Merge branch 'master' of git://git.lttng.org/lttng-tools
[lttng-tools.git] / liblttng-ht / 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 * Copyright 2011 - Lai Jiangshan <laijs@cn.fujitsu.com>
8 *
9 * This library is free software; you can redistribute it and/or
10 * modify it under the terms of the GNU Lesser General Public
11 * License as published by the Free Software Foundation; either
12 * version 2.1 of the License, or (at your option) any later version.
13 *
14 * This library is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
17 * Lesser General Public License for more details.
18 *
19 * You should have received a copy of the GNU Lesser General Public
20 * License along with this library; if not, write to the Free Software
21 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
22 */
23
24 /*
25 * Based on the following articles:
26 * - Ori Shalev and Nir Shavit. Split-ordered lists: Lock-free
27 * extensible hash tables. J. ACM 53, 3 (May 2006), 379-405.
28 * - Michael, M. M. High performance dynamic lock-free hash tables
29 * and list-based sets. In Proceedings of the fourteenth annual ACM
30 * symposium on Parallel algorithms and architectures, ACM Press,
31 * (2002), 73-82.
32 *
33 * Some specificities of this Lock-Free Resizable RCU Hash Table
34 * implementation:
35 *
36 * - RCU read-side critical section allows readers to perform hash
37 * table lookups and use the returned objects safely by delaying
38 * memory reclaim of a grace period.
39 * - Add and remove operations are lock-free, and do not need to
40 * allocate memory. They need to be executed within RCU read-side
41 * critical section to ensure the objects they read are valid and to
42 * deal with the cmpxchg ABA problem.
43 * - add and add_unique operations are supported. add_unique checks if
44 * the node key already exists in the hash table. It ensures no key
45 * duplicata exists.
46 * - The resize operation executes concurrently with add/remove/lookup.
47 * - Hash table nodes are contained within a split-ordered list. This
48 * list is ordered by incrementing reversed-bits-hash value.
49 * - An index of bucket nodes is kept. These bucket nodes are the hash
50 * table "buckets", and they are also chained together in the
51 * split-ordered list, which allows recursive expansion.
52 * - The resize operation for small tables only allows expanding the hash table.
53 * It is triggered automatically by detecting long chains in the add
54 * operation.
55 * - The resize operation for larger tables (and available through an
56 * API) allows both expanding and shrinking the hash table.
57 * - Split-counters are used to keep track of the number of
58 * nodes within the hash table for automatic resize triggering.
59 * - Resize operation initiated by long chain detection is executed by a
60 * call_rcu thread, which keeps lock-freedom of add and remove.
61 * - Resize operations are protected by a mutex.
62 * - The removal operation is split in two parts: first, a "removed"
63 * flag is set in the next pointer within the node to remove. Then,
64 * a "garbage collection" is performed in the bucket containing the
65 * removed node (from the start of the bucket up to the removed node).
66 * All encountered nodes with "removed" flag set in their next
67 * pointers are removed from the linked-list. If the cmpxchg used for
68 * removal fails (due to concurrent garbage-collection or concurrent
69 * add), we retry from the beginning of the bucket. This ensures that
70 * the node with "removed" flag set is removed from the hash table
71 * (not visible to lookups anymore) before the RCU read-side critical
72 * section held across removal ends. Furthermore, this ensures that
73 * the node with "removed" flag set is removed from the linked-list
74 * before its memory is reclaimed. Only the thread which removal
75 * successfully set the "removed" flag (with a cmpxchg) into a node's
76 * next pointer is considered to have succeeded its removal (and thus
77 * owns the node to reclaim). Because we garbage-collect starting from
78 * an invariant node (the start-of-bucket bucket node) up to the
79 * "removed" node (or find a reverse-hash that is higher), we are sure
80 * that a successful traversal of the chain leads to a chain that is
81 * present in the linked-list (the start node is never removed) and
82 * that is does not contain the "removed" node anymore, even if
83 * concurrent delete/add operations are changing the structure of the
84 * list concurrently.
85 * - The add operation performs gargage collection of buckets if it
86 * encounters nodes with removed flag set in the bucket where it wants
87 * to add its new node. This ensures lock-freedom of add operation by
88 * helping the remover unlink nodes from the list rather than to wait
89 * for it do to so.
90 * - A RCU "order table" indexed by log2(hash index) is copied and
91 * expanded by the resize operation. This order table allows finding
92 * the "bucket node" tables.
93 * - There is one bucket node table per hash index order. The size of
94 * each bucket node table is half the number of hashes contained in
95 * this order (except for order 0).
96 * - synchronzie_rcu is used to garbage-collect the old bucket node table.
97 * - The per-order bucket node tables contain a compact version of the
98 * hash table nodes. These tables are invariant after they are
99 * populated into the hash table.
100 *
101 * Bucket node tables:
102 *
103 * hash table hash table the last all bucket node tables
104 * order size bucket node 0 1 2 3 4 5 6(index)
105 * table size
106 * 0 1 1 1
107 * 1 2 1 1 1
108 * 2 4 2 1 1 2
109 * 3 8 4 1 1 2 4
110 * 4 16 8 1 1 2 4 8
111 * 5 32 16 1 1 2 4 8 16
112 * 6 64 32 1 1 2 4 8 16 32
113 *
114 * When growing/shrinking, we only focus on the last bucket node table
115 * which size is (!order ? 1 : (1 << (order -1))).
116 *
117 * Example for growing/shrinking:
118 * grow hash table from order 5 to 6: init the index=6 bucket node table
119 * shrink hash table from order 6 to 5: fini the index=6 bucket node table
120 *
121 * A bit of ascii art explanation:
122 *
123 * Order index is the off-by-one compare to the actual power of 2 because
124 * we use index 0 to deal with the 0 special-case.
125 *
126 * This shows the nodes for a small table ordered by reversed bits:
127 *
128 * bits reverse
129 * 0 000 000
130 * 4 100 001
131 * 2 010 010
132 * 6 110 011
133 * 1 001 100
134 * 5 101 101
135 * 3 011 110
136 * 7 111 111
137 *
138 * This shows the nodes in order of non-reversed bits, linked by
139 * reversed-bit order.
140 *
141 * order bits reverse
142 * 0 0 000 000
143 * 1 | 1 001 100 <-
144 * 2 | | 2 010 010 <- |
145 * | | | 3 011 110 | <- |
146 * 3 -> | | | 4 100 001 | |
147 * -> | | 5 101 101 |
148 * -> | 6 110 011
149 * -> 7 111 111
150 */
151
152 #define _LGPL_SOURCE
153 #include <stdlib.h>
154 #include <errno.h>
155 #include <assert.h>
156 #include <stdio.h>
157 #include <stdint.h>
158 #include <string.h>
159
160 #include "config.h"
161 #include <urcu.h>
162 #include <urcu-call-rcu.h>
163 #include <urcu/arch.h>
164 #include <urcu/uatomic.h>
165 #include <urcu/compiler.h>
166 #include <stdio.h>
167 #include <pthread.h>
168
169 #include "rculfhash.h"
170 #include "rculfhash-internal.h"
171 #include "urcu-flavor.h"
172
173 /*
174 * Split-counters lazily update the global counter each 1024
175 * addition/removal. It automatically keeps track of resize required.
176 * We use the bucket length as indicator for need to expand for small
177 * tables and machines lacking per-cpu data suppport.
178 */
179 #define COUNT_COMMIT_ORDER 10
180 #define DEFAULT_SPLIT_COUNT_MASK 0xFUL
181 #define CHAIN_LEN_TARGET 1
182 #define CHAIN_LEN_RESIZE_THRESHOLD 3
183
184 /*
185 * Define the minimum table size.
186 */
187 #define MIN_TABLE_ORDER 0
188 #define MIN_TABLE_SIZE (1UL << MIN_TABLE_ORDER)
189
190 /*
191 * Minimum number of bucket nodes to touch per thread to parallelize grow/shrink.
192 */
193 #define MIN_PARTITION_PER_THREAD_ORDER 12
194 #define MIN_PARTITION_PER_THREAD (1UL << MIN_PARTITION_PER_THREAD_ORDER)
195
196 /*
197 * The removed flag needs to be updated atomically with the pointer.
198 * It indicates that no node must attach to the node scheduled for
199 * removal, and that node garbage collection must be performed.
