Fix: mixup between URCU_WORKQUEUE_RT and URCU_CALL_RCU_RT
[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"
95747f9e 269#include <urcu-pointer.h>
abc490a1 270#include <urcu-call-rcu.h>
7b17c13e 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"
44395fb7 282
f8994aee 283/*
4c42f1b8 284 * Split-counters lazily update the global counter each 1024
f8994aee
MD
285 * addition/removal. It automatically keeps track of resize required.
286 * We use the bucket length as indicator for need to expand for small
ffa11a18 287 * tables and machines lacking per-cpu data support.
f8994aee
MD
288 */
289#define COUNT_COMMIT_ORDER 10
4ddbb355 290#define DEFAULT_SPLIT_COUNT_MASK 0xFUL
6ea6bc67
MD
291#define CHAIN_LEN_TARGET 1
292#define CHAIN_LEN_RESIZE_THRESHOLD 3
2ed95849 293
cd95516d 294/*
76a73da8 295 * Define the minimum table size.
cd95516d 296 */
d0d8f9aa
LJ
297#define MIN_TABLE_ORDER 0
298#define MIN_TABLE_SIZE (1UL << MIN_TABLE_ORDER)
cd95516d 299
b7d619b0 300/*
1ee8f000 301 * Minimum number of bucket nodes to touch per thread to parallelize grow/shrink.
b7d619b0 302 */
6083a889
MD
303#define MIN_PARTITION_PER_THREAD_ORDER 12
304#define MIN_PARTITION_PER_THREAD (1UL << MIN_PARTITION_PER_THREAD_ORDER)
b7d619b0 305
d95bd160
MD
306/*
307 * The removed flag needs to be updated atomically with the pointer.
48ed1c18 308 * It indicates that no node must attach to the node scheduled for
b198f0fd 309 * removal, and that node garbage collection must be performed.
1ee8f000 310 * The bucket flag does not require to be updated atomically with the
d95bd160 311 * pointer, but it is added as a pointer low bit flag to save space.
1f67ba50
MD
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.
d95bd160 317 */
d37166c6 318#define REMOVED_FLAG (1UL << 0)
1ee8f000 319#define BUCKET_FLAG (1UL << 1)
db00ccc3
MD
320#define REMOVAL_OWNER_FLAG (1UL << 2)
321#define FLAGS_MASK ((1UL << 3) - 1)
d37166c6 322
bb7b2f26 323/* Value of the end pointer. Should not interact with flags. */
f9c80341 324#define END_VALUE NULL
bb7b2f26 325
7f52427b
MD
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 */
df44348d 336struct ht_items_count {
860d07e8 337 unsigned long add, del;
df44348d
MD
338} __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
339
7f52427b 340/*
d0ec0ed2 341 * resize_work: Contains arguments passed to worker thread
7f52427b
MD
342 * responsible for performing lazy resize.
343 */
d0ec0ed2
MD
344struct resize_work {
345 struct urcu_work work;
14044b37 346 struct cds_lfht *ht;
abc490a1 347};
2ed95849 348
7f52427b
MD
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 */
b7d619b0 354struct partition_resize_work {
1af6e26e 355 pthread_t thread_id;
b7d619b0
MD
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
d0ec0ed2
MD
362static struct urcu_workqueue *cds_lfht_workqueue;
363static unsigned long cds_lfht_workqueue_user_count;
364
365/*
366 * Mutex ensuring mutual exclusion between workqueue initialization and
367 * fork handlers. cds_lfht_fork_mutex nests inside call_rcu_mutex.
368 */
369static pthread_mutex_t cds_lfht_fork_mutex = PTHREAD_MUTEX_INITIALIZER;
370
371static struct urcu_atfork cds_lfht_atfork;
372
373/*
374 * atfork handler nesting counters. Handle being registered to many urcu
375 * flavors, thus being possibly invoked more than once in the
376 * pthread_atfork list of callbacks.
377 */
378static int cds_lfht_workqueue_atfork_nesting;
379
380static void cds_lfht_init_worker(const struct rcu_flavor_struct *flavor);
381static void cds_lfht_fini_worker(const struct rcu_flavor_struct *flavor);
382
abc490a1
MD
383/*
384 * Algorithm to reverse bits in a word by lookup table, extended to
385 * 64-bit words.
f9830efd 386 * Source:
abc490a1 387 * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable
f9830efd 388 * Originally from Public Domain.
abc490a1
MD
389 */
390
67ecffc0 391static const uint8_t BitReverseTable256[256] =
2ed95849 392{
abc490a1
MD
393#define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64
394#define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16)
395#define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 )
396 R6(0), R6(2), R6(1), R6(3)
397};
398#undef R2
399#undef R4
400#undef R6
2ed95849 401
abc490a1
MD
402static
403uint8_t bit_reverse_u8(uint8_t v)
404{
405 return BitReverseTable256[v];
406}
ab7d5fc6 407
95bc7fb9
MD
408#if (CAA_BITS_PER_LONG == 32)
409static
abc490a1
MD
410uint32_t bit_reverse_u32(uint32_t v)
411{
67ecffc0
MD
412 return ((uint32_t) bit_reverse_u8(v) << 24) |
413 ((uint32_t) bit_reverse_u8(v >> 8) << 16) |
414 ((uint32_t) bit_reverse_u8(v >> 16) << 8) |
abc490a1 415 ((uint32_t) bit_reverse_u8(v >> 24));
2ed95849 416}
95bc7fb9
MD
417#else
418static
abc490a1 419uint64_t bit_reverse_u64(uint64_t v)
2ed95849 420{
67ecffc0
MD
421 return ((uint64_t) bit_reverse_u8(v) << 56) |
422 ((uint64_t) bit_reverse_u8(v >> 8) << 48) |
abc490a1
MD
423 ((uint64_t) bit_reverse_u8(v >> 16) << 40) |
424 ((uint64_t) bit_reverse_u8(v >> 24) << 32) |
67ecffc0
MD
425 ((uint64_t) bit_reverse_u8(v >> 32) << 24) |
426 ((uint64_t) bit_reverse_u8(v >> 40) << 16) |
abc490a1
MD
427 ((uint64_t) bit_reverse_u8(v >> 48) << 8) |
428 ((uint64_t) bit_reverse_u8(v >> 56));
429}
95bc7fb9 430#endif
abc490a1
MD
431
432static
433unsigned long bit_reverse_ulong(unsigned long v)
434{
435#if (CAA_BITS_PER_LONG == 32)
436 return bit_reverse_u32(v);
437#else
438 return bit_reverse_u64(v);
439#endif
440}
441
f9830efd 442/*
24365af7
MD
443 * fls: returns the position of the most significant bit.
444 * Returns 0 if no bit is set, else returns the position of the most
445 * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit).
f9830efd 446 */
24365af7
MD
447#if defined(__i386) || defined(__x86_64)
448static inline
449unsigned int fls_u32(uint32_t x)
f9830efd 450{
24365af7
MD
451 int r;
452
e1789ce2 453 __asm__ ("bsrl %1,%0\n\t"
24365af7
MD
454 "jnz 1f\n\t"
455 "movl $-1,%0\n\t"
456 "1:\n\t"
457 : "=r" (r) : "rm" (x));
458 return r + 1;
459}
460#define HAS_FLS_U32
461#endif
462
463#if defined(__x86_64)
464static inline
465unsigned int fls_u64(uint64_t x)
466{
467 long r;
468
e1789ce2 469 __asm__ ("bsrq %1,%0\n\t"
24365af7
MD
470 "jnz 1f\n\t"
471 "movq $-1,%0\n\t"
472 "1:\n\t"
473 : "=r" (r) : "rm" (x));
474 return r + 1;
475}
476#define HAS_FLS_U64
477#endif
478
479#ifndef HAS_FLS_U64
480static __attribute__((unused))
481unsigned int fls_u64(uint64_t x)
482{
483 unsigned int r = 64;
484
485 if (!x)
486 return 0;
487
488 if (!(x & 0xFFFFFFFF00000000ULL)) {
489 x <<= 32;
490 r -= 32;
491 }
492 if (!(x & 0xFFFF000000000000ULL)) {
493 x <<= 16;
494 r -= 16;
495 }
496 if (!(x & 0xFF00000000000000ULL)) {
497 x <<= 8;
498 r -= 8;
499 }
500 if (!(x & 0xF000000000000000ULL)) {
501 x <<= 4;
502 r -= 4;
503 }
504 if (!(x & 0xC000000000000000ULL)) {
505 x <<= 2;
506 r -= 2;
507 }
508 if (!(x & 0x8000000000000000ULL)) {
509 x <<= 1;
510 r -= 1;
511 }
512 return r;
513}
514#endif
515
516#ifndef HAS_FLS_U32
517static __attribute__((unused))
518unsigned int fls_u32(uint32_t x)
519{
520 unsigned int r = 32;
f9830efd 521
24365af7
MD
522 if (!x)
523 return 0;
524 if (!(x & 0xFFFF0000U)) {
525 x <<= 16;
526 r -= 16;
527 }
528 if (!(x & 0xFF000000U)) {
529 x <<= 8;
530 r -= 8;
531 }
532 if (!(x & 0xF0000000U)) {
533 x <<= 4;
534 r -= 4;
535 }
536 if (!(x & 0xC0000000U)) {
537 x <<= 2;
538 r -= 2;
539 }
540 if (!(x & 0x80000000U)) {
541 x <<= 1;
542 r -= 1;
543 }
544 return r;
545}
546#endif
547
5bc6b66f 548unsigned int cds_lfht_fls_ulong(unsigned long x)
f9830efd 549{
6887cc5e 550#if (CAA_BITS_PER_LONG == 32)
24365af7
MD
551 return fls_u32(x);
552#else
553 return fls_u64(x);
554#endif
555}
f9830efd 556
920f8ef6
LJ
557/*
558 * Return the minimum order for which x <= (1UL << order).
559 * Return -1 if x is 0.
560 */
5bc6b66f 561int cds_lfht_get_count_order_u32(uint32_t x)
24365af7 562{
920f8ef6
LJ
563 if (!x)
564 return -1;
24365af7 565
920f8ef6 566 return fls_u32(x - 1);
24365af7
MD
567}
568
920f8ef6
LJ
569/*
570 * Return the minimum order for which x <= (1UL << order).
