Cleanup: enable signed/unsigned compare compiler warning
[urcu.git] / src / rculfhash.c
CommitLineData
5e28c532 1/*
abc490a1
MD
2 * rculfhash.c
3 *
1475579c 4 * Userspace RCU library - Lock-Free Resizable RCU Hash Table
abc490a1
MD
5 *
6 * Copyright 2010-2011 - Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
0dcf4847 7 * Copyright 2011 - Lai Jiangshan <laijs@cn.fujitsu.com>
abc490a1
MD
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
5e28c532
MD
22 */
23
e753ff5a
MD
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 *
1475579c 33 * Some specificities of this Lock-Free Resizable RCU Hash Table
e753ff5a
MD
34 * implementation:
35 *
36 * - RCU read-side critical section allows readers to perform hash
1f67ba50
MD
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.
e753ff5a
MD
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
1f67ba50
MD
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.
e753ff5a
MD
50 * - Hash table nodes are contained within a split-ordered list. This
51 * list is ordered by incrementing reversed-bits-hash value.
1ee8f000 52 * - An index of bucket nodes is kept. These bucket nodes are the hash
1f67ba50
MD
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.
1475579c
MD
62 * - The resize operation for larger tables (and available through an
63 * API) allows both expanding and shrinking the hash table.
4c42f1b8 64 * - Split-counters are used to keep track of the number of
1475579c 65 * nodes within the hash table for automatic resize triggering.
e753ff5a 66 * - Resize operation initiated by long chain detection is executed by a
d0ec0ed2 67 * worker thread, which keeps lock-freedom of add and remove.
e753ff5a
MD
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
5c4ca589
MD
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
1f67ba50 89 * linked-list (the start node is never removed) and that it does not
5c4ca589
MD
90 * contain the "removed" node anymore, even if concurrent delete/add
91 * operations are changing the structure of the list concurrently.
1f67ba50
MD
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
29e669f6
MD
95 * helping the remover unlink nodes from the list rather than to wait
96 * for it do to so.
1f67ba50
MD
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.
5f177b1c 113 * - synchronize_rcu is used to garbage-collect the old bucket node table.
93d46c39 114 *
7f949215 115 * Ordering Guarantees:
0f5543cb 116 *
7f949215
MD
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
67ecffc0 120 * cds_lfht_add_unique (failure).
7f949215
MD
121 *
122 * We define "read traversal" operation as any of the following
123 * group of operations
0f5543cb 124 * - cds_lfht_lookup followed by iteration with cds_lfht_next_duplicate
7f949215
MD
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).
0f5543cb 131 *
bf09adc7 132 * We define "write" operations as any of cds_lfht_add, cds_lfht_replace,
7f949215
MD
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.
0f5543cb 206 *
1ee8f000 207 * Bucket node tables:
93d46c39 208 *
1ee8f000
LJ
209 * hash table hash table the last all bucket node tables
210 * order size bucket node 0 1 2 3 4 5 6(index)
93d46c39
LJ
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 *
1ee8f000 220 * When growing/shrinking, we only focus on the last bucket node table
93d46c39
LJ
221 * which size is (!order ? 1 : (1 << (order -1))).
222 *
223 * Example for growing/shrinking:
1ee8f000
LJ
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
93d46c39 226 *
1475579c 227 * A bit of ascii art explanation:
67ecffc0 228 *
1f67ba50
MD
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.
67ecffc0 231 *
1475579c 232 * This shows the nodes for a small table ordered by reversed bits:
67ecffc0 233 *
1475579c
MD
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
67ecffc0
MD
243 *
244 * This shows the nodes in order of non-reversed bits, linked by
1475579c 245 * reversed-bit order.
67ecffc0 246 *
1475579c
MD
247 * order bits reverse
248 * 0 0 000 000
0adc36a8
LJ
249 * 1 | 1 001 100 <-
250 * 2 | | 2 010 010 <- |
f6fdd688 251 * | | | 3 011 110 | <- |
1475579c
MD
252 * 3 -> | | | 4 100 001 | |
253 * -> | | 5 101 101 |
254 * -> | 6 110 011
255 * -> 7 111 111
e753ff5a
MD
256 */
257
2ed95849
MD
258#define _LGPL_SOURCE
259#include <stdlib.h>
e0ba718a
MD
260#include <errno.h>
261#include <assert.h>
262#include <stdio.h>
abc490a1 263#include <stdint.h>
f000907d 264#include <string.h>
125f41db 265#include <sched.h>
95747f9e 266#include <unistd.h>
e0ba718a 267
a47dd11c 268#include "compat-getcpu.h"
6cd23d47
MD
269#include <urcu/pointer.h>
270#include <urcu/call-rcu.h>
271#include <urcu/flavor.h>
a42cc659
MD
272#include <urcu/arch.h>
273#include <urcu/uatomic.h>
a42cc659 274#include <urcu/compiler.h>
abc490a1 275#include <urcu/rculfhash.h>
0b6aa001 276#include <rculfhash-internal.h>
5e28c532 277#include <stdio.h>
464a1ec9 278#include <pthread.h>
d0ec0ed2
MD
279#include <signal.h>
280#include "workqueue.h"
281#include "urcu-die.h"
83e334d0 282#include "urcu-utils.h"
44395fb7 283
f8994aee 284/*
4c42f1b8 285 * Split-counters lazily update the global counter each 1024
f8994aee
MD
286 * addition/removal. It automatically keeps track of resize required.
287 * We use the bucket length as indicator for need to expand for small
ffa11a18 288 * tables and machines lacking per-cpu data support.
f8994aee
MD
289 */
290#define COUNT_COMMIT_ORDER 10
4ddbb355 291#define DEFAULT_SPLIT_COUNT_MASK 0xFUL
6ea6bc67
MD
292#define CHAIN_LEN_TARGET 1
293#define CHAIN_LEN_RESIZE_THRESHOLD 3
2ed95849 294
cd95516d 295/*
76a73da8 296 * Define the minimum table size.
cd95516d 297 */
d0d8f9aa
LJ
298#define MIN_TABLE_ORDER 0
299#define MIN_TABLE_SIZE (1UL << MIN_TABLE_ORDER)
cd95516d 300
b7d619b0 301/*
1ee8f000 302 * Minimum number of bucket nodes to touch per thread to parallelize grow/shrink.
b7d619b0 303 */
6083a889
MD
304#define MIN_PARTITION_PER_THREAD_ORDER 12
305#define MIN_PARTITION_PER_THREAD (1UL << MIN_PARTITION_PER_THREAD_ORDER)
b7d619b0 306
d95bd160
MD
307/*
308 * The removed flag needs to be updated atomically with the pointer.
48ed1c18 309 * It indicates that no node must attach to the node scheduled for
b198f0fd 310 * removal, and that node garbage collection must be performed.
1ee8f000 311 * The bucket flag does not require to be updated atomically with the
d95bd160 312 * pointer, but it is added as a pointer low bit flag to save space.
1f67ba50
MD
313 * The "removal owner" flag is used to detect which of the "del"
314 * operation that has set the "removed flag" gets to return the removed
315 * node to its caller. Note that the replace operation does not need to
316 * iteract with the "removal owner" flag, because it validates that
317 * the "removed" flag is not set before performing its cmpxchg.
d95bd160 318 */
d37166c6 319#define REMOVED_FLAG (1UL << 0)
1ee8f000 320#define BUCKET_FLAG (1UL << 1)
db00ccc3
MD
321#define REMOVAL_OWNER_FLAG (1UL << 2)
322#define FLAGS_MASK ((1UL << 3) - 1)
d37166c6 323
bb7b2f26 324/* Value of the end pointer. Should not interact with flags. */
f9c80341 325#define END_VALUE NULL
bb7b2f26 326
7f52427b
MD
327/*
328 * ht_items_count: Split-counters counting the number of node addition
329 * and removal in the table. Only used if the CDS_LFHT_ACCOUNTING flag
330 * is set at hash table creation.
331 *
332 * These are free-running counters, never reset to zero. They count the
333 * number of add/remove, and trigger every (1 << COUNT_COMMIT_ORDER)
334 * operations to update the global counter. We choose a power-of-2 value
335 * for the trigger to deal with 32 or 64-bit overflow of the counter.
336 */
df44348d 337struct ht_items_count {
860d07e8 338 unsigned long add, del;
df44348d
MD
339} __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
340
7f52427b 341/*
d0ec0ed2 342 * resize_work: Contains arguments passed to worker thread
7f52427b
MD
343 * responsible for performing lazy resize.
344 */
d0ec0ed2
MD
345struct resize_work {
346 struct urcu_work work;
14044b37 347 struct cds_lfht *ht;
abc490a1 348};
2ed95849 349
7f52427b
MD
350/*
351 * partition_resize_work: Contains arguments passed to worker threads
352 * executing the hash table resize on partitions of the hash table
353 * assigned to each processor's worker thread.
354 */
b7d619b0 355struct partition_resize_work {
1af6e26e 356 pthread_t thread_id;
b7d619b0
MD
357 struct cds_lfht *ht;
358 unsigned long i, start, len;
359 void (*fct)(struct cds_lfht *ht, unsigned long i,
360 unsigned long start, unsigned long len);
361};
362
d0ec0ed2
MD
363static struct urcu_workqueue *cds_lfht_workqueue;
364static unsigned long cds_lfht_workqueue_user_count;
365
366/*
367 * Mutex ensuring mutual exclusion between workqueue initialization and
368 * fork handlers. cds_lfht_fork_mutex nests inside call_rcu_mutex.
369 */
370static pthread_mutex_t cds_lfht_fork_mutex = PTHREAD_MUTEX_INITIALIZER;
371
372static struct urcu_atfork cds_lfht_atfork;
373
374/*
375 * atfork handler nesting counters. Handle being registered to many urcu
376 * flavors, thus being possibly invoked more than once in the
377 * pthread_atfork list of callbacks.
378 */
379static int cds_lfht_workqueue_atfork_nesting;
380
381static void cds_lfht_init_worker(const struct rcu_flavor_struct *flavor);
382static void cds_lfht_fini_worker(const struct rcu_flavor_struct *flavor);
383
d7c76f85
MD
384#ifdef CONFIG_CDS_LFHT_ITER_DEBUG
385
386static
387void cds_lfht_iter_debug_set_ht(struct cds_lfht *ht, struct cds_lfht_iter *iter)
388{
389 iter->lfht = ht;
390}
391
392#define cds_lfht_iter_debug_assert(...) assert(__VA_ARGS__)
393
394#else
395
396static
397void cds_lfht_iter_debug_set_ht(struct cds_lfht *ht, struct cds_lfht_iter *iter)
398{
399}
400
401#define cds_lfht_iter_debug_assert(...)
