9baf4079877393a1289fdcca8a518286e530cddb
[lttng-tools.git] / src / common / 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 * 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, as well as traversals, and use the returned objects
38 * safely by allowing memory reclaim to take place only after a grace
39 * period.
40 * - Add and remove operations are lock-free, and do not need to
41 * allocate memory. They need to be executed within RCU read-side
42 * critical section to ensure the objects they read are valid and to
43 * deal with the cmpxchg ABA problem.
44 * - add and add_unique operations are supported. add_unique checks if
45 * the node key already exists in the hash table. It ensures not to
46 * populate a duplicate key if the node key already exists in the hash
47 * table.
48 * - The resize operation executes concurrently with
49 * add/add_unique/add_replace/remove/lookup/traversal.
50 * - Hash table nodes are contained within a split-ordered list. This
51 * list is ordered by incrementing reversed-bits-hash value.
52 * - An index of bucket nodes is kept. These bucket nodes are the hash
53 * table "buckets". These buckets are internal nodes that allow to
54 * perform a fast hash lookup, similarly to a skip list. These
55 * buckets are chained together in the split-ordered list, which
56 * allows recursive expansion by inserting new buckets between the
57 * existing buckets. The split-ordered list allows adding new buckets
58 * between existing buckets as the table needs to grow.
59 * - The resize operation for small tables only allows expanding the
60 * hash table. It is triggered automatically by detecting long chains
61 * in the add operation.
62 * - The resize operation for larger tables (and available through an
63 * API) allows both expanding and shrinking the hash table.
64 * - Split-counters are used to keep track of the number of
65 * nodes within the hash table for automatic resize triggering.
66 * - Resize operation initiated by long chain detection is executed by a
67 * call_rcu thread, which keeps lock-freedom of add and remove.
68 * - Resize operations are protected by a mutex.
69 * - The removal operation is split in two parts: first, a "removed"
70 * flag is set in the next pointer within the node to remove. Then,
71 * a "garbage collection" is performed in the bucket containing the
72 * removed node (from the start of the bucket up to the removed node).
73 * All encountered nodes with "removed" flag set in their next
74 * pointers are removed from the linked-list. If the cmpxchg used for
75 * removal fails (due to concurrent garbage-collection or concurrent
76 * add), we retry from the beginning of the bucket. This ensures that
77 * the node with "removed" flag set is removed from the hash table
78 * (not visible to lookups anymore) before the RCU read-side critical
79 * section held across removal ends. Furthermore, this ensures that
80 * the node with "removed" flag set is removed from the linked-list
81 * before its memory is reclaimed. After setting the "removal" flag,
82 * only the thread which removal is the first to set the "removal
83 * owner" flag (with an xchg) into a node's next pointer is considered
84 * to have succeeded its removal (and thus owns the node to reclaim).
85 * Because we garbage-collect starting from an invariant node (the
86 * start-of-bucket bucket node) up to the "removed" node (or find a
87 * reverse-hash that is higher), we are sure that a successful
88 * traversal of the chain leads to a chain that is present in the
89 * linked-list (the start node is never removed) and that it does not
90 * contain the "removed" node anymore, even if concurrent delete/add
91 * operations are changing the structure of the list concurrently.
92 * - The add operations perform garbage collection of buckets if they
93 * encounter nodes with removed flag set in the bucket where they want
94 * to add their new node. This ensures lock-freedom of add operation by
95 * helping the remover unlink nodes from the list rather than to wait
96 * for it do to so.
97 * - There are three memory backends for the hash table buckets: the
98 * "order table", the "chunks", and the "mmap".
99 * - These bucket containers contain a compact version of the hash table
100 * nodes.
101 * - The RCU "order table":
102 * - has a first level table indexed by log2(hash index) which is
103 * copied and expanded by the resize operation. This order table
104 * allows finding the "bucket node" tables.
105 * - There is one bucket node table per hash index order. The size of
106 * each bucket node table is half the number of hashes contained in
107 * this order (except for order 0).
108 * - The RCU "chunks" is best suited for close interaction with a page
109 * allocator. It uses a linear array as index to "chunks" containing
110 * each the same number of buckets.
111 * - The RCU "mmap" memory backend uses a single memory map to hold
112 * all buckets.
113 * - synchronize_rcu is used to garbage-collect the old bucket node table.
114 *
115 * Ordering Guarantees:
116 *
117 * To discuss these guarantees, we first define "read" operation as any
118 * of the the basic cds_lfht_lookup, cds_lfht_next_duplicate,
119 * cds_lfht_first, cds_lfht_next operation, as well as
120 * cds_lfht_add_unique (failure).
121 *
122 * We define "read traversal" operation as any of the following
123 * group of operations
124 * - cds_lfht_lookup followed by iteration with cds_lfht_next_duplicate
125 * (and/or cds_lfht_next, although less common).
126 * - cds_lfht_add_unique (failure) followed by iteration with
127 * cds_lfht_next_duplicate (and/or cds_lfht_next, although less
128 * common).
129 * - cds_lfht_first followed iteration with cds_lfht_next (and/or
130 * cds_lfht_next_duplicate, although less common).
131 *
132 * We define "write" operations as any of cds_lfht_add,
133 * cds_lfht_add_unique (success), cds_lfht_add_replace, cds_lfht_del.
134 *
135 * When cds_lfht_add_unique succeeds (returns the node passed as
136 * parameter), it acts as a "write" operation. When cds_lfht_add_unique
137 * fails (returns a node different from the one passed as parameter), it
138 * acts as a "read" operation. A cds_lfht_add_unique failure is a
139 * cds_lfht_lookup "read" operation, therefore, any ordering guarantee
140 * referring to "lookup" imply any of "lookup" or cds_lfht_add_unique
141 * (failure).
142 *
143 * We define "prior" and "later" node as nodes observable by reads and
144 * read traversals respectively before and after a write or sequence of
145 * write operations.
146 *
147 * Hash-table operations are often cascaded, for example, the pointer
148 * returned by a cds_lfht_lookup() might be passed to a cds_lfht_next(),
149 * whose return value might in turn be passed to another hash-table
150 * operation. This entire cascaded series of operations must be enclosed
151 * by a pair of matching rcu_read_lock() and rcu_read_unlock()
152 * operations.
153 *
154 * The following ordering guarantees are offered by this hash table:
155 *
156 * A.1) "read" after "write": if there is ordering between a write and a
157 * later read, then the read is guaranteed to see the write or some
158 * later write.
159 * A.2) "read traversal" after "write": given that there is dependency
160 * ordering between reads in a "read traversal", if there is
161 * ordering between a write and the first read of the traversal,
162 * then the "read traversal" is guaranteed to see the write or
163 * some later write.
164 * B.1) "write" after "read": if there is ordering between a read and a
165 * later write, then the read will never see the write.
166 * B.2) "write" after "read traversal": given that there is dependency
167 * ordering between reads in a "read traversal", if there is
168 * ordering between the last read of the traversal and a later
169 * write, then the "read traversal" will never see the write.
170 * C) "write" while "read traversal": if a write occurs during a "read
171 * traversal", the traversal may, or may not, see the write.
172 * D.1) "write" after "write": if there is ordering between a write and
173 * a later write, then the later write is guaranteed to see the
174 * effects of the first write.
175 * D.2) Concurrent "write" pairs: The system will assign an arbitrary
176 * order to any pair of concurrent conflicting writes.
177 * Non-conflicting writes (for example, to different keys) are
178 * unordered.
179 * E) If a grace period separates a "del" or "replace" operation
180 * and a subsequent operation, then that subsequent operation is
181 * guaranteed not to see the removed item.
182 * F) Uniqueness guarantee: given a hash table that does not contain
183 * duplicate items for a given key, there will only be one item in
184 * the hash table after an arbitrary sequence of add_unique and/or
185 * add_replace operations. Note, however, that a pair of
186 * concurrent read operations might well access two different items
187 * with that key.
188 * G.1) If a pair of lookups for a given key are ordered (e.g. by a
189 * memory barrier), then the second lookup will return the same
190 * node as the previous lookup, or some later node.
191 * G.2) A "read traversal" that starts after the end of a prior "read
192 * traversal" (ordered by memory barriers) is guaranteed to see the
193 * same nodes as the previous traversal, or some later nodes.
194 * G.3) Concurrent "read" pairs: concurrent reads are unordered. For
195 * example, if a pair of reads to the same key run concurrently
196 * with an insertion of that same key, the reads remain unordered
197 * regardless of their return values. In other words, you cannot
198 * rely on the values returned by the reads to deduce ordering.
199 *
200 * Progress guarantees:
201 *
202 * * Reads are wait-free. These operations always move forward in the
203 * hash table linked list, and this list has no loop.
204 * * Writes are lock-free. Any retry loop performed by a write operation
205 * is triggered by progress made within another update operation.
206 *
207 * Bucket node tables:
208 *
209 * hash table hash table the last all bucket node tables
210 * order size bucket node 0 1 2 3 4 5 6(index)
211 * table size
212 * 0 1 1 1
213 * 1 2 1 1 1
214 * 2 4 2 1 1 2
215 * 3 8 4 1 1 2 4
216 * 4 16 8 1 1 2 4 8
217 * 5 32 16 1 1 2 4 8 16
218 * 6 64 32 1 1 2 4 8 16 32
219 *
220 * When growing/shrinking, we only focus on the last bucket node table
221 * which size is (!order ? 1 : (1 << (order -1))).
