U)) {
425 x <<= 4;
426 r -= 4;
427 }
428 if (!(x & 0xC0000000U)) {
429 x <<= 2;
430 r -= 2;
431 }
432 if (!(x & 0x80000000U)) {
433 x <<= 1;
434 r -= 1;
435 }
436 return r;
437 }
438 #endif
439
440 unsigned int fls_ulong(unsigned long x)
441 {
442 #if (CAA_BITS_PER_lONG == 32)
443 return fls_u32(x);
444 #else
445 return fls_u64(x);
446 #endif
447 }
448
449 int get_count_order_u32(uint32_t x)
450 {
451 int order;
452
453 order = fls_u32(x) - 1;
454 if (x & (x - 1))
455 order++;
456 return order;
457 }
458
459 int get_count_order_ulong(unsigned long x)
460 {
461 int order;
462
463 order = fls_ulong(x) - 1;
464 if (x & (x - 1))
465 order++;
466 return order;
467 }
468
469 #ifdef POISON_FREE
470 #define poison_free(ptr) \
471 do { \
472 memset(ptr, 0x42, sizeof(*(ptr))); \
473 free(ptr); \
474 } while (0)
475 #else
476 #define poison_free(ptr) free(ptr)
477 #endif
478
479 static
480 void cds_lfht_resize_lazy(struct cds_lfht *ht, unsigned long size, int growth);
481
482 /*
483 * If the sched_getcpu() and sysconf(_SC_NPROCESSORS_CONF) calls are
484 * available, then we support hash table item accounting.
485 * In the unfortunate event the number of CPUs reported would be
486 * inaccurate, we use modulo arithmetic on the number of CPUs we got.
487 */
488 #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
489
490 static
491 void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
492 unsigned long count);
493
494 static long nr_cpus_mask = -1;
495
496 static
497 struct ht_items_count *alloc_per_cpu_items_count(void)
498 {
499 struct ht_items_count *count;
500
501 switch (nr_cpus_mask) {
502 case -2:
503 return NULL;
504 case -1:
505 {
506 long maxcpus;
507
508 maxcpus = sysconf(_SC_NPROCESSORS_CONF);
509 if (maxcpus <= 0) {
510 nr_cpus_mask = -2;
511 return NULL;
512 }
513 /*
514 * round up number of CPUs to next power of two, so we
515 * can use & for modulo.
516 */
517 maxcpus = 1UL << get_count_order_ulong(maxcpus);
518 nr_cpus_mask = maxcpus - 1;
519 }
520 /* Fall-through */
521 default:
522 return calloc(nr_cpus_mask + 1, sizeof(*count));
523 }
524 }
525
526 static
527 void free_per_cpu_items_count(struct ht_items_count *count)
528 {
529 poison_free(count);
530 }
531
532 static
533 int ht_get_cpu(void)
534 {
535 int cpu;
536
537 assert(nr_cpus_mask >= 0);
538 cpu = sched_getcpu();
539 if (unlikely(cpu < 0))
540 return cpu;
541 else
542 return cpu & nr_cpus_mask;
543 }
544
545 static
546 void ht_count_add(struct cds_lfht *ht, unsigned long size)
547 {
548 unsigned long percpu_count;
549 int cpu;
550
551 if (unlikely(!ht->percpu_count))
552 return;
553 cpu = ht_get_cpu();
554 if (unlikely(cpu < 0))
555 return;
556 percpu_count = uatomic_add_return(&ht->percpu_count[cpu].add, 1);
557 if (unlikely(!(percpu_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) {
558 long count;
559
560 dbg_printf("add percpu %lu\n", percpu_count);
561 count = uatomic_add_return(&ht->count,
562 1UL << COUNT_COMMIT_ORDER);
563 /* If power of 2 */
564 if (!(count & (count - 1))) {
565 if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) < size)
566 return;
567 dbg_printf("add set global %ld\n", count);
568 cds_lfht_resize_lazy_count(ht, size,
569 count >> (CHAIN_LEN_TARGET - 1));
570 }
571 }
572 }
573
574 static
575 void ht_count_del(struct cds_lfht *ht, unsigned long size)
576 {
577 unsigned long percpu_count;
578 int cpu;
579
580 if (unlikely(!ht->percpu_count))
581 return;
582 cpu = ht_get_cpu();
583 if (unlikely(cpu < 0))
584 return;
585 percpu_count = uatomic_add_return(&ht->percpu_count[cpu].del, 1);
586 if (unlikely(!(percpu_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) {
587 long count;
588
589 dbg_printf("del percpu %lu\n", percpu_count);
590 count = uatomic_add_return(&ht->count,
591 -(1UL << COUNT_COMMIT_ORDER));
592 /* If power of 2 */
593 if (!(count & (count - 1))) {
594 if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) >= size)
595 return;
596 dbg_printf("del set global %ld\n", count);
597 /*
598 * Don't shrink table if the number of nodes is below a
599 * certain threshold.
600 */
601 if (count < (1UL << COUNT_COMMIT_ORDER) * (nr_cpus_mask + 1))
602 return;
603 cds_lfht_resize_lazy_count(ht, size,
604 count >> (CHAIN_LEN_TARGET - 1));
605 }
606 }
607 }
608
609 #else /* #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
610
611 static const long nr_cpus_mask = -1;
612
613 static
614 struct ht_items_count *alloc_per_cpu_items_count(void)
615 {
616 return NULL;
617 }
618
619 static
620 void free_per_cpu_items_count(struct ht_items_count *count)
621 {
622 }
623
624 static
625 void ht_count_add(struct cds_lfht *ht, unsigned long size)
626 {
627 }
628
629 static
630 void ht_count_del(struct cds_lfht *ht, unsigned long size)
631 {
632 }
633
634 #endif /* #else #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
635
636
637 static
638 void check_resize(struct cds_lfht *ht, unsigned long size, uint32_t chain_len)
639 {
640 unsigned long count;
641
642 if (!(ht->flags & CDS_LFHT_AUTO_RESIZE))
643 return;
644 count = uatomic_read(&ht->count);
645 /*
646 * Use bucket-local length for small table expand and for
647 * environments lacking per-cpu data support.