571 * Return -1 if x is 0.
572 */
5bc6b66f 573int cds_lfht_get_count_order_ulong(unsigned long x)
24365af7 574{
920f8ef6
LJ
575 if (!x)
576 return -1;
24365af7 577
5bc6b66f 578 return cds_lfht_fls_ulong(x - 1);
f9830efd
MD
579}
580
581static
ab65b890 582void cds_lfht_resize_lazy_grow(struct cds_lfht *ht, unsigned long size, int growth);
f9830efd 583
f8994aee 584static
4105056a 585void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
f8994aee
MD
586 unsigned long count);
587
5ffcaeef
MD
588static void mutex_lock(pthread_mutex_t *mutex)
589{
590 int ret;
591
592#ifndef DISTRUST_SIGNALS_EXTREME
593 ret = pthread_mutex_lock(mutex);
594 if (ret)
595 urcu_die(ret);
596#else /* #ifndef DISTRUST_SIGNALS_EXTREME */
597 while ((ret = pthread_mutex_trylock(mutex)) != 0) {
598 if (ret != EBUSY && ret != EINTR)
599 urcu_die(ret);
600 if (CMM_LOAD_SHARED(URCU_TLS(rcu_reader).need_mb)) {
601 cmm_smp_mb();
602 _CMM_STORE_SHARED(URCU_TLS(rcu_reader).need_mb, 0);
603 cmm_smp_mb();
604 }
605 (void) poll(NULL, 0, 10);
606 }
607#endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
608}
609
610static void mutex_unlock(pthread_mutex_t *mutex)
611{
612 int ret;
613
614 ret = pthread_mutex_unlock(mutex);
615 if (ret)
616 urcu_die(ret);
617}
618
df44348d 619static long nr_cpus_mask = -1;
4c42f1b8 620static long split_count_mask = -1;
e53ab1eb 621static int split_count_order = -1;
4c42f1b8 622
4ddbb355 623#if defined(HAVE_SYSCONF)
4c42f1b8
LJ
624static void ht_init_nr_cpus_mask(void)
625{
626 long maxcpus;
627
628 maxcpus = sysconf(_SC_NPROCESSORS_CONF);
629 if (maxcpus <= 0) {
630 nr_cpus_mask = -2;
631 return;
632 }
633 /*
634 * round up number of CPUs to next power of two, so we
635 * can use & for modulo.
636 */
5bc6b66f 637 maxcpus = 1UL << cds_lfht_get_count_order_ulong(maxcpus);
4c42f1b8
LJ
638 nr_cpus_mask = maxcpus - 1;
639}
4ddbb355
LJ
640#else /* #if defined(HAVE_SYSCONF) */
641static void ht_init_nr_cpus_mask(void)
642{
643 nr_cpus_mask = -2;
644}
645#endif /* #else #if defined(HAVE_SYSCONF) */
df44348d
MD
646
647static
5afadd12 648void alloc_split_items_count(struct cds_lfht *ht)
df44348d 649{
4c42f1b8
LJ
650 if (nr_cpus_mask == -1) {
651 ht_init_nr_cpus_mask();
4ddbb355
LJ
652 if (nr_cpus_mask < 0)
653 split_count_mask = DEFAULT_SPLIT_COUNT_MASK;
654 else
655 split_count_mask = nr_cpus_mask;
e53ab1eb
MD
656 split_count_order =
657 cds_lfht_get_count_order_ulong(split_count_mask + 1);
df44348d 658 }
4c42f1b8 659
4ddbb355 660 assert(split_count_mask >= 0);
5afadd12
LJ
661
662 if (ht->flags & CDS_LFHT_ACCOUNTING) {
95bc7fb9
MD
663 ht->split_count = calloc(split_count_mask + 1,
664 sizeof(struct ht_items_count));
5afadd12
LJ
665 assert(ht->split_count);
666 } else {
667 ht->split_count = NULL;
668 }
df44348d
MD
669}
670
671static
5afadd12 672void free_split_items_count(struct cds_lfht *ht)
df44348d 673{
5afadd12 674 poison_free(ht->split_count);
df44348d
MD
675}
676
677static
14360f1c 678int ht_get_split_count_index(unsigned long hash)
df44348d
MD
679{
680 int cpu;
681
4c42f1b8 682 assert(split_count_mask >= 0);
a47dd11c 683 cpu = urcu_sched_getcpu();
8ed51e04 684 if (caa_unlikely(cpu < 0))
14360f1c 685 return hash & split_count_mask;
df44348d 686 else
4c42f1b8 687 return cpu & split_count_mask;
df44348d
MD
688}
689
690static
14360f1c 691void ht_count_add(struct cds_lfht *ht, unsigned long size, unsigned long hash)
df44348d 692{
4c42f1b8
LJ
693 unsigned long split_count;
694 int index;
314558bf 695 long count;
df44348d 696
8ed51e04 697 if (caa_unlikely(!ht->split_count))
3171717f 698 return;
14360f1c 699 index = ht_get_split_count_index(hash);
4c42f1b8 700 split_count = uatomic_add_return(&ht->split_count[index].add, 1);
314558bf
MD
701 if (caa_likely(split_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))
702 return;
703 /* Only if number of add multiple of 1UL << COUNT_COMMIT_ORDER */
704
705 dbg_printf("add split count %lu\n", split_count);
706 count = uatomic_add_return(&ht->count,
707 1UL << COUNT_COMMIT_ORDER);
4c299dcb 708 if (caa_likely(count & (count - 1)))
314558bf
MD
709 return;
710 /* Only if global count is power of 2 */
711
712 if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) < size)
713 return;
714 dbg_printf("add set global %ld\n", count);
715 cds_lfht_resize_lazy_count(ht, size,
716 count >> (CHAIN_LEN_TARGET - 1));
df44348d
MD
717}
718
719static
14360f1c 720void ht_count_del(struct cds_lfht *ht, unsigned long size, unsigned long hash)
df44348d 721{
4c42f1b8
LJ
722 unsigned long split_count;
723 int index;
314558bf 724 long count;
df44348d 725
8ed51e04 726 if (caa_unlikely(!ht->split_count))
3171717f 727 return;
14360f1c 728 index = ht_get_split_count_index(hash);
4c42f1b8 729 split_count = uatomic_add_return(&ht->split_count[index].del, 1);
314558bf
MD
730 if (caa_likely(split_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))
731 return;
732 /* Only if number of deletes multiple of 1UL << COUNT_COMMIT_ORDER */
733
734 dbg_printf("del split count %lu\n", split_count);
735 count = uatomic_add_return(&ht->count,
736 -(1UL << COUNT_COMMIT_ORDER));
4c299dcb 737 if (caa_likely(count & (count - 1)))
314558bf
MD
738 return;
739 /* Only if global count is power of 2 */
740
741 if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) >= size)
742 return;
743 dbg_printf("del set global %ld\n", count);
744 /*
745 * Don't shrink table if the number of nodes is below a
746 * certain threshold.
747 */
748 if (count < (1UL << COUNT_COMMIT_ORDER) * (split_count_mask + 1))
749 return;
750 cds_lfht_resize_lazy_count(ht, size,
751 count >> (CHAIN_LEN_TARGET - 1));
df44348d
MD
752}
753
f9830efd 754static
4105056a 755void check_resize(struct cds_lfht *ht, unsigned long size, uint32_t chain_len)
f9830efd 756{
f8994aee
MD
757 unsigned long count;
758
b8af5011
MD
759 if (!(ht->flags & CDS_LFHT_AUTO_RESIZE))
760 return;
f8994aee
MD
761 count = uatomic_read(&ht->count);
762 /*
763 * Use bucket-local length for small table expand and for
764 * environments lacking per-cpu data support.
765 */
e53ab1eb 766 if (count >= (1UL << (COUNT_COMMIT_ORDER + split_count_order)))
f8994aee 767 return;
24365af7 768 if (chain_len > 100)
f0c29ed7 769 dbg_printf("WARNING: large chain length: %u.\n",
24365af7 770 chain_len);
e53ab1eb
MD
771 if (chain_len >= CHAIN_LEN_RESIZE_THRESHOLD) {
772 int growth;
773
774 /*
775 * Ideal growth calculated based on chain length.
776 */
777 growth = cds_lfht_get_count_order_u32(chain_len
778 - (CHAIN_LEN_TARGET - 1));
779 if ((ht->flags & CDS_LFHT_ACCOUNTING)
780 && (size << growth)
781 >= (1UL << (COUNT_COMMIT_ORDER
782 + split_count_order))) {
783 /*
784 * If ideal growth expands the hash table size
785 * beyond the "small hash table" sizes, use the
786 * maximum small hash table size to attempt
787 * expanding the hash table. This only applies
788 * when node accounting is available, otherwise
789 * the chain length is used to expand the hash
790 * table in every case.
