Use workqueue in rculfhash
[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
df44348d 588static long nr_cpus_mask = -1;
4c42f1b8 589static long split_count_mask = -1;
e53ab1eb 590static int split_count_order = -1;
4c42f1b8 591
4ddbb355 592#if defined(HAVE_SYSCONF)
4c42f1b8
LJ
593static void ht_init_nr_cpus_mask(void)
594{
595 long maxcpus;
596
597 maxcpus = sysconf(_SC_NPROCESSORS_CONF);
598 if (maxcpus <= 0) {
599 nr_cpus_mask = -2;
600 return;
601 }
602 /*
603 * round up number of CPUs to next power of two, so we
604 * can use & for modulo.
605 */
5bc6b66f 606 maxcpus = 1UL << cds_lfht_get_count_order_ulong(maxcpus);
4c42f1b8
LJ
607 nr_cpus_mask = maxcpus - 1;
608}
4ddbb355
LJ
609#else /* #if defined(HAVE_SYSCONF) */
610static void ht_init_nr_cpus_mask(void)
611{
612 nr_cpus_mask = -2;
613}
614#endif /* #else #if defined(HAVE_SYSCONF) */
df44348d
MD
615
616static
5afadd12 617void alloc_split_items_count(struct cds_lfht *ht)
df44348d 618{
4c42f1b8
LJ
619 if (nr_cpus_mask == -1) {
620 ht_init_nr_cpus_mask();
4ddbb355
LJ
621 if (nr_cpus_mask < 0)
622 split_count_mask = DEFAULT_SPLIT_COUNT_MASK;
623 else
624 split_count_mask = nr_cpus_mask;
e53ab1eb
MD
625 split_count_order =
626 cds_lfht_get_count_order_ulong(split_count_mask + 1);
df44348d 627 }
4c42f1b8 628
4ddbb355 629 assert(split_count_mask >= 0);
5afadd12
LJ
630
631 if (ht->flags & CDS_LFHT_ACCOUNTING) {
95bc7fb9
MD
632 ht->split_count = calloc(split_count_mask + 1,
633 sizeof(struct ht_items_count));
5afadd12
LJ
634 assert(ht->split_count);
635 } else {
636 ht->split_count = NULL;
637 }
df44348d
MD
638}
639
640static
5afadd12 641void free_split_items_count(struct cds_lfht *ht)
df44348d 642{
5afadd12 643 poison_free(ht->split_count);
df44348d
MD
644}
645
646static
14360f1c 647int ht_get_split_count_index(unsigned long hash)
df44348d
MD
648{
649 int cpu;
650
4c42f1b8 651 assert(split_count_mask >= 0);
a47dd11c 652 cpu = urcu_sched_getcpu();
8ed51e04 653 if (caa_unlikely(cpu < 0))
14360f1c 654 return hash & split_count_mask;
df44348d 655 else
4c42f1b8 656 return cpu & split_count_mask;
df44348d
MD
657}
658
659static
14360f1c 660void ht_count_add(struct cds_lfht *ht, unsigned long size, unsigned long hash)
df44348d 661{
4c42f1b8
LJ
662 unsigned long split_count;
663 int index;
314558bf 664 long count;
df44348d 665
8ed51e04 666 if (caa_unlikely(!ht->split_count))
3171717f 667 return;
14360f1c 668 index = ht_get_split_count_index(hash);
4c42f1b8 669 split_count = uatomic_add_return(&ht->split_count[index].add, 1);
314558bf
MD
670 if (caa_likely(split_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))
671 return;
672 /* Only if number of add multiple of 1UL << COUNT_COMMIT_ORDER */
673
674 dbg_printf("add split count %lu\n", split_count);
675 count = uatomic_add_return(&ht->count,
676 1UL << COUNT_COMMIT_ORDER);
4c299dcb 677 if (caa_likely(count & (count - 1)))
314558bf
MD
678 return;
679 /* Only if global count is power of 2 */
680
681 if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) < size)
682 return;
683 dbg_printf("add set global %ld\n", count);
684 cds_lfht_resize_lazy_count(ht, size,
685 count >> (CHAIN_LEN_TARGET - 1));
df44348d
MD
686}
687
688static
14360f1c 689void ht_count_del(struct cds_lfht *ht, unsigned long size, unsigned long hash)
df44348d 690{
4c42f1b8
LJ
691 unsigned long split_count;
692 int index;
314558bf 693 long count;
df44348d 694
8ed51e04 695 if (caa_unlikely(!ht->split_count))
3171717f 696 return;
14360f1c 697 index = ht_get_split_count_index(hash);
4c42f1b8 698 split_count = uatomic_add_return(&ht->split_count[index].del, 1);
314558bf
MD
699 if (caa_likely(split_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))
700 return;
701 /* Only if number of deletes multiple of 1UL << COUNT_COMMIT_ORDER */
702
703 dbg_printf("del split count %lu\n", split_count);
704 count = uatomic_add_return(&ht->count,
705 -(1UL << COUNT_COMMIT_ORDER));
4c299dcb 706 if (caa_likely(count & (count - 1)))
314558bf
MD
707 return;
708 /* Only if global count is power of 2 */
709
710 if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) >= size)
711 return;
712 dbg_printf("del set global %ld\n", count);
713 /*
714 * Don't shrink table if the number of nodes is below a
715 * certain threshold.
716 */
717 if (count < (1UL << COUNT_COMMIT_ORDER) * (split_count_mask + 1))
718 return;
719 cds_lfht_resize_lazy_count(ht, size,
720 count >> (CHAIN_LEN_TARGET - 1));
df44348d
MD
721}
722
f9830efd 723static
4105056a 724void check_resize(struct cds_lfht *ht, unsigned long size, uint32_t chain_len)
f9830efd 725{
f8994aee
MD
726 unsigned long count;
727
b8af5011
MD
728 if (!(ht->flags & CDS_LFHT_AUTO_RESIZE))
729 return;
f8994aee
MD
730 count = uatomic_read(&ht->count);
731 /*
732 * Use bucket-local length for small table expand and for
733 * environments lacking per-cpu data support.