402
403#endif
404
abc490a1
MD
405/*
406 * Algorithm to reverse bits in a word by lookup table, extended to
407 * 64-bit words.
f9830efd 408 * Source:
abc490a1 409 * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable
f9830efd 410 * Originally from Public Domain.
abc490a1
MD
411 */
412
67ecffc0 413static const uint8_t BitReverseTable256[256] =
2ed95849 414{
abc490a1
MD
415#define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64
416#define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16)
417#define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 )
418 R6(0), R6(2), R6(1), R6(3)
419};
420#undef R2
421#undef R4
422#undef R6
2ed95849 423
abc490a1
MD
424static
425uint8_t bit_reverse_u8(uint8_t v)
426{
427 return BitReverseTable256[v];
428}
ab7d5fc6 429
95bc7fb9
MD
430#if (CAA_BITS_PER_LONG == 32)
431static
abc490a1
MD
432uint32_t bit_reverse_u32(uint32_t v)
433{
67ecffc0
MD
434 return ((uint32_t) bit_reverse_u8(v) << 24) |
435 ((uint32_t) bit_reverse_u8(v >> 8) << 16) |
436 ((uint32_t) bit_reverse_u8(v >> 16) << 8) |
abc490a1 437 ((uint32_t) bit_reverse_u8(v >> 24));
2ed95849 438}
95bc7fb9
MD
439#else
440static
abc490a1 441uint64_t bit_reverse_u64(uint64_t v)
2ed95849 442{
67ecffc0
MD
443 return ((uint64_t) bit_reverse_u8(v) << 56) |
444 ((uint64_t) bit_reverse_u8(v >> 8) << 48) |
abc490a1
MD
445 ((uint64_t) bit_reverse_u8(v >> 16) << 40) |
446 ((uint64_t) bit_reverse_u8(v >> 24) << 32) |
67ecffc0
MD
447 ((uint64_t) bit_reverse_u8(v >> 32) << 24) |
448 ((uint64_t) bit_reverse_u8(v >> 40) << 16) |
abc490a1
MD
449 ((uint64_t) bit_reverse_u8(v >> 48) << 8) |
450 ((uint64_t) bit_reverse_u8(v >> 56));
451}
95bc7fb9 452#endif
abc490a1
MD
453
454static
455unsigned long bit_reverse_ulong(unsigned long v)
456{
457#if (CAA_BITS_PER_LONG == 32)
458 return bit_reverse_u32(v);
459#else
460 return bit_reverse_u64(v);
461#endif
462}
463
f9830efd 464/*
24365af7
MD
465 * fls: returns the position of the most significant bit.
466 * Returns 0 if no bit is set, else returns the position of the most
467 * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit).
f9830efd 468 */
24365af7
MD
469#if defined(__i386) || defined(__x86_64)
470static inline
471unsigned int fls_u32(uint32_t x)
f9830efd 472{
24365af7
MD
473 int r;
474
e1789ce2 475 __asm__ ("bsrl %1,%0\n\t"
24365af7
MD
476 "jnz 1f\n\t"
477 "movl $-1,%0\n\t"
478 "1:\n\t"
479 : "=r" (r) : "rm" (x));
480 return r + 1;
481}
482#define HAS_FLS_U32
483#endif
484
485#if defined(__x86_64)
486static inline
487unsigned int fls_u64(uint64_t x)
488{
489 long r;
490
e1789ce2 491 __asm__ ("bsrq %1,%0\n\t"
24365af7
MD
492 "jnz 1f\n\t"
493 "movq $-1,%0\n\t"
494 "1:\n\t"
495 : "=r" (r) : "rm" (x));
496 return r + 1;
497}
498#define HAS_FLS_U64
499#endif
500
501#ifndef HAS_FLS_U64
502static __attribute__((unused))
503unsigned int fls_u64(uint64_t x)
504{
505 unsigned int r = 64;
506
507 if (!x)
508 return 0;
509
510 if (!(x & 0xFFFFFFFF00000000ULL)) {
511 x <<= 32;
512 r -= 32;
513 }
514 if (!(x & 0xFFFF000000000000ULL)) {
515 x <<= 16;
516 r -= 16;
517 }
518 if (!(x & 0xFF00000000000000ULL)) {
519 x <<= 8;
520 r -= 8;
521 }
522 if (!(x & 0xF000000000000000ULL)) {
523 x <<= 4;
524 r -= 4;
525 }
526 if (!(x & 0xC000000000000000ULL)) {
527 x <<= 2;
528 r -= 2;
529 }
530 if (!(x & 0x8000000000000000ULL)) {
531 x <<= 1;
532 r -= 1;
533 }
534 return r;
535}
536#endif
537
538#ifndef HAS_FLS_U32
539static __attribute__((unused))
540unsigned int fls_u32(uint32_t x)
541{
542 unsigned int r = 32;
f9830efd 543
24365af7
MD
544 if (!x)
545 return 0;
546 if (!(x & 0xFFFF0000U)) {
547 x <<= 16;
548 r -= 16;
549 }
550 if (!(x & 0xFF000000U)) {
551 x <<= 8;
552 r -= 8;
553 }
554 if (!(x & 0xF0000000U)) {
555 x <<= 4;
556 r -= 4;
557 }
558 if (!(x & 0xC0000000U)) {
559 x <<= 2;
560 r -= 2;
561 }
562 if (!(x & 0x80000000U)) {
563 x <<= 1;
564 r -= 1;
565 }
566 return r;
567}
568#endif
569
5bc6b66f 570unsigned int cds_lfht_fls_ulong(unsigned long x)
f9830efd 571{
6887cc5e 572#if (CAA_BITS_PER_LONG == 32)
24365af7
MD
573 return fls_u32(x);
574#else
575 return fls_u64(x);
576#endif
577}
f9830efd 578
920f8ef6
LJ
579/*
580 * Return the minimum order for which x <= (1UL << order).
581 * Return -1 if x is 0.
582 */
5bc6b66f 583int cds_lfht_get_count_order_u32(uint32_t x)
24365af7 584{
920f8ef6
LJ
585 if (!x)
586 return -1;
24365af7 587
920f8ef6 588 return fls_u32(x - 1);
24365af7
MD
589}
590
920f8ef6
LJ
591/*
592 * Return the minimum order for which x <= (1UL << order).
593 * Return -1 if x is 0.
594 */
5bc6b66f 595int cds_lfht_get_count_order_ulong(unsigned long x)
24365af7 596{
920f8ef6
LJ
597 if (!x)
598 return -1;
24365af7 599
5bc6b66f 600 return cds_lfht_fls_ulong(x - 1);
f9830efd
MD
601}
602
603static
ab65b890 604void cds_lfht_resize_lazy_grow(struct cds_lfht *ht, unsigned long size, int growth);
f9830efd 605
f8994aee 606static
4105056a 607void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
f8994aee
MD
608 unsigned long count);
609
5ffcaeef
MD
610static void mutex_lock(pthread_mutex_t *mutex)
611{
612 int ret;
613
614#ifndef DISTRUST_SIGNALS_EXTREME
615 ret = pthread_mutex_lock(mutex);
616 if (ret)
617 urcu_die(ret);
618#else /* #ifndef DISTRUST_SIGNALS_EXTREME */
619 while ((ret = pthread_mutex_trylock(mutex)) != 0) {
620 if (ret != EBUSY && ret != EINTR)
621 urcu_die(ret);
622 if (CMM_LOAD_SHARED(URCU_TLS(rcu_reader).need_mb)) {
623 cmm_smp_mb();
624 _CMM_STORE_SHARED(URCU_TLS(rcu_reader).need_mb, 0);
625 cmm_smp_mb();
626 }
627 (void) poll(NULL, 0, 10);
628 }
629#endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
630}
631
632static void mutex_unlock(pthread_mutex_t *mutex)
633{
634 int ret;
635
636 ret = pthread_mutex_unlock(mutex);
637 if (ret)
638 urcu_die(ret);
639}
640
df44348d 641static long nr_cpus_mask = -1;
4c42f1b8 642static long split_count_mask = -1;
e53ab1eb 643static int split_count_order = -1;
4c42f1b8 644
4ddbb355 645#if defined(HAVE_SYSCONF)
4c42f1b8
LJ
646static void ht_init_nr_cpus_mask(void)
647{
648 long maxcpus;
649
650 maxcpus = sysconf(_SC_NPROCESSORS_CONF);
651 if (maxcpus <= 0) {
652 nr_cpus_mask = -2;
653 return;
654 }
655 /*
656 * round up number of CPUs to next power of two, so we
657 * can use & for modulo.
658 */
5bc6b66f 659 maxcpus = 1UL << cds_lfht_get_count_order_ulong(maxcpus);
4c42f1b8
LJ
660 nr_cpus_mask = maxcpus - 1;
661}
4ddbb355
LJ
662#else /* #if defined(HAVE_SYSCONF) */
663static void ht_init_nr_cpus_mask(void)
664{
665 nr_cpus_mask = -2;
666}
667#endif /* #else #if defined(HAVE_SYSCONF) */
df44348d
MD
668
669static
5afadd12 670void alloc_split_items_count(struct cds_lfht *ht)
df44348d 671{
4c42f1b8
LJ
672 if (nr_cpus_mask == -1) {
673 ht_init_nr_cpus_mask();
4ddbb355
LJ
674 if (nr_cpus_mask < 0)
675 split_count_mask = DEFAULT_SPLIT_COUNT_MASK;
676 else
677 split_count_mask = nr_cpus_mask;
e53ab1eb
MD
678 split_count_order =
679 cds_lfht_get_count_order_ulong(split_count_mask + 1);
df44348d 680 }
4c42f1b8 681
4ddbb355 682 assert(split_count_mask >= 0);
5afadd12
LJ
683
684 if (ht->flags & CDS_LFHT_ACCOUNTING) {
95bc7fb9
MD
685 ht->split_count = calloc(split_count_mask + 1,
686 sizeof(struct ht_items_count));
5afadd12
LJ
687 assert(ht->split_count);
688 } else {
689 ht->split_count = NULL;
690 }
df44348d
MD
691}
692
693static
5afadd12 694void free_split_items_count(struct cds_lfht *ht)
df44348d 695{
5afadd12 696 poison_free(ht->split_count);
df44348d
MD
697}
698
699static
14360f1c 700int ht_get_split_count_index(unsigned long hash)
df44348d
MD
701{
702 int cpu;
703
4c42f1b8 704 assert(split_count_mask >= 0);
a47dd11c 705 cpu = urcu_sched_getcpu();
8ed51e04 706 if (caa_unlikely(cpu < 0))
14360f1c 707 return hash & split_count_mask;
df44348d 708 else
4c42f1b8 709 return cpu & split_count_mask;
df44348d
MD
710}
711
712static
14360f1c 713void ht_count_add(struct cds_lfht *ht, unsigned long size, unsigned long hash)
df44348d 714{
83e334d0 715 unsigned long split_count, count;
4c42f1b8 716 int index;
df44348d 717
8ed51e04 718 if (caa_unlikely(!ht->split_count))
3171717f 719 return;
14360f1c 720 index = ht_get_split_count_index(hash);
4c42f1b8 721 split_count = uatomic_add_return(&ht->split_count[index].add, 1);
314558bf
MD
722 if (caa_likely(split_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))
723 return;
724 /* Only if number of add multiple of 1UL << COUNT_COMMIT_ORDER */
725
726 dbg_printf("add split count %lu\n", split_count);
727 count = uatomic_add_return(&ht->count,
728 1UL << COUNT_COMMIT_ORDER);
4c299dcb 729 if (caa_likely(count & (count - 1)))
314558bf
MD
730 return;
731 /* Only if global count is power of 2 */
732
733 if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) < size)
734 return;
83e334d0 735 dbg_printf("add set global %lu\n", count);
314558bf
MD
736 cds_lfht_resize_lazy_count(ht, size,
737 count >> (CHAIN_LEN_TARGET - 1));
df44348d
MD
738}
739
740static
14360f1c 741void ht_count_del(struct cds_lfht *ht, unsigned long size, unsigned long hash)
df44348d 742{
83e334d0 743 unsigned long split_count, count;
4c42f1b8 744 int index;
df44348d 745
8ed51e04 746 if (caa_unlikely(!ht->split_count))
3171717f 747 return;
14360f1c 748 index = ht_get_split_count_index(hash);
4c42f1b8 749 split_count = uatomic_add_return(&ht->split_count[index].del, 1);
314558bf
MD
750 if (caa_likely(split_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))
751 return;
752 /* Only if number of deletes multiple of 1UL << COUNT_COMMIT_ORDER */
753
754 dbg_printf("del split count %lu\n", split_count);
755 count = uatomic_add_return(&ht->count,
756 -(1UL << COUNT_COMMIT_ORDER));
4c299dcb 757 if (caa_likely(count & (count - 1)))
314558bf
MD
758 return;
759 /* Only if global count is power of 2 */
760
761 if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) >= size)
762 return;
763 dbg_printf("del set global %ld\n", count);
764 /*
765 * Don't shrink table if the number of nodes is below a
766 * certain threshold.