200 * The bucket flag does not require to be updated atomically with the
201 * pointer, but it is added as a pointer low bit flag to save space.
202 */
203 #define REMOVED_FLAG (1UL << 0)
204 #define BUCKET_FLAG (1UL << 1)
205 #define REMOVAL_OWNER_FLAG (1UL << 2)
206 #define FLAGS_MASK ((1UL << 3) - 1)
207
208 /* Value of the end pointer. Should not interact with flags. */
209 #define END_VALUE NULL
210
211 DEFINE_RCU_FLAVOR(rcu_flavor);
212
213 /*
214 * ht_items_count: Split-counters counting the number of node addition
215 * and removal in the table. Only used if the CDS_LFHT_ACCOUNTING flag
216 * is set at hash table creation.
217 *
218 * These are free-running counters, never reset to zero. They count the
219 * number of add/remove, and trigger every (1 << COUNT_COMMIT_ORDER)
220 * operations to update the global counter. We choose a power-of-2 value
221 * for the trigger to deal with 32 or 64-bit overflow of the counter.
222 */
223 struct ht_items_count {
224 unsigned long add, del;
225 } __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
226
227 /*
228 * rcu_resize_work: Contains arguments passed to RCU worker thread
229 * responsible for performing lazy resize.
230 */
231 struct rcu_resize_work {
232 struct rcu_head head;
233 struct cds_lfht *ht;
234 };
235
236 /*
237 * partition_resize_work: Contains arguments passed to worker threads
238 * executing the hash table resize on partitions of the hash table
239 * assigned to each processor's worker thread.
240 */
241 struct partition_resize_work {
242 pthread_t thread_id;
243 struct cds_lfht *ht;
244 unsigned long i, start, len;
245 void (*fct)(struct cds_lfht *ht, unsigned long i,
246 unsigned long start, unsigned long len);
247 };
248
249 /*
250 * Algorithm to reverse bits in a word by lookup table, extended to
251 * 64-bit words.
252 * Source:
253 * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable
254 * Originally from Public Domain.
255 */
256
257 static const uint8_t BitReverseTable256[256] =
258 {
259 #define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64
260 #define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16)
261 #define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 )
262 R6(0), R6(2), R6(1), R6(3)
263 };
264 #undef R2
265 #undef R4
266 #undef R6
267
268 static
269 uint8_t bit_reverse_u8(uint8_t v)
270 {
271 return BitReverseTable256[v];
272 }
273
274 static __attribute__((unused))
275 uint32_t bit_reverse_u32(uint32_t v)
276 {
277 return ((uint32_t) bit_reverse_u8(v) << 24) |
278 ((uint32_t) bit_reverse_u8(v >> 8) << 16) |
279 ((uint32_t) bit_reverse_u8(v >> 16) << 8) |
280 ((uint32_t) bit_reverse_u8(v >> 24));
281 }
282
283 static __attribute__((unused))
284 uint64_t bit_reverse_u64(uint64_t v)
285 {
286 return ((uint64_t) bit_reverse_u8(v) << 56) |
287 ((uint64_t) bit_reverse_u8(v >> 8) << 48) |
288 ((uint64_t) bit_reverse_u8(v >> 16) << 40) |
289 ((uint64_t) bit_reverse_u8(v >> 24) << 32) |
290 ((uint64_t) bit_reverse_u8(v >> 32) << 24) |
291 ((uint64_t) bit_reverse_u8(v >> 40) << 16) |
292 ((uint64_t) bit_reverse_u8(v >> 48) << 8) |
293 ((uint64_t) bit_reverse_u8(v >> 56));
294 }
295
296 static
297 unsigned long bit_reverse_ulong(unsigned long v)
298 {
299 #if (CAA_BITS_PER_LONG == 32)
300 return bit_reverse_u32(v);
301 #else
302 return bit_reverse_u64(v);
303 #endif
304 }
305
306 /*
307 * fls: returns the position of the most significant bit.
308 * Returns 0 if no bit is set, else returns the position of the most
309 * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit).
310 */
311 #if defined(__i386) || defined(__x86_64)
312 static inline
313 unsigned int fls_u32(uint32_t x)
314 {
315 int r;
316
317 asm("bsrl %1,%0\n\t"
318 "jnz 1f\n\t"
319 "movl $-1,%0\n\t"
320 "1:\n\t"
321 : "=r" (r) : "rm" (x));
322 return r + 1;
323 }
324 #define HAS_FLS_U32
325 #endif
326
327 #if defined(__x86_64)
328 static inline
329 unsigned int fls_u64(uint64_t x)
330 {
331 long r;
332
333 asm("bsrq %1,%0\n\t"
334 "jnz 1f\n\t"
335 "movq $-1,%0\n\t"
336 "1:\n\t"
337 : "=r" (r) : "rm" (x));
338 return r + 1;
339 }
340 #define HAS_FLS_U64
341 #endif
342
343 #ifndef HAS_FLS_U64
344 static __attribute__((unused))
345 unsigned int fls_u64(uint64_t x)
346 {
347 unsigned int r = 64;
348
349 if (!x)
350 return 0;
351
352 if (!(x & 0xFFFFFFFF00000000ULL)) {
353 x <<= 32;
354 r -= 32;
355 }
356 if (!(x & 0xFFFF000000000000ULL)) {
357 x <<= 16;
358 r -= 16;
359 }
360 if (!(x & 0xFF00000000000000ULL)) {
361 x <<= 8;
362 r -= 8;
363 }
364 if (!(x & 0xF000000000000000ULL)) {
365 x <<= 4;
366 r -= 4;
367 }
368 if (!(x & 0xC000000000000000ULL)) {
369 x <<= 2;
370 r -= 2;
371 }
372 if (!(x & 0x8000000000000000ULL)) {
373 x <<= 1;
374 r -= 1;
375 }
376 return r;
377 }
378 #endif
379
380 #ifndef HAS_FLS_U32
381 static __attribute__((unused))
382 unsigned int fls_u32(uint32_t x)
383 {
384 unsigned int r = 32;
385
386 if (!x)
387 return 0;
388 if (!(x & 0xFFFF0000U)) {
389 x <<= 16;
390 r -= 16;
391 }
392 if (!(x & 0xFF000000U)) {
393 x <<= 8;
394 r -= 8;
395 }
396 if (!(x & 0xF0000000U)) {
397 x <<= 4;
398 r -= 4;
399 }
400 if (!(x & 0xC0000000U)) {
401 x <<= 2;
402 r -= 2;
403 }
404 if (!(x & 0x80000000U)) {
405 x <<= 1;
406 r -= 1;
407 }
408 return r;
409 }
410 #endif
411
412 unsigned int cds_lfht_fls_ulong(unsigned long x)
413 {
414 #if (CAA_BITS_PER_LONG == 32)
415 return fls_u32(x);
416 #else
417 return fls_u64(x);
418 #endif
419 }
420
421 /*
422 * Return the minimum order for which x <= (1UL << order).
423 * Return -1 if x is 0.
424 */
425 int cds_lfht_get_count_order_u32(uint32_t x)
426 {
427 if (!x)
428 return -1;
429
430 return fls_u32(x - 1);
431 }
432
433 /*
434 * Return the minimum order for which x <= (1UL << order).
435 * Return -1 if x is 0.
436 */
437 int cds_lfht_get_count_order_ulong(unsigned long x)
438 {
439 if (!x)
440 return -1;
441
442 return cds_lfht_fls_ulong(x - 1);
443 }
444
445 static
446 void cds_lfht_resize_lazy_grow(struct cds_lfht *ht, unsigned long size, int growth);
447
448 static
449 void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
450 unsigned long count);
451
452 static long nr_cpus_mask = -1;
453 static long split_count_mask = -1;
454
455 #if defined(HAVE_SYSCONF)
456 static void ht_init_nr_cpus_mask(void)
457 {
458 long maxcpus;
459
460 maxcpus = sysconf(_SC_NPROCESSORS_CONF);
461 if (maxcpus <= 0) {
462 nr_cpus_mask = -2;
463 return;
464 }
465 /*
466 * round up number of CPUs to next power of two, so we
467 * can use & for modulo.