791 */
792 growth = COUNT_COMMIT_ORDER + split_count_order
793 - cds_lfht_get_count_order_ulong(size);
794 if (growth <= 0)
795 return;
796 }
797 cds_lfht_resize_lazy_grow(ht, size, growth);
798 }
f9830efd
MD
799}
800
abc490a1 801static
14044b37 802struct cds_lfht_node *clear_flag(struct cds_lfht_node *node)
abc490a1 803{
14044b37 804 return (struct cds_lfht_node *) (((unsigned long) node) & ~FLAGS_MASK);
abc490a1
MD
805}
806
807static
14044b37 808int is_removed(struct cds_lfht_node *node)
abc490a1 809{
d37166c6 810 return ((unsigned long) node) & REMOVED_FLAG;
abc490a1
MD
811}
812
f5596c94 813static
1ee8f000 814int is_bucket(struct cds_lfht_node *node)
f5596c94 815{
1ee8f000 816 return ((unsigned long) node) & BUCKET_FLAG;
f5596c94
MD
817}
818
819static
1ee8f000 820struct cds_lfht_node *flag_bucket(struct cds_lfht_node *node)
f5596c94 821{
1ee8f000 822 return (struct cds_lfht_node *) (((unsigned long) node) | BUCKET_FLAG);
f5596c94 823}
bb7b2f26 824
db00ccc3
MD
825static
826int is_removal_owner(struct cds_lfht_node *node)
827{
828 return ((unsigned long) node) & REMOVAL_OWNER_FLAG;
829}
830
831static
832struct cds_lfht_node *flag_removal_owner(struct cds_lfht_node *node)
833{
834 return (struct cds_lfht_node *) (((unsigned long) node) | REMOVAL_OWNER_FLAG);
835}
836
71bb3aca
MD
837static
838struct cds_lfht_node *flag_removed_or_removal_owner(struct cds_lfht_node *node)
839{
840 return (struct cds_lfht_node *) (((unsigned long) node) | REMOVED_FLAG | REMOVAL_OWNER_FLAG);
841}
842
bb7b2f26
MD
843static
844struct cds_lfht_node *get_end(void)
845{
846 return (struct cds_lfht_node *) END_VALUE;
847}
848
849static
850int is_end(struct cds_lfht_node *node)
851{
852 return clear_flag(node) == (struct cds_lfht_node *) END_VALUE;
853}
854
abc490a1 855static
ab65b890
LJ
856unsigned long _uatomic_xchg_monotonic_increase(unsigned long *ptr,
857 unsigned long v)
abc490a1
MD
858{
859 unsigned long old1, old2;
860
861 old1 = uatomic_read(ptr);
862 do {
863 old2 = old1;
864 if (old2 >= v)
f9830efd 865 return old2;
abc490a1 866 } while ((old1 = uatomic_cmpxchg(ptr, old2, v)) != old2);
ab65b890 867 return old2;
abc490a1
MD
868}
869
48f1b16d
LJ
870static
871void cds_lfht_alloc_bucket_table(struct cds_lfht *ht, unsigned long order)
872{
0b6aa001 873 return ht->mm->alloc_bucket_table(ht, order);
48f1b16d
LJ
874}
875
876/*
877 * cds_lfht_free_bucket_table() should be called with decreasing order.
878 * When cds_lfht_free_bucket_table(0) is called, it means the whole
879 * lfht is destroyed.
880 */
881static
882void cds_lfht_free_bucket_table(struct cds_lfht *ht, unsigned long order)
883{
0b6aa001 884 return ht->mm->free_bucket_table(ht, order);
48f1b16d
LJ
885}
886
9d72a73f
LJ
887static inline
888struct cds_lfht_node *bucket_at(struct cds_lfht *ht, unsigned long index)
f4a9cc0b 889{
0b6aa001 890 return ht->bucket_at(ht, index);
f4a9cc0b
LJ
891}
892
9d72a73f
LJ
893static inline
894struct cds_lfht_node *lookup_bucket(struct cds_lfht *ht, unsigned long size,
895 unsigned long hash)
896{
897 assert(size > 0);
898 return bucket_at(ht, hash & (size - 1));
899}
900
273399de
MD
901/*
902 * Remove all logically deleted nodes from a bucket up to a certain node key.
903 */
904static
1ee8f000 905void _cds_lfht_gc_bucket(struct cds_lfht_node *bucket, struct cds_lfht_node *node)
273399de 906{
14044b37 907 struct cds_lfht_node *iter_prev, *iter, *next, *new_next;
273399de 908
1ee8f000
LJ
909 assert(!is_bucket(bucket));
910 assert(!is_removed(bucket));
2f943cd7 911 assert(!is_removal_owner(bucket));
1ee8f000 912 assert(!is_bucket(node));
c90201ac 913 assert(!is_removed(node));
2f943cd7 914 assert(!is_removal_owner(node));
273399de 915 for (;;) {
1ee8f000
LJ
916 iter_prev = bucket;
917 /* We can always skip the bucket node initially */
04db56f8 918 iter = rcu_dereference(iter_prev->next);
b4cb483f 919 assert(!is_removed(iter));
2f943cd7 920 assert(!is_removal_owner(iter));
04db56f8 921 assert(iter_prev->reverse_hash <= node->reverse_hash);
bd4db153 922 /*
1ee8f000 923 * We should never be called with bucket (start of chain)
bd4db153
MD
924 * and logically removed node (end of path compression
925 * marker) being the actual same node. This would be a
926 * bug in the algorithm implementation.
927 */
1ee8f000 928 assert(bucket != node);
273399de 929 for (;;) {
8ed51e04 930 if (caa_unlikely(is_end(iter)))
f9c80341 931 return;
04db56f8 932 if (caa_likely(clear_flag(iter)->reverse_hash > node->reverse_hash))
f9c80341 933 return;
04db56f8 934 next = rcu_dereference(clear_flag(iter)->next);
8ed51e04 935 if (caa_likely(is_removed(next)))
273399de 936 break;
b453eae1 937 iter_prev = clear_flag(iter);
273399de
MD
938 iter = next;
939 }
b198f0fd 940 assert(!is_removed(iter));
2f943cd7 941 assert(!is_removal_owner(iter));
1ee8f000
LJ
942 if (is_bucket(iter))
943 new_next = flag_bucket(clear_flag(next));
f5596c94
MD
944 else
945 new_next = clear_flag(next);
04db56f8 946 (void) uatomic_cmpxchg(&iter_prev->next, iter, new_next);
273399de
MD
947 }
948}
949
9357c415
MD
950static
951int _cds_lfht_replace(struct cds_lfht *ht, unsigned long size,
952 struct cds_lfht_node *old_node,
3fb86f26 953 struct cds_lfht_node *old_next,
9357c415
MD
954 struct cds_lfht_node *new_node)
955{
04db56f8 956 struct cds_lfht_node *bucket, *ret_next;
9357c415
MD
957
958 if (!old_node) /* Return -ENOENT if asked to replace NULL node */
7801dadd 959 return -ENOENT;
9357c415
MD
960
961 assert(!is_removed(old_node));
2f943cd7 962 assert(!is_removal_owner(old_node));
1ee8f000 963 assert(!is_bucket(old_node));
9357c415 964 assert(!is_removed(new_node));
2f943cd7 965 assert(!is_removal_owner(new_node));
1ee8f000 966 assert(!is_bucket(new_node));
9357c415 967 assert(new_node != old_node);
3fb86f26 968 for (;;) {
9357c415 969 /* Insert after node to be replaced */
9357c415
MD
970 if (is_removed(old_next)) {
971 /*
972 * Too late, the old node has been removed under us
973 * between lookup and replace. Fail.
974 */
7801dadd 975 return -ENOENT;
9357c415 976 }
feda2722
LJ
977 assert(old_next == clear_flag(old_next));
978 assert(new_node != old_next);
71bb3aca
MD
979 /*
980 * REMOVAL_OWNER flag is _NEVER_ set before the REMOVED
981 * flag. It is either set atomically at the same time
982 * (replace) or after (del).
983 */
984 assert(!is_removal_owner(old_next));
feda2722 985 new_node->next = old_next;
9357c415
MD
986 /*
987 * Here is the whole trick for lock-free replace: we add
988 * the replacement node _after_ the node we want to
989 * replace by atomically setting its next pointer at the
990 * same time we set its removal flag. Given that
991 * the lookups/get next use an iterator aware of the
992 * next pointer, they will either skip the old node due
993 * to the removal flag and see the new node, or use
994 * the old node, but will not see the new one.
db00ccc3
MD
995 * This is a replacement of a node with another node
996 * that has the same value: we are therefore not
71bb3aca
MD
997 * removing a value from the hash table. We set both the
998 * REMOVED and REMOVAL_OWNER flags atomically so we own
999 * the node after successful cmpxchg.
9357c415 1000 */
04db56f8 1001 ret_next = uatomic_cmpxchg(&old_node->next,
71bb3aca 1002 old_next, flag_removed_or_removal_owner(new_node));
3fb86f26 1003 if (ret_next == old_next)
7801dadd 1004 break; /* We performed the replacement. */
3fb86f26
LJ
1005 old_next = ret_next;
1006 }
9357c415 1007
9357c415
MD
1008 /*
1009 * Ensure that the old node is not visible to readers anymore:
1010 * lookup for the node, and remove it (along with any other
1011 * logically removed node) if found.
1012 */
04db56f8
LJ
1013 bucket = lookup_bucket(ht, size, bit_reverse_ulong(old_node->reverse_hash));
1014 _cds_lfht_gc_bucket(bucket, new_node);
7801dadd 1015
a85eff52 1016 assert(is_removed(CMM_LOAD_SHARED(old_node->next)));
7801dadd 1017 return 0;
9357c415
MD
1018}
1019
83beee94
MD
1020/*
1021 * A non-NULL unique_ret pointer uses the "add unique" (or uniquify) add
1022 * mode. A NULL unique_ret allows creation of duplicate keys.
1023 */
abc490a1 1024static
83beee94 1025void _cds_lfht_add(struct cds_lfht *ht,
91a75cc5 1026 unsigned long hash,
0422d92c 1027 cds_lfht_match_fct match,
996ff57c 1028 const void *key,
83beee94
MD
1029 unsigned long size,
1030 struct cds_lfht_node *node,
1031 struct cds_lfht_iter *unique_ret,
1ee8f000 1032 int bucket_flag)
abc490a1 1033{
14044b37 1034 struct cds_lfht_node *iter_prev, *iter, *next, *new_node, *new_next,
960c9e4f 1035 *return_node;
04db56f8 1036 struct cds_lfht_node *bucket;
abc490a1 1037
1ee8f000 1038 assert(!is_bucket(node));
c90201ac 1039 assert(!is_removed(node));
2f943cd7 1040 assert(!is_removal_owner(node));
91a75cc5 1041 bucket = lookup_bucket(ht, size, hash);
abc490a1 1042 for (;;) {
adc0de68 1043 uint32_t chain_len = 0;
abc490a1 1044
11519af6
MD
1045 /*
1046 * iter_prev points to the non-removed node prior to the
1047 * insert location.