222 *
223 * Example for growing/shrinking:
224 * grow hash table from order 5 to 6: init the index=6 bucket node table
225 * shrink hash table from order 6 to 5: fini the index=6 bucket node table
226 *
227 * A bit of ascii art explanation:
228 *
229 * The order index is the off-by-one compared to the actual power of 2
230 * because we use index 0 to deal with the 0 special-case.
231 *
232 * This shows the nodes for a small table ordered by reversed bits:
233 *
234 * bits reverse
235 * 0 000 000
236 * 4 100 001
237 * 2 010 010
238 * 6 110 011
239 * 1 001 100
240 * 5 101 101
241 * 3 011 110
242 * 7 111 111
243 *
244 * This shows the nodes in order of non-reversed bits, linked by
245 * reversed-bit order.
246 *
247 * order bits reverse
248 * 0 0 000 000
249 * 1 | 1 001 100 <-
250 * 2 | | 2 010 010 <- |
251 * | | | 3 011 110 | <- |
252 * 3 -> | | | 4 100 001 | |
253 * -> | | 5 101 101 |
254 * -> | 6 110 011
255 * -> 7 111 111
256 */
257
258 #define _LGPL_SOURCE
259 #define _GNU_SOURCE
260 #include <stdlib.h>
261 #include <errno.h>
262 #include <assert.h>
263 #include <stdio.h>
264 #include <stdint.h>
265 #include <string.h>
266 #include <sched.h>
267
268 #include "config.h"
269 #include <urcu.h>
270 #include <urcu-call-rcu.h>
271 #include <urcu/arch.h>
272 #include <urcu/uatomic.h>
273 #include <urcu/compiler.h>
274 #include <stdio.h>
275 #include <pthread.h>
276
277 #include "rculfhash.h"
278 #include "rculfhash-internal.h"
279 #include "urcu-flavor.h"
280
281 #include <common/common.h>
282
283 /*
284 * Split-counters lazily update the global counter each 1024
285 * addition/removal. It automatically keeps track of resize required.
286 * We use the bucket length as indicator for need to expand for small
287 * tables and machines lacking per-cpu data suppport.
288 */
289 #define COUNT_COMMIT_ORDER 10
290 #define DEFAULT_SPLIT_COUNT_MASK 0xFUL
291 #define CHAIN_LEN_TARGET 1
292 #define CHAIN_LEN_RESIZE_THRESHOLD 3
293
294 /*
295 * Define the minimum table size.
296 */
297 #define MIN_TABLE_ORDER 0
298 #define MIN_TABLE_SIZE (1UL << MIN_TABLE_ORDER)
299
300 /*
301 * Minimum number of bucket nodes to touch per thread to parallelize grow/shrink.
302 */
303 #define MIN_PARTITION_PER_THREAD_ORDER 12
304 #define MIN_PARTITION_PER_THREAD (1UL << MIN_PARTITION_PER_THREAD_ORDER)
305
306 /*
307 * The removed flag needs to be updated atomically with the pointer.
308 * It indicates that no node must attach to the node scheduled for
309 * removal, and that node garbage collection must be performed.
310 * The bucket flag does not require to be updated atomically with the
311 * pointer, but it is added as a pointer low bit flag to save space.
312 * The "removal owner" flag is used to detect which of the "del"
313 * operation that has set the "removed flag" gets to return the removed
314 * node to its caller. Note that the replace operation does not need to
315 * iteract with the "removal owner" flag, because it validates that
316 * the "removed" flag is not set before performing its cmpxchg.
317 */
318 #define REMOVED_FLAG (1UL << 0)
319 #define BUCKET_FLAG (1UL << 1)
320 #define REMOVAL_OWNER_FLAG (1UL << 2)
321 #define FLAGS_MASK ((1UL << 3) - 1)
322
323 /* Value of the end pointer. Should not interact with flags. */
324 #define END_VALUE NULL
325
326 /*
327 * ht_items_count: Split-counters counting the number of node addition
328 * and removal in the table. Only used if the CDS_LFHT_ACCOUNTING flag
329 * is set at hash table creation.
330 *
331 * These are free-running counters, never reset to zero. They count the
332 * number of add/remove, and trigger every (1 << COUNT_COMMIT_ORDER)
333 * operations to update the global counter. We choose a power-of-2 value
334 * for the trigger to deal with 32 or 64-bit overflow of the counter.
335 */
336 struct ht_items_count {
337 unsigned long add, del;
338 } __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
339
340 /*
341 * rcu_resize_work: Contains arguments passed to RCU worker thread
342 * responsible for performing lazy resize.
343 */
344 struct rcu_resize_work {
345 struct rcu_head head;
346 struct cds_lfht *ht;
347 };
348
349 /*
350 * partition_resize_work: Contains arguments passed to worker threads
351 * executing the hash table resize on partitions of the hash table
352 * assigned to each processor's worker thread.
353 */
354 struct partition_resize_work {
355 pthread_t thread_id;
356 struct cds_lfht *ht;
357 unsigned long i, start, len;
358 void (*fct)(struct cds_lfht *ht, unsigned long i,
359 unsigned long start, unsigned long len);
360 };
361
362 /*
363 * Algorithm to reverse bits in a word by lookup table, extended to
364 * 64-bit words.
365 * Source:
366 * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable
367 * Originally from Public Domain.
368 */
369
370 static const uint8_t BitReverseTable256[256] =
371 {
372 #define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64
373 #define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16)
374 #define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 )
375 R6(0), R6(2), R6(1), R6(3)
376 };
377 #undef R2
378 #undef R4
379 #undef R6
380
381 static
382 uint8_t bit_reverse_u8(uint8_t v)
383 {
384 return BitReverseTable256[v];
385 }
386
387 static __attribute__((unused))
388 uint32_t bit_reverse_u32(uint32_t v)
389 {
390 return ((uint32_t) bit_reverse_u8(v) << 24) |
391 ((uint32_t) bit_reverse_u8(v >> 8) << 16) |
392 ((uint32_t) bit_reverse_u8(v >> 16) << 8) |
393 ((uint32_t) bit_reverse_u8(v >> 24));
394 }
395
396 static __attribute__((unused))
397 uint64_t bit_reverse_u64(uint64_t v)
398 {
399 return ((uint64_t) bit_reverse_u8(v) << 56) |
400 ((uint64_t) bit_reverse_u8(v >> 8) << 48) |
401 ((uint64_t) bit_reverse_u8(v >> 16) << 40) |
402 ((uint64_t) bit_reverse_u8(v >> 24) << 32) |
403 ((uint64_t) bit_reverse_u8(v >> 32) << 24) |
404 ((uint64_t) bit_reverse_u8(v >> 40) << 16) |
405 ((uint64_t) bit_reverse_u8(v >> 48) << 8) |
406 ((uint64_t) bit_reverse_u8(v >> 56));
407 }
408
409 static
410 unsigned long bit_reverse_ulong(unsigned long v)
411 {
412 #if (CAA_BITS_PER_LONG == 32)
413 return bit_reverse_u32(v);
414 #else
415 return bit_reverse_u64(v);
416 #endif
417 }
418
419 /*
420 * fls: returns the position of the most significant bit.
421 * Returns 0 if no bit is set, else returns the position of the most
422 * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit).
423 */
424 #if defined(__i386) || defined(__x86_64)
425 static inline
426 unsigned int fls_u32(uint32_t x)
427 {
428 int r;
429
430 asm("bsrl %1,%0\n\t"
431 "jnz 1f\n\t"
432 "movl $-1,%0\n\t"
433 "1:\n\t"
434 : "=r" (r) : "rm" (x));
435 return r + 1;
436 }
437 #define HAS_FLS_U32
438 #endif
439
440 #if defined(__x86_64)
441 static inline
442 unsigned int fls_u64(uint64_t x)
443 {
444 long r;
445
446 asm("bsrq %1,%0\n\t"
447 "jnz 1f\n\t"
448 "movq $-1,%0\n\t"
449 "1:\n\t"
450 : "=r" (r) : "rm" (x));
451 return r + 1;
452 }
453 #define HAS_FLS_U64
454 #endif
455
456 #ifndef HAS_FLS_U64
457 static __attribute__((unused))
458 unsigned int fls_u64(uint64_t x)
459 {
460 unsigned int r = 64;
461
462 if (!x)
463 return 0;
464
465 if (!(x & 0xFFFFFFFF00000000ULL)) {
466 x <<= 32;
467 r -= 32;
468 }
469 if (!(x & 0xFFFF000000000000ULL)) {
470 x <<= 16;
471 r -= 16;
472 }
473 if (!(x & 0xFF00000000000000ULL)) {
474 x <<= 8;
475 r -= 8;
476 }
477 if (!(x & 0xF000000000000000ULL)) {
478 x <<= 4;
479 r -= 4;
480 }
481 if (!(x & 0xC000000000000000ULL)) {
482 x <<= 2;
483 r -= 2;
484 }
485 if (!(x & 0x8000000000000000ULL)) {
486 x <<= 1;
487 r -= 1;
488 }
489 return r;
490 }
491 #endif
492
493 #ifndef HAS_FLS_U32
494 static __attribute__((unused))
495 unsigned int fls_u32(uint32_t x)
496 {
497 unsigned int r = 32;
498
499 if (!x)
500 return 0;
501 if (!(x & 0xFFFF0000U)) {
502 x <<= 16;
503 r -= 16;
504 }
505 if (!(x & 0xFF000000U)) {
506 x <<= 8;
507 r -= 8;
508 }
509 if (!(x & 0xF0000000U)) {
510 x <<= 4;
511 r -= 4;
512 }
513 if (!(x & 0xC0000000U)) {
514 x <<= 2;
515 r -= 2;
516 }
517 if (!(x & 0x80000000U)) {
518 x <<= 1;
519 r -= 1;
520 }
521 return r;
522 }
523 #endif
524
525 unsigned int cds_lfht_fls_ulong(unsigned long x)
526 {
527 #if (CAA_BITS_PER_LONG == 32)
528 return fls_u32(x);
529 #else
530 return fls_u64(x);
531 #endif
532 }
533
534 /*
535 * Return the minimum order for which x <= (1UL << order).