468 */
469 maxcpus = 1UL << cds_lfht_get_count_order_ulong(maxcpus);
470 nr_cpus_mask = maxcpus - 1;
471 }
472 #else /* #if defined(HAVE_SYSCONF) */
473 static void ht_init_nr_cpus_mask(void)
474 {
475 nr_cpus_mask = -2;
476 }
477 #endif /* #else #if defined(HAVE_SYSCONF) */
478
479 static
480 void alloc_split_items_count(struct cds_lfht *ht)
481 {
482 struct ht_items_count *count;
483
484 if (nr_cpus_mask == -1) {
485 ht_init_nr_cpus_mask();
486 if (nr_cpus_mask < 0)
487 split_count_mask = DEFAULT_SPLIT_COUNT_MASK;
488 else
489 split_count_mask = nr_cpus_mask;
490 }
491
492 assert(split_count_mask >= 0);
493
494 if (ht->flags & CDS_LFHT_ACCOUNTING) {
495 ht->split_count = calloc(split_count_mask + 1, sizeof(*count));
496 assert(ht->split_count);
497 } else {
498 ht->split_count = NULL;
499 }
500 }
501
502 static
503 void free_split_items_count(struct cds_lfht *ht)
504 {
505 poison_free(ht->split_count);
506 }
507
508 #if defined(HAVE_SCHED_GETCPU)
509 static
510 int ht_get_split_count_index(unsigned long hash)
511 {
512 int cpu;
513
514 assert(split_count_mask >= 0);
515 cpu = sched_getcpu();
516 if (caa_unlikely(cpu < 0))
517 return hash & split_count_mask;
518 else
519 return cpu & split_count_mask;
520 }
521 #else /* #if defined(HAVE_SCHED_GETCPU) */
522 static
523 int ht_get_split_count_index(unsigned long hash)
524 {
525 return hash & split_count_mask;
526 }
527 #endif /* #else #if defined(HAVE_SCHED_GETCPU) */
528
529 static
530 void ht_count_add(struct cds_lfht *ht, unsigned long size, unsigned long hash)
531 {
532 unsigned long split_count;
533 int index;
534 long count;
535
536 if (caa_unlikely(!ht->split_count))
537 return;
538 index = ht_get_split_count_index(hash);
539 split_count = uatomic_add_return(&ht->split_count[index].add, 1);
540 if (caa_likely(split_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))
541 return;
542 /* Only if number of add multiple of 1UL << COUNT_COMMIT_ORDER */
543
544 dbg_printf("add split count %lu\n", split_count);
545 count = uatomic_add_return(&ht->count,
546 1UL << COUNT_COMMIT_ORDER);
547 if (caa_likely(count & (count - 1)))
548 return;
549 /* Only if global count is power of 2 */
550
551 if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) < size)
552 return;
553 dbg_printf("add set global %ld\n", count);
554 cds_lfht_resize_lazy_count(ht, size,
555 count >> (CHAIN_LEN_TARGET - 1));
556 }
557
558 static
559 void ht_count_del(struct cds_lfht *ht, unsigned long size, unsigned long hash)
560 {
561 unsigned long split_count;
562 int index;
563 long count;
564
565 if (caa_unlikely(!ht->split_count))
566 return;
567 index = ht_get_split_count_index(hash);
568 split_count = uatomic_add_return(&ht->split_count[index].del, 1);
569 if (caa_likely(split_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))
570 return;
571 /* Only if number of deletes multiple of 1UL << COUNT_COMMIT_ORDER */
572
573 dbg_printf("del split count %lu\n", split_count);
574 count = uatomic_add_return(&ht->count,
575 -(1UL << COUNT_COMMIT_ORDER));
576 if (caa_likely(count & (count - 1)))
577 return;
578 /* Only if global count is power of 2 */
579
580 if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) >= size)
581 return;
582 dbg_printf("del set global %ld\n", count);
583 /*
584 * Don't shrink table if the number of nodes is below a
585 * certain threshold.
586 */
587 if (count < (1UL << COUNT_COMMIT_ORDER) * (split_count_mask + 1))
588 return;
589 cds_lfht_resize_lazy_count(ht, size,
590 count >> (CHAIN_LEN_TARGET - 1));
591 }
592
593 static
594 void check_resize(struct cds_lfht *ht, unsigned long size, uint32_t chain_len)
595 {
596 unsigned long count;
597
598 if (!(ht->flags & CDS_LFHT_AUTO_RESIZE))
599 return;
600 count = uatomic_read(&ht->count);
601 /*
602 * Use bucket-local length for small table expand and for
603 * environments lacking per-cpu data support.
604 */
605 if (count >= (1UL << COUNT_COMMIT_ORDER))
606 return;
607 if (chain_len > 100)
608 dbg_printf("WARNING: large chain length: %u.\n",
609 chain_len);
610 if (chain_len >= CHAIN_LEN_RESIZE_THRESHOLD)
611 cds_lfht_resize_lazy_grow(ht, size,
612 cds_lfht_get_count_order_u32(chain_len - (CHAIN_LEN_TARGET - 1)));
613 }
614
615 static
616 struct cds_lfht_node *clear_flag(struct cds_lfht_node *node)
617 {
618 return (struct cds_lfht_node *) (((unsigned long) node) & ~FLAGS_MASK);
619 }
620
621 static
622 int is_removed(struct cds_lfht_node *node)
623 {
624 return ((unsigned long) node) & REMOVED_FLAG;
625 }
626
627 static
628 struct cds_lfht_node *flag_removed(struct cds_lfht_node *node)
629 {
630 return (struct cds_lfht_node *) (((unsigned long) node) | REMOVED_FLAG);
631 }
632
633 static
634 int is_bucket(struct cds_lfht_node *node)
635 {
636 return ((unsigned long) node) & BUCKET_FLAG;
637 }
638
639 static
640 struct cds_lfht_node *flag_bucket(struct cds_lfht_node *node)
641 {
642 return (struct cds_lfht_node *) (((unsigned long) node) | BUCKET_FLAG);
643 }
644
645 static
646 int is_removal_owner(struct cds_lfht_node *node)
647 {
648 return ((unsigned long) node) & REMOVAL_OWNER_FLAG;
649 }
650
651 static
652 struct cds_lfht_node *flag_removal_owner(struct cds_lfht_node *node)
653 {
654 return (struct cds_lfht_node *) (((unsigned long) node) | REMOVAL_OWNER_FLAG);
655 }
656
657 static
658 struct cds_lfht_node *get_end(void)
659 {
660 return (struct cds_lfht_node *) END_VALUE;
661 }
662
663 static
664 int is_end(struct cds_lfht_node *node)
665 {
666 return clear_flag(node) == (struct cds_lfht_node *) END_VALUE;
667 }
668
669 static
670 unsigned long _uatomic_xchg_monotonic_increase(unsigned long *ptr,
671 unsigned long v)
672 {
673 unsigned long old1, old2;
674
675 old1 = uatomic_read(ptr);
676 do {
677 old2 = old1;
678 if (old2 >= v)
679 return old2;
680 } while ((old1 = uatomic_cmpxchg(ptr, old2, v)) != old2);
681 return old2;
682 }
683
684 static
685 void cds_lfht_alloc_bucket_table(struct cds_lfht *ht, unsigned long order)
686 {
687 return ht->mm->alloc_bucket_table(ht, order);
688 }
689
690 /*
691 * cds_lfht_free_bucket_table() should be called with decreasing order.
692 * When cds_lfht_free_bucket_table(0) is called, it means the whole
693 * lfht is destroyed.
694 */
695 static
696 void cds_lfht_free_bucket_table(struct cds_lfht *ht, unsigned long order)
697 {
698 return ht->mm->free_bucket_table(ht, order);
699 }
700
701 static inline
702 struct cds_lfht_node *bucket_at(struct cds_lfht *ht, unsigned long index)
703 {
704 return ht->bucket_at(ht, index);
705 }
706
707 static inline
708 struct cds_lfht_node *lookup_bucket(struct cds_lfht *ht, unsigned long size,
709 unsigned long hash)
710 {
711 assert(size > 0);
712 return bucket_at(ht, hash & (size - 1));
713 }
714
715 /*
716 * Remove all logically deleted nodes from a bucket up to a certain node key.