734 */
e53ab1eb 735 if (count >= (1UL << (COUNT_COMMIT_ORDER + split_count_order)))
f8994aee 736 return;
24365af7 737 if (chain_len > 100)
f0c29ed7 738 dbg_printf("WARNING: large chain length: %u.\n",
24365af7 739 chain_len);
e53ab1eb
MD
740 if (chain_len >= CHAIN_LEN_RESIZE_THRESHOLD) {
741 int growth;
742
743 /*
744 * Ideal growth calculated based on chain length.
745 */
746 growth = cds_lfht_get_count_order_u32(chain_len
747 - (CHAIN_LEN_TARGET - 1));
748 if ((ht->flags & CDS_LFHT_ACCOUNTING)
749 && (size << growth)
750 >= (1UL << (COUNT_COMMIT_ORDER
751 + split_count_order))) {
752 /*
753 * If ideal growth expands the hash table size
754 * beyond the "small hash table" sizes, use the
755 * maximum small hash table size to attempt
756 * expanding the hash table. This only applies
757 * when node accounting is available, otherwise
758 * the chain length is used to expand the hash
759 * table in every case.
760 */
761 growth = COUNT_COMMIT_ORDER + split_count_order
762 - cds_lfht_get_count_order_ulong(size);
763 if (growth <= 0)
764 return;
765 }
766 cds_lfht_resize_lazy_grow(ht, size, growth);
767 }
f9830efd
MD
768}
769
abc490a1 770static
14044b37 771struct cds_lfht_node *clear_flag(struct cds_lfht_node *node)
abc490a1 772{
14044b37 773 return (struct cds_lfht_node *) (((unsigned long) node) & ~FLAGS_MASK);
abc490a1
MD
774}
775
776static
14044b37 777int is_removed(struct cds_lfht_node *node)
abc490a1 778{
d37166c6 779 return ((unsigned long) node) & REMOVED_FLAG;
abc490a1
MD
780}
781
f5596c94 782static
1ee8f000 783int is_bucket(struct cds_lfht_node *node)
f5596c94 784{
1ee8f000 785 return ((unsigned long) node) & BUCKET_FLAG;
f5596c94
MD
786}
787
788static
1ee8f000 789struct cds_lfht_node *flag_bucket(struct cds_lfht_node *node)
f5596c94 790{
1ee8f000 791 return (struct cds_lfht_node *) (((unsigned long) node) | BUCKET_FLAG);
f5596c94 792}
bb7b2f26 793
db00ccc3
MD
794static
795int is_removal_owner(struct cds_lfht_node *node)
796{
797 return ((unsigned long) node) & REMOVAL_OWNER_FLAG;
798}
799
800static
801struct cds_lfht_node *flag_removal_owner(struct cds_lfht_node *node)
802{
803 return (struct cds_lfht_node *) (((unsigned long) node) | REMOVAL_OWNER_FLAG);
804}
805
71bb3aca
MD
806static
807struct cds_lfht_node *flag_removed_or_removal_owner(struct cds_lfht_node *node)
808{
809 return (struct cds_lfht_node *) (((unsigned long) node) | REMOVED_FLAG | REMOVAL_OWNER_FLAG);
810}
811
bb7b2f26
MD
812static
813struct cds_lfht_node *get_end(void)
814{
815 return (struct cds_lfht_node *) END_VALUE;
816}
817
818static
819int is_end(struct cds_lfht_node *node)
820{
821 return clear_flag(node) == (struct cds_lfht_node *) END_VALUE;
822}
823
abc490a1 824static
ab65b890
LJ
825unsigned long _uatomic_xchg_monotonic_increase(unsigned long *ptr,
826 unsigned long v)
abc490a1
MD
827{
828 unsigned long old1, old2;
829
830 old1 = uatomic_read(ptr);
831 do {
832 old2 = old1;
833 if (old2 >= v)
f9830efd 834 return old2;
abc490a1 835 } while ((old1 = uatomic_cmpxchg(ptr, old2, v)) != old2);
ab65b890 836 return old2;
abc490a1
MD
837}
838
48f1b16d
LJ
839static
840void cds_lfht_alloc_bucket_table(struct cds_lfht *ht, unsigned long order)
841{
0b6aa001 842 return ht->mm->alloc_bucket_table(ht, order);
48f1b16d
LJ
843}
844
845/*
846 * cds_lfht_free_bucket_table() should be called with decreasing order.
847 * When cds_lfht_free_bucket_table(0) is called, it means the whole
848 * lfht is destroyed.
849 */
850static
851void cds_lfht_free_bucket_table(struct cds_lfht *ht, unsigned long order)
852{
0b6aa001 853 return ht->mm->free_bucket_table(ht, order);
48f1b16d
LJ
854}
855
9d72a73f
LJ
856static inline
857struct cds_lfht_node *bucket_at(struct cds_lfht *ht, unsigned long index)
f4a9cc0b 858{
0b6aa001 859 return ht->bucket_at(ht, index);
f4a9cc0b
LJ
860}
861
9d72a73f
LJ
862static inline
863struct cds_lfht_node *lookup_bucket(struct cds_lfht *ht, unsigned long size,
864 unsigned long hash)
865{
866 assert(size > 0);
867 return bucket_at(ht, hash & (size - 1));
868}
869
273399de
MD
870/*
871 * Remove all logically deleted nodes from a bucket up to a certain node key.