648 */
649 if (count >= (1UL << COUNT_COMMIT_ORDER))
650 return;
651 if (chain_len > 100)
652 dbg_printf("WARNING: large chain length: %u.\n",
653 chain_len);
654 if (chain_len >= CHAIN_LEN_RESIZE_THRESHOLD)
655 cds_lfht_resize_lazy(ht, size,
656 get_count_order_u32(chain_len - (CHAIN_LEN_TARGET - 1)));
657 }
658
659 static
660 struct cds_lfht_node *clear_flag(struct cds_lfht_node *node)
661 {
662 return (struct cds_lfht_node *) (((unsigned long) node) & ~FLAGS_MASK);
663 }
664
665 static
666 int is_removed(struct cds_lfht_node *node)
667 {
668 return ((unsigned long) node) & REMOVED_FLAG;
669 }
670
671 static
672 struct cds_lfht_node *flag_removed(struct cds_lfht_node *node)
673 {
674 return (struct cds_lfht_node *) (((unsigned long) node) | REMOVED_FLAG);
675 }
676
677 static
678 int is_dummy(struct cds_lfht_node *node)
679 {
680 return ((unsigned long) node) & DUMMY_FLAG;
681 }
682
683 static
684 struct cds_lfht_node *flag_dummy(struct cds_lfht_node *node)
685 {
686 return (struct cds_lfht_node *) (((unsigned long) node) | DUMMY_FLAG);
687 }
688
689 static
690 struct cds_lfht_node *get_end(void)
691 {
692 return (struct cds_lfht_node *) END_VALUE;
693 }
694
695 static
696 int is_end(struct cds_lfht_node *node)
697 {
698 return clear_flag(node) == (struct cds_lfht_node *) END_VALUE;
699 }
700
701 static
702 unsigned long _uatomic_max(unsigned long *ptr, unsigned long v)
703 {
704 unsigned long old1, old2;
705
706 old1 = uatomic_read(ptr);
707 do {
708 old2 = old1;
709 if (old2 >= v)
710 return old2;
711 } while ((old1 = uatomic_cmpxchg(ptr, old2, v)) != old2);
712 return v;
713 }
714
715 static
716 void cds_lfht_free_level(struct rcu_head *head)
717 {
718 struct rcu_level *l =
719 caa_container_of(head, struct rcu_level, head);
720 poison_free(l);
721 }
722
723 /*
724 * Remove all logically deleted nodes from a bucket up to a certain node key.
725 */
726 static
727 void _cds_lfht_gc_bucket(struct cds_lfht_node *dummy, struct cds_lfht_node *node)
728 {
729 struct cds_lfht_node *iter_prev, *iter, *next, *new_next;
730
731 assert(!is_dummy(dummy));
732 assert(!is_removed(dummy));
733 assert(!is_dummy(node));
734 assert(!is_removed(node));
735 for (;;) {
736 iter_prev = dummy;
737 /* We can always skip the dummy node initially */
738 iter = rcu_dereference(iter_prev->p.next);
739 assert(iter_prev->p.reverse_hash <= node->p.reverse_hash);
740 /*
741 * We should never be called with dummy (start of chain)
742 * and logically removed node (end of path compression
743 * marker) being the actual same node. This would be a
744 * bug in the algorithm implementation.
745 */
746 assert(dummy != node);
747 for (;;) {
748 if (unlikely(is_end(iter)))
749 return;
750 if (likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash))
751 return;
752 next = rcu_dereference(clear_flag(iter)->p.next);
753 if (likely(is_removed(next)))
754 break;
755 iter_prev = clear_flag(iter);
756 iter = next;
757 }
758 assert(!is_removed(iter));
759 if (is_dummy(iter))
760 new_next = flag_dummy(clear_flag(next));
761 else
762 new_next = clear_flag(next);
763 if (is_removed(iter))
764 new_next = flag_removed(new_next);
765 (void) uatomic_cmpxchg(&iter_prev->p.next, iter, new_next);
766 }
767 return;
768 }
769
770 static
771 int _cds_lfht_replace(struct cds_lfht *ht, unsigned long size,
772 struct cds_lfht_node *old_node,
773 struct cds_lfht_node *ret_next,
774 struct cds_lfht_node *new_node)
775 {
776 struct cds_lfht_node *dummy, *old_next;
777 struct _cds_lfht_node *lookup;
778 int flagged = 0;
779 unsigned long hash, index, order;
780
781 if (!old_node) /* Return -ENOENT if asked to replace NULL node */
782 goto end;
783
784 assert(!is_removed(old_node));
785 assert(!is_dummy(old_node));
786 assert(!is_removed(new_node));
787 assert(!is_dummy(new_node));
788 assert(new_node != old_node);
789 do {
790 /* Insert after node to be replaced */
791 old_next = ret_next;
792 if (is_removed(old_next)) {
793 /*
794 * Too late, the old node has been removed under us
795 * between lookup and replace. Fail.
796 */
797 goto end;
798 }
799 assert(!is_dummy(old_next));
800 assert(new_node != clear_flag(old_next));
801 new_node->p.next = clear_flag(old_next);
802 /*
803 * Here is the whole trick for lock-free replace: we add
804 * the replacement node _after_ the node we want to
805 * replace by atomically setting its next pointer at the
806 * same time we set its removal flag. Given that
807 * the lookups/get next use an iterator aware of the
808 * next pointer, they will either skip the old node due
809 * to the removal flag and see the new node, or use
810 * the old node, but will not see the new one.