717 */
718 static
719 void _cds_lfht_gc_bucket(struct cds_lfht_node *bucket, struct cds_lfht_node *node)
720 {
721 struct cds_lfht_node *iter_prev, *iter, *next, *new_next;
722
723 assert(!is_bucket(bucket));
724 assert(!is_removed(bucket));
725 assert(!is_bucket(node));
726 assert(!is_removed(node));
727 for (;;) {
728 iter_prev = bucket;
729 /* We can always skip the bucket node initially */
730 iter = rcu_dereference(iter_prev->next);
731 assert(!is_removed(iter));
732 assert(iter_prev->reverse_hash <= node->reverse_hash);
733 /*
734 * We should never be called with bucket (start of chain)
735 * and logically removed node (end of path compression
736 * marker) being the actual same node. This would be a
737 * bug in the algorithm implementation.
738 */
739 assert(bucket != node);
740 for (;;) {
741 if (caa_unlikely(is_end(iter)))
742 return;
743 if (caa_likely(clear_flag(iter)->reverse_hash > node->reverse_hash))
744 return;
745 next = rcu_dereference(clear_flag(iter)->next);
746 if (caa_likely(is_removed(next)))
747 break;
748 iter_prev = clear_flag(iter);
749 iter = next;
750 }
751 assert(!is_removed(iter));
752 if (is_bucket(iter))
753 new_next = flag_bucket(clear_flag(next));
754 else
755 new_next = clear_flag(next);
756 (void) uatomic_cmpxchg(&iter_prev->next, iter, new_next);
757 }
758 }
759
760 static
761 int _cds_lfht_replace(struct cds_lfht *ht, unsigned long size,
762 struct cds_lfht_node *old_node,
763 struct cds_lfht_node *old_next,
764 struct cds_lfht_node *new_node)
765 {
766 struct cds_lfht_node *bucket, *ret_next;
767
768 if (!old_node) /* Return -ENOENT if asked to replace NULL node */
769 return -ENOENT;
770
771 assert(!is_removed(old_node));
772 assert(!is_bucket(old_node));
773 assert(!is_removed(new_node));
774 assert(!is_bucket(new_node));
775 assert(new_node != old_node);
776 for (;;) {
777 /* Insert after node to be replaced */
778 if (is_removed(old_next)) {
779 /*
780 * Too late, the old node has been removed under us
781 * between lookup and replace. Fail.
782 */
783 return -ENOENT;
784 }
785 assert(old_next == clear_flag(old_next));
786 assert(new_node != old_next);
787 new_node->next = old_next;
788 /*
789 * Here is the whole trick for lock-free replace: we add
790 * the replacement node _after_ the node we want to
791 * replace by atomically setting its next pointer at the
792 * same time we set its removal flag. Given that
793 * the lookups/get next use an iterator aware of the
794 * next pointer, they will either skip the old node due
795 * to the removal flag and see the new node, or use
796 * the old node, but will not see the new one.
797 * This is a replacement of a node with another node
798 * that has the same value: we are therefore not
799 * removing a value from the hash table.
800 */
801 ret_next = uatomic_cmpxchg(&old_node->next,
802 old_next, flag_removed(new_node));
803 if (ret_next == old_next)
804 break; /* We performed the replacement. */
805 old_next = ret_next;
806 }
807
808 /*
809 * Ensure that the old node is not visible to readers anymore:
810 * lookup for the node, and remove it (along with any other
811 * logically removed node) if found.
812 */
813 bucket = lookup_bucket(ht, size, bit_reverse_ulong(old_node->reverse_hash));
814 _cds_lfht_gc_bucket(bucket, new_node);
815
816 assert(is_removed(rcu_dereference(old_node->next)));
817 return 0;
818 }
819
820 /*
821 * A non-NULL unique_ret pointer uses the "add unique" (or uniquify) add
822 * mode. A NULL unique_ret allows creation of duplicate keys.
823 */
824 static
825 void _cds_lfht_add(struct cds_lfht *ht,
826 unsigned long hash,
827 cds_lfht_match_fct match,
828 const void *key,
829 unsigned long size,
830 struct cds_lfht_node *node,
831 struct cds_lfht_iter *unique_ret,
832 int bucket_flag)
833 {
834 struct cds_lfht_node *iter_prev, *iter, *next, *new_node, *new_next,
835 *return_node;
836 struct cds_lfht_node *bucket;
837
838 assert(!is_bucket(node));
839 assert(!is_removed(node));
840 bucket = lookup_bucket(ht, size, hash);
841 for (;;) {
842 uint32_t chain_len = 0;
843
844 /*
845 * iter_prev points to the non-removed node prior to the
846 * insert location.
847 */
848 iter_prev = bucket;
849 /* We can always skip the bucket node initially */
850 iter = rcu_dereference(iter_prev->next);
851 assert(iter_prev->reverse_hash <= node->reverse_hash);
852 for (;;) {
853 if (caa_unlikely(is_end(iter)))
854 goto insert;
855 if (caa_likely(clear_flag(iter)->reverse_hash > node->reverse_hash))
856 goto insert;
857
858 /* bucket node is the first node of the identical-hash-value chain */
859 if (bucket_flag && clear_flag(iter)->reverse_hash == node->reverse_hash)
860 goto insert;
861
862 next = rcu_dereference(clear_flag(iter)->next);
863 if (caa_unlikely(is_removed(next)))
864 goto gc_node;
865
866 /* uniquely add */
867 if (unique_ret
868 && !is_bucket(next)
869 && clear_flag(iter)->reverse_hash == node->reverse_hash) {
870 struct cds_lfht_iter d_iter = { .node = node, .next = iter, };
871
872 /*
873 * uniquely adding inserts the node as the first
874 * node of the identical-hash-value node chain.
875 *
876 * This semantic ensures no duplicated keys
877 * should ever be observable in the table
878 * (including observe one node by one node
879 * by forward iterations)
880 */
881 cds_lfht_next_duplicate(ht, match, key, &d_iter);
882 if (!d_iter.node)
883 goto insert;
884
885 *unique_ret = d_iter;
886 return;
887 }
888
889 /* Only account for identical reverse hash once */
890 if (iter_prev->reverse_hash != clear_flag(iter)->reverse_hash
891 && !is_bucket(next))
892 check_resize(ht, size, ++chain_len);
893 iter_prev = clear_flag(iter);
894 iter = next;
895 }
896
897 insert:
898 assert(node != clear_flag(iter));
899 assert(!is_removed(iter_prev));
900 assert(!is_removed(iter));
901 assert(iter_prev != node);
902 if (!bucket_flag)
903 node->next = clear_flag(iter);
904 else
905 node->next = flag_bucket(clear_flag(iter));
906 if (is_bucket(iter))
907 new_node = flag_bucket(node);
908 else
909 new_node = node;
910 if (uatomic_cmpxchg(&iter_prev->next, iter,
911 new_node) != iter) {
912 continue; /* retry */
913 } else {
914 return_node = node;
915 goto end;
916 }
917
918 gc_node:
919 assert(!is_removed(iter));
920 if (is_bucket(iter))
921 new_next = flag_bucket(clear_flag(next));
922 else
923 new_next = clear_flag(next);
924 (void) uatomic_cmpxchg(&iter_prev->next, iter, new_next);
925 /* retry */
926 }
927 end:
928 if (unique_ret) {
929 unique_ret->node = return_node;
930 /* unique_ret->next left unset, never used. */
931 }
932 }
933
934 static
935 int _cds_lfht_del(struct cds_lfht *ht, unsigned long size,
936 struct cds_lfht_node *node)
937 {
938 struct cds_lfht_node *bucket, *next;
939
940 if (!node) /* Return -ENOENT if asked to delete NULL node */
941 return -ENOENT;
942
943 /* logically delete the node */
944 assert(!is_bucket(node));
945 assert(!is_removed(node));
946 assert(!is_removal_owner(node));
947
948 /*
949 * We are first checking if the node had previously been
950 * logically removed (this check is not atomic with setting the
951 * logical removal flag). Return -ENOENT if the node had
952 * previously been removed.
953 */
954 next = rcu_dereference(node->next);
955 if (caa_unlikely(is_removed(next)))
956 return -ENOENT;
957 assert(!is_bucket(next));
958 /*
959 * We set the REMOVED_FLAG unconditionally. Note that there may
960 * be more than one concurrent thread setting this flag.