767 */
768 if (count < (1UL << COUNT_COMMIT_ORDER) * (split_count_mask + 1))
769 return;
770 cds_lfht_resize_lazy_count(ht, size,
771 count >> (CHAIN_LEN_TARGET - 1));
df44348d
MD
772}
773
f9830efd 774static
4105056a 775void check_resize(struct cds_lfht *ht, unsigned long size, uint32_t chain_len)
f9830efd 776{
f8994aee
MD
777 unsigned long count;
778
b8af5011
MD
779 if (!(ht->flags & CDS_LFHT_AUTO_RESIZE))
780 return;
f8994aee
MD
781 count = uatomic_read(&ht->count);
782 /*
783 * Use bucket-local length for small table expand and for
784 * environments lacking per-cpu data support.
785 */
e53ab1eb 786 if (count >= (1UL << (COUNT_COMMIT_ORDER + split_count_order)))
f8994aee 787 return;
24365af7 788 if (chain_len > 100)
f0c29ed7 789 dbg_printf("WARNING: large chain length: %u.\n",
24365af7 790 chain_len);
e53ab1eb
MD
791 if (chain_len >= CHAIN_LEN_RESIZE_THRESHOLD) {
792 int growth;
793
794 /*
795 * Ideal growth calculated based on chain length.
796 */
797 growth = cds_lfht_get_count_order_u32(chain_len
798 - (CHAIN_LEN_TARGET - 1));
799 if ((ht->flags & CDS_LFHT_ACCOUNTING)
800 && (size << growth)
801 >= (1UL << (COUNT_COMMIT_ORDER
802 + split_count_order))) {
803 /*
804 * If ideal growth expands the hash table size
805 * beyond the "small hash table" sizes, use the
806 * maximum small hash table size to attempt
807 * expanding the hash table. This only applies
808 * when node accounting is available, otherwise
809 * the chain length is used to expand the hash
810 * table in every case.
811 */
812 growth = COUNT_COMMIT_ORDER + split_count_order
813 - cds_lfht_get_count_order_ulong(size);
814 if (growth <= 0)
815 return;
816 }
817 cds_lfht_resize_lazy_grow(ht, size, growth);
818 }
f9830efd
MD
819}
820
abc490a1 821static
14044b37 822struct cds_lfht_node *clear_flag(struct cds_lfht_node *node)
abc490a1 823{
14044b37 824 return (struct cds_lfht_node *) (((unsigned long) node) & ~FLAGS_MASK);
abc490a1
MD
825}
826
827static
14044b37 828int is_removed(struct cds_lfht_node *node)
abc490a1 829{
d37166c6 830 return ((unsigned long) node) & REMOVED_FLAG;
abc490a1
MD
831}
832
f5596c94 833static
1ee8f000 834int is_bucket(struct cds_lfht_node *node)
f5596c94 835{
1ee8f000 836 return ((unsigned long) node) & BUCKET_FLAG;
f5596c94
MD
837}
838
839static
1ee8f000 840struct cds_lfht_node *flag_bucket(struct cds_lfht_node *node)
f5596c94 841{
1ee8f000 842 return (struct cds_lfht_node *) (((unsigned long) node) | BUCKET_FLAG);
f5596c94 843}
bb7b2f26 844
db00ccc3
MD
845static
846int is_removal_owner(struct cds_lfht_node *node)
847{
848 return ((unsigned long) node) & REMOVAL_OWNER_FLAG;
849}
850
851static
852struct cds_lfht_node *flag_removal_owner(struct cds_lfht_node *node)
853{
854 return (struct cds_lfht_node *) (((unsigned long) node) | REMOVAL_OWNER_FLAG);
855}
856
71bb3aca
MD
857static
858struct cds_lfht_node *flag_removed_or_removal_owner(struct cds_lfht_node *node)
859{
860 return (struct cds_lfht_node *) (((unsigned long) node) | REMOVED_FLAG | REMOVAL_OWNER_FLAG);
861}
862
bb7b2f26
MD
863static
864struct cds_lfht_node *get_end(void)
865{
866 return (struct cds_lfht_node *) END_VALUE;
867}
868
869static
870int is_end(struct cds_lfht_node *node)
871{
872 return clear_flag(node) == (struct cds_lfht_node *) END_VALUE;
873}
874
abc490a1 875static
ab65b890
LJ
876unsigned long _uatomic_xchg_monotonic_increase(unsigned long *ptr,
877 unsigned long v)
abc490a1
MD
878{
879 unsigned long old1, old2;
880
881 old1 = uatomic_read(ptr);
882 do {
883 old2 = old1;
884 if (old2 >= v)
f9830efd 885 return old2;
abc490a1 886 } while ((old1 = uatomic_cmpxchg(ptr, old2, v)) != old2);
ab65b890 887 return old2;
abc490a1
MD
888}
889
48f1b16d
LJ
890static
891void cds_lfht_alloc_bucket_table(struct cds_lfht *ht, unsigned long order)
892{
0b6aa001 893 return ht->mm->alloc_bucket_table(ht, order);
48f1b16d
LJ
894}
895
896/*
897 * cds_lfht_free_bucket_table() should be called with decreasing order.
898 * When cds_lfht_free_bucket_table(0) is called, it means the whole
899 * lfht is destroyed.
900 */
901static
902void cds_lfht_free_bucket_table(struct cds_lfht *ht, unsigned long order)
903{
0b6aa001 904 return ht->mm->free_bucket_table(ht, order);
48f1b16d
LJ
905}
906
9d72a73f
LJ
907static inline
908struct cds_lfht_node *bucket_at(struct cds_lfht *ht, unsigned long index)
f4a9cc0b 909{
0b6aa001 910 return ht->bucket_at(ht, index);
f4a9cc0b
LJ
911}
912
9d72a73f
LJ
913static inline
914struct cds_lfht_node *lookup_bucket(struct cds_lfht *ht, unsigned long size,
915 unsigned long hash)
916{
917 assert(size > 0);
918 return bucket_at(ht, hash & (size - 1));
919}
920
273399de
MD
921/*
922 * Remove all logically deleted nodes from a bucket up to a certain node key.
923 */
924static
1ee8f000 925void _cds_lfht_gc_bucket(struct cds_lfht_node *bucket, struct cds_lfht_node *node)
273399de 926{
14044b37 927 struct cds_lfht_node *iter_prev, *iter, *next, *new_next;
273399de 928
1ee8f000
LJ
929 assert(!is_bucket(bucket));
930 assert(!is_removed(bucket));
2f943cd7 931 assert(!is_removal_owner(bucket));
1ee8f000 932 assert(!is_bucket(node));
c90201ac 933 assert(!is_removed(node));
2f943cd7 934 assert(!is_removal_owner(node));
273399de 935 for (;;) {
1ee8f000
LJ
936 iter_prev = bucket;
937 /* We can always skip the bucket node initially */
04db56f8 938 iter = rcu_dereference(iter_prev->next);
b4cb483f 939 assert(!is_removed(iter));
2f943cd7 940 assert(!is_removal_owner(iter));
04db56f8 941 assert(iter_prev->reverse_hash <= node->reverse_hash);
bd4db153 942 /*
1ee8f000 943 * We should never be called with bucket (start of chain)
bd4db153
MD
944 * and logically removed node (end of path compression
945 * marker) being the actual same node. This would be a
946 * bug in the algorithm implementation.
947 */
1ee8f000 948 assert(bucket != node);
273399de 949 for (;;) {
8ed51e04 950 if (caa_unlikely(is_end(iter)))
f9c80341 951 return;
04db56f8 952 if (caa_likely(clear_flag(iter)->reverse_hash > node->reverse_hash))
f9c80341 953 return;
04db56f8 954 next = rcu_dereference(clear_flag(iter)->next);
8ed51e04 955 if (caa_likely(is_removed(next)))
273399de 956 break;
b453eae1 957 iter_prev = clear_flag(iter);
273399de
MD
958 iter = next;
959 }
b198f0fd 960 assert(!is_removed(iter));
2f943cd7 961 assert(!is_removal_owner(iter));
1ee8f000
LJ
962 if (is_bucket(iter))
963 new_next = flag_bucket(clear_flag(next));
f5596c94
MD
964 else
965 new_next = clear_flag(next);
04db56f8 966 (void) uatomic_cmpxchg(&iter_prev->next, iter, new_next);
273399de
MD
967 }
968}
969
9357c415
MD
970static
971int _cds_lfht_replace(struct cds_lfht *ht, unsigned long size,
972 struct cds_lfht_node *old_node,
3fb86f26 973 struct cds_lfht_node *old_next,
9357c415
MD
974 struct cds_lfht_node *new_node)
975{
04db56f8 976 struct cds_lfht_node *bucket, *ret_next;
9357c415
MD
977
978 if (!old_node) /* Return -ENOENT if asked to replace NULL node */
7801dadd 979 return -ENOENT;
9357c415
MD
980
981 assert(!is_removed(old_node));
2f943cd7 982 assert(!is_removal_owner(old_node));
1ee8f000 983 assert(!is_bucket(old_node));
9357c415 984 assert(!is_removed(new_node));
2f943cd7 985 assert(!is_removal_owner(new_node));
1ee8f000 986 assert(!is_bucket(new_node));
9357c415 987 assert(new_node != old_node);
3fb86f26 988 for (;;) {
9357c415 989 /* Insert after node to be replaced */
9357c415
MD
990 if (is_removed(old_next)) {
991 /*
992 * Too late, the old node has been removed under us
993 * between lookup and replace. Fail.
994 */
7801dadd 995 return -ENOENT;
9357c415 996 }
feda2722
LJ
997 assert(old_next == clear_flag(old_next));
998 assert(new_node != old_next);
71bb3aca
MD
999 /*
1000 * REMOVAL_OWNER flag is _NEVER_ set before the REMOVED
1001 * flag. It is either set atomically at the same time
1002 * (replace) or after (del).