872 */
873static
1ee8f000 874void _cds_lfht_gc_bucket(struct cds_lfht_node *bucket, struct cds_lfht_node *node)
273399de 875{
14044b37 876 struct cds_lfht_node *iter_prev, *iter, *next, *new_next;
273399de 877
1ee8f000
LJ
878 assert(!is_bucket(bucket));
879 assert(!is_removed(bucket));
2f943cd7 880 assert(!is_removal_owner(bucket));
1ee8f000 881 assert(!is_bucket(node));
c90201ac 882 assert(!is_removed(node));
2f943cd7 883 assert(!is_removal_owner(node));
273399de 884 for (;;) {
1ee8f000
LJ
885 iter_prev = bucket;
886 /* We can always skip the bucket node initially */
04db56f8 887 iter = rcu_dereference(iter_prev->next);
b4cb483f 888 assert(!is_removed(iter));
2f943cd7 889 assert(!is_removal_owner(iter));
04db56f8 890 assert(iter_prev->reverse_hash <= node->reverse_hash);
bd4db153 891 /*
1ee8f000 892 * We should never be called with bucket (start of chain)
bd4db153
MD
893 * and logically removed node (end of path compression
894 * marker) being the actual same node. This would be a
895 * bug in the algorithm implementation.
896 */
1ee8f000 897 assert(bucket != node);
273399de 898 for (;;) {
8ed51e04 899 if (caa_unlikely(is_end(iter)))
f9c80341 900 return;
04db56f8 901 if (caa_likely(clear_flag(iter)->reverse_hash > node->reverse_hash))
f9c80341 902 return;
04db56f8 903 next = rcu_dereference(clear_flag(iter)->next);
8ed51e04 904 if (caa_likely(is_removed(next)))
273399de 905 break;
b453eae1 906 iter_prev = clear_flag(iter);
273399de
MD
907 iter = next;
908 }
b198f0fd 909 assert(!is_removed(iter));
2f943cd7 910 assert(!is_removal_owner(iter));
1ee8f000
LJ
911 if (is_bucket(iter))
912 new_next = flag_bucket(clear_flag(next));
f5596c94
MD
913 else
914 new_next = clear_flag(next);
04db56f8 915 (void) uatomic_cmpxchg(&iter_prev->next, iter, new_next);
273399de
MD
916 }
917}
918
9357c415
MD
919static
920int _cds_lfht_replace(struct cds_lfht *ht, unsigned long size,
921 struct cds_lfht_node *old_node,
3fb86f26 922 struct cds_lfht_node *old_next,
9357c415
MD
923 struct cds_lfht_node *new_node)
924{
04db56f8 925 struct cds_lfht_node *bucket, *ret_next;
9357c415
MD
926
927 if (!old_node) /* Return -ENOENT if asked to replace NULL node */
7801dadd 928 return -ENOENT;
9357c415
MD
929
930 assert(!is_removed(old_node));
2f943cd7 931 assert(!is_removal_owner(old_node));
1ee8f000 932 assert(!is_bucket(old_node));
9357c415 933 assert(!is_removed(new_node));
2f943cd7 934 assert(!is_removal_owner(new_node));
1ee8f000 935 assert(!is_bucket(new_node));
9357c415 936 assert(new_node != old_node);
3fb86f26 937 for (;;) {
9357c415 938 /* Insert after node to be replaced */
9357c415
MD
939 if (is_removed(old_next)) {
940 /*
941 * Too late, the old node has been removed under us
942 * between lookup and replace. Fail.
943 */
7801dadd 944 return -ENOENT;
9357c415 945 }
feda2722
LJ
946 assert(old_next == clear_flag(old_next));
947 assert(new_node != old_next);
71bb3aca
MD
948 /*
949 * REMOVAL_OWNER flag is _NEVER_ set before the REMOVED
950 * flag. It is either set atomically at the same time
951 * (replace) or after (del).
952 */
953 assert(!is_removal_owner(old_next));
feda2722 954 new_node->next = old_next;
9357c415
MD
955 /*
956 * Here is the whole trick for lock-free replace: we add
957 * the replacement node _after_ the node we want to
958 * replace by atomically setting its next pointer at the
959 * same time we set its removal flag. Given that
960 * the lookups/get next use an iterator aware of the
961 * next pointer, they will either skip the old node due
962 * to the removal flag and see the new node, or use
963 * the old node, but will not see the new one.
db00ccc3
MD
964 * This is a replacement of a node with another node
965 * that has the same value: we are therefore not
71bb3aca
MD
966 * removing a value from the hash table. We set both the
967 * REMOVED and REMOVAL_OWNER flags atomically so we own
968 * the node after successful cmpxchg.
9357c415 969 */
04db56f8 970 ret_next = uatomic_cmpxchg(&old_node->next,
71bb3aca 971 old_next, flag_removed_or_removal_owner(new_node));
3fb86f26 972 if (ret_next == old_next)
7801dadd 973 break; /* We performed the replacement. */
3fb86f26
LJ
974 old_next = ret_next;
975 }
9357c415 976
9357c415
MD
977 /*
978 * Ensure that the old node is not visible to readers anymore:
979 * lookup for the node, and remove it (along with any other
980 * logically removed node) if found.
981 */
04db56f8
LJ
982 bucket = lookup_bucket(ht, size, bit_reverse_ulong(old_node->reverse_hash));
983 _cds_lfht_gc_bucket(bucket, new_node);
7801dadd 984
a85eff52 985 assert(is_removed(CMM_LOAD_SHARED(old_node->next)));
7801dadd 986 return 0;
9357c415
MD
987}
988
83beee94
MD
989/*
990 * A non-NULL unique_ret pointer uses the "add unique" (or uniquify) add
991 * mode. A NULL unique_ret allows creation of duplicate keys.