961 * Knowing which wins the race will be known after the garbage
962 * collection phase, stay tuned!
963 */
964 uatomic_or(&node->next, REMOVED_FLAG);
965 /* We performed the (logical) deletion. */
966
967 /*
968 * Ensure that the node is not visible to readers anymore: lookup for
969 * the node, and remove it (along with any other logically removed node)
970 * if found.
971 */
972 bucket = lookup_bucket(ht, size, bit_reverse_ulong(node->reverse_hash));
973 _cds_lfht_gc_bucket(bucket, node);
974
975 assert(is_removed(rcu_dereference(node->next)));
976 /*
977 * Last phase: atomically exchange node->next with a version
978 * having "REMOVAL_OWNER_FLAG" set. If the returned node->next
979 * pointer did _not_ have "REMOVAL_OWNER_FLAG" set, we now own
980 * the node and win the removal race.
981 * It is interesting to note that all "add" paths are forbidden
982 * to change the next pointer starting from the point where the
983 * REMOVED_FLAG is set, so here using a read, followed by a
984 * xchg() suffice to guarantee that the xchg() will ever only
985 * set the "REMOVAL_OWNER_FLAG" (or change nothing if the flag
986 * was already set).
987 */
988 if (!is_removal_owner(uatomic_xchg(&node->next,
989 flag_removal_owner(node->next))))
990 return 0;
991 else
992 return -ENOENT;
993 }
994
995 static
996 void *partition_resize_thread(void *arg)
997 {
998 struct partition_resize_work *work = arg;
999
1000 work->ht->flavor->register_thread();
1001 work->fct(work->ht, work->i, work->start, work->len);
1002 work->ht->flavor->unregister_thread();
1003 return NULL;
1004 }
1005
1006 static
1007 void partition_resize_helper(struct cds_lfht *ht, unsigned long i,
1008 unsigned long len,
1009 void (*fct)(struct cds_lfht *ht, unsigned long i,
1010 unsigned long start, unsigned long len))
1011 {
1012 unsigned long partition_len;
1013 struct partition_resize_work *work;
1014 int thread, ret;
1015 unsigned long nr_threads;
1016
1017 /*
1018 * Note: nr_cpus_mask + 1 is always power of 2.
1019 * We spawn just the number of threads we need to satisfy the minimum
1020 * partition size, up to the number of CPUs in the system.
1021 */
1022 if (nr_cpus_mask > 0) {
1023 nr_threads = min(nr_cpus_mask + 1,
1024 len >> MIN_PARTITION_PER_THREAD_ORDER);
1025 } else {
1026 nr_threads = 1;
1027 }
1028 partition_len = len >> cds_lfht_get_count_order_ulong(nr_threads);
1029 work = calloc(nr_threads, sizeof(*work));
1030 assert(work);
1031 for (thread = 0; thread < nr_threads; thread++) {
1032 work[thread].ht = ht;
1033 work[thread].i = i;
1034 work[thread].len = partition_len;
1035 work[thread].start = thread * partition_len;
1036 work[thread].fct = fct;
1037 ret = pthread_create(&(work[thread].thread_id), ht->resize_attr,
1038 partition_resize_thread, &work[thread]);
1039 assert(!ret);
1040 }
1041 for (thread = 0; thread < nr_threads; thread++) {
1042 ret = pthread_join(work[thread].thread_id, NULL);
1043 assert(!ret);
1044 }
1045 free(work);
1046 }
1047
1048 /*
1049 * Holding RCU read lock to protect _cds_lfht_add against memory
1050 * reclaim that could be performed by other call_rcu worker threads (ABA
1051 * problem).
1052 *
1053 * When we reach a certain length, we can split this population phase over
1054 * many worker threads, based on the number of CPUs available in the system.
1055 * This should therefore take care of not having the expand lagging behind too
1056 * many concurrent insertion threads by using the scheduler's ability to
1057 * schedule bucket node population fairly with insertions.
1058 */
1059 static
1060 void init_table_populate_partition(struct cds_lfht *ht, unsigned long i,
1061 unsigned long start, unsigned long len)
1062 {
1063 unsigned long j, size = 1UL << (i - 1);
1064
1065 assert(i > MIN_TABLE_ORDER);
1066 ht->flavor->read_lock();
1067 for (j = size + start; j < size + start + len; j++) {
1068 struct cds_lfht_node *new_node = bucket_at(ht, j);
1069
1070 assert(j >= size && j < (size << 1));
1071 dbg_printf("init populate: order %lu index %lu hash %lu\n",
1072 i, j, j);
1073 new_node->reverse_hash = bit_reverse_ulong(j);
1074 _cds_lfht_add(ht, j, NULL, NULL, size, new_node, NULL, 1);
1075 }
1076 ht->flavor->read_unlock();
1077 }
1078
1079 static
1080 void init_table_populate(struct cds_lfht *ht, unsigned long i,
1081 unsigned long len)
1082 {
1083 assert(nr_cpus_mask != -1);
1084 if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD) {
1085 ht->flavor->thread_online();
1086 init_table_populate_partition(ht, i, 0, len);
1087 ht->flavor->thread_offline();
1088 return;
1089 }
1090 partition_resize_helper(ht, i, len, init_table_populate_partition);
1091 }
1092
1093 static
1094 void init_table(struct cds_lfht *ht,
1095 unsigned long first_order, unsigned long last_order)
1096 {
1097 unsigned long i;
1098
1099 dbg_printf("init table: first_order %lu last_order %lu\n",
1100 first_order, last_order);
1101 assert(first_order > MIN_TABLE_ORDER);
1102 for (i = first_order; i <= last_order; i++) {
1103 unsigned long len;
1104
1105 len = 1UL << (i - 1);
1106 dbg_printf("init order %lu len: %lu\n", i, len);
1107
1108 /* Stop expand if the resize target changes under us */
1109 if (CMM_LOAD_SHARED(ht->resize_target) < (1UL << i))
1110 break;
1111
1112 cds_lfht_alloc_bucket_table(ht, i);
1113
1114 /*
1115 * Set all bucket nodes reverse hash values for a level and
1116 * link all bucket nodes into the table.
1117 */
1118 init_table_populate(ht, i, len);
1119
1120 /*
1121 * Update table size.
1122 */
1123 cmm_smp_wmb(); /* populate data before RCU size */
1124 CMM_STORE_SHARED(ht->size, 1UL << i);
1125
1126 dbg_printf("init new size: %lu\n", 1UL << i);
1127 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1128 break;
1129 }
1130 }
1131
1132 /*
1133 * Holding RCU read lock to protect _cds_lfht_remove against memory
1134 * reclaim that could be performed by other call_rcu worker threads (ABA
1135 * problem).
1136 * For a single level, we logically remove and garbage collect each node.
1137 *
1138 * As a design choice, we perform logical removal and garbage collection on a
1139 * node-per-node basis to simplify this algorithm. We also assume keeping good
1140 * cache locality of the operation would overweight possible performance gain
1141 * that could be achieved by batching garbage collection for multiple levels.
1142 * However, this would have to be justified by benchmarks.
1143 *
1144 * Concurrent removal and add operations are helping us perform garbage
1145 * collection of logically removed nodes. We guarantee that all logically
1146 * removed nodes have been garbage-collected (unlinked) before call_rcu is
1147 * invoked to free a hole level of bucket nodes (after a grace period).
1148 *
1149 * Logical removal and garbage collection can therefore be done in batch or on a
1150 * node-per-node basis, as long as the guarantee above holds.
1151 *
1152 * When we reach a certain length, we can split this removal over many worker
1153 * threads, based on the number of CPUs available in the system. This should
1154 * take care of not letting resize process lag behind too many concurrent
1155 * updater threads actively inserting into the hash table.