536 * Return -1 if x is 0.
537 */
538 int cds_lfht_get_count_order_u32(uint32_t x)
539 {
540 if (!x)
541 return -1;
542
543 return fls_u32(x - 1);
544 }
545
546 /*
547 * Return the minimum order for which x <= (1UL << order).
548 * Return -1 if x is 0.
549 */
550 int cds_lfht_get_count_order_ulong(unsigned long x)
551 {
552 if (!x)
553 return -1;
554
555 return cds_lfht_fls_ulong(x - 1);
556 }
557
558 static
559 void cds_lfht_resize_lazy_grow(struct cds_lfht *ht, unsigned long size, int growth);
560
561 static
562 void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
563 unsigned long count);
564
565 static long nr_cpus_mask = -1;
566 static long split_count_mask = -1;
567 static int split_count_order = -1;
568
569 #if defined(HAVE_SYSCONF)
570 static void ht_init_nr_cpus_mask(void)
571 {
572 long maxcpus;
573
574 maxcpus = sysconf(_SC_NPROCESSORS_CONF);
575 if (maxcpus <= 0) {
576 nr_cpus_mask = -2;
577 return;
578 }
579 /*
580 * round up number of CPUs to next power of two, so we
581 * can use & for modulo.
582 */
583 maxcpus = 1UL << cds_lfht_get_count_order_ulong(maxcpus);
584 nr_cpus_mask = maxcpus - 1;
585 }
586 #else /* #if defined(HAVE_SYSCONF) */
587 static void ht_init_nr_cpus_mask(void)
588 {
589 nr_cpus_mask = -2;
590 }
591 #endif /* #else #if defined(HAVE_SYSCONF) */
592
593 static
594 void alloc_split_items_count(struct cds_lfht *ht)
595 {
596 struct ht_items_count *count;
597
598 if (nr_cpus_mask == -1) {
599 ht_init_nr_cpus_mask();
600 if (nr_cpus_mask < 0)
601 split_count_mask = DEFAULT_SPLIT_COUNT_MASK;
602 else
603 split_count_mask = nr_cpus_mask;
604 split_count_order =
605 cds_lfht_get_count_order_ulong(split_count_mask + 1);
606 }
607
608 assert(split_count_mask >= 0);
609
610 if (ht->flags & CDS_LFHT_ACCOUNTING) {
611 ht->split_count = calloc(split_count_mask + 1, sizeof(*count));
612 assert(ht->split_count);
613 } else {
614 ht->split_count = NULL;
615 }
616 }
617
618 static
619 void free_split_items_count(struct cds_lfht *ht)
620 {
621 poison_free(ht->split_count);
622 }
623
624 #if defined(HAVE_SCHED_GETCPU) && !defined(VALGRIND)
625 static
626 int ht_get_split_count_index(unsigned long hash)
627 {
628 int cpu;
629
630 assert(split_count_mask >= 0);
631 cpu = sched_getcpu();
632 if (caa_unlikely(cpu < 0))
633 return hash & split_count_mask;
634 else
635 return cpu & split_count_mask;
636 }
637 #else /* #if defined(HAVE_SCHED_GETCPU) */
638 static
639 int ht_get_split_count_index(unsigned long hash)
640 {
641 return hash & split_count_mask;
642 }
643 #endif /* #else #if defined(HAVE_SCHED_GETCPU) */
644
645 static
646 void ht_count_add(struct cds_lfht *ht, unsigned long size, unsigned long hash)
647 {
648 unsigned long split_count;
649 int index;
650 long count;
651
652 if (caa_unlikely(!ht->split_count))
653 return;
654 index = ht_get_split_count_index(hash);
655 split_count = uatomic_add_return(&ht->split_count[index].add, 1);
656 if (caa_likely(split_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))
657 return;
658 /* Only if number of add multiple of 1UL << COUNT_COMMIT_ORDER */
659
660 dbg_printf("add split count %lu\n", split_count);
661 count = uatomic_add_return(&ht->count,
662 1UL << COUNT_COMMIT_ORDER);
663 if (caa_likely(count & (count - 1)))
664 return;
665 /* Only if global count is power of 2 */
666
667 if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) < size)
668 return;
669 dbg_printf("add set global %ld\n", count);
670 cds_lfht_resize_lazy_count(ht, size,
671 count >> (CHAIN_LEN_TARGET - 1));
672 }
673
674 static
675 void ht_count_del(struct cds_lfht *ht, unsigned long size, unsigned long hash)
676 {
677 unsigned long split_count;
678 int index;
679 long count;
680
681 if (caa_unlikely(!ht->split_count))
682 return;
683 index = ht_get_split_count_index(hash);
684 split_count = uatomic_add_return(&ht->split_count[index].del, 1);
685 if (caa_likely(split_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))
686 return;
687 /* Only if number of deletes multiple of 1UL << COUNT_COMMIT_ORDER */
688
689 dbg_printf("del split count %lu\n", split_count);
690 count = uatomic_add_return(&ht->count,
691 -(1UL << COUNT_COMMIT_ORDER));
692 if (caa_likely(count & (count - 1)))
693 return;
694 /* Only if global count is power of 2 */
695
696 if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) >= size)
697 return;
698 dbg_printf("del set global %ld\n", count);
699 /*
700 * Don't shrink table if the number of nodes is below a
701 * certain threshold.
702 */
703 if (count < (1UL << COUNT_COMMIT_ORDER) * (split_count_mask + 1))
704 return;
705 cds_lfht_resize_lazy_count(ht, size,
706 count >> (CHAIN_LEN_TARGET - 1));
707 }
708
709 static
710 void check_resize(struct cds_lfht *ht, unsigned long size, uint32_t chain_len)
711 {
712 unsigned long count;
713
714 if (!(ht->flags & CDS_LFHT_AUTO_RESIZE))
715 return;
716 count = uatomic_read(&ht->count);
717 /*
718 * Use bucket-local length for small table expand and for
719 * environments lacking per-cpu data support.
720 */
721 if (count >= (1UL << (COUNT_COMMIT_ORDER + split_count_order)))
722 return;
723 if (chain_len > 100)
724 dbg_printf("WARNING: large chain length: %u.\n",
725 chain_len);
726 if (chain_len >= CHAIN_LEN_RESIZE_THRESHOLD) {
727 int growth;
728
729 /*
730 * Ideal growth calculated based on chain length.
731 */
732 growth = cds_lfht_get_count_order_u32(chain_len
733 - (CHAIN_LEN_TARGET - 1));
734 if ((ht->flags & CDS_LFHT_ACCOUNTING)
735 && (size << growth)
736 >= (1UL << (COUNT_COMMIT_ORDER
737 + split_count_order))) {
738 /*
739 * If ideal growth expands the hash table size
740 * beyond the "small hash table" sizes, use the
741 * maximum small hash table size to attempt
742 * expanding the hash table. This only applies
743 * when node accounting is available, otherwise
744 * the chain length is used to expand the hash
745 * table in every case.