11519af6 1048 */
04db56f8 1049 iter_prev = bucket;
1ee8f000 1050 /* We can always skip the bucket node initially */
04db56f8
LJ
1051 iter = rcu_dereference(iter_prev->next);
1052 assert(iter_prev->reverse_hash <= node->reverse_hash);
abc490a1 1053 for (;;) {
8ed51e04 1054 if (caa_unlikely(is_end(iter)))
273399de 1055 goto insert;
04db56f8 1056 if (caa_likely(clear_flag(iter)->reverse_hash > node->reverse_hash))
273399de 1057 goto insert;
238cc06e 1058
1ee8f000
LJ
1059 /* bucket node is the first node of the identical-hash-value chain */
1060 if (bucket_flag && clear_flag(iter)->reverse_hash == node->reverse_hash)
194fdbd1 1061 goto insert;
238cc06e 1062
04db56f8 1063 next = rcu_dereference(clear_flag(iter)->next);
8ed51e04 1064 if (caa_unlikely(is_removed(next)))
9dba85be 1065 goto gc_node;
238cc06e
LJ
1066
1067 /* uniquely add */
83beee94 1068 if (unique_ret
1ee8f000 1069 && !is_bucket(next)
04db56f8 1070 && clear_flag(iter)->reverse_hash == node->reverse_hash) {
238cc06e
LJ
1071 struct cds_lfht_iter d_iter = { .node = node, .next = iter, };
1072
1073 /*
1074 * uniquely adding inserts the node as the first
1075 * node of the identical-hash-value node chain.
1076 *
1077 * This semantic ensures no duplicated keys
1078 * should ever be observable in the table
1f67ba50
MD
1079 * (including traversing the table node by
1080 * node by forward iterations)
238cc06e 1081 */
04db56f8 1082 cds_lfht_next_duplicate(ht, match, key, &d_iter);
238cc06e
LJ
1083 if (!d_iter.node)
1084 goto insert;
1085
1086 *unique_ret = d_iter;
83beee94 1087 return;
48ed1c18 1088 }
238cc06e 1089
11519af6 1090 /* Only account for identical reverse hash once */
04db56f8 1091 if (iter_prev->reverse_hash != clear_flag(iter)->reverse_hash
1ee8f000 1092 && !is_bucket(next))
4105056a 1093 check_resize(ht, size, ++chain_len);
11519af6 1094 iter_prev = clear_flag(iter);
273399de 1095 iter = next;
abc490a1 1096 }
48ed1c18 1097
273399de 1098 insert:
7ec59d3b 1099 assert(node != clear_flag(iter));
11519af6 1100 assert(!is_removed(iter_prev));
2f943cd7 1101 assert(!is_removal_owner(iter_prev));
c90201ac 1102 assert(!is_removed(iter));
2f943cd7 1103 assert(!is_removal_owner(iter));
f000907d 1104 assert(iter_prev != node);
1ee8f000 1105 if (!bucket_flag)
04db56f8 1106 node->next = clear_flag(iter);
f9c80341 1107 else
1ee8f000
LJ
1108 node->next = flag_bucket(clear_flag(iter));
1109 if (is_bucket(iter))
1110 new_node = flag_bucket(node);
f5596c94
MD
1111 else
1112 new_node = node;
04db56f8 1113 if (uatomic_cmpxchg(&iter_prev->next, iter,
48ed1c18 1114 new_node) != iter) {
273399de 1115 continue; /* retry */
48ed1c18 1116 } else {
83beee94 1117 return_node = node;
960c9e4f 1118 goto end;
48ed1c18
MD
1119 }
1120
9dba85be
MD
1121 gc_node:
1122 assert(!is_removed(iter));
2f943cd7 1123 assert(!is_removal_owner(iter));
1ee8f000
LJ
1124 if (is_bucket(iter))
1125 new_next = flag_bucket(clear_flag(next));
f5596c94
MD
1126 else
1127 new_next = clear_flag(next);
04db56f8 1128 (void) uatomic_cmpxchg(&iter_prev->next, iter, new_next);
273399de 1129 /* retry */
464a1ec9 1130 }
9357c415 1131end:
83beee94
MD
1132 if (unique_ret) {
1133 unique_ret->node = return_node;
1134 /* unique_ret->next left unset, never used. */
1135 }
abc490a1 1136}
464a1ec9 1137
abc490a1 1138static
860d07e8 1139int _cds_lfht_del(struct cds_lfht *ht, unsigned long size,
b65ec430 1140 struct cds_lfht_node *node)
abc490a1 1141{
db00ccc3 1142 struct cds_lfht_node *bucket, *next;
5e28c532 1143
9357c415 1144 if (!node) /* Return -ENOENT if asked to delete NULL node */
743f9143 1145 return -ENOENT;
9357c415 1146
7ec59d3b 1147 /* logically delete the node */
1ee8f000 1148 assert(!is_bucket(node));
c90201ac 1149 assert(!is_removed(node));
db00ccc3 1150 assert(!is_removal_owner(node));
48ed1c18 1151
db00ccc3
MD
1152 /*
1153 * We are first checking if the node had previously been
1154 * logically removed (this check is not atomic with setting the
1155 * logical removal flag). Return -ENOENT if the node had
1156 * previously been removed.
1157 */
a85eff52 1158 next = CMM_LOAD_SHARED(node->next); /* next is not dereferenced */
db00ccc3
MD
1159 if (caa_unlikely(is_removed(next)))
1160 return -ENOENT;
b65ec430 1161 assert(!is_bucket(next));
196f4fab
MD
1162 /*
1163 * The del operation semantic guarantees a full memory barrier
1164 * before the uatomic_or atomic commit of the deletion flag.
1165 */
1166 cmm_smp_mb__before_uatomic_or();
db00ccc3
MD
1167 /*
1168 * We set the REMOVED_FLAG unconditionally. Note that there may
1169 * be more than one concurrent thread setting this flag.
1170 * Knowing which wins the race will be known after the garbage
1171 * collection phase, stay tuned!
1172 */
1173 uatomic_or(&node->next, REMOVED_FLAG);
7ec59d3b 1174 /* We performed the (logical) deletion. */
7ec59d3b
MD
1175
1176 /*
1177 * Ensure that the node is not visible to readers anymore: lookup for
273399de
MD
1178 * the node, and remove it (along with any other logically removed node)
1179 * if found.
11519af6 1180 */
04db56f8
LJ
1181 bucket = lookup_bucket(ht, size, bit_reverse_ulong(node->reverse_hash));
1182 _cds_lfht_gc_bucket(bucket, node);
743f9143 1183
a85eff52 1184 assert(is_removed(CMM_LOAD_SHARED(node->next)));
db00ccc3
MD
1185 /*
1186 * Last phase: atomically exchange node->next with a version
1187 * having "REMOVAL_OWNER_FLAG" set. If the returned node->next
1188 * pointer did _not_ have "REMOVAL_OWNER_FLAG" set, we now own
1189 * the node and win the removal race.
1190 * It is interesting to note that all "add" paths are forbidden
1191 * to change the next pointer starting from the point where the
1192 * REMOVED_FLAG is set, so here using a read, followed by a
1193 * xchg() suffice to guarantee that the xchg() will ever only
1194 * set the "REMOVAL_OWNER_FLAG" (or change nothing if the flag
1195 * was already set).
1196 */
1197 if (!is_removal_owner(uatomic_xchg(&node->next,
1198 flag_removal_owner(node->next))))
1199 return 0;
1200 else
1201 return -ENOENT;
abc490a1 1202}
2ed95849 1203
b7d619b0
MD
1204static
1205void *partition_resize_thread(void *arg)
1206{
1207 struct partition_resize_work *work = arg;
1208
7b17c13e 1209 work->ht->flavor->register_thread();
b7d619b0 1210 work->fct(work->ht, work->i, work->start, work->len);
7b17c13e 1211 work->ht->flavor->unregister_thread();
b7d619b0
MD
1212 return NULL;
1213}
1214
1215static
1216void partition_resize_helper(struct cds_lfht *ht, unsigned long i,
1217 unsigned long len,
1218 void (*fct)(struct cds_lfht *ht, unsigned long i,
1219 unsigned long start, unsigned long len))
1220{
e54ec2f5 1221 unsigned long partition_len, start = 0;
b7d619b0 1222 struct partition_resize_work *work;
6083a889
MD
1223 int thread, ret;
1224 unsigned long nr_threads;
b7d619b0 1225
d7f3ba4c
EW
1226 assert(nr_cpus_mask != -1);
1227 if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD)
1228 goto fallback;
1229
6083a889
MD
1230 /*
1231 * Note: nr_cpus_mask + 1 is always power of 2.
1232 * We spawn just the number of threads we need to satisfy the minimum
1233 * partition size, up to the number of CPUs in the system.
1234 */
91452a6a
MD
1235 if (nr_cpus_mask > 0) {
1236 nr_threads = min(nr_cpus_mask + 1,
1237 len >> MIN_PARTITION_PER_THREAD_ORDER);
1238 } else {
1239 nr_threads = 1;
1240 }
5bc6b66f 1241 partition_len = len >> cds_lfht_get_count_order_ulong(nr_threads);
6083a889 1242 work = calloc(nr_threads, sizeof(*work));
7c75d498
EW
1243 if (!work) {
1244 dbg_printf("error allocating for resize, single-threading\n");
1245 goto fallback;
1246 }
6083a889
MD
1247 for (thread = 0; thread < nr_threads; thread++) {
1248 work[thread].ht = ht;
1249 work[thread].i = i;
1250 work[thread].len = partition_len;
1251 work[thread].start = thread * partition_len;
1252 work[thread].fct = fct;
1af6e26e 1253 ret = pthread_create(&(work[thread].thread_id), ht->resize_attr,
6083a889 1254 partition_resize_thread, &work[thread]);
e54ec2f5
EW
1255 if (ret == EAGAIN) {
1256 /*
1257 * Out of resources: wait and join the threads
1258 * we've created, then handle leftovers.
1259 */
1260 dbg_printf("error spawning for resize, single-threading\n");
1261 start = work[thread].start;
1262 len -= start;
1263 nr_threads = thread;
1264 break;
1265 }
b7d619b0
MD
1266 assert(!ret);
1267 }
6083a889 1268 for (thread = 0; thread < nr_threads; thread++) {
1af6e26e 1269 ret = pthread_join(work[thread].thread_id, NULL);
b7d619b0
MD
1270 assert(!ret);
1271 }
1272 free(work);
e54ec2f5
EW
1273
1274 /*
1275 * A pthread_create failure above will either lead in us having
1276 * no threads to join or starting at a non-zero offset,
1277 * fallback to single thread processing of leftovers.
1278 */
1279 if (start == 0 && nr_threads > 0)
1280 return;
7c75d498 1281fallback:
e54ec2f5 1282 fct(ht, i, start, len);
b7d619b0
MD
1283}
1284
e8de508e
MD
1285/*
1286 * Holding RCU read lock to protect _cds_lfht_add against memory
d0ec0ed2 1287 * reclaim that could be performed by other worker threads (ABA
e8de508e 1288 * problem).