811 */
812 ret_next = uatomic_cmpxchg(&old_node->p.next,
813 old_next, flag_removed(new_node));
814 } while (ret_next != old_next);
815
816 /* We performed the replacement. */
817 flagged = 1;
818
819 /*
820 * Ensure that the old node is not visible to readers anymore:
821 * lookup for the node, and remove it (along with any other
822 * logically removed node) if found.
823 */
824 hash = bit_reverse_ulong(old_node->p.reverse_hash);
825 assert(size > 0);
826 index = hash & (size - 1);
827 order = get_count_order_ulong(index + 1);
828 lookup = &ht->t.tbl[order]->nodes[index & (!order ? 0 : ((1UL << (order - 1)) - 1))];
829 dummy = (struct cds_lfht_node *) lookup;
830 _cds_lfht_gc_bucket(dummy, new_node);
831 end:
832 /*
833 * Only the flagging action indicated that we (and no other)
834 * replaced the node from the hash table.
835 */
836 if (flagged) {
837 assert(is_removed(rcu_dereference(old_node->p.next)));
838 return 0;
839 } else {
840 return -ENOENT;
841 }
842 }
843
844 static
845 struct cds_lfht_node *_cds_lfht_add(struct cds_lfht *ht,
846 unsigned long size,
847 struct cds_lfht_node *node,
848 enum add_mode mode, int dummy)
849 {
850 struct cds_lfht_node *iter_prev, *iter, *next, *new_node, *new_next,
851 *dummy_node, *return_node;
852 struct _cds_lfht_node *lookup;
853 unsigned long hash, index, order;
854
855 assert(!is_dummy(node));
856 assert(!is_removed(node));
857 if (!size) {
858 assert(dummy);
859 node->p.next = flag_dummy(get_end());
860 return node; /* Initial first add (head) */
861 }
862 hash = bit_reverse_ulong(node->p.reverse_hash);
863 for (;;) {
864 uint32_t chain_len = 0;
865
866 /*
867 * iter_prev points to the non-removed node prior to the
868 * insert location.
869 */
870 index = hash & (size - 1);
871 order = get_count_order_ulong(index + 1);
872 lookup = &ht->t.tbl[order]->nodes[index & ((!order ? 0 : (1UL << (order - 1))) - 1)];
873 iter_prev = (struct cds_lfht_node *) lookup;
874 /* We can always skip the dummy node initially */
875 iter = rcu_dereference(iter_prev->p.next);
876 assert(iter_prev->p.reverse_hash <= node->p.reverse_hash);
877 for (;;) {
878 if (unlikely(is_end(iter)))
879 goto insert;
880 if (likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash))
881 goto insert;
882 next = rcu_dereference(clear_flag(iter)->p.next);
883 if (unlikely(is_removed(next)))
884 goto gc_node;
885 if ((mode == ADD_UNIQUE || mode == ADD_REPLACE)
886 && !is_dummy(next)
887 && !ht->compare_fct(node->key, node->key_len,
888 clear_flag(iter)->key,
889 clear_flag(iter)->key_len)) {
890 if (mode == ADD_UNIQUE)
891 return clear_flag(iter);
892 else /* mode == ADD_REPLACE */
893 goto replace;
894 }
895 /* Only account for identical reverse hash once */
896 if (iter_prev->p.reverse_hash != clear_flag(iter)->p.reverse_hash
897 && !is_dummy(next))
898 check_resize(ht, size, ++chain_len);
899 iter_prev = clear_flag(iter);
900 iter = next;
901 }
902
903 insert:
904 assert(node != clear_flag(iter));
905 assert(!is_removed(iter_prev));
906 assert(!is_removed(iter));
907 assert(iter_prev != node);
908 if (!dummy)
909 node->p.next = clear_flag(iter);
910 else
911 node->p.next = flag_dummy(clear_flag(iter));
912 if (is_dummy(iter))
913 new_node = flag_dummy(node);
914 else
915 new_node = node;
916 if (uatomic_cmpxchg(&iter_prev->p.next, iter,
917 new_node) != iter) {
918 continue; /* retry */
919 } else {
920 if (mode == ADD_REPLACE)
921 return_node = NULL;
922 else /* ADD_DEFAULT and ADD_UNIQUE */
923 return_node = node;
924 goto gc_end;
925 }
926
927 replace:
928
929 if (!_cds_lfht_replace(ht, size, clear_flag(iter), next,
930 node)) {
931 return_node = clear_flag(iter);
932 goto end; /* gc already done */
933 } else {
934 continue; /* retry */
935 }
936
937 gc_node:
938 assert(!is_removed(iter));
939 if (is_dummy(iter))
940 new_next = flag_dummy(clear_flag(next));
941 else
942 new_next = clear_flag(next);
943 (void) uatomic_cmpxchg(&iter_prev->p.next, iter, new_next);
944 /* retry */
945 }
946 gc_end:
947 /* Garbage collect logically removed nodes in the bucket */
948 index = hash & (size - 1);
949 order = get_count_order_ulong(index + 1);
950 lookup = &ht->t.tbl[order]->nodes[index & (!order ? 0 : ((1UL << (order - 1)) - 1))];
951 dummy_node = (struct cds_lfht_node *) lookup;
952 _cds_lfht_gc_bucket(dummy_node, node);
953 end:
954 return return_node;
955 }
956
957 static
958 int _cds_lfht_del(struct cds_lfht *ht, unsigned long size,
959 struct cds_lfht_node *node,
960 int dummy_removal)
961 {
962 struct cds_lfht_node *dummy, *next, *old;
963 struct _cds_lfht_node *lookup;
964 int flagged = 0;
965 unsigned long hash, index, order;
966
967 if (!node) /* Return -ENOENT if asked to delete NULL node */
968 goto end;
969
970 /* logically delete the node */
971 assert(!is_dummy(node));
972 assert(!is_removed(node));
973 old = rcu_dereference(node->p.next);
974 do {
975 struct cds_lfht_node *new_next;
976
977 next = old;
978 if (unlikely(is_removed(next)))
979 goto end;
980 if (dummy_removal)
981 assert(is_dummy(next));
982 else
983 assert(!is_dummy(next));
984 new_next = flag_removed(next);
985 old = uatomic_cmpxchg(&node->p.next, next, new_next);
986 } while (old != next);
987
988 /* We performed the (logical) deletion. */
989 flagged = 1;
990
991 /*
992 * Ensure that the node is not visible to readers anymore: lookup for
993 * the node, and remove it (along with any other logically removed node)
994 * if found.