1003 */
1004 assert(!is_removal_owner(old_next));
feda2722 1005 new_node->next = old_next;
9357c415
MD
1006 /*
1007 * Here is the whole trick for lock-free replace: we add
1008 * the replacement node _after_ the node we want to
1009 * replace by atomically setting its next pointer at the
1010 * same time we set its removal flag. Given that
1011 * the lookups/get next use an iterator aware of the
1012 * next pointer, they will either skip the old node due
1013 * to the removal flag and see the new node, or use
1014 * the old node, but will not see the new one.
db00ccc3
MD
1015 * This is a replacement of a node with another node
1016 * that has the same value: we are therefore not
71bb3aca
MD
1017 * removing a value from the hash table. We set both the
1018 * REMOVED and REMOVAL_OWNER flags atomically so we own
1019 * the node after successful cmpxchg.
9357c415 1020 */
04db56f8 1021 ret_next = uatomic_cmpxchg(&old_node->next,
71bb3aca 1022 old_next, flag_removed_or_removal_owner(new_node));
3fb86f26 1023 if (ret_next == old_next)
7801dadd 1024 break; /* We performed the replacement. */
3fb86f26
LJ
1025 old_next = ret_next;
1026 }
9357c415 1027
9357c415
MD
1028 /*
1029 * Ensure that the old node is not visible to readers anymore:
1030 * lookup for the node, and remove it (along with any other
1031 * logically removed node) if found.
1032 */
04db56f8
LJ
1033 bucket = lookup_bucket(ht, size, bit_reverse_ulong(old_node->reverse_hash));
1034 _cds_lfht_gc_bucket(bucket, new_node);
7801dadd 1035
a85eff52 1036 assert(is_removed(CMM_LOAD_SHARED(old_node->next)));
7801dadd 1037 return 0;
9357c415
MD
1038}
1039
83beee94
MD
1040/*
1041 * A non-NULL unique_ret pointer uses the "add unique" (or uniquify) add
1042 * mode. A NULL unique_ret allows creation of duplicate keys.
1043 */
abc490a1 1044static
83beee94 1045void _cds_lfht_add(struct cds_lfht *ht,
91a75cc5 1046 unsigned long hash,
0422d92c 1047 cds_lfht_match_fct match,
996ff57c 1048 const void *key,
83beee94
MD
1049 unsigned long size,
1050 struct cds_lfht_node *node,
1051 struct cds_lfht_iter *unique_ret,
1ee8f000 1052 int bucket_flag)
abc490a1 1053{
14044b37 1054 struct cds_lfht_node *iter_prev, *iter, *next, *new_node, *new_next,
960c9e4f 1055 *return_node;
04db56f8 1056 struct cds_lfht_node *bucket;
abc490a1 1057
1ee8f000 1058 assert(!is_bucket(node));
c90201ac 1059 assert(!is_removed(node));
2f943cd7 1060 assert(!is_removal_owner(node));
91a75cc5 1061 bucket = lookup_bucket(ht, size, hash);
abc490a1 1062 for (;;) {
adc0de68 1063 uint32_t chain_len = 0;
abc490a1 1064
11519af6
MD
1065 /*
1066 * iter_prev points to the non-removed node prior to the
1067 * insert location.
11519af6 1068 */
04db56f8 1069 iter_prev = bucket;
1ee8f000 1070 /* We can always skip the bucket node initially */
04db56f8
LJ
1071 iter = rcu_dereference(iter_prev->next);
1072 assert(iter_prev->reverse_hash <= node->reverse_hash);
abc490a1 1073 for (;;) {
8ed51e04 1074 if (caa_unlikely(is_end(iter)))
273399de 1075 goto insert;
04db56f8 1076 if (caa_likely(clear_flag(iter)->reverse_hash > node->reverse_hash))
273399de 1077 goto insert;
238cc06e 1078
1ee8f000
LJ
1079 /* bucket node is the first node of the identical-hash-value chain */
1080 if (bucket_flag && clear_flag(iter)->reverse_hash == node->reverse_hash)
194fdbd1 1081 goto insert;
238cc06e 1082
04db56f8 1083 next = rcu_dereference(clear_flag(iter)->next);
8ed51e04 1084 if (caa_unlikely(is_removed(next)))
9dba85be 1085 goto gc_node;
238cc06e
LJ
1086
1087 /* uniquely add */
83beee94 1088 if (unique_ret
1ee8f000 1089 && !is_bucket(next)
04db56f8 1090 && clear_flag(iter)->reverse_hash == node->reverse_hash) {
d7c76f85
MD
1091 struct cds_lfht_iter d_iter = {
1092 .node = node,
1093 .next = iter,
1094#ifdef CONFIG_CDS_LFHT_ITER_DEBUG
1095 .lfht = ht,
1096#endif
1097 };
238cc06e
LJ
1098
1099 /*
1100 * uniquely adding inserts the node as the first
1101 * node of the identical-hash-value node chain.
1102 *
1103 * This semantic ensures no duplicated keys
1104 * should ever be observable in the table
1f67ba50
MD
1105 * (including traversing the table node by
1106 * node by forward iterations)
238cc06e 1107 */
04db56f8 1108 cds_lfht_next_duplicate(ht, match, key, &d_iter);
238cc06e
LJ
1109 if (!d_iter.node)
1110 goto insert;
1111
1112 *unique_ret = d_iter;
83beee94 1113 return;
48ed1c18 1114 }
238cc06e 1115
11519af6 1116 /* Only account for identical reverse hash once */
04db56f8 1117 if (iter_prev->reverse_hash != clear_flag(iter)->reverse_hash
1ee8f000 1118 && !is_bucket(next))
4105056a 1119 check_resize(ht, size, ++chain_len);
11519af6 1120 iter_prev = clear_flag(iter);
273399de 1121 iter = next;
abc490a1 1122 }
48ed1c18 1123
273399de 1124 insert:
7ec59d3b 1125 assert(node != clear_flag(iter));
11519af6 1126 assert(!is_removed(iter_prev));
2f943cd7 1127 assert(!is_removal_owner(iter_prev));
c90201ac 1128 assert(!is_removed(iter));
2f943cd7 1129 assert(!is_removal_owner(iter));
f000907d 1130 assert(iter_prev != node);
1ee8f000 1131 if (!bucket_flag)
04db56f8 1132 node->next = clear_flag(iter);
f9c80341 1133 else
1ee8f000
LJ
1134 node->next = flag_bucket(clear_flag(iter));
1135 if (is_bucket(iter))
1136 new_node = flag_bucket(node);
f5596c94
MD
1137 else
1138 new_node = node;
04db56f8 1139 if (uatomic_cmpxchg(&iter_prev->next, iter,
48ed1c18 1140 new_node) != iter) {
273399de 1141 continue; /* retry */
48ed1c18 1142 } else {
83beee94 1143 return_node = node;
960c9e4f 1144 goto end;
48ed1c18
MD
1145 }
1146
9dba85be
MD
1147 gc_node:
1148 assert(!is_removed(iter));
2f943cd7 1149 assert(!is_removal_owner(iter));
1ee8f000
LJ
1150 if (is_bucket(iter))
1151 new_next = flag_bucket(clear_flag(next));
f5596c94
MD
1152 else
1153 new_next = clear_flag(next);
04db56f8 1154 (void) uatomic_cmpxchg(&iter_prev->next, iter, new_next);
273399de 1155 /* retry */
464a1ec9 1156 }
9357c415 1157end:
83beee94
MD
1158 if (unique_ret) {
1159 unique_ret->node = return_node;
1160 /* unique_ret->next left unset, never used. */
1161 }
abc490a1 1162}
464a1ec9 1163
abc490a1 1164static
860d07e8 1165int _cds_lfht_del(struct cds_lfht *ht, unsigned long size,
b65ec430 1166 struct cds_lfht_node *node)
abc490a1 1167{
db00ccc3 1168 struct cds_lfht_node *bucket, *next;
5e28c532 1169
9357c415 1170 if (!node) /* Return -ENOENT if asked to delete NULL node */
743f9143 1171 return -ENOENT;
9357c415 1172
7ec59d3b 1173 /* logically delete the node */
1ee8f000 1174 assert(!is_bucket(node));
c90201ac 1175 assert(!is_removed(node));
db00ccc3 1176 assert(!is_removal_owner(node));
48ed1c18 1177
db00ccc3
MD
1178 /*
1179 * We are first checking if the node had previously been
1180 * logically removed (this check is not atomic with setting the
1181 * logical removal flag). Return -ENOENT if the node had
1182 * previously been removed.
1183 */
a85eff52 1184 next = CMM_LOAD_SHARED(node->next); /* next is not dereferenced */
db00ccc3
MD
1185 if (caa_unlikely(is_removed(next)))
1186 return -ENOENT;
b65ec430 1187 assert(!is_bucket(next));
196f4fab
MD
1188 /*
1189 * The del operation semantic guarantees a full memory barrier
1190 * before the uatomic_or atomic commit of the deletion flag.
1191 */
1192 cmm_smp_mb__before_uatomic_or();
db00ccc3
MD
1193 /*
1194 * We set the REMOVED_FLAG unconditionally. Note that there may
1195 * be more than one concurrent thread setting this flag.
1196 * Knowing which wins the race will be known after the garbage
1197 * collection phase, stay tuned!
1198 */
1199 uatomic_or(&node->next, REMOVED_FLAG);
7ec59d3b 1200 /* We performed the (logical) deletion. */
7ec59d3b
MD
1201
1202 /*
1203 * Ensure that the node is not visible to readers anymore: lookup for
273399de
MD
1204 * the node, and remove it (along with any other logically removed node)
1205 * if found.
11519af6 1206 */
04db56f8
LJ
1207 bucket = lookup_bucket(ht, size, bit_reverse_ulong(node->reverse_hash));
1208 _cds_lfht_gc_bucket(bucket, node);
743f9143 1209
a85eff52 1210 assert(is_removed(CMM_LOAD_SHARED(node->next)));
db00ccc3
MD
1211 /*
1212 * Last phase: atomically exchange node->next with a version
1213 * having "REMOVAL_OWNER_FLAG" set. If the returned node->next
1214 * pointer did _not_ have "REMOVAL_OWNER_FLAG" set, we now own
1215 * the node and win the removal race.
1216 * It is interesting to note that all "add" paths are forbidden
1217 * to change the next pointer starting from the point where the
1218 * REMOVED_FLAG is set, so here using a read, followed by a
1219 * xchg() suffice to guarantee that the xchg() will ever only
1220 * set the "REMOVAL_OWNER_FLAG" (or change nothing if the flag
1221 * was already set).
1222 */
1223 if (!is_removal_owner(uatomic_xchg(&node->next,
1224 flag_removal_owner(node->next))))
1225 return 0;
1226 else
1227 return -ENOENT;
abc490a1 1228}
2ed95849 1229
b7d619b0
MD
1230static
1231void *partition_resize_thread(void *arg)
1232{
1233 struct partition_resize_work *work = arg;
1234
7b17c13e 1235 work->ht->flavor->register_thread();
b7d619b0 1236 work->fct(work->ht, work->i, work->start, work->len);
7b17c13e 1237 work->ht->flavor->unregister_thread();
b7d619b0
MD
1238 return NULL;
1239}
1240
1241static
1242void partition_resize_helper(struct cds_lfht *ht, unsigned long i,
1243 unsigned long len,
1244 void (*fct)(struct cds_lfht *ht, unsigned long i,
1245 unsigned long start, unsigned long len))
1246{
e54ec2f5 1247 unsigned long partition_len, start = 0;
b7d619b0 1248 struct partition_resize_work *work;
83e334d0
MJ
1249 int ret;
1250 unsigned long thread, nr_threads;
b7d619b0 1251
d7f3ba4c
EW
1252 assert(nr_cpus_mask != -1);
1253 if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD)
1254 goto fallback;
1255
6083a889
MD
1256 /*
1257 * Note: nr_cpus_mask + 1 is always power of 2.