9ee0fc9a 1289 *
b7d619b0 1290 * When we reach a certain length, we can split this population phase over
9ee0fc9a
MD
1291 * many worker threads, based on the number of CPUs available in the system.
1292 * This should therefore take care of not having the expand lagging behind too
1293 * many concurrent insertion threads by using the scheduler's ability to
1ee8f000 1294 * schedule bucket node population fairly with insertions.
e8de508e 1295 */
4105056a 1296static
b7d619b0
MD
1297void init_table_populate_partition(struct cds_lfht *ht, unsigned long i,
1298 unsigned long start, unsigned long len)
4105056a 1299{
9d72a73f 1300 unsigned long j, size = 1UL << (i - 1);
4105056a 1301
d0d8f9aa 1302 assert(i > MIN_TABLE_ORDER);
7b17c13e 1303 ht->flavor->read_lock();
9d72a73f
LJ
1304 for (j = size + start; j < size + start + len; j++) {
1305 struct cds_lfht_node *new_node = bucket_at(ht, j);
1306
1307 assert(j >= size && j < (size << 1));
1308 dbg_printf("init populate: order %lu index %lu hash %lu\n",
1309 i, j, j);
1310 new_node->reverse_hash = bit_reverse_ulong(j);
91a75cc5 1311 _cds_lfht_add(ht, j, NULL, NULL, size, new_node, NULL, 1);
4105056a 1312 }
7b17c13e 1313 ht->flavor->read_unlock();
b7d619b0
MD
1314}
1315
1316static
1317void init_table_populate(struct cds_lfht *ht, unsigned long i,
1318 unsigned long len)
1319{
b7d619b0 1320 partition_resize_helper(ht, i, len, init_table_populate_partition);
4105056a
MD
1321}
1322
abc490a1 1323static
4105056a 1324void init_table(struct cds_lfht *ht,
93d46c39 1325 unsigned long first_order, unsigned long last_order)
24365af7 1326{
93d46c39 1327 unsigned long i;
24365af7 1328
93d46c39
LJ
1329 dbg_printf("init table: first_order %lu last_order %lu\n",
1330 first_order, last_order);
d0d8f9aa 1331 assert(first_order > MIN_TABLE_ORDER);
93d46c39 1332 for (i = first_order; i <= last_order; i++) {
4105056a 1333 unsigned long len;
24365af7 1334
4f6e90b7 1335 len = 1UL << (i - 1);
f0c29ed7 1336 dbg_printf("init order %lu len: %lu\n", i, len);
4d676753
MD
1337
1338 /* Stop expand if the resize target changes under us */
7b3893e4 1339 if (CMM_LOAD_SHARED(ht->resize_target) < (1UL << i))
4d676753
MD
1340 break;
1341
48f1b16d 1342 cds_lfht_alloc_bucket_table(ht, i);
4105056a 1343
4105056a 1344 /*
1ee8f000
LJ
1345 * Set all bucket nodes reverse hash values for a level and
1346 * link all bucket nodes into the table.
4105056a 1347 */
dc1da8f6 1348 init_table_populate(ht, i, len);
4105056a 1349
f9c80341
MD
1350 /*
1351 * Update table size.
1352 */
1353 cmm_smp_wmb(); /* populate data before RCU size */
7b3893e4 1354 CMM_STORE_SHARED(ht->size, 1UL << i);
f9c80341 1355
4f6e90b7 1356 dbg_printf("init new size: %lu\n", 1UL << i);
4105056a
MD
1357 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1358 break;
1359 }
1360}
1361
e8de508e
MD
1362/*
1363 * Holding RCU read lock to protect _cds_lfht_remove against memory
d0ec0ed2 1364 * reclaim that could be performed by other worker threads (ABA
e8de508e
MD
1365 * problem).
1366 * For a single level, we logically remove and garbage collect each node.
1367 *
1368 * As a design choice, we perform logical removal and garbage collection on a
1369 * node-per-node basis to simplify this algorithm. We also assume keeping good
1370 * cache locality of the operation would overweight possible performance gain
1371 * that could be achieved by batching garbage collection for multiple levels.
1372 * However, this would have to be justified by benchmarks.
1373 *
1374 * Concurrent removal and add operations are helping us perform garbage
1375 * collection of logically removed nodes. We guarantee that all logically
d0ec0ed2
MD
1376 * removed nodes have been garbage-collected (unlinked) before work
1377 * enqueue is invoked to free a hole level of bucket nodes (after a
1378 * grace period).
e8de508e 1379 *
1f67ba50
MD
1380 * Logical removal and garbage collection can therefore be done in batch
1381 * or on a node-per-node basis, as long as the guarantee above holds.
9ee0fc9a 1382 *
b7d619b0
MD
1383 * When we reach a certain length, we can split this removal over many worker
1384 * threads, based on the number of CPUs available in the system. This should
1385 * take care of not letting resize process lag behind too many concurrent
9ee0fc9a 1386 * updater threads actively inserting into the hash table.
e8de508e 1387 */
4105056a 1388static
b7d619b0
MD
1389void remove_table_partition(struct cds_lfht *ht, unsigned long i,
1390 unsigned long start, unsigned long len)
4105056a 1391{
9d72a73f 1392 unsigned long j, size = 1UL << (i - 1);
4105056a 1393
d0d8f9aa 1394 assert(i > MIN_TABLE_ORDER);
7b17c13e 1395 ht->flavor->read_lock();
9d72a73f 1396 for (j = size + start; j < size + start + len; j++) {
2e2ce1e9
LJ
1397 struct cds_lfht_node *fini_bucket = bucket_at(ht, j);
1398 struct cds_lfht_node *parent_bucket = bucket_at(ht, j - size);
9d72a73f
LJ
1399
1400 assert(j >= size && j < (size << 1));
1401 dbg_printf("remove entry: order %lu index %lu hash %lu\n",
1402 i, j, j);
2e2ce1e9
LJ
1403 /* Set the REMOVED_FLAG to freeze the ->next for gc */
1404 uatomic_or(&fini_bucket->next, REMOVED_FLAG);
1405 _cds_lfht_gc_bucket(parent_bucket, fini_bucket);
abc490a1 1406 }
7b17c13e 1407 ht->flavor->read_unlock();
b7d619b0
MD
1408}
1409
1410static
1411void remove_table(struct cds_lfht *ht, unsigned long i, unsigned long len)
1412{
b7d619b0 1413 partition_resize_helper(ht, i, len, remove_table_partition);
2ed95849
MD
1414}
1415
61adb337
MD
1416/*
1417 * fini_table() is never called for first_order == 0, which is why
1418 * free_by_rcu_order == 0 can be used as criterion to know if free must
1419 * be called.
1420 */
1475579c 1421static
4105056a 1422void fini_table(struct cds_lfht *ht,
93d46c39 1423 unsigned long first_order, unsigned long last_order)
1475579c 1424{
93d46c39 1425 long i;
48f1b16d 1426 unsigned long free_by_rcu_order = 0;
1475579c 1427
93d46c39
LJ
1428 dbg_printf("fini table: first_order %lu last_order %lu\n",
1429 first_order, last_order);
d0d8f9aa 1430 assert(first_order > MIN_TABLE_ORDER);
93d46c39 1431 for (i = last_order; i >= first_order; i--) {
4105056a 1432 unsigned long len;
1475579c 1433
4f6e90b7 1434 len = 1UL << (i - 1);
e15df1cc 1435 dbg_printf("fini order %ld len: %lu\n", i, len);
4105056a 1436
4d676753 1437 /* Stop shrink if the resize target changes under us */
7b3893e4 1438 if (CMM_LOAD_SHARED(ht->resize_target) > (1UL << (i - 1)))
4d676753
MD
1439 break;
1440
1441 cmm_smp_wmb(); /* populate data before RCU size */
7b3893e4 1442 CMM_STORE_SHARED(ht->size, 1UL << (i - 1));
4d676753
MD
1443
1444 /*
1445 * We need to wait for all add operations to reach Q.S. (and
1446 * thus use the new table for lookups) before we can start
1ee8f000 1447 * releasing the old bucket nodes. Otherwise their lookup will
4d676753
MD
1448 * return a logically removed node as insert position.
1449 */
7b17c13e 1450 ht->flavor->update_synchronize_rcu();
48f1b16d
LJ
1451 if (free_by_rcu_order)
1452 cds_lfht_free_bucket_table(ht, free_by_rcu_order);
4d676753 1453
21263e21 1454 /*
1ee8f000
LJ
1455 * Set "removed" flag in bucket nodes about to be removed.
1456 * Unlink all now-logically-removed bucket node pointers.
4105056a
MD
1457 * Concurrent add/remove operation are helping us doing
1458 * the gc.
21263e21 1459 */
4105056a
MD
1460 remove_table(ht, i, len);
1461
48f1b16d 1462 free_by_rcu_order = i;
4105056a
MD
1463
1464 dbg_printf("fini new size: %lu\n", 1UL << i);
1475579c
MD
1465 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1466 break;
1467 }
0d14ceb2 1468
48f1b16d 1469 if (free_by_rcu_order) {
7b17c13e 1470 ht->flavor->update_synchronize_rcu();
48f1b16d 1471 cds_lfht_free_bucket_table(ht, free_by_rcu_order);
0d14ceb2 1472 }
1475579c
MD
1473}
1474
ff0d69de 1475static
1ee8f000 1476void cds_lfht_create_bucket(struct cds_lfht *ht, unsigned long size)
ff0d69de 1477{
04db56f8 1478 struct cds_lfht_node *prev, *node;
9d72a73f 1479 unsigned long order, len, i;
ff0d69de 1480
48f1b16d 1481 cds_lfht_alloc_bucket_table(ht, 0);
ff0d69de 1482
9d72a73f
LJ
1483 dbg_printf("create bucket: order 0 index 0 hash 0\n");
1484 node = bucket_at(ht, 0);
1485 node->next = flag_bucket(get_end());
1486 node->reverse_hash = 0;
ff0d69de 1487
5bc6b66f 1488 for (order = 1; order < cds_lfht_get_count_order_ulong(size) + 1; order++) {
ff0d69de 1489 len = 1UL << (order - 1);
48f1b16d 1490 cds_lfht_alloc_bucket_table(ht, order);
ff0d69de 1491
9d72a73f
LJ
1492 for (i = 0; i < len; i++) {
1493 /*
1494 * Now, we are trying to init the node with the
1495 * hash=(len+i) (which is also a bucket with the
1496 * index=(len+i)) and insert it into the hash table,
1497 * so this node has to be inserted after the bucket
1498 * with the index=(len+i)&(len-1)=i. And because there
1499 * is no other non-bucket node nor bucket node with
1500 * larger index/hash inserted, so the bucket node
1501 * being inserted should be inserted directly linked
1502 * after the bucket node with index=i.