1156 */
1157 static
1158 void remove_table_partition(struct cds_lfht *ht, unsigned long i,
1159 unsigned long start, unsigned long len)
1160 {
1161 unsigned long j, size = 1UL << (i - 1);
1162
1163 assert(i > MIN_TABLE_ORDER);
1164 ht->flavor->read_lock();
1165 for (j = size + start; j < size + start + len; j++) {
1166 struct cds_lfht_node *fini_bucket = bucket_at(ht, j);
1167 struct cds_lfht_node *parent_bucket = bucket_at(ht, j - size);
1168
1169 assert(j >= size && j < (size << 1));
1170 dbg_printf("remove entry: order %lu index %lu hash %lu\n",
1171 i, j, j);
1172 /* Set the REMOVED_FLAG to freeze the ->next for gc */
1173 uatomic_or(&fini_bucket->next, REMOVED_FLAG);
1174 _cds_lfht_gc_bucket(parent_bucket, fini_bucket);
1175 }
1176 ht->flavor->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->flavor->thread_online();
1186 remove_table_partition(ht, i, 0, len);
1187 ht->flavor->thread_offline();
1188 return;
1189 }
1190 partition_resize_helper(ht, i, len, remove_table_partition);
1191 }
1192
1193 /*
1194 * fini_table() is never called for first_order == 0, which is why
1195 * free_by_rcu_order == 0 can be used as criterion to know if free must
1196 * be called.
1197 */
1198 static
1199 void fini_table(struct cds_lfht *ht,
1200 unsigned long first_order, unsigned long last_order)
1201 {
1202 long i;
1203 unsigned long free_by_rcu_order = 0;
1204
1205 dbg_printf("fini table: first_order %lu last_order %lu\n",
1206 first_order, last_order);
1207 assert(first_order > MIN_TABLE_ORDER);
1208 for (i = last_order; i >= first_order; i--) {
1209 unsigned long len;
1210
1211 len = 1UL << (i - 1);
1212 dbg_printf("fini order %lu len: %lu\n", i, len);
1213
1214 /* Stop shrink if the resize target changes under us */
1215 if (CMM_LOAD_SHARED(ht->resize_target) > (1UL << (i - 1)))
1216 break;
1217
1218 cmm_smp_wmb(); /* populate data before RCU size */
1219 CMM_STORE_SHARED(ht->size, 1UL << (i - 1));
1220
1221 /*
1222 * We need to wait for all add operations to reach Q.S. (and
1223 * thus use the new table for lookups) before we can start
1224 * releasing the old bucket nodes. Otherwise their lookup will
1225 * return a logically removed node as insert position.
1226 */
1227 ht->flavor->update_synchronize_rcu();
1228 if (free_by_rcu_order)
1229 cds_lfht_free_bucket_table(ht, free_by_rcu_order);
1230
1231 /*
1232 * Set "removed" flag in bucket nodes about to be removed.
1233 * Unlink all now-logically-removed bucket node pointers.
1234 * Concurrent add/remove operation are helping us doing
1235 * the gc.
1236 */
1237 remove_table(ht, i, len);
1238
1239 free_by_rcu_order = i;
1240
1241 dbg_printf("fini new size: %lu\n", 1UL << i);
1242 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1243 break;
1244 }
1245
1246 if (free_by_rcu_order) {
1247 ht->flavor->update_synchronize_rcu();
1248 cds_lfht_free_bucket_table(ht, free_by_rcu_order);
1249 }
1250 }
1251
1252 static
1253 void cds_lfht_create_bucket(struct cds_lfht *ht, unsigned long size)
1254 {
1255 struct cds_lfht_node *prev, *node;
1256 unsigned long order, len, i;
1257
1258 cds_lfht_alloc_bucket_table(ht, 0);
1259
1260 dbg_printf("create bucket: order 0 index 0 hash 0\n");
1261 node = bucket_at(ht, 0);
1262 node->next = flag_bucket(get_end());
1263 node->reverse_hash = 0;
1264
1265 for (order = 1; order < cds_lfht_get_count_order_ulong(size) + 1; order++) {
1266 len = 1UL << (order - 1);
1267 cds_lfht_alloc_bucket_table(ht, order);
1268
1269 for (i = 0; i < len; i++) {
1270 /*
1271 * Now, we are trying to init the node with the
1272 * hash=(len+i) (which is also a bucket with the
1273 * index=(len+i)) and insert it into the hash table,
1274 * so this node has to be inserted after the bucket
1275 * with the index=(len+i)&(len-1)=i. And because there
1276 * is no other non-bucket node nor bucket node with
1277 * larger index/hash inserted, so the bucket node
1278 * being inserted should be inserted directly linked
1279 * after the bucket node with index=i.
1280 */
1281 prev = bucket_at(ht, i);
1282 node = bucket_at(ht, len + i);
1283
1284 dbg_printf("create bucket: order %lu index %lu hash %lu\n",
1285 order, len + i, len + i);
1286 node->reverse_hash = bit_reverse_ulong(len + i);
1287
1288 /* insert after prev */
1289 assert(is_bucket(prev->next));
1290 node->next = prev->next;
1291 prev->next = flag_bucket(node);
1292 }
1293 }
1294 }
1295
1296 struct cds_lfht *_cds_lfht_new(unsigned long init_size,
1297 unsigned long min_nr_alloc_buckets,
1298 unsigned long max_nr_buckets,
1299 int flags,
1300 const struct cds_lfht_mm_type *mm,
1301 const struct rcu_flavor_struct *flavor,
1302 pthread_attr_t *attr)
1303 {
1304 struct cds_lfht *ht;
1305 unsigned long order;
1306
1307 /* min_nr_alloc_buckets must be power of two */
1308 if (!min_nr_alloc_buckets || (min_nr_alloc_buckets & (min_nr_alloc_buckets - 1)))
1309 return NULL;
1310
1311 /* init_size must be power of two */
1312 if (!init_size || (init_size & (init_size - 1)))
1313 return NULL;
1314
1315 /*
1316 * Memory management plugin default.
1317 */
1318 if (!mm) {
1319 if (CAA_BITS_PER_LONG > 32
1320 && max_nr_buckets
1321 && max_nr_buckets <= (1ULL << 32)) {
1322 /*
1323 * For 64-bit architectures, with max number of
1324 * buckets small enough not to use the entire
1325 * 64-bit memory mapping space (and allowing a
1326 * fair number of hash table instances), use the
1327 * mmap allocator, which is faster than the
1328 * order allocator.
1329 */
1330 mm = &cds_lfht_mm_mmap;
1331 } else {
1332 /*
1333 * The fallback is to use the order allocator.