1258 * We spawn just the number of threads we need to satisfy the minimum
1259 * partition size, up to the number of CPUs in the system.
1260 */
91452a6a 1261 if (nr_cpus_mask > 0) {
83e334d0 1262 nr_threads = min_t(unsigned long, nr_cpus_mask + 1,
91452a6a
MD
1263 len >> MIN_PARTITION_PER_THREAD_ORDER);
1264 } else {
1265 nr_threads = 1;
1266 }
5bc6b66f 1267 partition_len = len >> cds_lfht_get_count_order_ulong(nr_threads);
6083a889 1268 work = calloc(nr_threads, sizeof(*work));
7c75d498
EW
1269 if (!work) {
1270 dbg_printf("error allocating for resize, single-threading\n");
1271 goto fallback;
1272 }
6083a889
MD
1273 for (thread = 0; thread < nr_threads; thread++) {
1274 work[thread].ht = ht;
1275 work[thread].i = i;
1276 work[thread].len = partition_len;
1277 work[thread].start = thread * partition_len;
1278 work[thread].fct = fct;
1af6e26e 1279 ret = pthread_create(&(work[thread].thread_id), ht->resize_attr,
6083a889 1280 partition_resize_thread, &work[thread]);
e54ec2f5
EW
1281 if (ret == EAGAIN) {
1282 /*
1283 * Out of resources: wait and join the threads
1284 * we've created, then handle leftovers.
1285 */
1286 dbg_printf("error spawning for resize, single-threading\n");
1287 start = work[thread].start;
1288 len -= start;
1289 nr_threads = thread;
1290 break;
1291 }
b7d619b0
MD
1292 assert(!ret);
1293 }
6083a889 1294 for (thread = 0; thread < nr_threads; thread++) {
1af6e26e 1295 ret = pthread_join(work[thread].thread_id, NULL);
b7d619b0
MD
1296 assert(!ret);
1297 }
1298 free(work);
e54ec2f5
EW
1299
1300 /*
1301 * A pthread_create failure above will either lead in us having
1302 * no threads to join or starting at a non-zero offset,
1303 * fallback to single thread processing of leftovers.
1304 */
1305 if (start == 0 && nr_threads > 0)
1306 return;
7c75d498 1307fallback:
e54ec2f5 1308 fct(ht, i, start, len);
b7d619b0
MD
1309}
1310
e8de508e
MD
1311/*
1312 * Holding RCU read lock to protect _cds_lfht_add against memory
d0ec0ed2 1313 * reclaim that could be performed by other worker threads (ABA
e8de508e 1314 * problem).
9ee0fc9a 1315 *
b7d619b0 1316 * When we reach a certain length, we can split this population phase over
9ee0fc9a
MD
1317 * many worker threads, based on the number of CPUs available in the system.
1318 * This should therefore take care of not having the expand lagging behind too
1319 * many concurrent insertion threads by using the scheduler's ability to
1ee8f000 1320 * schedule bucket node population fairly with insertions.
e8de508e 1321 */
4105056a 1322static
b7d619b0
MD
1323void init_table_populate_partition(struct cds_lfht *ht, unsigned long i,
1324 unsigned long start, unsigned long len)
4105056a 1325{
9d72a73f 1326 unsigned long j, size = 1UL << (i - 1);
4105056a 1327
d0d8f9aa 1328 assert(i > MIN_TABLE_ORDER);
7b17c13e 1329 ht->flavor->read_lock();
9d72a73f
LJ
1330 for (j = size + start; j < size + start + len; j++) {
1331 struct cds_lfht_node *new_node = bucket_at(ht, j);
1332
1333 assert(j >= size && j < (size << 1));
1334 dbg_printf("init populate: order %lu index %lu hash %lu\n",
1335 i, j, j);
1336 new_node->reverse_hash = bit_reverse_ulong(j);
91a75cc5 1337 _cds_lfht_add(ht, j, NULL, NULL, size, new_node, NULL, 1);
4105056a 1338 }
7b17c13e 1339 ht->flavor->read_unlock();
b7d619b0
MD
1340}
1341
1342static
1343void init_table_populate(struct cds_lfht *ht, unsigned long i,
1344 unsigned long len)
1345{
b7d619b0 1346 partition_resize_helper(ht, i, len, init_table_populate_partition);
4105056a
MD
1347}
1348
abc490a1 1349static
4105056a 1350void init_table(struct cds_lfht *ht,
93d46c39 1351 unsigned long first_order, unsigned long last_order)
24365af7 1352{
93d46c39 1353 unsigned long i;
24365af7 1354
93d46c39
LJ
1355 dbg_printf("init table: first_order %lu last_order %lu\n",
1356 first_order, last_order);
d0d8f9aa 1357 assert(first_order > MIN_TABLE_ORDER);
93d46c39 1358 for (i = first_order; i <= last_order; i++) {
4105056a 1359 unsigned long len;
24365af7 1360
4f6e90b7 1361 len = 1UL << (i - 1);
f0c29ed7 1362 dbg_printf("init order %lu len: %lu\n", i, len);
4d676753
MD
1363
1364 /* Stop expand if the resize target changes under us */
7b3893e4 1365 if (CMM_LOAD_SHARED(ht->resize_target) < (1UL << i))
4d676753
MD
1366 break;
1367
48f1b16d 1368 cds_lfht_alloc_bucket_table(ht, i);
4105056a 1369
4105056a 1370 /*
1ee8f000
LJ
1371 * Set all bucket nodes reverse hash values for a level and
1372 * link all bucket nodes into the table.
4105056a 1373 */
dc1da8f6 1374 init_table_populate(ht, i, len);
4105056a 1375
f9c80341
MD
1376 /*
1377 * Update table size.
1378 */
1379 cmm_smp_wmb(); /* populate data before RCU size */
7b3893e4 1380 CMM_STORE_SHARED(ht->size, 1UL << i);
f9c80341 1381
4f6e90b7 1382 dbg_printf("init new size: %lu\n", 1UL << i);
4105056a
MD
1383 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1384 break;
1385 }
1386}
1387
e8de508e
MD
1388/*
1389 * Holding RCU read lock to protect _cds_lfht_remove against memory
d0ec0ed2 1390 * reclaim that could be performed by other worker threads (ABA
e8de508e
MD
1391 * problem).
1392 * For a single level, we logically remove and garbage collect each node.
1393 *
1394 * As a design choice, we perform logical removal and garbage collection on a
1395 * node-per-node basis to simplify this algorithm. We also assume keeping good
1396 * cache locality of the operation would overweight possible performance gain
1397 * that could be achieved by batching garbage collection for multiple levels.
1398 * However, this would have to be justified by benchmarks.
1399 *
1400 * Concurrent removal and add operations are helping us perform garbage
1401 * collection of logically removed nodes. We guarantee that all logically
d0ec0ed2
MD
1402 * removed nodes have been garbage-collected (unlinked) before work
1403 * enqueue is invoked to free a hole level of bucket nodes (after a
1404 * grace period).
e8de508e 1405 *
1f67ba50
MD
1406 * Logical removal and garbage collection can therefore be done in batch
1407 * or on a node-per-node basis, as long as the guarantee above holds.
9ee0fc9a 1408 *
b7d619b0
MD
1409 * When we reach a certain length, we can split this removal over many worker
1410 * threads, based on the number of CPUs available in the system. This should
1411 * take care of not letting resize process lag behind too many concurrent
9ee0fc9a 1412 * updater threads actively inserting into the hash table.
e8de508e 1413 */
4105056a 1414static
b7d619b0
MD
1415void remove_table_partition(struct cds_lfht *ht, unsigned long i,
1416 unsigned long start, unsigned long len)
4105056a 1417{
9d72a73f 1418 unsigned long j, size = 1UL << (i - 1);
4105056a 1419
d0d8f9aa 1420 assert(i > MIN_TABLE_ORDER);
7b17c13e 1421 ht->flavor->read_lock();
9d72a73f 1422 for (j = size + start; j < size + start + len; j++) {
2e2ce1e9
LJ
1423 struct cds_lfht_node *fini_bucket = bucket_at(ht, j);
1424 struct cds_lfht_node *parent_bucket = bucket_at(ht, j - size);
9d72a73f
LJ
1425
1426 assert(j >= size && j < (size << 1));
1427 dbg_printf("remove entry: order %lu index %lu hash %lu\n",
1428 i, j, j);
2e2ce1e9
LJ
1429 /* Set the REMOVED_FLAG to freeze the ->next for gc */
1430 uatomic_or(&fini_bucket->next, REMOVED_FLAG);
1431 _cds_lfht_gc_bucket(parent_bucket, fini_bucket);
abc490a1 1432 }
7b17c13e 1433 ht->flavor->read_unlock();
b7d619b0
MD
1434}
1435
1436static
1437void remove_table(struct cds_lfht *ht, unsigned long i, unsigned long len)
1438{
b7d619b0 1439 partition_resize_helper(ht, i, len, remove_table_partition);
2ed95849
MD
1440}
1441
61adb337
MD
1442/*
1443 * fini_table() is never called for first_order == 0, which is why
1444 * free_by_rcu_order == 0 can be used as criterion to know if free must
1445 * be called.
1446 */
1475579c 1447static
4105056a 1448void fini_table(struct cds_lfht *ht,
93d46c39 1449 unsigned long first_order, unsigned long last_order)
1475579c 1450{
83e334d0 1451 unsigned long free_by_rcu_order = 0, i;
1475579c 1452
93d46c39
LJ
1453 dbg_printf("fini table: first_order %lu last_order %lu\n",
1454 first_order, last_order);
d0d8f9aa 1455 assert(first_order > MIN_TABLE_ORDER);
93d46c39 1456 for (i = last_order; i >= first_order; i--) {
4105056a 1457 unsigned long len;
1475579c 1458
4f6e90b7 1459 len = 1UL << (i - 1);
e15df1cc 1460 dbg_printf("fini order %ld len: %lu\n", i, len);
4105056a 1461
4d676753 1462 /* Stop shrink if the resize target changes under us */
7b3893e4 1463 if (CMM_LOAD_SHARED(ht->resize_target) > (1UL << (i - 1)))
4d676753
MD
1464 break;
1465
1466 cmm_smp_wmb(); /* populate data before RCU size */
7b3893e4 1467 CMM_STORE_SHARED(ht->size, 1UL << (i - 1));
4d676753
MD
1468
1469 /*
1470 * We need to wait for all add operations to reach Q.S. (and
1471 * thus use the new table for lookups) before we can start
1ee8f000 1472 * releasing the old bucket nodes. Otherwise their lookup will
4d676753
MD
1473 * return a logically removed node as insert position.