995 */
996 hash = bit_reverse_ulong(node->p.reverse_hash);
997 assert(size > 0);
998 index = hash & (size - 1);
999 order = get_count_order_ulong(index + 1);
1000 lookup = &ht->t.tbl[order]->nodes[index & (!order ? 0 : ((1UL << (order - 1)) - 1))];
1001 dummy = (struct cds_lfht_node *) lookup;
1002 _cds_lfht_gc_bucket(dummy, node);
1003 end:
1004 /*
1005 * Only the flagging action indicated that we (and no other)
1006 * removed the node from the hash.
1007 */
1008 if (flagged) {
1009 assert(is_removed(rcu_dereference(node->p.next)));
1010 return 0;
1011 } else {
1012 return -ENOENT;
1013 }
1014 }
1015
1016 static
1017 void *partition_resize_thread(void *arg)
1018 {
1019 struct partition_resize_work *work = arg;
1020
1021 work->ht->cds_lfht_rcu_register_thread();
1022 work->fct(work->ht, work->i, work->start, work->len);
1023 work->ht->cds_lfht_rcu_unregister_thread();
1024 return NULL;
1025 }
1026
1027 static
1028 void partition_resize_helper(struct cds_lfht *ht, unsigned long i,
1029 unsigned long len,
1030 void (*fct)(struct cds_lfht *ht, unsigned long i,
1031 unsigned long start, unsigned long len))
1032 {
1033 unsigned long partition_len;
1034 struct partition_resize_work *work;
1035 int thread, ret;
1036 unsigned long nr_threads;
1037 pthread_t *thread_id;
1038
1039 /*
1040 * Note: nr_cpus_mask + 1 is always power of 2.
1041 * We spawn just the number of threads we need to satisfy the minimum
1042 * partition size, up to the number of CPUs in the system.
1043 */
1044 nr_threads = min(nr_cpus_mask + 1,
1045 len >> MIN_PARTITION_PER_THREAD_ORDER);
1046 partition_len = len >> get_count_order_ulong(nr_threads);
1047 work = calloc(nr_threads, sizeof(*work));
1048 thread_id = calloc(nr_threads, sizeof(*thread_id));
1049 assert(work);
1050 for (thread = 0; thread < nr_threads; thread++) {
1051 work[thread].ht = ht;
1052 work[thread].i = i;
1053 work[thread].len = partition_len;
1054 work[thread].start = thread * partition_len;
1055 work[thread].fct = fct;
1056 ret = pthread_create(&thread_id[thread], ht->resize_attr,
1057 partition_resize_thread, &work[thread]);
1058 assert(!ret);
1059 }
1060 for (thread = 0; thread < nr_threads; thread++) {
1061 ret = pthread_join(thread_id[thread], NULL);
1062 assert(!ret);
1063 }
1064 free(work);
1065 free(thread_id);
1066 }
1067
1068 /*
1069 * Holding RCU read lock to protect _cds_lfht_add against memory
1070 * reclaim that could be performed by other call_rcu worker threads (ABA
1071 * problem).
1072 *
1073 * When we reach a certain length, we can split this population phase over
1074 * many worker threads, based on the number of CPUs available in the system.
1075 * This should therefore take care of not having the expand lagging behind too
1076 * many concurrent insertion threads by using the scheduler's ability to
1077 * schedule dummy node population fairly with insertions.
1078 */
1079 static
1080 void init_table_populate_partition(struct cds_lfht *ht, unsigned long i,
1081 unsigned long start, unsigned long len)
1082 {
1083 unsigned long j;
1084
1085 ht->cds_lfht_rcu_read_lock();
1086 for (j = start; j < start + len; j++) {
1087 struct cds_lfht_node *new_node =
1088 (struct cds_lfht_node *) &ht->t.tbl[i]->nodes[j];
1089
1090 dbg_printf("init populate: i %lu j %lu hash %lu\n",
1091 i, j, !i ? 0 : (1UL << (i - 1)) + j);
1092 new_node->p.reverse_hash =
1093 bit_reverse_ulong(!i ? 0 : (1UL << (i - 1)) + j);
1094 (void) _cds_lfht_add(ht, !i ? 0 : (1UL << (i - 1)),
1095 new_node, ADD_DEFAULT, 1);
1096 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1097 break;
1098 }
1099 ht->cds_lfht_rcu_read_unlock();
1100 }
1101
1102 static
1103 void init_table_populate(struct cds_lfht *ht, unsigned long i,
1104 unsigned long len)
1105 {
1106 assert(nr_cpus_mask != -1);
1107 if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD) {
1108 ht->cds_lfht_rcu_thread_online();
1109 init_table_populate_partition(ht, i, 0, len);
1110 ht->cds_lfht_rcu_thread_offline();
1111 return;
1112 }
1113 partition_resize_helper(ht, i, len, init_table_populate_partition);
1114 }
1115
1116 static
1117 void init_table(struct cds_lfht *ht,
1118 unsigned long first_order, unsigned long len_order)
1119 {
1120 unsigned long i, end_order;
1121
1122 dbg_printf("init table: first_order %lu end_order %lu\n",
1123 first_order, first_order + len_order);
1124 end_order = first_order + len_order;
1125 for (i = first_order; i < end_order; i++) {
1126 unsigned long len;
1127
1128 len = !i ? 1 : 1UL << (i - 1);
1129 dbg_printf("init order %lu len: %lu\n", i, len);
1130
1131 /* Stop expand if the resize target changes under us */
1132 if (CMM_LOAD_SHARED(ht->t.resize_target) < (!i ? 1 : (1UL << i)))
1133 break;
1134
1135 ht->t.tbl[i] = calloc(1, sizeof(struct rcu_level)
1136 + (len * sizeof(struct _cds_lfht_node)));
1137 assert(ht->t.tbl[i]);
1138
1139 /*
1140 * Set all dummy nodes reverse hash values for a level and
1141 * link all dummy nodes into the table.