1334 */
1335 mm = &cds_lfht_mm_order;
1336 }
1337 }
1338
1339 /* max_nr_buckets == 0 for order based mm means infinite */
1340 if (mm == &cds_lfht_mm_order && !max_nr_buckets)
1341 max_nr_buckets = 1UL << (MAX_TABLE_ORDER - 1);
1342
1343 /* max_nr_buckets must be power of two */
1344 if (!max_nr_buckets || (max_nr_buckets & (max_nr_buckets - 1)))
1345 return NULL;
1346
1347 min_nr_alloc_buckets = max(min_nr_alloc_buckets, MIN_TABLE_SIZE);
1348 init_size = max(init_size, MIN_TABLE_SIZE);
1349 max_nr_buckets = max(max_nr_buckets, min_nr_alloc_buckets);
1350 init_size = min(init_size, max_nr_buckets);
1351
1352 ht = mm->alloc_cds_lfht(min_nr_alloc_buckets, max_nr_buckets);
1353 assert(ht);
1354 assert(ht->mm == mm);
1355 assert(ht->bucket_at == mm->bucket_at);
1356
1357 ht->flags = flags;
1358 ht->flavor = flavor;
1359 ht->resize_attr = attr;
1360 alloc_split_items_count(ht);
1361 /* this mutex should not nest in read-side C.S. */
1362 pthread_mutex_init(&ht->resize_mutex, NULL);
1363 order = cds_lfht_get_count_order_ulong(init_size);
1364 ht->resize_target = 1UL << order;
1365 cds_lfht_create_bucket(ht, 1UL << order);
1366 ht->size = 1UL << order;
1367 return ht;
1368 }
1369
1370 void cds_lfht_lookup(struct cds_lfht *ht, unsigned long hash,
1371 cds_lfht_match_fct match, const void *key,
1372 struct cds_lfht_iter *iter)
1373 {
1374 struct cds_lfht_node *node, *next, *bucket;
1375 unsigned long reverse_hash, size;
1376
1377 reverse_hash = bit_reverse_ulong(hash);
1378
1379 size = rcu_dereference(ht->size);
1380 bucket = lookup_bucket(ht, size, hash);
1381 /* We can always skip the bucket node initially */
1382 node = rcu_dereference(bucket->next);
1383 node = clear_flag(node);
1384 for (;;) {
1385 if (caa_unlikely(is_end(node))) {
1386 node = next = NULL;
1387 break;
1388 }
1389 if (caa_unlikely(node->reverse_hash > reverse_hash)) {
1390 node = next = NULL;
1391 break;
1392 }
1393 next = rcu_dereference(node->next);
1394 assert(node == clear_flag(node));
1395 if (caa_likely(!is_removed(next))
1396 && !is_bucket(next)
1397 && node->reverse_hash == reverse_hash
1398 && caa_likely(match(node, key))) {
1399 break;
1400 }
1401 node = clear_flag(next);
1402 }
1403 assert(!node || !is_bucket(rcu_dereference(node->next)));
1404 iter->node = node;
1405 iter->next = next;
1406 }
1407
1408 void cds_lfht_next_duplicate(struct cds_lfht *ht, cds_lfht_match_fct match,
1409 const void *key, struct cds_lfht_iter *iter)
1410 {
1411 struct cds_lfht_node *node, *next;
1412 unsigned long reverse_hash;
1413
1414 node = iter->node;
1415 reverse_hash = node->reverse_hash;
1416 next = iter->next;
1417 node = clear_flag(next);
1418
1419 for (;;) {
1420 if (caa_unlikely(is_end(node))) {
1421 node = next = NULL;
1422 break;
1423 }
1424 if (caa_unlikely(node->reverse_hash > reverse_hash)) {
1425 node = next = NULL;
1426 break;
1427 }
1428 next = rcu_dereference(node->next);
1429 if (caa_likely(!is_removed(next))
1430 && !is_bucket(next)
1431 && caa_likely(match(node, key))) {
1432 break;
1433 }
1434 node = clear_flag(next);
1435 }
1436 assert(!node || !is_bucket(rcu_dereference(node->next)));
1437 iter->node = node;
1438 iter->next = next;
1439 }
1440
1441 void cds_lfht_next(struct cds_lfht *ht, struct cds_lfht_iter *iter)
1442 {
1443 struct cds_lfht_node *node, *next;
1444
1445 node = clear_flag(iter->next);
1446 for (;;) {
1447 if (caa_unlikely(is_end(node))) {
1448 node = next = NULL;
1449 break;
1450 }
1451 next = rcu_dereference(node->next);
1452 if (caa_likely(!is_removed(next))
1453 && !is_bucket(next)) {
1454 break;
1455 }
1456 node = clear_flag(next);
1457 }
1458 assert(!node || !is_bucket(rcu_dereference(node->next)));
1459 iter->node = node;
1460 iter->next = next;
1461 }
1462
1463 void cds_lfht_first(struct cds_lfht *ht, struct cds_lfht_iter *iter)
1464 {
1465 /*
1466 * Get next after first bucket node. The first bucket node is the
1467 * first node of the linked list.
1468 */
1469 iter->next = bucket_at(ht, 0)->next;
1470 cds_lfht_next(ht, iter);
1471 }
1472
1473 void cds_lfht_add(struct cds_lfht *ht, unsigned long hash,
1474 struct cds_lfht_node *node)
1475 {
1476 unsigned long size;
1477
1478 node->reverse_hash = bit_reverse_ulong(hash);
1479 size = rcu_dereference(ht->size);
1480 _cds_lfht_add(ht, hash, NULL, NULL, size, node, NULL, 0);
1481 ht_count_add(ht, size, hash);
1482 }
1483
1484 struct cds_lfht_node *cds_lfht_add_unique(struct cds_lfht *ht,
1485 unsigned long hash,
1486 cds_lfht_match_fct match,
1487 const void *key,
1488 struct cds_lfht_node *node)
1489 {
1490 unsigned long size;
1491 struct cds_lfht_iter iter;
1492
1493 node->reverse_hash = bit_reverse_ulong(hash);
1494 size = rcu_dereference(ht->size);
1495 _cds_lfht_add(ht, hash, match, key, size, node, &iter, 0);
1496 if (iter.node == node)
1497 ht_count_add(ht, size, hash);
1498 return iter.node;
1499 }
1500
1501 struct cds_lfht_node *cds_lfht_add_replace(struct cds_lfht *ht,
1502 unsigned long hash,
1503 cds_lfht_match_fct match,
1504 const void *key,
1505 struct cds_lfht_node *node)
1506 {
1507 unsigned long size;
1508 struct cds_lfht_iter iter;
1509
1510 node->reverse_hash = bit_reverse_ulong(hash);
1511 size = rcu_dereference(ht->size);
1512 for (;;) {
1513 _cds_lfht_add(ht, hash, match, key, size, node, &iter, 0);
1514 if (iter.node == node) {
1515 ht_count_add(ht, size, hash);
1516 return NULL;
1517 }
1518
1519 if (!_cds_lfht_replace(ht, size, iter.node, iter.next, node))
1520 return iter.node;
1521 }
1522 }
1523
1524 int cds_lfht_replace(struct cds_lfht *ht,
1525 struct cds_lfht_iter *old_iter,
1526 unsigned long hash,
1527 cds_lfht_match_fct match,
1528 const void *key,
1529 struct cds_lfht_node *new_node)
1530 {
1531 unsigned long size;
1532
1533 new_node->reverse_hash = bit_reverse_ulong(hash);
1534 if (!old_iter->node)
1535 return -ENOENT;
1536 if (caa_unlikely(old_iter->node->reverse_hash != new_node->reverse_hash))
1537 return -EINVAL;
1538 if (caa_unlikely(!match(old_iter->node, key)))
1539 return -EINVAL;
1540 size = rcu_dereference(ht->size);
1541 return _cds_lfht_replace(ht, size, old_iter->node, old_iter->next,
1542 new_node);
1543 }
1544
1545 int cds_lfht_del(struct cds_lfht *ht, struct cds_lfht_node *node)
1546 {
1547 unsigned long size, hash;
1548 int ret;
1549
1550 size = rcu_dereference(ht->size);
1551 ret = _cds_lfht_del(ht, size, node);
1552 if (!ret) {
1553 hash = bit_reverse_ulong(node->reverse_hash);
1554 ht_count_del(ht, size, hash);
1555 }
1556 return ret;
1557 }
1558
1559 static
1560 int cds_lfht_delete_bucket(struct cds_lfht *ht)
1561 {
1562 struct cds_lfht_node *node;
1563 unsigned long order, i, size;
1564
1565 /* Check that the table is empty */
1566 node = bucket_at(ht, 0);
1567 do {
1568 node = clear_flag(node)->next;
1569 if (!is_bucket(node))
1570 return -EPERM;
1571 assert(!is_removed(node));
1572 } while (!is_end(node));
1573 /*
1574 * size accessed without rcu_dereference because hash table is
1575 * being destroyed.
1576 */
1577 size = ht->size;
1578 /* Internal sanity check: all nodes left should be bucket */
1579 for (i = 0; i < size; i++) {
1580 node = bucket_at(ht, i);
1581 dbg_printf("delete bucket: index %lu expected hash %lu hash %lu\n",
1582 i, i, bit_reverse_ulong(node->reverse_hash));
1583 assert(is_bucket(node->next));
1584 }
1585
1586 for (order = cds_lfht_get_count_order_ulong(size); (long)order >= 0; order--)
1587 cds_lfht_free_bucket_table(ht, order);
1588
1589 return 0;
1590 }
1591
1592 /*
1593 * Should only be called when no more concurrent readers nor writers can
1594 * possibly access the table.