746 */
747 growth = COUNT_COMMIT_ORDER + split_count_order
748 - cds_lfht_get_count_order_ulong(size);
749 if (growth <= 0)
750 return;
751 }
752 cds_lfht_resize_lazy_grow(ht, size, growth);
753 }
754 }
755
756 static
757 struct cds_lfht_node *clear_flag(struct cds_lfht_node *node)
758 {
759 return (struct cds_lfht_node *) (((unsigned long) node) & ~FLAGS_MASK);
760 }
761
762 static
763 int is_removed(struct cds_lfht_node *node)
764 {
765 return ((unsigned long) node) & REMOVED_FLAG;
766 }
767
768 static
769 int is_bucket(struct cds_lfht_node *node)
770 {
771 return ((unsigned long) node) & BUCKET_FLAG;
772 }
773
774 static
775 struct cds_lfht_node *flag_bucket(struct cds_lfht_node *node)
776 {
777 return (struct cds_lfht_node *) (((unsigned long) node) | BUCKET_FLAG);
778 }
779
780 static
781 int is_removal_owner(struct cds_lfht_node *node)
782 {
783 return ((unsigned long) node) & REMOVAL_OWNER_FLAG;
784 }
785
786 static
787 struct cds_lfht_node *flag_removal_owner(struct cds_lfht_node *node)
788 {
789 return (struct cds_lfht_node *) (((unsigned long) node) | REMOVAL_OWNER_FLAG);
790 }
791
792 static
793 struct cds_lfht_node *flag_removed_or_removal_owner(struct cds_lfht_node *node)
794 {
795 return (struct cds_lfht_node *) (((unsigned long) node) | REMOVED_FLAG | REMOVAL_OWNER_FLAG);
796 }
797
798 static
799 struct cds_lfht_node *get_end(void)
800 {
801 return (struct cds_lfht_node *) END_VALUE;
802 }
803
804 static
805 int is_end(struct cds_lfht_node *node)
806 {
807 return clear_flag(node) == (struct cds_lfht_node *) END_VALUE;
808 }
809
810 static
811 unsigned long _uatomic_xchg_monotonic_increase(unsigned long *ptr,
812 unsigned long v)
813 {
814 unsigned long old1, old2;
815
816 old1 = uatomic_read(ptr);
817 do {
818 old2 = old1;
819 if (old2 >= v)
820 return old2;
821 } while ((old1 = uatomic_cmpxchg(ptr, old2, v)) != old2);
822 return old2;
823 }
824
825 static
826 void cds_lfht_alloc_bucket_table(struct cds_lfht *ht, unsigned long order)
827 {
828 return ht->mm->alloc_bucket_table(ht, order);
829 }
830
831 /*
832 * cds_lfht_free_bucket_table() should be called with decreasing order.
833 * When cds_lfht_free_bucket_table(0) is called, it means the whole
834 * lfht is destroyed.
835 */
836 static
837 void cds_lfht_free_bucket_table(struct cds_lfht *ht, unsigned long order)
838 {
839 return ht->mm->free_bucket_table(ht, order);
840 }
841
842 static inline
843 struct cds_lfht_node *bucket_at(struct cds_lfht *ht, unsigned long index)
844 {
845 return ht->bucket_at(ht, index);
846 }
847
848 static inline
849 struct cds_lfht_node *lookup_bucket(struct cds_lfht *ht, unsigned long size,
850 unsigned long hash)
851 {
852 assert(size > 0);
853 return bucket_at(ht, hash & (size - 1));
854 }
855
856 /*
857 * Remove all logically deleted nodes from a bucket up to a certain node key.
858 */
859 static
860 void _cds_lfht_gc_bucket(struct cds_lfht_node *bucket, struct cds_lfht_node *node)
861 {
862 struct cds_lfht_node *iter_prev, *iter, *next, *new_next;
863
864 assert(!is_bucket(bucket));
865 assert(!is_removed(bucket));
866 assert(!is_bucket(node));
867 assert(!is_removed(node));
868 for (;;) {
869 iter_prev = bucket;
870 /* We can always skip the bucket node initially */
871 iter = rcu_dereference(iter_prev->next);
872 assert(!is_removed(iter));
873 assert(iter_prev->reverse_hash <= node->reverse_hash);
874 /*
875 * We should never be called with bucket (start of chain)
876 * and logically removed node (end of path compression
877 * marker) being the actual same node. This would be a
878 * bug in the algorithm implementation.
879 */
880 assert(bucket != node);
881 for (;;) {
882 if (caa_unlikely(is_end(iter)))
883 return;
884 if (caa_likely(clear_flag(iter)->reverse_hash > node->reverse_hash))
885 return;
886 next = rcu_dereference(clear_flag(iter)->next);
887 if (caa_likely(is_removed(next)))
888 break;
889 iter_prev = clear_flag(iter);
890 iter = next;
891 }
892 assert(!is_removed(iter));
893 if (is_bucket(iter))
894 new_next = flag_bucket(clear_flag(next));
895 else
896 new_next = clear_flag(next);
897 (void) uatomic_cmpxchg(&iter_prev->next, iter, new_next);
898 }
899 }
900
901 static
902 int _cds_lfht_replace(struct cds_lfht *ht, unsigned long size,
903 struct cds_lfht_node *old_node,
904 struct cds_lfht_node *old_next,
905 struct cds_lfht_node *new_node)
906 {
907 struct cds_lfht_node *bucket, *ret_next;
908
909 if (!old_node) /* Return -ENOENT if asked to replace NULL node */
910 return -ENOENT;
911
912 assert(!is_removed(old_node));
913 assert(!is_bucket(old_node));
914 assert(!is_removed(new_node));
915 assert(!is_bucket(new_node));
916 assert(new_node != old_node);
917 for (;;) {
918 /* Insert after node to be replaced */
919 if (is_removed(old_next)) {
920 /*
921 * Too late, the old node has been removed under us
922 * between lookup and replace. Fail.
923 */
924 return -ENOENT;
925 }
926 assert(old_next == clear_flag(old_next));
927 assert(new_node != old_next);
928 /*
929 * REMOVAL_OWNER flag is _NEVER_ set before the REMOVED
930 * flag. It is either set atomically at the same time
931 * (replace) or after (del).
932 */
933 assert(!is_removal_owner(old_next));
934 new_node->next = old_next;
935 /*
936 * Here is the whole trick for lock-free replace: we add
937 * the replacement node _after_ the node we want to
938 * replace by atomically setting its next pointer at the
939 * same time we set its removal flag. Given that
940 * the lookups/get next use an iterator aware of the
941 * next pointer, they will either skip the old node due
942 * to the removal flag and see the new node, or use
943 * the old node, but will not see the new one.
944 * This is a replacement of a node with another node
945 * that has the same value: we are therefore not
946 * removing a value from the hash table. We set both the
947 * REMOVED and REMOVAL_OWNER flags atomically so we own
948 * the node after successful cmpxchg.
949 */
950 ret_next = uatomic_cmpxchg(&old_node->next,
951 old_next, flag_removed_or_removal_owner(new_node));
952 if (ret_next == old_next)
953 break; /* We performed the replacement. */
954 old_next = ret_next;
955 }
956
957 /*
958 * Ensure that the old node is not visible to readers anymore:
959 * lookup for the node, and remove it (along with any other
960 * logically removed node) if found.
961 */
962 bucket = lookup_bucket(ht, size, bit_reverse_ulong(old_node->reverse_hash));
963 _cds_lfht_gc_bucket(bucket, new_node);
964
965 assert(is_removed(CMM_LOAD_SHARED(old_node->next)));
966 return 0;
967 }
968
969 /*
970 * A non-NULL unique_ret pointer uses the "add unique" (or uniquify) add
971 * mode. A NULL unique_ret allows creation of duplicate keys.
972 */
973 static
974 void _cds_lfht_add(struct cds_lfht *ht,
975 unsigned long hash,
976 cds_lfht_match_fct match,
977 const void *key,
978 unsigned long size,
979 struct cds_lfht_node *node,
980 struct cds_lfht_iter *unique_ret,
981 int bucket_flag)
982 {
983 struct cds_lfht_node *iter_prev, *iter, *next, *new_node, *new_next,
984 *return_node;
985 struct cds_lfht_node *bucket;
986
987 assert(!is_bucket(node));
988 assert(!is_removed(node));
989 bucket = lookup_bucket(ht, size, hash);
990 for (;;) {
991 uint32_t chain_len = 0;
992
993 /*
994 * iter_prev points to the non-removed node prior to the
995 * insert location.
996 */
997 iter_prev = bucket;
998 /* We can always skip the bucket node initially */
999 iter = rcu_dereference(iter_prev->next);
1000 assert(iter_prev->reverse_hash <= node->reverse_hash);
1001 for (;;) {
1002 if (caa_unlikely(is_end(iter)))
1003 goto insert;
1004 if (caa_likely(clear_flag(iter)->reverse_hash > node->reverse_hash))
1005 goto insert;
1006
1007 /* bucket node is the first node of the identical-hash-value chain */
1008 if (bucket_flag && clear_flag(iter)->reverse_hash == node->reverse_hash)
1009 goto insert;
1010
1011 next = rcu_dereference(clear_flag(iter)->next);
1012 if (caa_unlikely(is_removed(next)))
1013 goto gc_node;
1014
1015 /* uniquely add */
1016 if (unique_ret
1017 && !is_bucket(next)
1018 && clear_flag(iter)->reverse_hash == node->reverse_hash) {
1019 struct cds_lfht_iter d_iter = { .node = node, .next = iter, };
1020
1021 /*
1022 * uniquely adding inserts the node as the first
1023 * node of the identical-hash-value node chain.