1142 */
1143 init_table_populate(ht, i, len);
1144
1145 /*
1146 * Update table size.
1147 */
1148 cmm_smp_wmb(); /* populate data before RCU size */
1149 CMM_STORE_SHARED(ht->t.size, !i ? 1 : (1UL << i));
1150
1151 dbg_printf("init new size: %lu\n", !i ? 1 : (1UL << i));
1152 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1153 break;
1154 }
1155 }
1156
1157 /*
1158 * Holding RCU read lock to protect _cds_lfht_remove against memory
1159 * reclaim that could be performed by other call_rcu worker threads (ABA
1160 * problem).
1161 * For a single level, we logically remove and garbage collect each node.
1162 *
1163 * As a design choice, we perform logical removal and garbage collection on a
1164 * node-per-node basis to simplify this algorithm. We also assume keeping good
1165 * cache locality of the operation would overweight possible performance gain
1166 * that could be achieved by batching garbage collection for multiple levels.
1167 * However, this would have to be justified by benchmarks.
1168 *
1169 * Concurrent removal and add operations are helping us perform garbage
1170 * collection of logically removed nodes. We guarantee that all logically
1171 * removed nodes have been garbage-collected (unlinked) before call_rcu is
1172 * invoked to free a hole level of dummy nodes (after a grace period).
1173 *
1174 * Logical removal and garbage collection can therefore be done in batch or on a
1175 * node-per-node basis, as long as the guarantee above holds.
1176 *
1177 * When we reach a certain length, we can split this removal over many worker
1178 * threads, based on the number of CPUs available in the system. This should
1179 * take care of not letting resize process lag behind too many concurrent
1180 * updater threads actively inserting into the hash table.
1181 */
1182 static
1183 void remove_table_partition(struct cds_lfht *ht, unsigned long i,
1184 unsigned long start, unsigned long len)
1185 {
1186 unsigned long j;
1187
1188 ht->cds_lfht_rcu_read_lock();
1189 for (j = start; j < start + len; j++) {
1190 struct cds_lfht_node *fini_node =
1191 (struct cds_lfht_node *) &ht->t.tbl[i]->nodes[j];
1192
1193 dbg_printf("remove entry: i %lu j %lu hash %lu\n",
1194 i, j, !i ? 0 : (1UL << (i - 1)) + j);
1195 fini_node->p.reverse_hash =
1196 bit_reverse_ulong(!i ? 0 : (1UL << (i - 1)) + j);
1197 (void) _cds_lfht_del(ht, !i ? 0 : (1UL << (i - 1)),
1198 fini_node, 1);
1199 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1200 break;
1201 }
1202 ht->cds_lfht_rcu_read_unlock();
1203 }
1204
1205 static
1206 void remove_table(struct cds_lfht *ht, unsigned long i, unsigned long len)
1207 {
1208
1209 assert(nr_cpus_mask != -1);
1210 if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD) {
1211 ht->cds_lfht_rcu_thread_online();
1212 remove_table_partition(ht, i, 0, len);
1213 ht->cds_lfht_rcu_thread_offline();
1214 return;
1215 }
1216 partition_resize_helper(ht, i, len, remove_table_partition);
1217 }
1218
1219 static
1220 void fini_table(struct cds_lfht *ht,
1221 unsigned long first_order, unsigned long len_order)
1222 {
1223 long i, end_order;
1224
1225 dbg_printf("fini table: first_order %lu end_order %lu\n",
1226 first_order, first_order + len_order);
1227 end_order = first_order + len_order;
1228 assert(first_order > 0);
1229 for (i = end_order - 1; i >= first_order; i--) {
1230 unsigned long len;
1231
1232 len = !i ? 1 : 1UL << (i - 1);
1233 dbg_printf("fini order %lu len: %lu\n", i, len);
1234
1235 /* Stop shrink if the resize target changes under us */
1236 if (CMM_LOAD_SHARED(ht->t.resize_target) > (1UL << (i - 1)))
1237 break;
1238
1239 cmm_smp_wmb(); /* populate data before RCU size */
1240 CMM_STORE_SHARED(ht->t.size, 1UL << (i - 1));
1241
1242 /*
1243 * We need to wait for all add operations to reach Q.S. (and
1244 * thus use the new table for lookups) before we can start
1245 * releasing the old dummy nodes. Otherwise their lookup will
1246 * return a logically removed node as insert position.
1247 */
1248 ht->cds_lfht_synchronize_rcu();
1249
1250 /*
1251 * Set "removed" flag in dummy nodes about to be removed.
1252 * Unlink all now-logically-removed dummy node pointers.
1253 * Concurrent add/remove operation are helping us doing
1254 * the gc.