1503 */
1504 prev = bucket_at(ht, i);
1505 node = bucket_at(ht, len + i);
ff0d69de 1506
1ee8f000 1507 dbg_printf("create bucket: order %lu index %lu hash %lu\n",
9d72a73f
LJ
1508 order, len + i, len + i);
1509 node->reverse_hash = bit_reverse_ulong(len + i);
1510
1511 /* insert after prev */
1512 assert(is_bucket(prev->next));
ff0d69de 1513 node->next = prev->next;
1ee8f000 1514 prev->next = flag_bucket(node);
ff0d69de
LJ
1515 }
1516 }
1517}
1518
0422d92c 1519struct cds_lfht *_cds_lfht_new(unsigned long init_size,
0722081a 1520 unsigned long min_nr_alloc_buckets,
747d725c 1521 unsigned long max_nr_buckets,
b8af5011 1522 int flags,
0b6aa001 1523 const struct cds_lfht_mm_type *mm,
7b17c13e 1524 const struct rcu_flavor_struct *flavor,
b7d619b0 1525 pthread_attr_t *attr)
abc490a1 1526{
14044b37 1527 struct cds_lfht *ht;
24365af7 1528 unsigned long order;
abc490a1 1529
0722081a
LJ
1530 /* min_nr_alloc_buckets must be power of two */
1531 if (!min_nr_alloc_buckets || (min_nr_alloc_buckets & (min_nr_alloc_buckets - 1)))
5488222b 1532 return NULL;
747d725c 1533
8129be4e 1534 /* init_size must be power of two */
5488222b 1535 if (!init_size || (init_size & (init_size - 1)))
8129be4e 1536 return NULL;
747d725c 1537
c1888f3a
MD
1538 /*
1539 * Memory management plugin default.
1540 */
1541 if (!mm) {
5a2141a7
MD
1542 if (CAA_BITS_PER_LONG > 32
1543 && max_nr_buckets
c1888f3a
MD
1544 && max_nr_buckets <= (1ULL << 32)) {
1545 /*
1546 * For 64-bit architectures, with max number of
1547 * buckets small enough not to use the entire
1548 * 64-bit memory mapping space (and allowing a
1549 * fair number of hash table instances), use the
1550 * mmap allocator, which is faster than the
1551 * order allocator.
1552 */
1553 mm = &cds_lfht_mm_mmap;
1554 } else {
1555 /*
1556 * The fallback is to use the order allocator.
1557 */
1558 mm = &cds_lfht_mm_order;
1559 }
1560 }
1561
0b6aa001
LJ
1562 /* max_nr_buckets == 0 for order based mm means infinite */
1563 if (mm == &cds_lfht_mm_order && !max_nr_buckets)
747d725c
LJ
1564 max_nr_buckets = 1UL << (MAX_TABLE_ORDER - 1);
1565
1566 /* max_nr_buckets must be power of two */
1567 if (!max_nr_buckets || (max_nr_buckets & (max_nr_buckets - 1)))
1568 return NULL;
1569
d0ec0ed2
MD
1570 if (flags & CDS_LFHT_AUTO_RESIZE)
1571 cds_lfht_init_worker(flavor);
1572
0722081a 1573 min_nr_alloc_buckets = max(min_nr_alloc_buckets, MIN_TABLE_SIZE);
d0d8f9aa 1574 init_size = max(init_size, MIN_TABLE_SIZE);
747d725c
LJ
1575 max_nr_buckets = max(max_nr_buckets, min_nr_alloc_buckets);
1576 init_size = min(init_size, max_nr_buckets);
0b6aa001
LJ
1577
1578 ht = mm->alloc_cds_lfht(min_nr_alloc_buckets, max_nr_buckets);
b7d619b0 1579 assert(ht);
0b6aa001
LJ
1580 assert(ht->mm == mm);
1581 assert(ht->bucket_at == mm->bucket_at);
1582
b5d6b20f 1583 ht->flags = flags;
7b17c13e 1584 ht->flavor = flavor;
b7d619b0 1585 ht->resize_attr = attr;
5afadd12 1586 alloc_split_items_count(ht);
abc490a1
MD
1587 /* this mutex should not nest in read-side C.S. */
1588 pthread_mutex_init(&ht->resize_mutex, NULL);
5bc6b66f 1589 order = cds_lfht_get_count_order_ulong(init_size);
7b3893e4 1590 ht->resize_target = 1UL << order;
1ee8f000 1591 cds_lfht_create_bucket(ht, 1UL << order);
7b3893e4 1592 ht->size = 1UL << order;
abc490a1
MD
1593 return ht;
1594}
1595
6f554439 1596void cds_lfht_lookup(struct cds_lfht *ht, unsigned long hash,
996ff57c 1597 cds_lfht_match_fct match, const void *key,
6f554439 1598 struct cds_lfht_iter *iter)
2ed95849 1599{
04db56f8 1600 struct cds_lfht_node *node, *next, *bucket;
0422d92c 1601 unsigned long reverse_hash, size;
2ed95849 1602
abc490a1 1603 reverse_hash = bit_reverse_ulong(hash);
464a1ec9 1604
7b3893e4 1605 size = rcu_dereference(ht->size);
04db56f8 1606 bucket = lookup_bucket(ht, size, hash);
1ee8f000 1607 /* We can always skip the bucket node initially */
04db56f8 1608 node = rcu_dereference(bucket->next);
bb7b2f26 1609 node = clear_flag(node);
2ed95849 1610 for (;;) {
8ed51e04 1611 if (caa_unlikely(is_end(node))) {
96ad1112 1612 node = next = NULL;
abc490a1 1613 break;
bb7b2f26 1614 }
04db56f8 1615 if (caa_unlikely(node->reverse_hash > reverse_hash)) {
96ad1112 1616 node = next = NULL;
abc490a1 1617 break;
2ed95849 1618 }
04db56f8 1619 next = rcu_dereference(node->next);
7f52427b 1620 assert(node == clear_flag(node));
8ed51e04 1621 if (caa_likely(!is_removed(next))
1ee8f000 1622 && !is_bucket(next)
04db56f8 1623 && node->reverse_hash == reverse_hash
0422d92c 1624 && caa_likely(match(node, key))) {
273399de 1625 break;
2ed95849 1626 }
1b81fe1a 1627 node = clear_flag(next);
2ed95849 1628 }
a85eff52 1629 assert(!node || !is_bucket(CMM_LOAD_SHARED(node->next)));
adc0de68
MD
1630 iter->node = node;
1631 iter->next = next;
abc490a1 1632}
e0ba718a 1633
0422d92c 1634void cds_lfht_next_duplicate(struct cds_lfht *ht, cds_lfht_match_fct match,
996ff57c 1635 const void *key, struct cds_lfht_iter *iter)
a481e5ff 1636{
adc0de68 1637 struct cds_lfht_node *node, *next;
a481e5ff 1638 unsigned long reverse_hash;
a481e5ff 1639
adc0de68 1640 node = iter->node;
04db56f8 1641 reverse_hash = node->reverse_hash;
adc0de68 1642 next = iter->next;
a481e5ff
MD
1643 node = clear_flag(next);
1644
1645 for (;;) {
8ed51e04 1646 if (caa_unlikely(is_end(node))) {
96ad1112 1647 node = next = NULL;
a481e5ff 1648 break;
bb7b2f26 1649 }
04db56f8 1650 if (caa_unlikely(node->reverse_hash > reverse_hash)) {
96ad1112 1651 node = next = NULL;
a481e5ff
MD
1652 break;
1653 }
04db56f8 1654 next = rcu_dereference(node->next);
8ed51e04 1655 if (caa_likely(!is_removed(next))
1ee8f000 1656 && !is_bucket(next)
04db56f8 1657 && caa_likely(match(node, key))) {
a481e5ff
MD
1658 break;
1659 }
1660 node = clear_flag(next);
1661 }
a85eff52 1662 assert(!node || !is_bucket(CMM_LOAD_SHARED(node->next)));
adc0de68
MD
1663 iter->node = node;
1664 iter->next = next;
a481e5ff
MD
1665}
1666
4e9b9fbf
MD
1667void cds_lfht_next(struct cds_lfht *ht, struct cds_lfht_iter *iter)
1668{
1669 struct cds_lfht_node *node, *next;
1670
853395e1 1671 node = clear_flag(iter->next);
4e9b9fbf 1672 for (;;) {
8ed51e04 1673 if (caa_unlikely(is_end(node))) {
4e9b9fbf
MD
1674 node = next = NULL;
1675 break;
1676 }
04db56f8 1677 next = rcu_dereference(node->next);
8ed51e04 1678 if (caa_likely(!is_removed(next))
1ee8f000 1679 && !is_bucket(next)) {
4e9b9fbf
MD
1680 break;
1681 }
1682 node = clear_flag(next);
1683 }
a85eff52 1684 assert(!node || !is_bucket(CMM_LOAD_SHARED(node->next)));
4e9b9fbf
MD
1685 iter->node = node;
1686 iter->next = next;
1687}
1688
1689void cds_lfht_first(struct cds_lfht *ht, struct cds_lfht_iter *iter)
1690{
4e9b9fbf 1691 /*
1ee8f000 1692 * Get next after first bucket node. The first bucket node is the
4e9b9fbf
MD
1693 * first node of the linked list.