992 */
abc490a1 993static
83beee94 994void _cds_lfht_add(struct cds_lfht *ht,
91a75cc5 995 unsigned long hash,
0422d92c 996 cds_lfht_match_fct match,
996ff57c 997 const void *key,
83beee94
MD
998 unsigned long size,
999 struct cds_lfht_node *node,
1000 struct cds_lfht_iter *unique_ret,
1ee8f000 1001 int bucket_flag)
abc490a1 1002{
14044b37 1003 struct cds_lfht_node *iter_prev, *iter, *next, *new_node, *new_next,
960c9e4f 1004 *return_node;
04db56f8 1005 struct cds_lfht_node *bucket;
abc490a1 1006
1ee8f000 1007 assert(!is_bucket(node));
c90201ac 1008 assert(!is_removed(node));
2f943cd7 1009 assert(!is_removal_owner(node));
91a75cc5 1010 bucket = lookup_bucket(ht, size, hash);
abc490a1 1011 for (;;) {
adc0de68 1012 uint32_t chain_len = 0;
abc490a1 1013
11519af6
MD
1014 /*
1015 * iter_prev points to the non-removed node prior to the
1016 * insert location.
11519af6 1017 */
04db56f8 1018 iter_prev = bucket;
1ee8f000 1019 /* We can always skip the bucket node initially */
04db56f8
LJ
1020 iter = rcu_dereference(iter_prev->next);
1021 assert(iter_prev->reverse_hash <= node->reverse_hash);
abc490a1 1022 for (;;) {
8ed51e04 1023 if (caa_unlikely(is_end(iter)))
273399de 1024 goto insert;
04db56f8 1025 if (caa_likely(clear_flag(iter)->reverse_hash > node->reverse_hash))
273399de 1026 goto insert;
238cc06e 1027
1ee8f000
LJ
1028 /* bucket node is the first node of the identical-hash-value chain */
1029 if (bucket_flag && clear_flag(iter)->reverse_hash == node->reverse_hash)
194fdbd1 1030 goto insert;
238cc06e 1031
04db56f8 1032 next = rcu_dereference(clear_flag(iter)->next);
8ed51e04 1033 if (caa_unlikely(is_removed(next)))
9dba85be 1034 goto gc_node;
238cc06e
LJ
1035
1036 /* uniquely add */
83beee94 1037 if (unique_ret
1ee8f000 1038 && !is_bucket(next)
04db56f8 1039 && clear_flag(iter)->reverse_hash == node->reverse_hash) {
238cc06e
LJ
1040 struct cds_lfht_iter d_iter = { .node = node, .next = iter, };
1041
1042 /*
1043 * uniquely adding inserts the node as the first
1044 * node of the identical-hash-value node chain.
1045 *
1046 * This semantic ensures no duplicated keys
1047 * should ever be observable in the table
1f67ba50
MD
1048 * (including traversing the table node by
1049 * node by forward iterations)
238cc06e 1050 */
04db56f8 1051 cds_lfht_next_duplicate(ht, match, key, &d_iter);
238cc06e
LJ
1052 if (!d_iter.node)
1053 goto insert;
1054
1055 *unique_ret = d_iter;
83beee94 1056 return;
48ed1c18 1057 }
238cc06e 1058
11519af6 1059 /* Only account for identical reverse hash once */
04db56f8 1060 if (iter_prev->reverse_hash != clear_flag(iter)->reverse_hash
1ee8f000 1061 && !is_bucket(next))
4105056a 1062 check_resize(ht, size, ++chain_len);
11519af6 1063 iter_prev = clear_flag(iter);
273399de 1064 iter = next;
abc490a1 1065 }
48ed1c18 1066
273399de 1067 insert:
7ec59d3b 1068 assert(node != clear_flag(iter));
11519af6 1069 assert(!is_removed(iter_prev));
2f943cd7 1070 assert(!is_removal_owner(iter_prev));
c90201ac 1071 assert(!is_removed(iter));
2f943cd7 1072 assert(!is_removal_owner(iter));
f000907d 1073 assert(iter_prev != node);
1ee8f000 1074 if (!bucket_flag)
04db56f8 1075 node->next = clear_flag(iter);
f9c80341 1076 else
1ee8f000
LJ
1077 node->next = flag_bucket(clear_flag(iter));
1078 if (is_bucket(iter))
1079 new_node = flag_bucket(node);
f5596c94
MD
1080 else
1081 new_node = node;
04db56f8 1082 if (uatomic_cmpxchg(&iter_prev->next, iter,
48ed1c18 1083 new_node) != iter) {
273399de 1084 continue; /* retry */
48ed1c18 1085 } else {
83beee94 1086 return_node = node;
960c9e4f 1087 goto end;
48ed1c18
MD
1088 }
1089
9dba85be
MD
1090 gc_node:
1091 assert(!is_removed(iter));
2f943cd7 1092 assert(!is_removal_owner(iter));
1ee8f000
LJ
1093 if (is_bucket(iter))
1094 new_next = flag_bucket(clear_flag(next));
f5596c94
MD
1095 else
1096 new_next = clear_flag(next);
04db56f8 1097 (void) uatomic_cmpxchg(&iter_prev->next, iter, new_next);
273399de 1098 /* retry */
464a1ec9 1099 }
9357c415 1100end:
83beee94
MD
1101 if (unique_ret) {
1102 unique_ret->node = return_node;
1103 /* unique_ret->next left unset, never used. */
1104 }
abc490a1 1105}
464a1ec9 1106
abc490a1 1107static
860d07e8 1108int _cds_lfht_del(struct cds_lfht *ht, unsigned long size,
b65ec430 1109 struct cds_lfht_node *node)
abc490a1 1110{
db00ccc3 1111 struct cds_lfht_node *bucket, *next;
5e28c532 1112
9357c415 1113 if (!node) /* Return -ENOENT if asked to delete NULL node */
743f9143 1114 return -ENOENT;
9357c415 1115
7ec59d3b 1116 /* logically delete the node */
1ee8f000 1117 assert(!is_bucket(node));
c90201ac 1118 assert(!is_removed(node));
db00ccc3 1119 assert(!is_removal_owner(node));
48ed1c18 1120
db00ccc3
MD
1121 /*
1122 * We are first checking if the node had previously been
1123 * logically removed (this check is not atomic with setting the
1124 * logical removal flag). Return -ENOENT if the node had
1125 * previously been removed.