1255 */
1256 remove_table(ht, i, len);
1257
1258 ht->cds_lfht_call_rcu(&ht->t.tbl[i]->head, cds_lfht_free_level);
1259
1260 dbg_printf("fini new size: %lu\n", 1UL << i);
1261 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1262 break;
1263 }
1264 }
1265
1266 struct cds_lfht *_cds_lfht_new(cds_lfht_hash_fct hash_fct,
1267 cds_lfht_compare_fct compare_fct,
1268 unsigned long hash_seed,
1269 unsigned long init_size,
1270 int flags,
1271 void (*cds_lfht_call_rcu)(struct rcu_head *head,
1272 void (*func)(struct rcu_head *head)),
1273 void (*cds_lfht_synchronize_rcu)(void),
1274 void (*cds_lfht_rcu_read_lock)(void),
1275 void (*cds_lfht_rcu_read_unlock)(void),
1276 void (*cds_lfht_rcu_thread_offline)(void),
1277 void (*cds_lfht_rcu_thread_online)(void),
1278 void (*cds_lfht_rcu_register_thread)(void),
1279 void (*cds_lfht_rcu_unregister_thread)(void),
1280 pthread_attr_t *attr)
1281 {
1282 struct cds_lfht *ht;
1283 unsigned long order;
1284
1285 /* init_size must be power of two */
1286 if (init_size && (init_size & (init_size - 1)))
1287 return NULL;
1288 ht = calloc(1, sizeof(struct cds_lfht));
1289 assert(ht);
1290 ht->hash_fct = hash_fct;
1291 ht->compare_fct = compare_fct;
1292 ht->hash_seed = hash_seed;
1293 ht->cds_lfht_call_rcu = cds_lfht_call_rcu;
1294 ht->cds_lfht_synchronize_rcu = cds_lfht_synchronize_rcu;
1295 ht->cds_lfht_rcu_read_lock = cds_lfht_rcu_read_lock;
1296 ht->cds_lfht_rcu_read_unlock = cds_lfht_rcu_read_unlock;
1297 ht->cds_lfht_rcu_thread_offline = cds_lfht_rcu_thread_offline;
1298 ht->cds_lfht_rcu_thread_online = cds_lfht_rcu_thread_online;
1299 ht->cds_lfht_rcu_register_thread = cds_lfht_rcu_register_thread;
1300 ht->cds_lfht_rcu_unregister_thread = cds_lfht_rcu_unregister_thread;
1301 ht->resize_attr = attr;
1302 ht->percpu_count = alloc_per_cpu_items_count();
1303 /* this mutex should not nest in read-side C.S. */
1304 pthread_mutex_init(&ht->resize_mutex, NULL);
1305 order = get_count_order_ulong(max(init_size, MIN_TABLE_SIZE)) + 1;
1306 ht->flags = flags;
1307 ht->cds_lfht_rcu_thread_offline();
1308 pthread_mutex_lock(&ht->resize_mutex);
1309 ht->t.resize_target = 1UL << (order - 1);
1310 init_table(ht, 0, order);
1311 pthread_mutex_unlock(&ht->resize_mutex);
1312 ht->cds_lfht_rcu_thread_online();
1313 return ht;
1314 }
1315
1316 void cds_lfht_lookup(struct cds_lfht *ht, void *key, size_t key_len,
1317 struct cds_lfht_iter *iter)
1318 {
1319 struct cds_lfht_node *node, *next, *dummy_node;
1320 struct _cds_lfht_node *lookup;
1321 unsigned long hash, reverse_hash, index, order, size;
1322
1323 hash = ht->hash_fct(key, key_len, ht->hash_seed);
1324 reverse_hash = bit_reverse_ulong(hash);
1325
1326 size = rcu_dereference(ht->t.size);
1327 index = hash & (size - 1);
1328 order = get_count_order_ulong(index + 1);
1329 lookup = &ht->t.tbl[order]->nodes[index & (!order ? 0 : ((1UL << (order - 1))) - 1)];
1330 dbg_printf("lookup hash %lu index %lu order %lu aridx %lu\n",
1331 hash, index, order, index & (!order ? 0 : ((1UL << (order - 1)) - 1)));
1332 dummy_node = (struct cds_lfht_node *) lookup;
1333 /* We can always skip the dummy node initially */
1334 node = rcu_dereference(dummy_node->p.next);
1335 node = clear_flag(node);
1336 for (;;) {
1337 if (unlikely(is_end(node))) {
1338 node = next = NULL;
1339 break;
1340 }
1341 if (unlikely(node->p.reverse_hash > reverse_hash)) {
1342 node = next = NULL;
1343 break;
1344 }
1345 next = rcu_dereference(node->p.next);
1346 if (likely(!is_removed(next))
1347 && !is_dummy(next)
1348 && likely(!ht->compare_fct(node->key, node->key_len, key, key_len))) {
1349 break;
1350 }
1351 node = clear_flag(next);
1352 }
1353 assert(!node || !is_dummy(rcu_dereference(node->p.next)));
1354 iter->node = node;
1355 iter->next = next;
1356 }
1357
1358 void cds_lfht_next(struct cds_lfht *ht, struct cds_lfht_iter *iter)
1359 {
1360 struct cds_lfht_node *node, *next;
1361 unsigned long reverse_hash;
1362 void *key;
1363 size_t key_len;
1364
1365 node = iter->node;
1366 reverse_hash = node->p.reverse_hash;
1367 key = node->key;
1368 key_len = node->key_len;
1369 next = iter->next;
1370 node = clear_flag(next);
1371
1372 for (;;) {
1373 if (unlikely(is_end(node))) {
1374 node = next = NULL;
1375 break;
1376 }
1377 if (unlikely(node->p.reverse_hash > reverse_hash)) {
1378 node = next = NULL;
1379 break;
1380 }
1381 next = rcu_dereference(node->p.next);
1382 if (likely(!is_removed(next))
1383 && !is_dummy(next)
1384 && likely(!ht->compare_fct(node->key, node->key_len, key, key_len))) {
1385 break;
1386 }
1387 node = clear_flag(next);
1388 }
1389 assert(!node || !is_dummy(rcu_dereference(node->p.next)));
1390 iter->node = node;
1391 iter->next = next;
1392 }
1393
1394 void cds_lfht_add(struct cds_lfht *ht, struct cds_lfht_node *node)
1395 {
1396 unsigned long hash, size;
1397
1398 hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed);
1399 node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash);
1400
1401 size = rcu_dereference(ht->t.size);
1402 (void) _cds_lfht_add(ht, size, node, ADD_DEFAULT, 0);
1403 ht_count_add(ht, size);
1404 }
1405
1406 struct cds_lfht_node *cds_lfht_add_unique(struct cds_lfht *ht,
1407 struct cds_lfht_node *node)
1408 {
1409 unsigned long hash, size;
1410 struct cds_lfht_node *ret;
1411
1412 hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed);
1413 node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash);
1414
1415 size = rcu_dereference(ht->t.size);
1416 ret = _cds_lfht_add(ht, size, node, ADD_UNIQUE, 0);
1417 if (ret == node)
1418 ht_count_add(ht, size);
1419 return ret;
1420 }