1024 *
1025 * This semantic ensures no duplicated keys
1026 * should ever be observable in the table
1027 * (including traversing the table node by
1028 * node by forward iterations)
1029 */
1030 cds_lfht_next_duplicate(ht, match, key, &d_iter);
1031 if (!d_iter.node)
1032 goto insert;
1033
1034 *unique_ret = d_iter;
1035 return;
1036 }
1037
1038 /* Only account for identical reverse hash once */
1039 if (iter_prev->reverse_hash != clear_flag(iter)->reverse_hash
1040 && !is_bucket(next))
1041 check_resize(ht, size, ++chain_len);
1042 iter_prev = clear_flag(iter);
1043 iter = next;
1044 }
1045
1046 insert:
1047 assert(node != clear_flag(iter));
1048 assert(!is_removed(iter_prev));
1049 assert(!is_removed(iter));
1050 assert(iter_prev != node);
1051 if (!bucket_flag)
1052 node->next = clear_flag(iter);
1053 else
1054 node->next = flag_bucket(clear_flag(iter));
1055 if (is_bucket(iter))
1056 new_node = flag_bucket(node);
1057 else
1058 new_node = node;
1059 if (uatomic_cmpxchg(&iter_prev->next, iter,
1060 new_node) != iter) {
1061 continue; /* retry */
1062 } else {
1063 return_node = node;
1064 goto end;
1065 }
1066
1067 gc_node:
1068 assert(!is_removed(iter));
1069 if (is_bucket(iter))
1070 new_next = flag_bucket(clear_flag(next));
1071 else
1072 new_next = clear_flag(next);
1073 (void) uatomic_cmpxchg(&iter_prev->next, iter, new_next);
1074 /* retry */
1075 }
1076 end:
1077 if (unique_ret) {
1078 unique_ret->node = return_node;
1079 /* unique_ret->next left unset, never used. */
1080 }
1081 }
1082
1083 static
1084 int _cds_lfht_del(struct cds_lfht *ht, unsigned long size,
1085 struct cds_lfht_node *node)
1086 {
1087 struct cds_lfht_node *bucket, *next;
1088
1089 if (!node) /* Return -ENOENT if asked to delete NULL node */
1090 return -ENOENT;
1091
1092 /* logically delete the node */
1093 assert(!is_bucket(node));
1094 assert(!is_removed(node));
1095 assert(!is_removal_owner(node));
1096
1097 /*
1098 * We are first checking if the node had previously been
1099 * logically removed (this check is not atomic with setting the
1100 * logical removal flag). Return -ENOENT if the node had
1101 * previously been removed.
1102 */
1103 next = CMM_LOAD_SHARED(node->next); /* next is not dereferenced */
1104 if (caa_unlikely(is_removed(next)))
1105 return -ENOENT;
1106 assert(!is_bucket(next));
1107 /*
1108 * The del operation semantic guarantees a full memory barrier
1109 * before the uatomic_or atomic commit of the deletion flag.
1110 */
1111 cmm_smp_mb__before_uatomic_or();
1112 /*
1113 * We set the REMOVED_FLAG unconditionally. Note that there may
1114 * be more than one concurrent thread setting this flag.
1115 * Knowing which wins the race will be known after the garbage
1116 * collection phase, stay tuned!
1117 */
1118 uatomic_or(&node->next, REMOVED_FLAG);
1119 /* We performed the (logical) deletion. */
1120
1121 /*
1122 * Ensure that the node is not visible to readers anymore: lookup for
1123 * the node, and remove it (along with any other logically removed node)
1124 * if found.
1125 */
1126 bucket = lookup_bucket(ht, size, bit_reverse_ulong(node->reverse_hash));
1127 _cds_lfht_gc_bucket(bucket, node);
1128
1129 assert(is_removed(CMM_LOAD_SHARED(node->next)));
1130 /*
1131 * Last phase: atomically exchange node->next with a version
1132 * having "REMOVAL_OWNER_FLAG" set. If the returned node->next
1133 * pointer did _not_ have "REMOVAL_OWNER_FLAG" set, we now own
1134 * the node and win the removal race.
1135 * It is interesting to note that all "add" paths are forbidden
1136 * to change the next pointer starting from the point where the
1137 * REMOVED_FLAG is set, so here using a read, followed by a
1138 * xchg() suffice to guarantee that the xchg() will ever only
1139 * set the "REMOVAL_OWNER_FLAG" (or change nothing if the flag
1140 * was already set).
1141 */
1142 if (!is_removal_owner(uatomic_xchg(&node->next,
1143 flag_removal_owner(node->next))))
1144 return 0;
1145 else
1146 return -ENOENT;
1147 }
1148
1149 static
1150 void *partition_resize_thread(void *arg)
1151 {
1152 struct partition_resize_work *work = arg;
1153
1154 work->ht->flavor->register_thread();
1155 work->fct(work->ht, work->i, work->start, work->len);
1156 work->ht->flavor->unregister_thread();
1157 return NULL;
1158 }
1159
1160 static
1161 void partition_resize_helper(struct cds_lfht *ht, unsigned long i,
1162 unsigned long len,
1163 void (*fct)(struct cds_lfht *ht, unsigned long i,
1164 unsigned long start, unsigned long len))
1165 {
1166 unsigned long partition_len;
1167 struct partition_resize_work *work;
1168 int thread, ret;
1169 unsigned long nr_threads;
1170
1171 /*
1172 * Note: nr_cpus_mask + 1 is always power of 2.
1173 * We spawn just the number of threads we need to satisfy the minimum
1174 * partition size, up to the number of CPUs in the system.
1175 */
1176 if (nr_cpus_mask > 0) {
1177 nr_threads = min(nr_cpus_mask + 1,
1178 len >> MIN_PARTITION_PER_THREAD_ORDER);
1179 } else {
1180 nr_threads = 1;
1181 }
1182 partition_len = len >> cds_lfht_get_count_order_ulong(nr_threads);
1183 work = calloc(nr_threads, sizeof(*work));
1184 assert(work);
1185 for (thread = 0; thread < nr_threads; thread++) {
1186 work[thread].ht = ht;
1187 work[thread].i = i;
1188 work[thread].len = partition_len;
1189 work[thread].start = thread * partition_len;
1190 work[thread].fct = fct;
1191 ret = pthread_create(&(work[thread].thread_id), ht->resize_attr,
1192 partition_resize_thread, &work[thread]);
1193 assert(!ret);
1194 }
1195 for (thread = 0; thread < nr_threads; thread++) {
1196 ret = pthread_join(work[thread].thread_id, NULL);
1197 assert(!ret);
1198 }
1199 free(work);
1200 }
1201
1202 /*
1203 * Holding RCU read lock to protect _cds_lfht_add against memory
1204 * reclaim that could be performed by other call_rcu worker threads (ABA
1205 * problem).
1206 *
1207 * When we reach a certain length, we can split this population phase over
1208 * many worker threads, based on the number of CPUs available in the system.
1209 * This should therefore take care of not having the expand lagging behind too
1210 * many concurrent insertion threads by using the scheduler's ability to
1211 * schedule bucket node population fairly with insertions.
1212 */
1213 static
1214 void init_table_populate_partition(struct cds_lfht *ht, unsigned long i,
1215 unsigned long start, unsigned long len)
1216 {
1217 unsigned long j, size = 1UL << (i - 1);
1218
1219 assert(i > MIN_TABLE_ORDER);
1220 ht->flavor->read_lock();
1221 for (j = size + start; j < size + start + len; j++) {
1222 struct cds_lfht_node *new_node = bucket_at(ht, j);
1223
1224 assert(j >= size && j < (size << 1));
1225 dbg_printf("init populate: order %lu index %lu hash %lu\n",
1226 i, j, j);
1227 new_node->reverse_hash = bit_reverse_ulong(j);
1228 _cds_lfht_add(ht, j, NULL, NULL, size, new_node, NULL, 1);
1229 }
1230 ht->flavor->read_unlock();
1231 }
1232
1233 static
1234 void init_table_populate(struct cds_lfht *ht, unsigned long i,
1235 unsigned long len)
1236 {
1237 assert(nr_cpus_mask != -1);
1238 if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD) {
1239 ht->flavor->thread_online();
1240 init_table_populate_partition(ht, i, 0, len);
1241 ht->flavor->thread_offline();
1242 return;
1243 }
1244 partition_resize_helper(ht, i, len, init_table_populate_partition);
1245 }
1246
1247 static
1248 void init_table(struct cds_lfht *ht,
1249 unsigned long first_order, unsigned long last_order)
1250 {
1251 unsigned long i;
1252
1253 dbg_printf("init table: first_order %lu last_order %lu\n",
1254 first_order, last_order);
1255 assert(first_order > MIN_TABLE_ORDER);
1256 for (i = first_order; i <= last_order; i++) {
1257 unsigned long len;
1258
1259 len = 1UL << (i - 1);
1260 dbg_printf("init order %lu len: %lu\n", i, len);
1261
1262 /* Stop expand if the resize target changes under us */
1263 if (CMM_LOAD_SHARED(ht->resize_target) < (1UL << i))
1264 break;
1265
1266 cds_lfht_alloc_bucket_table(ht, i);
1267
1268 /*
1269 * Set all bucket nodes reverse hash values for a level and
1270 * link all bucket nodes into the table.
1271 */
1272 init_table_populate(ht, i, len);
1273
1274 /*
1275 * Update table size.
1276 */
1277 cmm_smp_wmb(); /* populate data before RCU size */
1278 CMM_STORE_SHARED(ht->size, 1UL << i);
1279
1280 dbg_printf("init new size: %lu\n", 1UL << i);
1281 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1282 break;
1283 }
1284 }
1285
1286 /*
1287 * Holding RCU read lock to protect _cds_lfht_remove against memory
1288 * reclaim that could be performed by other call_rcu worker threads (ABA
1289 * problem).
1290 * For a single level, we logically remove and garbage collect each node.
1291 *
1292 * As a design choice, we perform logical removal and garbage collection on a
1293 * node-per-node basis to simplify this algorithm. We also assume keeping good
1294 * cache locality of the operation would overweight possible performance gain
1295 * that could be achieved by batching garbage collection for multiple levels.
1296 * However, this would have to be justified by benchmarks.