1126 */
a85eff52 1127 next = CMM_LOAD_SHARED(node->next); /* next is not dereferenced */
db00ccc3
MD
1128 if (caa_unlikely(is_removed(next)))
1129 return -ENOENT;
b65ec430 1130 assert(!is_bucket(next));
196f4fab
MD
1131 /*
1132 * The del operation semantic guarantees a full memory barrier
1133 * before the uatomic_or atomic commit of the deletion flag.
1134 */
1135 cmm_smp_mb__before_uatomic_or();
db00ccc3
MD
1136 /*
1137 * We set the REMOVED_FLAG unconditionally. Note that there may
1138 * be more than one concurrent thread setting this flag.
1139 * Knowing which wins the race will be known after the garbage
1140 * collection phase, stay tuned!
1141 */
1142 uatomic_or(&node->next, REMOVED_FLAG);
7ec59d3b 1143 /* We performed the (logical) deletion. */
7ec59d3b
MD
1144
1145 /*
1146 * Ensure that the node is not visible to readers anymore: lookup for
273399de
MD
1147 * the node, and remove it (along with any other logically removed node)
1148 * if found.
11519af6 1149 */
04db56f8
LJ
1150 bucket = lookup_bucket(ht, size, bit_reverse_ulong(node->reverse_hash));
1151 _cds_lfht_gc_bucket(bucket, node);
743f9143 1152
a85eff52 1153 assert(is_removed(CMM_LOAD_SHARED(node->next)));
db00ccc3
MD
1154 /*
1155 * Last phase: atomically exchange node->next with a version
1156 * having "REMOVAL_OWNER_FLAG" set. If the returned node->next
1157 * pointer did _not_ have "REMOVAL_OWNER_FLAG" set, we now own
1158 * the node and win the removal race.
1159 * It is interesting to note that all "add" paths are forbidden
1160 * to change the next pointer starting from the point where the
1161 * REMOVED_FLAG is set, so here using a read, followed by a
1162 * xchg() suffice to guarantee that the xchg() will ever only
1163 * set the "REMOVAL_OWNER_FLAG" (or change nothing if the flag
1164 * was already set).
1165 */
1166 if (!is_removal_owner(uatomic_xchg(&node->next,
1167 flag_removal_owner(node->next))))
1168 return 0;
1169 else
1170 return -ENOENT;
abc490a1 1171}
2ed95849 1172
b7d619b0
MD
1173static
1174void *partition_resize_thread(void *arg)
1175{
1176 struct partition_resize_work *work = arg;
1177
7b17c13e 1178 work->ht->flavor->register_thread();
b7d619b0 1179 work->fct(work->ht, work->i, work->start, work->len);
7b17c13e 1180 work->ht->flavor->unregister_thread();
b7d619b0
MD
1181 return NULL;
1182}
1183
1184static
1185void partition_resize_helper(struct cds_lfht *ht, unsigned long i,
1186 unsigned long len,
1187 void (*fct)(struct cds_lfht *ht, unsigned long i,
1188 unsigned long start, unsigned long len))
1189{
e54ec2f5 1190 unsigned long partition_len, start = 0;
b7d619b0 1191 struct partition_resize_work *work;
6083a889
MD
1192 int thread, ret;
1193 unsigned long nr_threads;
b7d619b0 1194
d7f3ba4c
EW
1195 assert(nr_cpus_mask != -1);
1196 if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD)
1197 goto fallback;
1198
6083a889
MD
1199 /*
1200 * Note: nr_cpus_mask + 1 is always power of 2.
1201 * We spawn just the number of threads we need to satisfy the minimum
1202 * partition size, up to the number of CPUs in the system.
1203 */
91452a6a
MD
1204 if (nr_cpus_mask > 0) {
1205 nr_threads = min(nr_cpus_mask + 1,
1206 len >> MIN_PARTITION_PER_THREAD_ORDER);
1207 } else {
1208 nr_threads = 1;
1209 }
5bc6b66f 1210 partition_len = len >> cds_lfht_get_count_order_ulong(nr_threads);
6083a889 1211 work = calloc(nr_threads, sizeof(*work));
7c75d498
EW
1212 if (!work) {
1213 dbg_printf("error allocating for resize, single-threading\n");
1214 goto fallback;
1215 }
6083a889
MD
1216 for (thread = 0; thread < nr_threads; thread++) {
1217 work[thread].ht = ht;
1218 work[thread].i = i;
1219 work[thread].len = partition_len;
1220 work[thread].start = thread * partition_len;
1221 work[thread].fct = fct;
1af6e26e 1222 ret = pthread_create(&(work[thread].thread_id), ht->resize_attr,
6083a889 1223 partition_resize_thread, &work[thread]);
e54ec2f5
EW
1224 if (ret == EAGAIN) {
1225 /*
1226 * Out of resources: wait and join the threads
1227 * we've created, then handle leftovers.