1421
1422 struct cds_lfht_node *cds_lfht_add_replace(struct cds_lfht *ht,
1423 struct cds_lfht_node *node)
1424 {
1425 unsigned long hash, size;
1426 struct cds_lfht_node *ret;
1427
1428 hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed);
1429 node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash);
1430
1431 size = rcu_dereference(ht->t.size);
1432 ret = _cds_lfht_add(ht, size, node, ADD_REPLACE, 0);
1433 if (ret == NULL)
1434 ht_count_add(ht, size);
1435 return ret;
1436 }
1437
1438 int cds_lfht_replace(struct cds_lfht *ht, struct cds_lfht_iter *old_iter,
1439 struct cds_lfht_node *new_node)
1440 {
1441 unsigned long size;
1442
1443 size = rcu_dereference(ht->t.size);
1444 return _cds_lfht_replace(ht, size, old_iter->node, old_iter->next,
1445 new_node);
1446 }
1447
1448 int cds_lfht_del(struct cds_lfht *ht, struct cds_lfht_iter *iter)
1449 {
1450 unsigned long size;
1451 int ret;
1452
1453 size = rcu_dereference(ht->t.size);
1454 ret = _cds_lfht_del(ht, size, iter->node, 0);
1455 if (!ret)
1456 ht_count_del(ht, size);
1457 return ret;
1458 }
1459
1460 static
1461 int cds_lfht_delete_dummy(struct cds_lfht *ht)
1462 {
1463 struct cds_lfht_node *node;
1464 struct _cds_lfht_node *lookup;
1465 unsigned long order, i, size;
1466
1467 /* Check that the table is empty */
1468 lookup = &ht->t.tbl[0]->nodes[0];
1469 node = (struct cds_lfht_node *) lookup;
1470 do {
1471 node = clear_flag(node)->p.next;
1472 if (!is_dummy(node))
1473 return -EPERM;
1474 assert(!is_removed(node));
1475 } while (!is_end(node));
1476 /*
1477 * size accessed without rcu_dereference because hash table is
1478 * being destroyed.
1479 */
1480 size = ht->t.size;
1481 /* Internal sanity check: all nodes left should be dummy */
1482 for (order = 0; order < get_count_order_ulong(size) + 1; order++) {
1483 unsigned long len;
1484
1485 len = !order ? 1 : 1UL << (order - 1);
1486 for (i = 0; i < len; i++) {
1487 dbg_printf("delete order %lu i %lu hash %lu\n",
1488 order, i,
1489 bit_reverse_ulong(ht->t.tbl[order]->nodes[i].reverse_hash));
1490 assert(is_dummy(ht->t.tbl[order]->nodes[i].next));
1491 }
1492 poison_free(ht->t.tbl[order]);
1493 }
1494 return 0;
1495 }
1496
1497 /*
1498 * Should only be called when no more concurrent readers nor writers can
1499 * possibly access the table.
1500 */
1501 int cds_lfht_destroy(struct cds_lfht *ht, pthread_attr_t **attr)
1502 {
1503 int ret;
1504
1505 /* Wait for in-flight resize operations to complete */
1506 CMM_STORE_SHARED(ht->in_progress_destroy, 1);
1507 while (uatomic_read(&ht->in_progress_resize))
1508 poll(NULL, 0, 100); /* wait for 100ms */
1509 ret = cds_lfht_delete_dummy(ht);
1510 if (ret)
1511 return ret;
1512 free_per_cpu_items_count(ht->percpu_count);
1513 if (attr)
1514 *attr = ht->resize_attr;
1515 poison_free(ht);
1516 return ret;
1517 }
1518
1519 void cds_lfht_count_nodes(struct cds_lfht *ht,
1520 long *approx_before,
1521 unsigned long *count,
1522 unsigned long *removed,
1523 long *approx_after)
1524 {
1525 struct cds_lfht_node *node, *next;
1526 struct _cds_lfht_node *lookup;
1527 unsigned long nr_dummy = 0;
1528
1529 *approx_before = 0;
1530 if (nr_cpus_mask >= 0) {
1531 int i;
1532
1533 for (i = 0; i < nr_cpus_mask + 1; i++) {
1534 *approx_before += uatomic_read(&ht->percpu_count[i].add);
1535 *approx_before -= uatomic_read(&ht->percpu_count[i].del);
1536 }
1537 }
1538
1539 *count = 0;
1540 *removed = 0;
1541
1542 /* Count non-dummy nodes in the table */
1543 lookup = &ht->t.tbl[0]->nodes[0];
1544 node = (struct cds_lfht_node *) lookup;
1545 do {
1546 next = rcu_dereference(node->p.next);
1547 if (is_removed(next)) {
1548 if (!is_dummy(next))
1549 (*removed)++;
1550 else
1551 (nr_dummy)++;
1552 } else if (!is_dummy(next))
1553 (*count)++;
1554 else
1555 (nr_dummy)++;
1556 node = clear_flag(next);
1557 } while (!is_end(node));
1558 dbg_printf("number of dummy nodes: %lu\n", nr_dummy);
1559 *approx_after = 0;
1560 if (nr_cpus_mask >= 0) {
1561 int i;
1562
1563 for (i = 0; i < nr_cpus_mask + 1; i++) {
1564 *approx_after += uatomic_read(&ht->percpu_count[i].add);
1565 *approx_after -= uatomic_read(&ht->percpu_count[i].del);
1566 }
1567 }
1568 }
1569
1570 /* called with resize mutex held */
1571 static
1572 void _do_cds_lfht_grow(struct cds_lfht *ht,
1573 unsigned long old_size, unsigned long new_size)
1574 {
1575 unsigned long old_order, new_order;
1576
1577 old_order = get_count_order_ulong(old_size) + 1;
1578 new_order = get_count_order_ulong(new_size) + 1;
1579 printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1580 old_size, old_order, new_size, new_order);
1581 assert(new_size > old_size);
1582 init_table(ht, old_order, new_order - old_order);
1583 }
1584
1585 /* called with resize mutex held */
1586 static
1587 void _do_cds_lfht_shrink(struct cds_lfht *ht,
1588 unsigned long old_size, unsigned long new_size)
1589 {
1590 unsigned long old_order, new_order;
1591
1592 new_size = max(new_size, MIN_TABLE_SIZE);
1593 old_order = get_count_order_ulong(old_size) + 1;
1594 new_order = get_count_order_ulong(new_size) + 1;
1595 printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1596 old_size, old_order, new_size, new_order);
1597 assert(new_size < old_size);
1598
1599 /* Remove and unlink all dummy nodes to remove. */
1600 fini_table(ht, new_order, old_order - new_order);
1601 }
1602
1603
1604 /* called with resize mutex held */
1605 static
1606 void _do_cds_lfht_resize(struct cds_lfht *ht)
1607 {
1608 unsigned long new_size, old_size;
1609
1610 /*
1611 * Resize table, re-do if the target size has changed under us.