1595 */
1596 int cds_lfht_destroy(struct cds_lfht *ht, pthread_attr_t **attr)
1597 {
1598 int ret;
1599
1600 /* Wait for in-flight resize operations to complete */
1601 _CMM_STORE_SHARED(ht->in_progress_destroy, 1);
1602 cmm_smp_mb(); /* Store destroy before load resize */
1603 while (uatomic_read(&ht->in_progress_resize))
1604 poll(NULL, 0, 100); /* wait for 100ms */
1605 ret = cds_lfht_delete_bucket(ht);
1606 if (ret)
1607 return ret;
1608 free_split_items_count(ht);
1609 if (attr)
1610 *attr = ht->resize_attr;
1611 poison_free(ht);
1612 return ret;
1613 }
1614
1615 void cds_lfht_count_nodes(struct cds_lfht *ht,
1616 long *approx_before,
1617 unsigned long *count,
1618 long *approx_after)
1619 {
1620 struct cds_lfht_node *node, *next;
1621 unsigned long nr_bucket = 0, nr_removed = 0;
1622
1623 *approx_before = 0;
1624 if (ht->split_count) {
1625 int i;
1626
1627 for (i = 0; i < split_count_mask + 1; i++) {
1628 *approx_before += uatomic_read(&ht->split_count[i].add);
1629 *approx_before -= uatomic_read(&ht->split_count[i].del);
1630 }
1631 }
1632
1633 *count = 0;
1634
1635 /* Count non-bucket nodes in the table */
1636 node = bucket_at(ht, 0);
1637 do {
1638 next = rcu_dereference(node->next);
1639 if (is_removed(next)) {
1640 if (!is_bucket(next))
1641 (nr_removed)++;
1642 else
1643 (nr_bucket)++;
1644 } else if (!is_bucket(next))
1645 (*count)++;
1646 else
1647 (nr_bucket)++;
1648 node = clear_flag(next);
1649 } while (!is_end(node));
1650 dbg_printf("number of logically removed nodes: %lu\n", nr_removed);
1651 dbg_printf("number of bucket nodes: %lu\n", nr_bucket);
1652 *approx_after = 0;
1653 if (ht->split_count) {
1654 int i;
1655
1656 for (i = 0; i < split_count_mask + 1; i++) {
1657 *approx_after += uatomic_read(&ht->split_count[i].add);
1658 *approx_after -= uatomic_read(&ht->split_count[i].del);
1659 }
1660 }
1661 }
1662
1663 /* called with resize mutex held */
1664 static
1665 void _do_cds_lfht_grow(struct cds_lfht *ht,
1666 unsigned long old_size, unsigned long new_size)
1667 {
1668 unsigned long old_order, new_order;
1669
1670 old_order = cds_lfht_get_count_order_ulong(old_size);
1671 new_order = cds_lfht_get_count_order_ulong(new_size);
1672 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1673 old_size, old_order, new_size, new_order);
1674 assert(new_size > old_size);
1675 init_table(ht, old_order + 1, new_order);
1676 }
1677
1678 /* called with resize mutex held */
1679 static
1680 void _do_cds_lfht_shrink(struct cds_lfht *ht,
1681 unsigned long old_size, unsigned long new_size)
1682 {
1683 unsigned long old_order, new_order;
1684
1685 new_size = max(new_size, MIN_TABLE_SIZE);
1686 old_order = cds_lfht_get_count_order_ulong(old_size);
1687 new_order = cds_lfht_get_count_order_ulong(new_size);
1688 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1689 old_size, old_order, new_size, new_order);
1690 assert(new_size < old_size);
1691
1692 /* Remove and unlink all bucket nodes to remove. */
1693 fini_table(ht, new_order + 1, old_order);
1694 }
1695
1696
1697 /* called with resize mutex held */
1698 static
1699 void _do_cds_lfht_resize(struct cds_lfht *ht)
1700 {
1701 unsigned long new_size, old_size;
1702
1703 /*
1704 * Resize table, re-do if the target size has changed under us.
1705 */
1706 do {
1707 assert(uatomic_read(&ht->in_progress_resize));
1708 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1709 break;
1710 ht->resize_initiated = 1;
1711 old_size = ht->size;
1712 new_size = CMM_LOAD_SHARED(ht->resize_target);
1713 if (old_size < new_size)
1714 _do_cds_lfht_grow(ht, old_size, new_size);
1715 else if (old_size > new_size)
1716 _do_cds_lfht_shrink(ht, old_size, new_size);
1717 ht->resize_initiated = 0;
1718 /* write resize_initiated before read resize_target */
1719 cmm_smp_mb();
1720 } while (ht->size != CMM_LOAD_SHARED(ht->resize_target));
1721 }
1722
1723 static
1724 unsigned long resize_target_grow(struct cds_lfht *ht, unsigned long new_size)
1725 {
1726 return _uatomic_xchg_monotonic_increase(&ht->resize_target, new_size);
1727 }
1728
1729 static
1730 void resize_target_update_count(struct cds_lfht *ht,
1731 unsigned long count)
1732 {
1733 count = max(count, MIN_TABLE_SIZE);
1734 count = min(count, ht->max_nr_buckets);
1735 uatomic_set(&ht->resize_target, count);
1736 }
1737
1738 void cds_lfht_resize(struct cds_lfht *ht, unsigned long new_size)
1739 {
1740 resize_target_update_count(ht, new_size);
1741 CMM_STORE_SHARED(ht->resize_initiated, 1);
1742 ht->flavor->thread_offline();
1743 pthread_mutex_lock(&ht->resize_mutex);
1744 _do_cds_lfht_resize(ht);
1745 pthread_mutex_unlock(&ht->resize_mutex);
1746 ht->flavor->thread_online();
1747 }
1748
1749 static
1750 void do_resize_cb(struct rcu_head *head)
1751 {
1752 struct rcu_resize_work *work =
1753 caa_container_of(head, struct rcu_resize_work, head);
1754 struct cds_lfht *ht = work->ht;
1755
1756 ht->flavor->thread_offline();
1757 pthread_mutex_lock(&ht->resize_mutex);
1758 _do_cds_lfht_resize(ht);
1759 pthread_mutex_unlock(&ht->resize_mutex);
1760 ht->flavor->thread_online();
1761 poison_free(work);
1762 cmm_smp_mb(); /* finish resize before decrement */
1763 uatomic_dec(&ht->in_progress_resize);
1764 }
1765
1766 static
1767 void __cds_lfht_resize_lazy_launch(struct cds_lfht *ht)
1768 {
1769 struct rcu_resize_work *work;
1770
1771 /* Store resize_target before read resize_initiated */
1772 cmm_smp_mb();
1773 if (!CMM_LOAD_SHARED(ht->resize_initiated)) {
1774 uatomic_inc(&ht->in_progress_resize);
1775 cmm_smp_mb(); /* increment resize count before load destroy */
1776 if (CMM_LOAD_SHARED(ht->in_progress_destroy)) {
1777 uatomic_dec(&ht->in_progress_resize);
1778 return;
1779 }
1780 work = malloc(sizeof(*work));
1781 work->ht = ht;
1782 ht->flavor->update_call_rcu(&work->head, do_resize_cb);
1783 CMM_STORE_SHARED(ht->resize_initiated, 1);
1784 }
1785 }
1786
1787 static
1788 void cds_lfht_resize_lazy_grow(struct cds_lfht *ht, unsigned long size, int growth)
1789 {
1790 unsigned long target_size = size << growth;
1791
1792 target_size = min(target_size, ht->max_nr_buckets);
1793 if (resize_target_grow(ht, target_size) >= target_size)
1794 return;
1795
1796 __cds_lfht_resize_lazy_launch(ht);
1797 }
1798
1799 /*
1800 * We favor grow operations over shrink. A shrink operation never occurs
1801 * if a grow operation is queued for lazy execution. A grow operation
1802 * cancels any pending shrink lazy execution.
1803 */
1804 static
1805 void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
1806 unsigned long count)
1807 {
1808 if (!(ht->flags & CDS_LFHT_AUTO_RESIZE))
1809 return;
1810 count = max(count, MIN_TABLE_SIZE);
1811 count = min(count, ht->max_nr_buckets);
1812 if (count == size)
1813 return; /* Already the right size, no resize needed */
1814 if (count > size) { /* lazy grow */
1815 if (resize_target_grow(ht, count) >= count)
1816 return;
1817 } else { /* lazy shrink */
1818 for (;;) {
1819 unsigned long s;
1820
1821 s = uatomic_cmpxchg(&ht->resize_target, size, count);
1822 if (s == size)
1823 break; /* no resize needed */
1824 if (s > size)
1825 return; /* growing is/(was just) in progress */
1826 if (s <= count)
1827 return; /* some other thread do shrink */
1828 size = s;
1829 }
1830 }
1831 __cds_lfht_resize_lazy_launch(ht);
1832 }
This page took 0.116785 seconds and 4 git commands to generate.