1228 */
1229 dbg_printf("error spawning for resize, single-threading\n");
1230 start = work[thread].start;
1231 len -= start;
1232 nr_threads = thread;
1233 break;
1234 }
b7d619b0
MD
1235 assert(!ret);
1236 }
6083a889 1237 for (thread = 0; thread < nr_threads; thread++) {
1af6e26e 1238 ret = pthread_join(work[thread].thread_id, NULL);
b7d619b0
MD
1239 assert(!ret);
1240 }
1241 free(work);
e54ec2f5
EW
1242
1243 /*
1244 * A pthread_create failure above will either lead in us having
1245 * no threads to join or starting at a non-zero offset,
1246 * fallback to single thread processing of leftovers.
1247 */
1248 if (start == 0 && nr_threads > 0)
1249 return;
7c75d498 1250fallback:
e54ec2f5 1251 fct(ht, i, start, len);
b7d619b0
MD
1252}
1253
e8de508e
MD
1254/*
1255 * Holding RCU read lock to protect _cds_lfht_add against memory
d0ec0ed2 1256 * reclaim that could be performed by other worker threads (ABA
e8de508e 1257 * problem).
9ee0fc9a 1258 *
b7d619b0 1259 * When we reach a certain length, we can split this population phase over
9ee0fc9a
MD
1260 * many worker threads, based on the number of CPUs available in the system.
1261 * This should therefore take care of not having the expand lagging behind too
1262 * many concurrent insertion threads by using the scheduler's ability to
1ee8f000 1263 * schedule bucket node population fairly with insertions.
e8de508e 1264 */
4105056a 1265static
b7d619b0
MD
1266void init_table_populate_partition(struct cds_lfht *ht, unsigned long i,
1267 unsigned long start, unsigned long len)
4105056a 1268{
9d72a73f 1269 unsigned long j, size = 1UL << (i - 1);
4105056a 1270
d0d8f9aa 1271 assert(i > MIN_TABLE_ORDER);
7b17c13e 1272 ht->flavor->read_lock();
9d72a73f
LJ
1273 for (j = size + start; j < size + start + len; j++) {
1274 struct cds_lfht_node *new_node = bucket_at(ht, j);
1275
1276 assert(j >= size && j < (size << 1));
1277 dbg_printf("init populate: order %lu index %lu hash %lu\n",
1278 i, j, j);
1279 new_node->reverse_hash = bit_reverse_ulong(j);
91a75cc5 1280 _cds_lfht_add(ht, j, NULL, NULL, size, new_node, NULL, 1);
4105056a 1281 }
7b17c13e 1282 ht->flavor->read_unlock();
b7d619b0
MD
1283}
1284
1285static
1286void init_table_populate(struct cds_lfht *ht, unsigned long i,
1287 unsigned long len)
1288{
b7d619b0 1289 partition_resize_helper(ht, i, len, init_table_populate_partition);
4105056a
MD
1290}
1291
abc490a1 1292static
4105056a 1293void init_table(struct cds_lfht *ht,
93d46c39 1294 unsigned long first_order, unsigned long last_order)
24365af7 1295{
93d46c39 1296 unsigned long i;
24365af7 1297
93d46c39
LJ
1298 dbg_printf("init table: first_order %lu last_order %lu\n",
1299 first_order, last_order);
d0d8f9aa 1300 assert(first_order > MIN_TABLE_ORDER);
93d46c39 1301 for (i = first_order; i <= last_order; i++) {
4105056a 1302 unsigned long len;
24365af7 1303
4f6e90b7 1304 len = 1UL << (i - 1);
f0c29ed7 1305 dbg_printf("init order %lu len: %lu\n", i, len);
4d676753
MD
1306
1307 /* Stop expand if the resize target changes under us */
7b3893e4 1308 if (CMM_LOAD_SHARED(ht->resize_target) < (1UL << i))
4d676753
MD
1309 break;
1310
48f1b16d 1311 cds_lfht_alloc_bucket_table(ht, i);
4105056a 1312
4105056a 1313 /*
1ee8f000
LJ
1314 * Set all bucket nodes reverse hash values for a level and
1315 * link all bucket nodes into the table.
4105056a 1316 */
dc1da8f6 1317 init_table_populate(ht, i, len);
4105056a 1318
f9c80341
MD
1319 /*
1320 * Update table size.
1321 */
1322 cmm_smp_wmb(); /* populate data before RCU size */
7b3893e4 1323 CMM_STORE_SHARED(ht->size, 1UL << i);
f9c80341 1324
4f6e90b7 1325 dbg_printf("init new size: %lu\n", 1UL << i);
4105056a
MD
1326 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1327 break;
1328 }
1329}
1330
e8de508e
MD
1331/*
1332 * Holding RCU read lock to protect _cds_lfht_remove against memory
d0ec0ed2 1333 * reclaim that could be performed by other worker threads (ABA
e8de508e
MD
1334 * problem).
1335 * For a single level, we logically remove and garbage collect each node.
1336 *
1337 * As a design choice, we perform logical removal and garbage collection on a
1338 * node-per-node basis to simplify this algorithm. We also assume keeping good
1339 * cache locality of the operation would overweight possible performance gain
1340 * that could be achieved by batching garbage collection for multiple levels.
1341 * However, this would have to be justified by benchmarks.
1342 *
1343 * Concurrent removal and add operations are helping us perform garbage
1344 * collection of logically removed nodes. We guarantee that all logically
d0ec0ed2
MD
1345 * removed nodes have been garbage-collected (unlinked) before work
1346 * enqueue is invoked to free a hole level of bucket nodes (after a
1347 * grace period).
e8de508e 1348 *
1f67ba50
MD
1349 * Logical removal and garbage collection can therefore be done in batch
1350 * or on a node-per-node basis, as long as the guarantee above holds.