1474 */
7b17c13e 1475 ht->flavor->update_synchronize_rcu();
48f1b16d
LJ
1476 if (free_by_rcu_order)
1477 cds_lfht_free_bucket_table(ht, free_by_rcu_order);
4d676753 1478
21263e21 1479 /*
1ee8f000
LJ
1480 * Set "removed" flag in bucket nodes about to be removed.
1481 * Unlink all now-logically-removed bucket node pointers.
4105056a
MD
1482 * Concurrent add/remove operation are helping us doing
1483 * the gc.
21263e21 1484 */
4105056a
MD
1485 remove_table(ht, i, len);
1486
48f1b16d 1487 free_by_rcu_order = i;
4105056a
MD
1488
1489 dbg_printf("fini new size: %lu\n", 1UL << i);
1475579c
MD
1490 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1491 break;
1492 }
0d14ceb2 1493
48f1b16d 1494 if (free_by_rcu_order) {
7b17c13e 1495 ht->flavor->update_synchronize_rcu();
48f1b16d 1496 cds_lfht_free_bucket_table(ht, free_by_rcu_order);
0d14ceb2 1497 }
1475579c
MD
1498}
1499
83e334d0
MJ
1500/*
1501 * Never called with size < 1.
1502 */
ff0d69de 1503static
1ee8f000 1504void cds_lfht_create_bucket(struct cds_lfht *ht, unsigned long size)
ff0d69de 1505{
04db56f8 1506 struct cds_lfht_node *prev, *node;
9d72a73f 1507 unsigned long order, len, i;
83e334d0 1508 int bucket_order;
ff0d69de 1509
48f1b16d 1510 cds_lfht_alloc_bucket_table(ht, 0);
ff0d69de 1511
9d72a73f
LJ
1512 dbg_printf("create bucket: order 0 index 0 hash 0\n");
1513 node = bucket_at(ht, 0);
1514 node->next = flag_bucket(get_end());
1515 node->reverse_hash = 0;
ff0d69de 1516
83e334d0
MJ
1517 bucket_order = cds_lfht_get_count_order_ulong(size);
1518 assert(bucket_order >= 0);
1519
1520 for (order = 1; order < (unsigned long) bucket_order + 1; order++) {
ff0d69de 1521 len = 1UL << (order - 1);
48f1b16d 1522 cds_lfht_alloc_bucket_table(ht, order);
ff0d69de 1523
9d72a73f
LJ
1524 for (i = 0; i < len; i++) {
1525 /*
1526 * Now, we are trying to init the node with the
1527 * hash=(len+i) (which is also a bucket with the
1528 * index=(len+i)) and insert it into the hash table,
1529 * so this node has to be inserted after the bucket
1530 * with the index=(len+i)&(len-1)=i. And because there
1531 * is no other non-bucket node nor bucket node with
1532 * larger index/hash inserted, so the bucket node
1533 * being inserted should be inserted directly linked
1534 * after the bucket node with index=i.
1535 */
1536 prev = bucket_at(ht, i);
1537 node = bucket_at(ht, len + i);
ff0d69de 1538
1ee8f000 1539 dbg_printf("create bucket: order %lu index %lu hash %lu\n",
9d72a73f
LJ
1540 order, len + i, len + i);
1541 node->reverse_hash = bit_reverse_ulong(len + i);
1542
1543 /* insert after prev */
1544 assert(is_bucket(prev->next));
ff0d69de 1545 node->next = prev->next;
1ee8f000 1546 prev->next = flag_bucket(node);
ff0d69de
LJ
1547 }
1548 }
1549}
1550
99ab1528
MJ
1551#if (CAA_BITS_PER_LONG > 32)
1552/*
1553 * For 64-bit architectures, with max number of buckets small enough not to
1554 * use the entire 64-bit memory mapping space (and allowing a fair number of
1555 * hash table instances), use the mmap allocator, which is faster. Otherwise,
1556 * fallback to the order allocator.
1557 */
1558static
1559const struct cds_lfht_mm_type *get_mm_type(unsigned long max_nr_buckets)
1560{
1561 if (max_nr_buckets && max_nr_buckets <= (1ULL << 32))
1562 return &cds_lfht_mm_mmap;
1563 else
1564 return &cds_lfht_mm_order;
1565}
1566#else
1567/*
1568 * For 32-bit architectures, use the order allocator.
1569 */
1570static
1571const struct cds_lfht_mm_type *get_mm_type(unsigned long max_nr_buckets)
1572{
1573 return &cds_lfht_mm_order;
1574}
1575#endif
1576
0422d92c 1577struct cds_lfht *_cds_lfht_new(unsigned long init_size,
0722081a 1578 unsigned long min_nr_alloc_buckets,
747d725c 1579 unsigned long max_nr_buckets,
b8af5011 1580 int flags,
0b6aa001 1581 const struct cds_lfht_mm_type *mm,
7b17c13e 1582 const struct rcu_flavor_struct *flavor,
b7d619b0 1583 pthread_attr_t *attr)
abc490a1 1584{
14044b37 1585 struct cds_lfht *ht;
24365af7 1586 unsigned long order;
abc490a1 1587
0722081a
LJ
1588 /* min_nr_alloc_buckets must be power of two */
1589 if (!min_nr_alloc_buckets || (min_nr_alloc_buckets & (min_nr_alloc_buckets - 1)))
5488222b 1590 return NULL;
747d725c 1591
8129be4e 1592 /* init_size must be power of two */
5488222b 1593 if (!init_size || (init_size & (init_size - 1)))
8129be4e 1594 return NULL;
747d725c 1595
c1888f3a
MD
1596 /*
1597 * Memory management plugin default.
1598 */
99ab1528
MJ
1599 if (!mm)
1600 mm = get_mm_type(max_nr_buckets);
c1888f3a 1601
0b6aa001
LJ
1602 /* max_nr_buckets == 0 for order based mm means infinite */
1603 if (mm == &cds_lfht_mm_order && !max_nr_buckets)
747d725c
LJ
1604 max_nr_buckets = 1UL << (MAX_TABLE_ORDER - 1);
1605
1606 /* max_nr_buckets must be power of two */
1607 if (!max_nr_buckets || (max_nr_buckets & (max_nr_buckets - 1)))
1608 return NULL;
1609
d0ec0ed2
MD
1610 if (flags & CDS_LFHT_AUTO_RESIZE)
1611 cds_lfht_init_worker(flavor);
1612
0722081a 1613 min_nr_alloc_buckets = max(min_nr_alloc_buckets, MIN_TABLE_SIZE);
d0d8f9aa 1614 init_size = max(init_size, MIN_TABLE_SIZE);
747d725c
LJ
1615 max_nr_buckets = max(max_nr_buckets, min_nr_alloc_buckets);
1616 init_size = min(init_size, max_nr_buckets);
0b6aa001
LJ
1617
1618 ht = mm->alloc_cds_lfht(min_nr_alloc_buckets, max_nr_buckets);
b7d619b0 1619 assert(ht);
0b6aa001
LJ
1620 assert(ht->mm == mm);
1621 assert(ht->bucket_at == mm->bucket_at);
1622
b5d6b20f 1623 ht->flags = flags;
7b17c13e 1624 ht->flavor = flavor;
b7d619b0 1625 ht->resize_attr = attr;
5afadd12 1626 alloc_split_items_count(ht);
abc490a1
MD
1627 /* this mutex should not nest in read-side C.S. */
1628 pthread_mutex_init(&ht->resize_mutex, NULL);
5bc6b66f 1629 order = cds_lfht_get_count_order_ulong(init_size);
7b3893e4 1630 ht->resize_target = 1UL << order;
1ee8f000 1631 cds_lfht_create_bucket(ht, 1UL << order);
7b3893e4 1632 ht->size = 1UL << order;
abc490a1
MD
1633 return ht;
1634}
1635
6f554439 1636void cds_lfht_lookup(struct cds_lfht *ht, unsigned long hash,
996ff57c 1637 cds_lfht_match_fct match, const void *key,
6f554439 1638 struct cds_lfht_iter *iter)
2ed95849 1639{
04db56f8 1640 struct cds_lfht_node *node, *next, *bucket;
0422d92c 1641 unsigned long reverse_hash, size;
2ed95849 1642
d7c76f85
MD
1643 cds_lfht_iter_debug_set_ht(ht, iter);
1644
abc490a1 1645 reverse_hash = bit_reverse_ulong(hash);
464a1ec9 1646
7b3893e4 1647 size = rcu_dereference(ht->size);
04db56f8 1648 bucket = lookup_bucket(ht, size, hash);
1ee8f000 1649 /* We can always skip the bucket node initially */
04db56f8 1650 node = rcu_dereference(bucket->next);
bb7b2f26 1651 node = clear_flag(node);
2ed95849 1652 for (;;) {
8ed51e04 1653 if (caa_unlikely(is_end(node))) {
96ad1112 1654 node = next = NULL;
abc490a1 1655 break;
bb7b2f26 1656 }
04db56f8 1657 if (caa_unlikely(node->reverse_hash > reverse_hash)) {
96ad1112 1658 node = next = NULL;
abc490a1 1659 break;
2ed95849 1660 }
04db56f8 1661 next = rcu_dereference(node->next);
7f52427b 1662 assert(node == clear_flag(node));
8ed51e04 1663 if (caa_likely(!is_removed(next))
1ee8f000 1664 && !is_bucket(next)
04db56f8 1665 && node->reverse_hash == reverse_hash
0422d92c 1666 && caa_likely(match(node, key))) {
273399de 1667 break;
2ed95849 1668 }
1b81fe1a 1669 node = clear_flag(next);
2ed95849 1670 }
a85eff52 1671 assert(!node || !is_bucket(CMM_LOAD_SHARED(node->next)));
adc0de68
MD
1672 iter->node = node;
1673 iter->next = next;
abc490a1 1674}
e0ba718a 1675
0422d92c 1676void cds_lfht_next_duplicate(struct cds_lfht *ht, cds_lfht_match_fct match,
996ff57c 1677 const void *key, struct cds_lfht_iter *iter)
a481e5ff 1678{
adc0de68 1679 struct cds_lfht_node *node, *next;
a481e5ff 1680 unsigned long reverse_hash;
a481e5ff 1681
d7c76f85 1682 cds_lfht_iter_debug_assert(ht == iter->lfht);
adc0de68 1683 node = iter->node;
04db56f8 1684 reverse_hash = node->reverse_hash;
adc0de68 1685 next = iter->next;
a481e5ff
MD
1686 node = clear_flag(next);
1687
1688 for (;;) {
8ed51e04 1689 if (caa_unlikely(is_end(node))) {
96ad1112 1690 node = next = NULL;
a481e5ff 1691 break;
bb7b2f26 1692 }
04db56f8 1693 if (caa_unlikely(node->reverse_hash > reverse_hash)) {
96ad1112 1694 node = next = NULL;
a481e5ff
MD
1695 break;
1696 }
04db56f8 1697 next = rcu_dereference(node->next);
8ed51e04 1698 if (caa_likely(!is_removed(next))
1ee8f000 1699 && !is_bucket(next)
04db56f8 1700 && caa_likely(match(node, key))) {
a481e5ff
MD
1701 break;
1702 }
1703 node = clear_flag(next);
1704 }
a85eff52 1705 assert(!node || !is_bucket(CMM_LOAD_SHARED(node->next)));
adc0de68
MD
1706 iter->node = node;
1707 iter->next = next;
a481e5ff
MD
1708}
1709
4e9b9fbf
MD
1710void cds_lfht_next(struct cds_lfht *ht, struct cds_lfht_iter *iter)
1711{
1712 struct cds_lfht_node *node, *next;
1713
d7c76f85 1714 cds_lfht_iter_debug_assert(ht == iter->lfht);
853395e1 1715 node = clear_flag(iter->next);
4e9b9fbf 1716 for (;;) {
8ed51e04 1717 if (caa_unlikely(is_end(node))) {
4e9b9fbf
MD
1718 node = next = NULL;
1719 break;
1720 }
04db56f8 1721 next = rcu_dereference(node->next);
8ed51e04 1722 if (caa_likely(!is_removed(next))
1ee8f000 1723 && !is_bucket(next)) {
4e9b9fbf
MD
1724 break;
1725 }
1726 node = clear_flag(next);
1727 }
a85eff52 1728 assert(!node || !is_bucket(CMM_LOAD_SHARED(node->next)));
4e9b9fbf
MD
1729 iter->node = node;
1730 iter->next = next;
1731}
1732
1733void cds_lfht_first(struct cds_lfht *ht, struct cds_lfht_iter *iter)
1734{
d7c76f85 1735 cds_lfht_iter_debug_set_ht(ht, iter);
4e9b9fbf 1736 /*
1ee8f000 1737 * Get next after first bucket node. The first bucket node is the
4e9b9fbf
MD
1738 * first node of the linked list.