1694 */
9d72a73f 1695 iter->next = bucket_at(ht, 0)->next;
4e9b9fbf
MD
1696 cds_lfht_next(ht, iter);
1697}
1698
0422d92c
MD
1699void cds_lfht_add(struct cds_lfht *ht, unsigned long hash,
1700 struct cds_lfht_node *node)
abc490a1 1701{
0422d92c 1702 unsigned long size;
ab7d5fc6 1703
709bacf9 1704 node->reverse_hash = bit_reverse_ulong(hash);
7b3893e4 1705 size = rcu_dereference(ht->size);
91a75cc5 1706 _cds_lfht_add(ht, hash, NULL, NULL, size, node, NULL, 0);
14360f1c 1707 ht_count_add(ht, size, hash);
3eca1b8c
MD
1708}
1709
14044b37 1710struct cds_lfht_node *cds_lfht_add_unique(struct cds_lfht *ht,
6f554439 1711 unsigned long hash,
0422d92c 1712 cds_lfht_match_fct match,
996ff57c 1713 const void *key,
48ed1c18 1714 struct cds_lfht_node *node)
3eca1b8c 1715{
0422d92c 1716 unsigned long size;
83beee94 1717 struct cds_lfht_iter iter;
3eca1b8c 1718
709bacf9 1719 node->reverse_hash = bit_reverse_ulong(hash);
7b3893e4 1720 size = rcu_dereference(ht->size);
91a75cc5 1721 _cds_lfht_add(ht, hash, match, key, size, node, &iter, 0);
83beee94 1722 if (iter.node == node)
14360f1c 1723 ht_count_add(ht, size, hash);
83beee94 1724 return iter.node;
2ed95849
MD
1725}
1726
9357c415 1727struct cds_lfht_node *cds_lfht_add_replace(struct cds_lfht *ht,
6f554439 1728 unsigned long hash,
0422d92c 1729 cds_lfht_match_fct match,
996ff57c 1730 const void *key,
48ed1c18
MD
1731 struct cds_lfht_node *node)
1732{
0422d92c 1733 unsigned long size;
83beee94 1734 struct cds_lfht_iter iter;
48ed1c18 1735
709bacf9 1736 node->reverse_hash = bit_reverse_ulong(hash);
7b3893e4 1737 size = rcu_dereference(ht->size);
83beee94 1738 for (;;) {
91a75cc5 1739 _cds_lfht_add(ht, hash, match, key, size, node, &iter, 0);
83beee94 1740 if (iter.node == node) {
14360f1c 1741 ht_count_add(ht, size, hash);
83beee94
MD
1742 return NULL;
1743 }
1744
1745 if (!_cds_lfht_replace(ht, size, iter.node, iter.next, node))
1746 return iter.node;
1747 }
48ed1c18
MD
1748}
1749
2e79c445
MD
1750int cds_lfht_replace(struct cds_lfht *ht,
1751 struct cds_lfht_iter *old_iter,
1752 unsigned long hash,
1753 cds_lfht_match_fct match,
1754 const void *key,
9357c415
MD
1755 struct cds_lfht_node *new_node)
1756{
1757 unsigned long size;
1758
709bacf9 1759 new_node->reverse_hash = bit_reverse_ulong(hash);
2e79c445
MD
1760 if (!old_iter->node)
1761 return -ENOENT;
1762 if (caa_unlikely(old_iter->node->reverse_hash != new_node->reverse_hash))
1763 return -EINVAL;
1764 if (caa_unlikely(!match(old_iter->node, key)))
1765 return -EINVAL;
7b3893e4 1766 size = rcu_dereference(ht->size);
9357c415
MD
1767 return _cds_lfht_replace(ht, size, old_iter->node, old_iter->next,
1768 new_node);
1769}
1770
bc8c3c74 1771int cds_lfht_del(struct cds_lfht *ht, struct cds_lfht_node *node)
2ed95849 1772{
95bc7fb9 1773 unsigned long size;
df44348d 1774 int ret;
abc490a1 1775
7b3893e4 1776 size = rcu_dereference(ht->size);
bc8c3c74 1777 ret = _cds_lfht_del(ht, size, node);
14360f1c 1778 if (!ret) {
95bc7fb9
MD
1779 unsigned long hash;
1780
bc8c3c74 1781 hash = bit_reverse_ulong(node->reverse_hash);
14360f1c
LJ
1782 ht_count_del(ht, size, hash);
1783 }
df44348d 1784 return ret;
2ed95849 1785}
ab7d5fc6 1786
df55172a
MD
1787int cds_lfht_is_node_deleted(struct cds_lfht_node *node)
1788{
a85eff52 1789 return is_removed(CMM_LOAD_SHARED(node->next));
df55172a
MD
1790}
1791
abc490a1 1792static
1ee8f000 1793int cds_lfht_delete_bucket(struct cds_lfht *ht)
674f7a69 1794{
14044b37 1795 struct cds_lfht_node *node;
4105056a 1796 unsigned long order, i, size;
674f7a69 1797
abc490a1 1798 /* Check that the table is empty */
9d72a73f 1799 node = bucket_at(ht, 0);
abc490a1 1800 do {
04db56f8 1801 node = clear_flag(node)->next;
1ee8f000 1802 if (!is_bucket(node))
abc490a1 1803 return -EPERM;
273399de 1804 assert(!is_removed(node));
2f943cd7 1805 assert(!is_removal_owner(node));
bb7b2f26 1806 } while (!is_end(node));
4105056a
MD
1807 /*
1808 * size accessed without rcu_dereference because hash table is
1809 * being destroyed.
1810 */
7b3893e4 1811 size = ht->size;
1f67ba50 1812 /* Internal sanity check: all nodes left should be buckets */
48f1b16d
LJ
1813 for (i = 0; i < size; i++) {
1814 node = bucket_at(ht, i);
1815 dbg_printf("delete bucket: index %lu expected hash %lu hash %lu\n",
1816 i, i, bit_reverse_ulong(node->reverse_hash));
1817 assert(is_bucket(node->next));
1818 }
24365af7 1819
5bc6b66f 1820 for (order = cds_lfht_get_count_order_ulong(size); (long)order >= 0; order--)
48f1b16d 1821 cds_lfht_free_bucket_table(ht, order);
5488222b 1822
abc490a1 1823 return 0;
674f7a69
MD
1824}
1825
1826/*
1827 * Should only be called when no more concurrent readers nor writers can
1828 * possibly access the table.
1829 */
b7d619b0 1830int cds_lfht_destroy(struct cds_lfht *ht, pthread_attr_t **attr)
674f7a69 1831{
d0ec0ed2
MD
1832 int ret;
1833
1834 if (ht->flags & CDS_LFHT_AUTO_RESIZE) {
1835 /* Cancel ongoing resize operations. */
1836 _CMM_STORE_SHARED(ht->in_progress_destroy, 1);
1837 /* Wait for in-flight resize operations to complete */
1838 urcu_workqueue_flush_queued_work(cds_lfht_workqueue);
10e68472 1839 }
1ee8f000 1840 ret = cds_lfht_delete_bucket(ht);
abc490a1
MD
1841 if (ret)
1842 return ret;
5afadd12 1843 free_split_items_count(ht);
b7d619b0
MD
1844 if (attr)
1845 *attr = ht->resize_attr;
59629f09
MD
1846 ret = pthread_mutex_destroy(&ht->resize_mutex);
1847 if (ret)
1848 ret = -EBUSY;
d0ec0ed2
MD
1849 if (ht->flags & CDS_LFHT_AUTO_RESIZE)
1850 cds_lfht_fini_worker(ht->flavor);
98808fb1 1851 poison_free(ht);
5e28c532 1852 return ret;
674f7a69
MD
1853}
1854
14044b37 1855void cds_lfht_count_nodes(struct cds_lfht *ht,
d933dd0e 1856 long *approx_before,
273399de 1857 unsigned long *count,
d933dd0e 1858 long *approx_after)
273399de 1859{
14044b37 1860 struct cds_lfht_node *node, *next;
caf3653d 1861 unsigned long nr_bucket = 0, nr_removed = 0;
273399de 1862
7ed7682f 1863 *approx_before = 0;
5afadd12 1864 if (ht->split_count) {
973e5e1b
MD
1865 int i;
1866
4c42f1b8
LJ
1867 for (i = 0; i < split_count_mask + 1; i++) {
1868 *approx_before += uatomic_read(&ht->split_count[i].add);
1869 *approx_before -= uatomic_read(&ht->split_count[i].del);
973e5e1b
MD
1870 }
1871 }
1872
273399de 1873 *count = 0;
273399de 1874
1ee8f000 1875 /* Count non-bucket nodes in the table */
9d72a73f 1876 node = bucket_at(ht, 0);
273399de 1877 do {
04db56f8 1878 next = rcu_dereference(node->next);
b198f0fd 1879 if (is_removed(next)) {
1ee8f000 1880 if (!is_bucket(next))
caf3653d 1881 (nr_removed)++;
973e5e1b 1882 else
1ee8f000
LJ
1883 (nr_bucket)++;
1884 } else if (!is_bucket(next))
273399de 1885 (*count)++;
24365af7 1886 else
1ee8f000 1887 (nr_bucket)++;
273399de 1888 node = clear_flag(next);
bb7b2f26 1889 } while (!is_end(node));
caf3653d 1890 dbg_printf("number of logically removed nodes: %lu\n", nr_removed);
1ee8f000 1891 dbg_printf("number of bucket nodes: %lu\n", nr_bucket);
7ed7682f 1892 *approx_after = 0;
5afadd12 1893 if (ht->split_count) {
973e5e1b
MD
1894 int i;
1895
4c42f1b8
LJ
1896 for (i = 0; i < split_count_mask + 1; i++) {
1897 *approx_after += uatomic_read(&ht->split_count[i].add);
1898 *approx_after -= uatomic_read(&ht->split_count[i].del);
973e5e1b
MD
1899 }
1900 }
273399de
MD
1901}
1902
1475579c 1903/* called with resize mutex held */
abc490a1 1904static
4105056a 1905void _do_cds_lfht_grow(struct cds_lfht *ht,
1475579c 1906 unsigned long old_size, unsigned long new_size)
abc490a1 1907{
1475579c 1908 unsigned long old_order, new_order;
1475579c 1909
5bc6b66f
MD
1910 old_order = cds_lfht_get_count_order_ulong(old_size);
1911 new_order = cds_lfht_get_count_order_ulong(new_size);
1a401918
LJ
1912 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1913 old_size, old_order, new_size, new_order);
1475579c 1914 assert(new_size > old_size);
93d46c39 1915 init_table(ht, old_order + 1, new_order);
abc490a1
MD
1916}
1917
1918/* called with resize mutex held */
1919static
4105056a 1920void _do_cds_lfht_shrink(struct cds_lfht *ht,
1475579c 1921 unsigned long old_size, unsigned long new_size)
464a1ec9 1922{
1475579c 1923 unsigned long old_order, new_order;
464a1ec9 1924
d0d8f9aa 1925 new_size = max(new_size, MIN_TABLE_SIZE);
5bc6b66f
MD
1926 old_order = cds_lfht_get_count_order_ulong(old_size);
1927 new_order = cds_lfht_get_count_order_ulong(new_size);
1a401918
LJ
1928 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1929 old_size, old_order, new_size, new_order);
1475579c 1930 assert(new_size < old_size);
1475579c 1931
1ee8f000 1932 /* Remove and unlink all bucket nodes to remove. */
93d46c39 1933 fini_table(ht, new_order + 1, old_order);
464a1ec9
MD
1934}
1935
1475579c
MD
1936
1937/* called with resize mutex held */
1938static
1939void _do_cds_lfht_resize(struct cds_lfht *ht)
1940{
1941 unsigned long new_size, old_size;
4105056a
MD
1942
1943 /*
1944 * Resize table, re-do if the target size has changed under us.