1297 *
1298 * Concurrent removal and add operations are helping us perform garbage
1299 * collection of logically removed nodes. We guarantee that all logically
1300 * removed nodes have been garbage-collected (unlinked) before call_rcu is
1301 * invoked to free a hole level of bucket nodes (after a grace period).
1302 *
1303 * Logical removal and garbage collection can therefore be done in batch
1304 * or on a node-per-node basis, as long as the guarantee above holds.
1305 *
1306 * When we reach a certain length, we can split this removal over many worker
1307 * threads, based on the number of CPUs available in the system. This should
1308 * take care of not letting resize process lag behind too many concurrent
1309 * updater threads actively inserting into the hash table.
1310 */
1311 static
1312 void remove_table_partition(struct cds_lfht *ht, unsigned long i,
1313 unsigned long start, unsigned long len)
1314 {
1315 unsigned long j, size = 1UL << (i - 1);
1316
1317 assert(i > MIN_TABLE_ORDER);
1318 ht->flavor->read_lock();
1319 for (j = size + start; j < size + start + len; j++) {
1320 struct cds_lfht_node *fini_bucket = bucket_at(ht, j);
1321 struct cds_lfht_node *parent_bucket = bucket_at(ht, j - size);
1322
1323 assert(j >= size && j < (size << 1));
1324 dbg_printf("remove entry: order %lu index %lu hash %lu\n",
1325 i, j, j);
1326 /* Set the REMOVED_FLAG to freeze the ->next for gc */
1327 uatomic_or(&fini_bucket->next, REMOVED_FLAG);
1328 _cds_lfht_gc_bucket(parent_bucket, fini_bucket);
1329 }
1330 ht->flavor->read_unlock();
1331 }
1332
1333 static
1334 void remove_table(struct cds_lfht *ht, unsigned long i, unsigned long len)
1335 {
1336
1337 assert(nr_cpus_mask != -1);
1338 if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD) {
1339 ht->flavor->thread_online();
1340 remove_table_partition(ht, i, 0, len);
1341 ht->flavor->thread_offline();
1342 return;
1343 }
1344 partition_resize_helper(ht, i, len, remove_table_partition);
1345 }
1346
1347 /*
1348 * fini_table() is never called for first_order == 0, which is why
1349 * free_by_rcu_order == 0 can be used as criterion to know if free must
1350 * be called.
1351 */
1352 static
1353 void fini_table(struct cds_lfht *ht,
1354 unsigned long first_order, unsigned long last_order)
1355 {
1356 long i;
1357 unsigned long free_by_rcu_order = 0;
1358
1359 dbg_printf("fini table: first_order %lu last_order %lu\n",
1360 first_order, last_order);
1361 assert(first_order > MIN_TABLE_ORDER);
1362 for (i = last_order; i >= first_order; i--) {
1363 unsigned long len;
1364
1365 len = 1UL << (i - 1);
1366 dbg_printf("fini order %lu len: %lu\n", i, len);
1367
1368 /* Stop shrink if the resize target changes under us */
1369 if (CMM_LOAD_SHARED(ht->resize_target) > (1UL << (i - 1)))
1370 break;
1371
1372 cmm_smp_wmb(); /* populate data before RCU size */
1373 CMM_STORE_SHARED(ht->size, 1UL << (i - 1));
1374
1375 /*
1376 * We need to wait for all add operations to reach Q.S. (and
1377 * thus use the new table for lookups) before we can start
1378 * releasing the old bucket nodes. Otherwise their lookup will
1379 * return a logically removed node as insert position.
1380 */
1381 ht->flavor->update_synchronize_rcu();
1382 if (free_by_rcu_order)
1383 cds_lfht_free_bucket_table(ht, free_by_rcu_order);
1384
1385 /*
1386 * Set "removed" flag in bucket nodes about to be removed.
1387 * Unlink all now-logically-removed bucket node pointers.
1388 * Concurrent add/remove operation are helping us doing
1389 * the gc.
1390 */
1391 remove_table(ht, i, len);
1392
1393 free_by_rcu_order = i;
1394
1395 dbg_printf("fini new size: %lu\n", 1UL << i);
1396 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1397 break;
1398 }
1399
1400 if (free_by_rcu_order) {
1401 ht->flavor->update_synchronize_rcu();
1402 cds_lfht_free_bucket_table(ht, free_by_rcu_order);
1403 }
1404 }
1405
1406 static
1407 void cds_lfht_create_bucket(struct cds_lfht *ht, unsigned long size)
1408 {
1409 struct cds_lfht_node *prev, *node;
1410 unsigned long order, len, i;
1411
1412 cds_lfht_alloc_bucket_table(ht, 0);
1413
1414 dbg_printf("create bucket: order 0 index 0 hash 0\n");
1415 node = bucket_at(ht, 0);
1416 node->next = flag_bucket(get_end());
1417 node->reverse_hash = 0;
1418
1419 for (order = 1; order < cds_lfht_get_count_order_ulong(size) + 1; order++) {
1420 len = 1UL << (order - 1);
1421 cds_lfht_alloc_bucket_table(ht, order);
1422
1423 for (i = 0; i < len; i++) {
1424 /*
1425 * Now, we are trying to init the node with the
1426 * hash=(len+i) (which is also a bucket with the
1427 * index=(len+i)) and insert it into the hash table,
1428 * so this node has to be inserted after the bucket
1429 * with the index=(len+i)&(len-1)=i. And because there
1430 * is no other non-bucket node nor bucket node with
1431 * larger index/hash inserted, so the bucket node
1432 * being inserted should be inserted directly linked
1433 * after the bucket node with index=i.
1434 */
1435 prev = bucket_at(ht, i);
1436 node = bucket_at(ht, len + i);
1437
1438 dbg_printf("create bucket: order %lu index %lu hash %lu\n",
1439 order, len + i, len + i);
1440 node->reverse_hash = bit_reverse_ulong(len + i);
1441
1442 /* insert after prev */
1443 assert(is_bucket(prev->next));
1444 node->next = prev->next;
1445 prev->next = flag_bucket(node);
1446 }
1447 }
1448 }
1449
1450 struct cds_lfht *_cds_lfht_new(unsigned long init_size,
1451 unsigned long min_nr_alloc_buckets,
1452 unsigned long max_nr_buckets,
1453 int flags,
1454 const struct cds_lfht_mm_type *mm,
1455 const struct rcu_flavor_struct *flavor,
1456 pthread_attr_t *attr)
1457 {
1458 struct cds_lfht *ht;
1459 unsigned long order;
1460
1461 /* min_nr_alloc_buckets must be power of two */
1462 if (!min_nr_alloc_buckets || (min_nr_alloc_buckets & (min_nr_alloc_buckets - 1)))
1463 return NULL;
1464
1465 /* init_size must be power of two */
1466 if (!init_size || (init_size & (init_size - 1)))
1467 return NULL;
1468
1469 /*
1470 * Memory management plugin default.
1471 */
1472 if (!mm) {
1473 if (CAA_BITS_PER_LONG > 32
1474 && max_nr_buckets
1475 && max_nr_buckets <= (1ULL << 32)) {
1476 /*
1477 * For 64-bit architectures, with max number of
1478 * buckets small enough not to use the entire
1479 * 64-bit memory mapping space (and allowing a
1480 * fair number of hash table instances), use the
1481 * mmap allocator, which is faster than the
1482 * order allocator.
1483 */
1484 mm = &cds_lfht_mm_mmap;
1485 } else {
1486 /*
1487 * The fallback is to use the order allocator.