1739 */
9d72a73f 1740 iter->next = bucket_at(ht, 0)->next;
4e9b9fbf
MD
1741 cds_lfht_next(ht, iter);
1742}
1743
0422d92c
MD
1744void cds_lfht_add(struct cds_lfht *ht, unsigned long hash,
1745 struct cds_lfht_node *node)
abc490a1 1746{
0422d92c 1747 unsigned long size;
ab7d5fc6 1748
709bacf9 1749 node->reverse_hash = bit_reverse_ulong(hash);
7b3893e4 1750 size = rcu_dereference(ht->size);
91a75cc5 1751 _cds_lfht_add(ht, hash, NULL, NULL, size, node, NULL, 0);
14360f1c 1752 ht_count_add(ht, size, hash);
3eca1b8c
MD
1753}
1754
14044b37 1755struct cds_lfht_node *cds_lfht_add_unique(struct cds_lfht *ht,
6f554439 1756 unsigned long hash,
0422d92c 1757 cds_lfht_match_fct match,
996ff57c 1758 const void *key,
48ed1c18 1759 struct cds_lfht_node *node)
3eca1b8c 1760{
0422d92c 1761 unsigned long size;
83beee94 1762 struct cds_lfht_iter iter;
3eca1b8c 1763
709bacf9 1764 node->reverse_hash = bit_reverse_ulong(hash);
7b3893e4 1765 size = rcu_dereference(ht->size);
91a75cc5 1766 _cds_lfht_add(ht, hash, match, key, size, node, &iter, 0);
83beee94 1767 if (iter.node == node)
14360f1c 1768 ht_count_add(ht, size, hash);
83beee94 1769 return iter.node;
2ed95849
MD
1770}
1771
9357c415 1772struct cds_lfht_node *cds_lfht_add_replace(struct cds_lfht *ht,
6f554439 1773 unsigned long hash,
0422d92c 1774 cds_lfht_match_fct match,
996ff57c 1775 const void *key,
48ed1c18
MD
1776 struct cds_lfht_node *node)
1777{
0422d92c 1778 unsigned long size;
83beee94 1779 struct cds_lfht_iter iter;
48ed1c18 1780
709bacf9 1781 node->reverse_hash = bit_reverse_ulong(hash);
7b3893e4 1782 size = rcu_dereference(ht->size);
83beee94 1783 for (;;) {
91a75cc5 1784 _cds_lfht_add(ht, hash, match, key, size, node, &iter, 0);
83beee94 1785 if (iter.node == node) {
14360f1c 1786 ht_count_add(ht, size, hash);
83beee94
MD
1787 return NULL;
1788 }
1789
1790 if (!_cds_lfht_replace(ht, size, iter.node, iter.next, node))
1791 return iter.node;
1792 }
48ed1c18
MD
1793}
1794
2e79c445
MD
1795int cds_lfht_replace(struct cds_lfht *ht,
1796 struct cds_lfht_iter *old_iter,
1797 unsigned long hash,
1798 cds_lfht_match_fct match,
1799 const void *key,
9357c415
MD
1800 struct cds_lfht_node *new_node)
1801{
1802 unsigned long size;
1803
709bacf9 1804 new_node->reverse_hash = bit_reverse_ulong(hash);
2e79c445
MD
1805 if (!old_iter->node)
1806 return -ENOENT;
1807 if (caa_unlikely(old_iter->node->reverse_hash != new_node->reverse_hash))
1808 return -EINVAL;
1809 if (caa_unlikely(!match(old_iter->node, key)))
1810 return -EINVAL;
7b3893e4 1811 size = rcu_dereference(ht->size);
9357c415
MD
1812 return _cds_lfht_replace(ht, size, old_iter->node, old_iter->next,
1813 new_node);
1814}
1815
bc8c3c74 1816int cds_lfht_del(struct cds_lfht *ht, struct cds_lfht_node *node)
2ed95849 1817{
95bc7fb9 1818 unsigned long size;
df44348d 1819 int ret;
abc490a1 1820
7b3893e4 1821 size = rcu_dereference(ht->size);
bc8c3c74 1822 ret = _cds_lfht_del(ht, size, node);
14360f1c 1823 if (!ret) {
95bc7fb9
MD
1824 unsigned long hash;
1825
bc8c3c74 1826 hash = bit_reverse_ulong(node->reverse_hash);
14360f1c
LJ
1827 ht_count_del(ht, size, hash);
1828 }
df44348d 1829 return ret;
2ed95849 1830}
ab7d5fc6 1831
df55172a
MD
1832int cds_lfht_is_node_deleted(struct cds_lfht_node *node)
1833{
a85eff52 1834 return is_removed(CMM_LOAD_SHARED(node->next));
df55172a
MD
1835}
1836
abc490a1 1837static
1ee8f000 1838int cds_lfht_delete_bucket(struct cds_lfht *ht)
674f7a69 1839{
14044b37 1840 struct cds_lfht_node *node;
4105056a 1841 unsigned long order, i, size;
674f7a69 1842
abc490a1 1843 /* Check that the table is empty */
9d72a73f 1844 node = bucket_at(ht, 0);
abc490a1 1845 do {
04db56f8 1846 node = clear_flag(node)->next;
1ee8f000 1847 if (!is_bucket(node))
abc490a1 1848 return -EPERM;
273399de 1849 assert(!is_removed(node));
2f943cd7 1850 assert(!is_removal_owner(node));
bb7b2f26 1851 } while (!is_end(node));
4105056a
MD
1852 /*
1853 * size accessed without rcu_dereference because hash table is
1854 * being destroyed.
1855 */
7b3893e4 1856 size = ht->size;
1f67ba50 1857 /* Internal sanity check: all nodes left should be buckets */
48f1b16d
LJ
1858 for (i = 0; i < size; i++) {
1859 node = bucket_at(ht, i);
1860 dbg_printf("delete bucket: index %lu expected hash %lu hash %lu\n",
1861 i, i, bit_reverse_ulong(node->reverse_hash));
1862 assert(is_bucket(node->next));
1863 }
24365af7 1864
5bc6b66f 1865 for (order = cds_lfht_get_count_order_ulong(size); (long)order >= 0; order--)
48f1b16d 1866 cds_lfht_free_bucket_table(ht, order);
5488222b 1867
abc490a1 1868 return 0;
674f7a69
MD
1869}
1870
1871/*
1872 * Should only be called when no more concurrent readers nor writers can
1873 * possibly access the table.
1874 */
b7d619b0 1875int cds_lfht_destroy(struct cds_lfht *ht, pthread_attr_t **attr)
674f7a69 1876{
d0ec0ed2
MD
1877 int ret;
1878
1879 if (ht->flags & CDS_LFHT_AUTO_RESIZE) {
1880 /* Cancel ongoing resize operations. */
1881 _CMM_STORE_SHARED(ht->in_progress_destroy, 1);
1882 /* Wait for in-flight resize operations to complete */
1883 urcu_workqueue_flush_queued_work(cds_lfht_workqueue);
10e68472 1884 }
1ee8f000 1885 ret = cds_lfht_delete_bucket(ht);
abc490a1
MD
1886 if (ret)
1887 return ret;
5afadd12 1888 free_split_items_count(ht);
b7d619b0
MD
1889 if (attr)
1890 *attr = ht->resize_attr;
59629f09
MD
1891 ret = pthread_mutex_destroy(&ht->resize_mutex);
1892 if (ret)
1893 ret = -EBUSY;
d0ec0ed2
MD
1894 if (ht->flags & CDS_LFHT_AUTO_RESIZE)
1895 cds_lfht_fini_worker(ht->flavor);
98808fb1 1896 poison_free(ht);
5e28c532 1897 return ret;
674f7a69
MD
1898}
1899
14044b37 1900void cds_lfht_count_nodes(struct cds_lfht *ht,
d933dd0e 1901 long *approx_before,
273399de 1902 unsigned long *count,
d933dd0e 1903 long *approx_after)
273399de 1904{
14044b37 1905 struct cds_lfht_node *node, *next;
caf3653d 1906 unsigned long nr_bucket = 0, nr_removed = 0;
273399de 1907
7ed7682f 1908 *approx_before = 0;
5afadd12 1909 if (ht->split_count) {
973e5e1b
MD
1910 int i;
1911
4c42f1b8
LJ
1912 for (i = 0; i < split_count_mask + 1; i++) {
1913 *approx_before += uatomic_read(&ht->split_count[i].add);
1914 *approx_before -= uatomic_read(&ht->split_count[i].del);
973e5e1b
MD
1915 }
1916 }
1917
273399de 1918 *count = 0;
273399de 1919
1ee8f000 1920 /* Count non-bucket nodes in the table */
9d72a73f 1921 node = bucket_at(ht, 0);
273399de 1922 do {
04db56f8 1923 next = rcu_dereference(node->next);
b198f0fd 1924 if (is_removed(next)) {
1ee8f000 1925 if (!is_bucket(next))
caf3653d 1926 (nr_removed)++;
973e5e1b 1927 else
1ee8f000
LJ
1928 (nr_bucket)++;
1929 } else if (!is_bucket(next))
273399de 1930 (*count)++;
24365af7 1931 else
1ee8f000 1932 (nr_bucket)++;
273399de 1933 node = clear_flag(next);
bb7b2f26 1934 } while (!is_end(node));
caf3653d 1935 dbg_printf("number of logically removed nodes: %lu\n", nr_removed);
1ee8f000 1936 dbg_printf("number of bucket nodes: %lu\n", nr_bucket);
7ed7682f 1937 *approx_after = 0;
5afadd12 1938 if (ht->split_count) {
973e5e1b
MD
1939 int i;
1940
4c42f1b8
LJ
1941 for (i = 0; i < split_count_mask + 1; i++) {
1942 *approx_after += uatomic_read(&ht->split_count[i].add);
1943 *approx_after -= uatomic_read(&ht->split_count[i].del);
973e5e1b
MD
1944 }
1945 }
273399de
MD
1946}
1947
1475579c 1948/* called with resize mutex held */
abc490a1 1949static
4105056a 1950void _do_cds_lfht_grow(struct cds_lfht *ht,
1475579c 1951 unsigned long old_size, unsigned long new_size)
abc490a1 1952{
1475579c 1953 unsigned long old_order, new_order;
1475579c 1954
5bc6b66f
MD
1955 old_order = cds_lfht_get_count_order_ulong(old_size);
1956 new_order = cds_lfht_get_count_order_ulong(new_size);
1a401918
LJ
1957 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1958 old_size, old_order, new_size, new_order);
1475579c 1959 assert(new_size > old_size);
93d46c39 1960 init_table(ht, old_order + 1, new_order);
abc490a1
MD
1961}
1962
1963/* called with resize mutex held */
1964static
4105056a 1965void _do_cds_lfht_shrink(struct cds_lfht *ht,
1475579c 1966 unsigned long old_size, unsigned long new_size)
464a1ec9 1967{
1475579c 1968 unsigned long old_order, new_order;
464a1ec9 1969
d0d8f9aa 1970 new_size = max(new_size, MIN_TABLE_SIZE);
5bc6b66f
MD
1971 old_order = cds_lfht_get_count_order_ulong(old_size);
1972 new_order = cds_lfht_get_count_order_ulong(new_size);
1a401918
LJ
1973 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1974 old_size, old_order, new_size, new_order);
1475579c 1975 assert(new_size < old_size);
1475579c 1976
1ee8f000 1977 /* Remove and unlink all bucket nodes to remove. */
93d46c39 1978 fini_table(ht, new_order + 1, old_order);
464a1ec9
MD
1979}
1980
1475579c
MD
1981
1982/* called with resize mutex held */
1983static
1984void _do_cds_lfht_resize(struct cds_lfht *ht)
1985{
1986 unsigned long new_size, old_size;
4105056a
MD
1987
1988 /*
1989 * Resize table, re-do if the target size has changed under us.