9ee0fc9a 1351 *
b7d619b0
MD
1352 * When we reach a certain length, we can split this removal over many worker
1353 * threads, based on the number of CPUs available in the system. This should
1354 * take care of not letting resize process lag behind too many concurrent
9ee0fc9a 1355 * updater threads actively inserting into the hash table.
e8de508e 1356 */
4105056a 1357static
b7d619b0
MD
1358void remove_table_partition(struct cds_lfht *ht, unsigned long i,
1359 unsigned long start, unsigned long len)
4105056a 1360{
9d72a73f 1361 unsigned long j, size = 1UL << (i - 1);
4105056a 1362
d0d8f9aa 1363 assert(i > MIN_TABLE_ORDER);
7b17c13e 1364 ht->flavor->read_lock();
9d72a73f 1365 for (j = size + start; j < size + start + len; j++) {
2e2ce1e9
LJ
1366 struct cds_lfht_node *fini_bucket = bucket_at(ht, j);
1367 struct cds_lfht_node *parent_bucket = bucket_at(ht, j - size);
9d72a73f
LJ
1368
1369 assert(j >= size && j < (size << 1));
1370 dbg_printf("remove entry: order %lu index %lu hash %lu\n",
1371 i, j, j);
2e2ce1e9
LJ
1372 /* Set the REMOVED_FLAG to freeze the ->next for gc */
1373 uatomic_or(&fini_bucket->next, REMOVED_FLAG);
1374 _cds_lfht_gc_bucket(parent_bucket, fini_bucket);
abc490a1 1375 }
7b17c13e 1376 ht->flavor->read_unlock();
b7d619b0
MD
1377}
1378
1379static
1380void remove_table(struct cds_lfht *ht, unsigned long i, unsigned long len)
1381{
b7d619b0 1382 partition_resize_helper(ht, i, len, remove_table_partition);
2ed95849
MD
1383}
1384
61adb337
MD
1385/*
1386 * fini_table() is never called for first_order == 0, which is why
1387 * free_by_rcu_order == 0 can be used as criterion to know if free must
1388 * be called.
1389 */
1475579c 1390static
4105056a 1391void fini_table(struct cds_lfht *ht,
93d46c39 1392 unsigned long first_order, unsigned long last_order)
1475579c 1393{
93d46c39 1394 long i;
48f1b16d 1395 unsigned long free_by_rcu_order = 0;
1475579c 1396
93d46c39
LJ
1397 dbg_printf("fini table: first_order %lu last_order %lu\n",
1398 first_order, last_order);
d0d8f9aa 1399 assert(first_order > MIN_TABLE_ORDER);
93d46c39 1400 for (i = last_order; i >= first_order; i--) {
4105056a 1401 unsigned long len;
1475579c 1402
4f6e90b7 1403 len = 1UL << (i - 1);
e15df1cc 1404 dbg_printf("fini order %ld len: %lu\n", i, len);
4105056a 1405
4d676753 1406 /* Stop shrink if the resize target changes under us */
7b3893e4 1407 if (CMM_LOAD_SHARED(ht->resize_target) > (1UL << (i - 1)))
4d676753
MD
1408 break;
1409
1410 cmm_smp_wmb(); /* populate data before RCU size */
7b3893e4 1411 CMM_STORE_SHARED(ht->size, 1UL << (i - 1));
4d676753
MD
1412
1413 /*
1414 * We need to wait for all add operations to reach Q.S. (and
1415 * thus use the new table for lookups) before we can start
1ee8f000 1416 * releasing the old bucket nodes. Otherwise their lookup will
4d676753
MD
1417 * return a logically removed node as insert position.
1418 */
7b17c13e 1419 ht->flavor->update_synchronize_rcu();
48f1b16d
LJ
1420 if (free_by_rcu_order)
1421 cds_lfht_free_bucket_table(ht, free_by_rcu_order);
4d676753 1422
21263e21 1423 /*
1ee8f000
LJ
1424 * Set "removed" flag in bucket nodes about to be removed.
1425 * Unlink all now-logically-removed bucket node pointers.
4105056a
MD
1426 * Concurrent add/remove operation are helping us doing
1427 * the gc.
21263e21 1428 */
4105056a
MD
1429 remove_table(ht, i, len);
1430
48f1b16d 1431 free_by_rcu_order = i;
4105056a
MD
1432
1433 dbg_printf("fini new size: %lu\n", 1UL << i);
1475579c
MD
1434 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1435 break;
1436 }
0d14ceb2 1437
48f1b16d 1438 if (free_by_rcu_order) {
7b17c13e 1439 ht->flavor->update_synchronize_rcu();
48f1b16d 1440 cds_lfht_free_bucket_table(ht, free_by_rcu_order);
0d14ceb2 1441 }
1475579c
MD
1442}
1443
ff0d69de 1444static
1ee8f000 1445void cds_lfht_create_bucket(struct cds_lfht *ht, unsigned long size)
ff0d69de 1446{
04db56f8 1447 struct cds_lfht_node *prev, *node;
9d72a73f 1448 unsigned long order, len, i;
ff0d69de 1449
48f1b16d 1450 cds_lfht_alloc_bucket_table(ht, 0);
ff0d69de 1451
9d72a73f
LJ
1452 dbg_printf("create bucket: order 0 index 0 hash 0\n");
1453 node = bucket_at(ht, 0);
1454 node->next = flag_bucket(get_end());
1455 node->reverse_hash = 0;
ff0d69de 1456
5bc6b66f 1457 for (order = 1; order < cds_lfht_get_count_order_ulong(size) + 1; order++) {
ff0d69de 1458 len = 1UL << (order - 1);
48f1b16d 1459 cds_lfht_alloc_bucket_table(ht, order);
ff0d69de 1460
9d72a73f
LJ
1461 for (i = 0; i < len; i++) {
1462 /*
1463 * Now, we are trying to init the node with the
1464 * hash=(len+i) (which is also a bucket with the
1465 * index=(len+i)) and insert it into the hash table,
1466 * so this node has to be inserted after the bucket
1467 * with the index=(len+i)&(len-1)=i. And because there
1468 * is no other non-bucket node nor bucket node with
1469 * larger index/hash inserted, so the bucket node
1470 * being inserted should be inserted directly linked
1471 * after the bucket node with index=i.