1488 */
1489 mm = &cds_lfht_mm_order;
1490 }
1491 }
1492
1493 /* max_nr_buckets == 0 for order based mm means infinite */
1494 if (mm == &cds_lfht_mm_order && !max_nr_buckets)
1495 max_nr_buckets = 1UL << (MAX_TABLE_ORDER - 1);
1496
1497 /* max_nr_buckets must be power of two */
1498 if (!max_nr_buckets || (max_nr_buckets & (max_nr_buckets - 1)))
1499 return NULL;
1500
1501 min_nr_alloc_buckets = max(min_nr_alloc_buckets, MIN_TABLE_SIZE);
1502 init_size = max(init_size, MIN_TABLE_SIZE);
1503 max_nr_buckets = max(max_nr_buckets, min_nr_alloc_buckets);
1504 init_size = min(init_size, max_nr_buckets);
1505
1506 ht = mm->alloc_cds_lfht(min_nr_alloc_buckets, max_nr_buckets);
1507 assert(ht);
1508 assert(ht->mm == mm);
1509 assert(ht->bucket_at == mm->bucket_at);
1510
1511 ht->flags = flags;
1512 ht->flavor = flavor;
1513 ht->resize_attr = attr;
1514 alloc_split_items_count(ht);
1515 /* this mutex should not nest in read-side C.S. */
1516 pthread_mutex_init(&ht->resize_mutex, NULL);
1517 order = cds_lfht_get_count_order_ulong(init_size);
1518 ht->resize_target = 1UL << order;
1519 cds_lfht_create_bucket(ht, 1UL << order);
1520 ht->size = 1UL << order;
1521 return ht;
1522 }
1523
1524 void cds_lfht_lookup(struct cds_lfht *ht, unsigned long hash,
1525 cds_lfht_match_fct match, const void *key,
1526 struct cds_lfht_iter *iter)
1527 {
1528 struct cds_lfht_node *node, *next, *bucket;
1529 unsigned long reverse_hash, size;
1530
1531 reverse_hash = bit_reverse_ulong(hash);
1532
1533 size = rcu_dereference(ht->size);
1534 bucket = lookup_bucket(ht, size, hash);
1535 /* We can always skip the bucket node initially */
1536 node = rcu_dereference(bucket->next);
1537 node = clear_flag(node);
1538 for (;;) {
1539 if (caa_unlikely(is_end(node))) {
1540 node = next = NULL;
1541 break;
1542 }
1543 if (caa_unlikely(node->reverse_hash > reverse_hash)) {
1544 node = next = NULL;
1545 break;
1546 }
1547 next = rcu_dereference(node->next);
1548 assert(node == clear_flag(node));
1549 if (caa_likely(!is_removed(next))
1550 && !is_bucket(next)
1551 && node->reverse_hash == reverse_hash
1552 && caa_likely(match(node, key))) {
1553 break;
1554 }
1555 node = clear_flag(next);
1556 }
1557 assert(!node || !is_bucket(CMM_LOAD_SHARED(node->next)));
1558 iter->node = node;
1559 iter->next = next;
1560 }
1561
1562 void cds_lfht_next_duplicate(struct cds_lfht *ht, cds_lfht_match_fct match,
1563 const void *key, struct cds_lfht_iter *iter)
1564 {
1565 struct cds_lfht_node *node, *next;
1566 unsigned long reverse_hash;
1567
1568 node = iter->node;
1569 reverse_hash = node->reverse_hash;
1570 next = iter->next;
1571 node = clear_flag(next);
1572
1573 for (;;) {
1574 if (caa_unlikely(is_end(node))) {
1575 node = next = NULL;
1576 break;
1577 }
1578 if (caa_unlikely(node->reverse_hash > reverse_hash)) {
1579 node = next = NULL;
1580 break;
1581 }
1582 next = rcu_dereference(node->next);
1583 if (caa_likely(!is_removed(next))
1584 && !is_bucket(next)
1585 && caa_likely(match(node, key))) {
1586 break;
1587 }
1588 node = clear_flag(next);
1589 }
1590 assert(!node || !is_bucket(CMM_LOAD_SHARED(node->next)));
1591 iter->node = node;
1592 iter->next = next;
1593 }
1594
1595 void cds_lfht_next(struct cds_lfht *ht, struct cds_lfht_iter *iter)
1596 {
1597 struct cds_lfht_node *node, *next;
1598
1599 node = clear_flag(iter->next);
1600 for (;;) {
1601 if (caa_unlikely(is_end(node))) {
1602 node = next = NULL;
1603 break;
1604 }
1605 next = rcu_dereference(node->next);
1606 if (caa_likely(!is_removed(next))
1607 && !is_bucket(next)) {
1608 break;
1609 }
1610 node = clear_flag(next);
1611 }
1612 assert(!node || !is_bucket(CMM_LOAD_SHARED(node->next)));
1613 iter->node = node;
1614 iter->next = next;
1615 }
1616
1617 void cds_lfht_first(struct cds_lfht *ht, struct cds_lfht_iter *iter)
1618 {
1619 /*
1620 * Get next after first bucket node. The first bucket node is the
1621 * first node of the linked list.
1622 */
1623 iter->next = bucket_at(ht, 0)->next;
1624 cds_lfht_next(ht, iter);
1625 }
1626
1627 void cds_lfht_add(struct cds_lfht *ht, unsigned long hash,
1628 struct cds_lfht_node *node)
1629 {
1630 unsigned long size;
1631
1632 node->reverse_hash = bit_reverse_ulong(hash);
1633 size = rcu_dereference(ht->size);
1634 _cds_lfht_add(ht, hash, NULL, NULL, size, node, NULL, 0);
1635 ht_count_add(ht, size, hash);
1636 }
1637
1638 struct cds_lfht_node *cds_lfht_add_unique(struct cds_lfht *ht,
1639 unsigned long hash,
1640 cds_lfht_match_fct match,
1641 const void *key,
1642 struct cds_lfht_node *node)
1643 {
1644 unsigned long size;
1645 struct cds_lfht_iter iter;
1646
1647 node->reverse_hash = bit_reverse_ulong(hash);
1648 size = rcu_dereference(ht->size);
1649 _cds_lfht_add(ht, hash, match, key, size, node, &iter, 0);
1650 if (iter.node == node)
1651 ht_count_add(ht, size, hash);
1652 return iter.node;
1653 }
1654
1655 struct cds_lfht_node *cds_lfht_add_replace(struct cds_lfht *ht,
1656 unsigned long hash,
1657 cds_lfht_match_fct match,
1658 const void *key,
1659 struct cds_lfht_node *node)
1660 {
1661 unsigned long size;
1662 struct cds_lfht_iter iter;
1663
1664 node->reverse_hash = bit_reverse_ulong(hash);
1665 size = rcu_dereference(ht->size);
1666 for (;;) {
1667 _cds_lfht_add(ht, hash, match, key, size, node, &iter, 0);
1668 if (iter.node == node) {
1669 ht_count_add(ht, size, hash);
1670 return NULL;
1671 }
1672
1673 if (!_cds_lfht_replace(ht, size, iter.node, iter.next, node))
1674 return iter.node;
1675 }
1676 }
1677
1678 int cds_lfht_replace(struct cds_lfht *ht,
1679 struct cds_lfht_iter *old_iter,
1680 unsigned long hash,
1681 cds_lfht_match_fct match,
1682 const void *key,
1683 struct cds_lfht_node *new_node)
1684 {
1685 unsigned long size;
1686
1687 new_node->reverse_hash = bit_reverse_ulong(hash);
1688 if (!old_iter->node)
1689 return -ENOENT;
1690 if (caa_unlikely(old_iter->node->reverse_hash != new_node->reverse_hash))
1691 return -EINVAL;
1692 if (caa_unlikely(!match(old_iter->node, key)))
1693 return -EINVAL;
1694 size = rcu_dereference(ht->size);
1695 return _cds_lfht_replace(ht, size, old_iter->node, old_iter->next,
1696 new_node);
1697 }
1698
1699 int cds_lfht_del(struct cds_lfht *ht, struct cds_lfht_node *node)
1700 {
1701 unsigned long size, hash;
1702 int ret;
1703
1704 size = rcu_dereference(ht->size);
1705 ret = _cds_lfht_del(ht, size, node);
1706 if (!ret) {
1707 hash = bit_reverse_ulong(node->reverse_hash);
1708 ht_count_del(ht, size, hash);
1709 }
1710 return ret;
1711 }
1712
1713 int cds_lfht_is_node_deleted(struct cds_lfht_node *node)
1714 {
1715 return is_removed(CMM_LOAD_SHARED(node->next));
1716 }
1717
1718 static
1719 int cds_lfht_delete_bucket(struct cds_lfht *ht)
1720 {
1721 struct cds_lfht_node *node;
1722 unsigned long order, i, size;
1723
1724 /* Check that the table is empty */
1725 node = bucket_at(ht, 0);
1726 do {
1727 node = clear_flag(node)->next;
1728 if (!is_bucket(node))
1729 return -EPERM;
1730 assert(!is_removed(node));
1731 } while (!is_end(node));
1732 /*
1733 * size accessed without rcu_dereference because hash table is
1734 * being destroyed.
1735 */
1736 size = ht->size;
1737 /* Internal sanity check: all nodes left should be buckets */
1738 for (i = 0; i < size; i++) {
1739 node = bucket_at(ht, i);
1740 dbg_printf("delete bucket: index %lu expected hash %lu hash %lu\n",
1741 i, i, bit_reverse_ulong(node->reverse_hash));
1742 assert(is_bucket(node->next));
1743 }
1744
1745 for (order = cds_lfht_get_count_order_ulong(size); (long)order >= 0; order--)
1746 cds_lfht_free_bucket_table(ht, order);
1747
1748 return 0;
1749 }
1750
1751 /*
1752 * Should only be called when no more concurrent readers nor writers can
1753 * possibly access the table.