1612 */
1613 do {
1614 ht->t.resize_initiated = 1;
1615 old_size = ht->t.size;
1616 new_size = CMM_LOAD_SHARED(ht->t.resize_target);
1617 if (old_size < new_size)
1618 _do_cds_lfht_grow(ht, old_size, new_size);
1619 else if (old_size > new_size)
1620 _do_cds_lfht_shrink(ht, old_size, new_size);
1621 ht->t.resize_initiated = 0;
1622 /* write resize_initiated before read resize_target */
1623 cmm_smp_mb();
1624 } while (ht->t.size != CMM_LOAD_SHARED(ht->t.resize_target));
1625 }
1626
1627 static
1628 unsigned long resize_target_update(struct cds_lfht *ht, unsigned long size,
1629 int growth_order)
1630 {
1631 return _uatomic_max(&ht->t.resize_target,
1632 size << growth_order);
1633 }
1634
1635 static
1636 void resize_target_update_count(struct cds_lfht *ht,
1637 unsigned long count)
1638 {
1639 count = max(count, MIN_TABLE_SIZE);
1640 uatomic_set(&ht->t.resize_target, count);
1641 }
1642
1643 void cds_lfht_resize(struct cds_lfht *ht, unsigned long new_size)
1644 {
1645 resize_target_update_count(ht, new_size);
1646 CMM_STORE_SHARED(ht->t.resize_initiated, 1);
1647 ht->cds_lfht_rcu_thread_offline();
1648 pthread_mutex_lock(&ht->resize_mutex);
1649 _do_cds_lfht_resize(ht);
1650 pthread_mutex_unlock(&ht->resize_mutex);
1651 ht->cds_lfht_rcu_thread_online();
1652 }
1653
1654 static
1655 void do_resize_cb(struct rcu_head *head)
1656 {
1657 struct rcu_resize_work *work =
1658 caa_container_of(head, struct rcu_resize_work, head);
1659 struct cds_lfht *ht = work->ht;
1660
1661 ht->cds_lfht_rcu_thread_offline();
1662 pthread_mutex_lock(&ht->resize_mutex);
1663 _do_cds_lfht_resize(ht);
1664 pthread_mutex_unlock(&ht->resize_mutex);
1665 ht->cds_lfht_rcu_thread_online();
1666 poison_free(work);
1667 cmm_smp_mb(); /* finish resize before decrement */
1668 uatomic_dec(&ht->in_progress_resize);
1669 }
1670
1671 static
1672 void cds_lfht_resize_lazy(struct cds_lfht *ht, unsigned long size, int growth)
1673 {
1674 struct rcu_resize_work *work;
1675 unsigned long target_size;
1676
1677 target_size = resize_target_update(ht, size, growth);
1678 /* Store resize_target before read resize_initiated */
1679 cmm_smp_mb();
1680 if (!CMM_LOAD_SHARED(ht->t.resize_initiated) && size < target_size) {
1681 uatomic_inc(&ht->in_progress_resize);
1682 cmm_smp_mb(); /* increment resize count before calling it */
1683 work = malloc(sizeof(*work));
1684 work->ht = ht;
1685 ht->cds_lfht_call_rcu(&work->head, do_resize_cb);
1686 CMM_STORE_SHARED(ht->t.resize_initiated, 1);
1687 }
1688 }
1689
1690 #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
1691
1692 static
1693 void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
1694 unsigned long count)
1695 {
1696 struct rcu_resize_work *work;
1697
1698 if (!(ht->flags & CDS_LFHT_AUTO_RESIZE))
1699 return;
1700 resize_target_update_count(ht, count);
1701 /* Store resize_target before read resize_initiated */
1702 cmm_smp_mb();
1703 if (!CMM_LOAD_SHARED(ht->t.resize_initiated)) {
1704 uatomic_inc(&ht->in_progress_resize);
1705 cmm_smp_mb(); /* increment resize count before calling it */
1706 work = malloc(sizeof(*work));
1707 work->ht = ht;
1708 ht->cds_lfht_call_rcu(&work->head, do_resize_cb);
1709 CMM_STORE_SHARED(ht->t.resize_initiated, 1);
1710 }
1711 }
1712
1713 #endif
This page took 0.074721 seconds and 4 git commands to generate.