1945 */
1946 do {
d2be3620
MD
1947 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1948 break;
7b3893e4
LJ
1949 ht->resize_initiated = 1;
1950 old_size = ht->size;
1951 new_size = CMM_LOAD_SHARED(ht->resize_target);
4105056a
MD
1952 if (old_size < new_size)
1953 _do_cds_lfht_grow(ht, old_size, new_size);
1954 else if (old_size > new_size)
1955 _do_cds_lfht_shrink(ht, old_size, new_size);
7b3893e4 1956 ht->resize_initiated = 0;
4105056a
MD
1957 /* write resize_initiated before read resize_target */
1958 cmm_smp_mb();
7b3893e4 1959 } while (ht->size != CMM_LOAD_SHARED(ht->resize_target));
1475579c
MD
1960}
1961
abc490a1 1962static
ab65b890 1963unsigned long resize_target_grow(struct cds_lfht *ht, unsigned long new_size)
464a1ec9 1964{
7b3893e4 1965 return _uatomic_xchg_monotonic_increase(&ht->resize_target, new_size);
464a1ec9
MD
1966}
1967
1475579c 1968static
4105056a 1969void resize_target_update_count(struct cds_lfht *ht,
b8af5011 1970 unsigned long count)
1475579c 1971{
d0d8f9aa 1972 count = max(count, MIN_TABLE_SIZE);
747d725c 1973 count = min(count, ht->max_nr_buckets);
7b3893e4 1974 uatomic_set(&ht->resize_target, count);
1475579c
MD
1975}
1976
1977void cds_lfht_resize(struct cds_lfht *ht, unsigned long new_size)
464a1ec9 1978{
10e68472
MD
1979 resize_target_update_count(ht, new_size);
1980 CMM_STORE_SHARED(ht->resize_initiated, 1);
5ffcaeef 1981 mutex_lock(&ht->resize_mutex);
1475579c 1982 _do_cds_lfht_resize(ht);
5ffcaeef 1983 mutex_unlock(&ht->resize_mutex);
abc490a1 1984}
464a1ec9 1985
abc490a1 1986static
d0ec0ed2 1987void do_resize_cb(struct urcu_work *work)
abc490a1 1988{
d0ec0ed2
MD
1989 struct resize_work *resize_work =
1990 caa_container_of(work, struct resize_work, work);
1991 struct cds_lfht *ht = resize_work->ht;
abc490a1 1992
d0ec0ed2 1993 ht->flavor->register_thread();
5ffcaeef 1994 mutex_lock(&ht->resize_mutex);
14044b37 1995 _do_cds_lfht_resize(ht);
5ffcaeef 1996 mutex_unlock(&ht->resize_mutex);
d0ec0ed2 1997 ht->flavor->unregister_thread();
98808fb1 1998 poison_free(work);
464a1ec9
MD
1999}
2000
abc490a1 2001static
f1f119ee 2002void __cds_lfht_resize_lazy_launch(struct cds_lfht *ht)
ab7d5fc6 2003{
d0ec0ed2 2004 struct resize_work *work;
abc490a1 2005
4105056a
MD
2006 /* Store resize_target before read resize_initiated */
2007 cmm_smp_mb();
7b3893e4 2008 if (!CMM_LOAD_SHARED(ht->resize_initiated)) {
ed35e6d8 2009 if (CMM_LOAD_SHARED(ht->in_progress_destroy)) {
59290e9d 2010 return;
ed35e6d8 2011 }
f9830efd 2012 work = malloc(sizeof(*work));
741f378e
MD
2013 if (work == NULL) {
2014 dbg_printf("error allocating resize work, bailing out\n");
741f378e
MD
2015 return;
2016 }
f9830efd 2017 work->ht = ht;
d0ec0ed2
MD
2018 urcu_workqueue_queue_work(cds_lfht_workqueue,
2019 &work->work, do_resize_cb);
7b3893e4 2020 CMM_STORE_SHARED(ht->resize_initiated, 1);
f9830efd 2021 }
ab7d5fc6 2022}
3171717f 2023
f1f119ee
LJ
2024static
2025void cds_lfht_resize_lazy_grow(struct cds_lfht *ht, unsigned long size, int growth)
2026{
2027 unsigned long target_size = size << growth;
2028
747d725c 2029 target_size = min(target_size, ht->max_nr_buckets);
f1f119ee
LJ
2030 if (resize_target_grow(ht, target_size) >= target_size)
2031 return;
2032
2033 __cds_lfht_resize_lazy_launch(ht);
2034}
2035
89bb121d
LJ
2036/*
2037 * We favor grow operations over shrink. A shrink operation never occurs
2038 * if a grow operation is queued for lazy execution. A grow operation
2039 * cancels any pending shrink lazy execution.
2040 */
3171717f 2041static
4105056a 2042void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
3171717f
MD
2043 unsigned long count)
2044{
b8af5011
MD
2045 if (!(ht->flags & CDS_LFHT_AUTO_RESIZE))
2046 return;
d0d8f9aa 2047 count = max(count, MIN_TABLE_SIZE);
747d725c 2048 count = min(count, ht->max_nr_buckets);
89bb121d
LJ
2049 if (count == size)
2050 return; /* Already the right size, no resize needed */
2051 if (count > size) { /* lazy grow */
2052 if (resize_target_grow(ht, count) >= count)
2053 return;
2054 } else { /* lazy shrink */
2055 for (;;) {
2056 unsigned long s;
2057
7b3893e4 2058 s = uatomic_cmpxchg(&ht->resize_target, size, count);
89bb121d
LJ
2059 if (s == size)
2060 break; /* no resize needed */
2061 if (s > size)
2062 return; /* growing is/(was just) in progress */
2063 if (s <= count)
2064 return; /* some other thread do shrink */
2065 size = s;
2066 }
2067 }
f1f119ee 2068 __cds_lfht_resize_lazy_launch(ht);
3171717f 2069}
d0ec0ed2 2070
d0ec0ed2
MD
2071static void cds_lfht_before_fork(void *priv)
2072{
2073 if (cds_lfht_workqueue_atfork_nesting++)
2074 return;
2075 mutex_lock(&cds_lfht_fork_mutex);
2076 if (!cds_lfht_workqueue)
2077 return;
2078 urcu_workqueue_pause_worker(cds_lfht_workqueue);
2079}
2080
2081static void cds_lfht_after_fork_parent(void *priv)
2082{
2083 if (--cds_lfht_workqueue_atfork_nesting)
2084 return;
2085 if (!cds_lfht_workqueue)
2086 goto end;
2087 urcu_workqueue_resume_worker(cds_lfht_workqueue);
2088end:
2089 mutex_unlock(&cds_lfht_fork_mutex);
2090}
2091
2092static void cds_lfht_after_fork_child(void *priv)
2093{
2094 if (--cds_lfht_workqueue_atfork_nesting)
2095 return;
2096 if (!cds_lfht_workqueue)
2097 goto end;
2098 urcu_workqueue_create_worker(cds_lfht_workqueue);
2099end:
2100 mutex_unlock(&cds_lfht_fork_mutex);
2101}
2102
2103static struct urcu_atfork cds_lfht_atfork = {
2104 .before_fork = cds_lfht_before_fork,
2105 .after_fork_parent = cds_lfht_after_fork_parent,
2106 .after_fork_child = cds_lfht_after_fork_child,
2107};
2108
2109/* Block all signals to ensure we don't disturb the application. */
2110static void cds_lfht_worker_init(struct urcu_workqueue *workqueue,
2111 void *priv)
2112{
2113 int ret;
2114 sigset_t mask;
2115
2116 /* Block signal for entire process, so only our thread processes it. */
2117 ret = sigfillset(&mask);
2118 if (ret)
2119 urcu_die(errno);
2120 ret = pthread_sigmask(SIG_BLOCK, &mask, NULL);
2121 if (ret)
2122 urcu_die(ret);
2123}
2124
2125static void cds_lfht_init_worker(const struct rcu_flavor_struct *flavor)
2126{
2127 flavor->register_rculfhash_atfork(&cds_lfht_atfork);
2128
2129 mutex_lock(&cds_lfht_fork_mutex);
2130 if (cds_lfht_workqueue_user_count++)
2131 goto end;
2132 cds_lfht_workqueue = urcu_workqueue_create(0, -1, NULL,
2133 NULL, cds_lfht_worker_init, NULL, NULL, NULL, NULL, NULL);
2134end:
2135 mutex_unlock(&cds_lfht_fork_mutex);
2136}
2137
2138static void cds_lfht_fini_worker(const struct rcu_flavor_struct *flavor)
2139{
2140 mutex_lock(&cds_lfht_fork_mutex);
2141 if (--cds_lfht_workqueue_user_count)
2142 goto end;
2143 urcu_workqueue_destroy(cds_lfht_workqueue);
2144 cds_lfht_workqueue = NULL;
2145end:
2146 mutex_unlock(&cds_lfht_fork_mutex);
2147
2148 flavor->unregister_rculfhash_atfork(&cds_lfht_atfork);
2149}
This page took 0.197765 seconds and 4 git commands to generate.