1990 */
1991 do {
d2be3620
MD
1992 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1993 break;
7b3893e4
LJ
1994 ht->resize_initiated = 1;
1995 old_size = ht->size;
1996 new_size = CMM_LOAD_SHARED(ht->resize_target);
4105056a
MD
1997 if (old_size < new_size)
1998 _do_cds_lfht_grow(ht, old_size, new_size);
1999 else if (old_size > new_size)
2000 _do_cds_lfht_shrink(ht, old_size, new_size);
7b3893e4 2001 ht->resize_initiated = 0;
4105056a
MD
2002 /* write resize_initiated before read resize_target */
2003 cmm_smp_mb();
7b3893e4 2004 } while (ht->size != CMM_LOAD_SHARED(ht->resize_target));
1475579c
MD
2005}
2006
abc490a1 2007static
ab65b890 2008unsigned long resize_target_grow(struct cds_lfht *ht, unsigned long new_size)
464a1ec9 2009{
7b3893e4 2010 return _uatomic_xchg_monotonic_increase(&ht->resize_target, new_size);
464a1ec9
MD
2011}
2012
1475579c 2013static
4105056a 2014void resize_target_update_count(struct cds_lfht *ht,
b8af5011 2015 unsigned long count)
1475579c 2016{
d0d8f9aa 2017 count = max(count, MIN_TABLE_SIZE);
747d725c 2018 count = min(count, ht->max_nr_buckets);
7b3893e4 2019 uatomic_set(&ht->resize_target, count);
1475579c
MD
2020}
2021
2022void cds_lfht_resize(struct cds_lfht *ht, unsigned long new_size)
464a1ec9 2023{
10e68472
MD
2024 resize_target_update_count(ht, new_size);
2025 CMM_STORE_SHARED(ht->resize_initiated, 1);
5ffcaeef 2026 mutex_lock(&ht->resize_mutex);
1475579c 2027 _do_cds_lfht_resize(ht);
5ffcaeef 2028 mutex_unlock(&ht->resize_mutex);
abc490a1 2029}
464a1ec9 2030
abc490a1 2031static
d0ec0ed2 2032void do_resize_cb(struct urcu_work *work)
abc490a1 2033{
d0ec0ed2
MD
2034 struct resize_work *resize_work =
2035 caa_container_of(work, struct resize_work, work);
2036 struct cds_lfht *ht = resize_work->ht;
abc490a1 2037
d0ec0ed2 2038 ht->flavor->register_thread();
5ffcaeef 2039 mutex_lock(&ht->resize_mutex);
14044b37 2040 _do_cds_lfht_resize(ht);
5ffcaeef 2041 mutex_unlock(&ht->resize_mutex);
d0ec0ed2 2042 ht->flavor->unregister_thread();
98808fb1 2043 poison_free(work);
464a1ec9
MD
2044}
2045
abc490a1 2046static
f1f119ee 2047void __cds_lfht_resize_lazy_launch(struct cds_lfht *ht)
ab7d5fc6 2048{
d0ec0ed2 2049 struct resize_work *work;
abc490a1 2050
4105056a
MD
2051 /* Store resize_target before read resize_initiated */
2052 cmm_smp_mb();
7b3893e4 2053 if (!CMM_LOAD_SHARED(ht->resize_initiated)) {
ed35e6d8 2054 if (CMM_LOAD_SHARED(ht->in_progress_destroy)) {
59290e9d 2055 return;
ed35e6d8 2056 }
f9830efd 2057 work = malloc(sizeof(*work));
741f378e
MD
2058 if (work == NULL) {
2059 dbg_printf("error allocating resize work, bailing out\n");
741f378e
MD
2060 return;
2061 }
f9830efd 2062 work->ht = ht;
d0ec0ed2
MD
2063 urcu_workqueue_queue_work(cds_lfht_workqueue,
2064 &work->work, do_resize_cb);
7b3893e4 2065 CMM_STORE_SHARED(ht->resize_initiated, 1);
f9830efd 2066 }
ab7d5fc6 2067}
3171717f 2068
f1f119ee
LJ
2069static
2070void cds_lfht_resize_lazy_grow(struct cds_lfht *ht, unsigned long size, int growth)
2071{
2072 unsigned long target_size = size << growth;
2073
747d725c 2074 target_size = min(target_size, ht->max_nr_buckets);
f1f119ee
LJ
2075 if (resize_target_grow(ht, target_size) >= target_size)
2076 return;
2077
2078 __cds_lfht_resize_lazy_launch(ht);
2079}
2080
89bb121d
LJ
2081/*
2082 * We favor grow operations over shrink. A shrink operation never occurs
2083 * if a grow operation is queued for lazy execution. A grow operation
2084 * cancels any pending shrink lazy execution.
2085 */
3171717f 2086static
4105056a 2087void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
3171717f
MD
2088 unsigned long count)
2089{
b8af5011
MD
2090 if (!(ht->flags & CDS_LFHT_AUTO_RESIZE))
2091 return;
d0d8f9aa 2092 count = max(count, MIN_TABLE_SIZE);
747d725c 2093 count = min(count, ht->max_nr_buckets);
89bb121d
LJ
2094 if (count == size)
2095 return; /* Already the right size, no resize needed */
2096 if (count > size) { /* lazy grow */
2097 if (resize_target_grow(ht, count) >= count)
2098 return;
2099 } else { /* lazy shrink */
2100 for (;;) {
2101 unsigned long s;
2102
7b3893e4 2103 s = uatomic_cmpxchg(&ht->resize_target, size, count);
89bb121d
LJ
2104 if (s == size)
2105 break; /* no resize needed */
2106 if (s > size)
2107 return; /* growing is/(was just) in progress */
2108 if (s <= count)
2109 return; /* some other thread do shrink */
2110 size = s;
2111 }
2112 }
f1f119ee 2113 __cds_lfht_resize_lazy_launch(ht);
3171717f 2114}
d0ec0ed2 2115
d0ec0ed2
MD
2116static void cds_lfht_before_fork(void *priv)
2117{
2118 if (cds_lfht_workqueue_atfork_nesting++)
2119 return;
2120 mutex_lock(&cds_lfht_fork_mutex);
2121 if (!cds_lfht_workqueue)
2122 return;
2123 urcu_workqueue_pause_worker(cds_lfht_workqueue);
2124}
2125
2126static void cds_lfht_after_fork_parent(void *priv)
2127{
2128 if (--cds_lfht_workqueue_atfork_nesting)
2129 return;
2130 if (!cds_lfht_workqueue)
2131 goto end;
2132 urcu_workqueue_resume_worker(cds_lfht_workqueue);
2133end:
2134 mutex_unlock(&cds_lfht_fork_mutex);
2135}
2136
2137static void cds_lfht_after_fork_child(void *priv)
2138{
2139 if (--cds_lfht_workqueue_atfork_nesting)
2140 return;
2141 if (!cds_lfht_workqueue)
2142 goto end;
2143 urcu_workqueue_create_worker(cds_lfht_workqueue);
2144end:
2145 mutex_unlock(&cds_lfht_fork_mutex);
2146}
2147
2148static struct urcu_atfork cds_lfht_atfork = {
2149 .before_fork = cds_lfht_before_fork,
2150 .after_fork_parent = cds_lfht_after_fork_parent,
2151 .after_fork_child = cds_lfht_after_fork_child,
2152};
2153
2154/* Block all signals to ensure we don't disturb the application. */
2155static void cds_lfht_worker_init(struct urcu_workqueue *workqueue,
2156 void *priv)
2157{
2158 int ret;
2159 sigset_t mask;
2160
2161 /* Block signal for entire process, so only our thread processes it. */
2162 ret = sigfillset(&mask);
2163 if (ret)
2164 urcu_die(errno);
2165 ret = pthread_sigmask(SIG_BLOCK, &mask, NULL);
2166 if (ret)
2167 urcu_die(ret);
2168}
2169
2170static void cds_lfht_init_worker(const struct rcu_flavor_struct *flavor)
2171{
2172 flavor->register_rculfhash_atfork(&cds_lfht_atfork);
2173
2174 mutex_lock(&cds_lfht_fork_mutex);
2175 if (cds_lfht_workqueue_user_count++)
2176 goto end;
2177 cds_lfht_workqueue = urcu_workqueue_create(0, -1, NULL,
2178 NULL, cds_lfht_worker_init, NULL, NULL, NULL, NULL, NULL);
2179end:
2180 mutex_unlock(&cds_lfht_fork_mutex);
2181}
2182
2183static void cds_lfht_fini_worker(const struct rcu_flavor_struct *flavor)
2184{
2185 mutex_lock(&cds_lfht_fork_mutex);
2186 if (--cds_lfht_workqueue_user_count)
2187 goto end;
2188 urcu_workqueue_destroy(cds_lfht_workqueue);
2189 cds_lfht_workqueue = NULL;
2190end:
2191 mutex_unlock(&cds_lfht_fork_mutex);
2192
2193 flavor->unregister_rculfhash_atfork(&cds_lfht_atfork);
2194}
This page took 0.284162 seconds and 4 git commands to generate.