1754 */
1755 int cds_lfht_destroy(struct cds_lfht *ht, pthread_attr_t **attr)
1756 {
1757 int ret;
1758 #ifdef rcu_read_ongoing_mb
1759 int was_online;
1760 #endif
1761
1762 /* Wait for in-flight resize operations to complete */
1763 _CMM_STORE_SHARED(ht->in_progress_destroy, 1);
1764 cmm_smp_mb(); /* Store destroy before load resize */
1765 #ifdef rcu_read_ongoing_mb
1766 was_online = ht->flavor->read_ongoing();
1767 if (was_online)
1768 ht->flavor->thread_offline();
1769 /* Calling with RCU read-side held is an error. */
1770 if (ht->flavor->read_ongoing()) {
1771 ret = -EINVAL;
1772 if (was_online)
1773 ht->flavor->thread_online();
1774 goto end;
1775 }
1776 #endif
1777 while (uatomic_read(&ht->in_progress_resize))
1778 (void) poll(NULL, 0, 100); /* wait for 100ms */
1779 ret = cds_lfht_delete_bucket(ht);
1780 if (ret)
1781 return ret;
1782 free_split_items_count(ht);
1783 if (attr)
1784 *attr = ht->resize_attr;
1785 poison_free(ht);
1786 #ifdef rcu_read_ongoing_mb
1787 end:
1788 #endif
1789 return ret;
1790 }
1791
1792 void cds_lfht_count_nodes(struct cds_lfht *ht,
1793 long *approx_before,
1794 unsigned long *count,
1795 long *approx_after)
1796 {
1797 struct cds_lfht_node *node, *next;
1798 unsigned long nr_bucket = 0, nr_removed = 0;
1799
1800 *approx_before = 0;
1801 if (ht->split_count) {
1802 int i;
1803
1804 for (i = 0; i < split_count_mask + 1; i++) {
1805 *approx_before += uatomic_read(&ht->split_count[i].add);
1806 *approx_before -= uatomic_read(&ht->split_count[i].del);
1807 }
1808 }
1809
1810 *count = 0;
1811
1812 /* Count non-bucket nodes in the table */
1813 node = bucket_at(ht, 0);
1814 do {
1815 next = rcu_dereference(node->next);
1816 if (is_removed(next)) {
1817 if (!is_bucket(next))
1818 (nr_removed)++;
1819 else
1820 (nr_bucket)++;
1821 } else if (!is_bucket(next))
1822 (*count)++;
1823 else
1824 (nr_bucket)++;
1825 node = clear_flag(next);
1826 } while (!is_end(node));
1827 dbg_printf("number of logically removed nodes: %lu\n", nr_removed);
1828 dbg_printf("number of bucket nodes: %lu\n", nr_bucket);
1829 *approx_after = 0;
1830 if (ht->split_count) {
1831 int i;
1832
1833 for (i = 0; i < split_count_mask + 1; i++) {
1834 *approx_after += uatomic_read(&ht->split_count[i].add);
1835 *approx_after -= uatomic_read(&ht->split_count[i].del);
1836 }
1837 }
1838 }
1839
1840 /* called with resize mutex held */
1841 static
1842 void _do_cds_lfht_grow(struct cds_lfht *ht,
1843 unsigned long old_size, unsigned long new_size)
1844 {
1845 unsigned long old_order, new_order;
1846
1847 old_order = cds_lfht_get_count_order_ulong(old_size);
1848 new_order = cds_lfht_get_count_order_ulong(new_size);
1849 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1850 old_size, old_order, new_size, new_order);
1851 assert(new_size > old_size);
1852 init_table(ht, old_order + 1, new_order);
1853 }
1854
1855 /* called with resize mutex held */
1856 static
1857 void _do_cds_lfht_shrink(struct cds_lfht *ht,
1858 unsigned long old_size, unsigned long new_size)
1859 {
1860 unsigned long old_order, new_order;
1861
1862 new_size = max(new_size, MIN_TABLE_SIZE);
1863 old_order = cds_lfht_get_count_order_ulong(old_size);
1864 new_order = cds_lfht_get_count_order_ulong(new_size);
1865 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1866 old_size, old_order, new_size, new_order);
1867 assert(new_size < old_size);
1868
1869 /* Remove and unlink all bucket nodes to remove. */
1870 fini_table(ht, new_order + 1, old_order);
1871 }
1872
1873
1874 /* called with resize mutex held */
1875 static
1876 void _do_cds_lfht_resize(struct cds_lfht *ht)
1877 {
1878 unsigned long new_size, old_size;
1879
1880 /*
1881 * Resize table, re-do if the target size has changed under us.
1882 */
1883 do {
1884 assert(uatomic_read(&ht->in_progress_resize));
1885 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1886 break;
1887 ht->resize_initiated = 1;
1888 old_size = ht->size;
1889 new_size = CMM_LOAD_SHARED(ht->resize_target);
1890 if (old_size < new_size)
1891 _do_cds_lfht_grow(ht, old_size, new_size);
1892 else if (old_size > new_size)
1893 _do_cds_lfht_shrink(ht, old_size, new_size);
1894 ht->resize_initiated = 0;
1895 /* write resize_initiated before read resize_target */
1896 cmm_smp_mb();
1897 } while (ht->size != CMM_LOAD_SHARED(ht->resize_target));
1898 }
1899
1900 static
1901 unsigned long resize_target_grow(struct cds_lfht *ht, unsigned long new_size)
1902 {
1903 return _uatomic_xchg_monotonic_increase(&ht->resize_target, new_size);
1904 }
1905
1906 static
1907 void resize_target_update_count(struct cds_lfht *ht,
1908 unsigned long count)
1909 {
1910 count = max(count, MIN_TABLE_SIZE);
1911 count = min(count, ht->max_nr_buckets);
1912 uatomic_set(&ht->resize_target, count);
1913 }
1914
1915 void cds_lfht_resize(struct cds_lfht *ht, unsigned long new_size)
1916 {
1917 #ifdef rcu_read_ongoing_mb
1918 int was_online;
1919
1920 was_online = ht->flavor->read_ongoing();
1921 if (was_online)
1922 ht->flavor->thread_offline();
1923 /* Calling with RCU read-side held is an error. */
1924 if (ht->flavor->read_ongoing()) {
1925 static int print_once;
1926
1927 if (!CMM_LOAD_SHARED(print_once))
1928 fprintf(stderr, "[error] rculfhash: cds_lfht_resize "
1929 "called with RCU read-side lock held.\n");
1930 CMM_STORE_SHARED(print_once, 1);
1931 assert(0);
1932 goto end;
1933 }
1934 #endif
1935 resize_target_update_count(ht, new_size);
1936 CMM_STORE_SHARED(ht->resize_initiated, 1);
1937 pthread_mutex_lock(&ht->resize_mutex);
1938 _do_cds_lfht_resize(ht);
1939 pthread_mutex_unlock(&ht->resize_mutex);
1940 #ifdef rcu_read_ongoing_mb
1941 end:
1942 if (was_online)
1943 ht->flavor->thread_online();
1944 #endif
1945 }
1946
1947 static
1948 void do_resize_cb(struct rcu_head *head)
1949 {
1950 struct rcu_resize_work *work =
1951 caa_container_of(head, struct rcu_resize_work, head);
1952 struct cds_lfht *ht = work->ht;
1953
1954 ht->flavor->thread_offline();
1955 pthread_mutex_lock(&ht->resize_mutex);
1956 _do_cds_lfht_resize(ht);
1957 pthread_mutex_unlock(&ht->resize_mutex);
1958 ht->flavor->thread_online();
1959 poison_free(work);
1960 cmm_smp_mb(); /* finish resize before decrement */
1961 uatomic_dec(&ht->in_progress_resize);
1962 }
1963
1964 static
1965 void __cds_lfht_resize_lazy_launch(struct cds_lfht *ht)
1966 {
1967 struct rcu_resize_work *work;
1968
1969 /* Store resize_target before read resize_initiated */
1970 cmm_smp_mb();
1971 if (!CMM_LOAD_SHARED(ht->resize_initiated)) {
1972 uatomic_inc(&ht->in_progress_resize);
1973 cmm_smp_mb(); /* increment resize count before load destroy */
1974 if (CMM_LOAD_SHARED(ht->in_progress_destroy)) {
1975 uatomic_dec(&ht->in_progress_resize);
1976 return;
1977 }
1978 work = zmalloc(sizeof(*work));
1979 if (work == NULL) {
1980 dbg_printf("error allocating resize work, bailing out\n");
1981 uatomic_dec(&ht->in_progress_resize);
1982 return;
1983 }
1984 work->ht = ht;
1985 ht->flavor->update_call_rcu(&work->head, do_resize_cb);
1986 CMM_STORE_SHARED(ht->resize_initiated, 1);
1987 }
1988 }
1989
1990 static
1991 void cds_lfht_resize_lazy_grow(struct cds_lfht *ht, unsigned long size, int growth)
1992 {
1993 unsigned long target_size = size << growth;
1994
1995 target_size = min(target_size, ht->max_nr_buckets);
1996 if (resize_target_grow(ht, target_size) >= target_size)
1997 return;
1998
1999 __cds_lfht_resize_lazy_launch(ht);
2000 }
2001
2002 /*
2003 * We favor grow operations over shrink. A shrink operation never occurs
2004 * if a grow operation is queued for lazy execution. A grow operation
2005 * cancels any pending shrink lazy execution.
2006 */
2007 static
2008 void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
2009 unsigned long count)
2010 {
2011 if (!(ht->flags & CDS_LFHT_AUTO_RESIZE))
2012 return;
2013 count = max(count, MIN_TABLE_SIZE);
2014 count = min(count, ht->max_nr_buckets);
2015 if (count == size)
2016 return; /* Already the right size, no resize needed */
2017 if (count > size) { /* lazy grow */
2018 if (resize_target_grow(ht, count) >= count)
2019 return;
2020 } else { /* lazy shrink */
2021 for (;;) {
2022 unsigned long s;
2023
2024 s = uatomic_cmpxchg(&ht->resize_target, size, count);
2025 if (s == size)
2026 break; /* no resize needed */
2027 if (s > size)
2028 return; /* growing is/(was just) in progress */
2029 if (s <= count)
2030 return; /* some other thread do shrink */
2031 size = s;
2032 }
2033 }
2034 __cds_lfht_resize_lazy_launch(ht);
2035 }
This page took 0.118008 seconds and 3 git commands to generate.