1472 */
1473 prev = bucket_at(ht, i);
1474 node = bucket_at(ht, len + i);
ff0d69de 1475
1ee8f000 1476 dbg_printf("create bucket: order %lu index %lu hash %lu\n",
9d72a73f
LJ
1477 order, len + i, len + i);
1478 node->reverse_hash = bit_reverse_ulong(len + i);
1479
1480 /* insert after prev */
1481 assert(is_bucket(prev->next));
ff0d69de 1482 node->next = prev->next;
1ee8f000 1483 prev->next = flag_bucket(node);
ff0d69de
LJ
1484 }
1485 }
1486}
1487
0422d92c 1488struct cds_lfht *_cds_lfht_new(unsigned long init_size,
0722081a 1489 unsigned long min_nr_alloc_buckets,
747d725c 1490 unsigned long max_nr_buckets,
b8af5011 1491 int flags,
0b6aa001 1492 const struct cds_lfht_mm_type *mm,
7b17c13e 1493 const struct rcu_flavor_struct *flavor,
b7d619b0 1494 pthread_attr_t *attr)
abc490a1 1495{
14044b37 1496 struct cds_lfht *ht;
24365af7 1497 unsigned long order;
abc490a1 1498
0722081a
LJ
1499 /* min_nr_alloc_buckets must be power of two */
1500 if (!min_nr_alloc_buckets || (min_nr_alloc_buckets & (min_nr_alloc_buckets - 1)))
5488222b 1501 return NULL;
747d725c 1502
8129be4e 1503 /* init_size must be power of two */
5488222b 1504 if (!init_size || (init_size & (init_size - 1)))
8129be4e 1505 return NULL;
747d725c 1506
c1888f3a
MD
1507 /*
1508 * Memory management plugin default.
1509 */
1510 if (!mm) {
5a2141a7
MD
1511 if (CAA_BITS_PER_LONG > 32
1512 && max_nr_buckets
c1888f3a
MD
1513 && max_nr_buckets <= (1ULL << 32)) {
1514 /*
1515 * For 64-bit architectures, with max number of
1516 * buckets small enough not to use the entire
1517 * 64-bit memory mapping space (and allowing a
1518 * fair number of hash table instances), use the
1519 * mmap allocator, which is faster than the
1520 * order allocator.
1521 */
1522 mm = &cds_lfht_mm_mmap;
1523 } else {
1524 /*
1525 * The fallback is to use the order allocator.
1526 */
1527 mm = &cds_lfht_mm_order;
1528 }
1529 }
1530
0b6aa001
LJ
1531 /* max_nr_buckets == 0 for order based mm means infinite */
1532 if (mm == &cds_lfht_mm_order && !max_nr_buckets)
747d725c
LJ
1533 max_nr_buckets = 1UL << (MAX_TABLE_ORDER - 1);
1534
1535 /* max_nr_buckets must be power of two */
1536 if (!max_nr_buckets || (max_nr_buckets & (max_nr_buckets - 1)))
1537 return NULL;
1538
d0ec0ed2
MD
1539 if (flags & CDS_LFHT_AUTO_RESIZE)
1540 cds_lfht_init_worker(flavor);
1541
0722081a 1542 min_nr_alloc_buckets = max(min_nr_alloc_buckets, MIN_TABLE_SIZE);
d0d8f9aa 1543 init_size = max(init_size, MIN_TABLE_SIZE);
747d725c
LJ
1544 max_nr_buckets = max(max_nr_buckets, min_nr_alloc_buckets);
1545 init_size = min(init_size, max_nr_buckets);
0b6aa001
LJ
1546
1547 ht = mm->alloc_cds_lfht(min_nr_alloc_buckets, max_nr_buckets);
b7d619b0 1548 assert(ht);
0b6aa001
LJ
1549 assert(ht->mm == mm);
1550 assert(ht->bucket_at == mm->bucket_at);
1551
b5d6b20f 1552 ht->flags = flags;
7b17c13e 1553 ht->flavor = flavor;
b7d619b0 1554 ht->resize_attr = attr;
5afadd12 1555 alloc_split_items_count(ht);
abc490a1
MD
1556 /* this mutex should not nest in read-side C.S. */
1557 pthread_mutex_init(&ht->resize_mutex, NULL);
5bc6b66f 1558 order = cds_lfht_get_count_order_ulong(init_size);
7b3893e4 1559 ht->resize_target = 1UL << order;
1ee8f000 1560 cds_lfht_create_bucket(ht, 1UL << order);
7b3893e4 1561 ht->size = 1UL << order;
abc490a1
MD
1562 return ht;
1563}
1564
6f554439 1565void cds_lfht_lookup(struct cds_lfht *ht, unsigned long hash,
996ff57c 1566 cds_lfht_match_fct match, const void *key,
6f554439 1567 struct cds_lfht_iter *iter)
2ed95849 1568{
04db56f8 1569 struct cds_lfht_node *node, *next, *bucket;
0422d92c 1570 unsigned long reverse_hash, size;
2ed95849 1571
abc490a1 1572 reverse_hash = bit_reverse_ulong(hash);
464a1ec9 1573
7b3893e4 1574 size = rcu_dereference(ht->size);
04db56f8 1575 bucket = lookup_bucket(ht, size, hash);
1ee8f000 1576 /* We can always skip the bucket node initially */
04db56f8 1577 node = rcu_dereference(bucket->next);
bb7b2f26 1578 node = clear_flag(node);
2ed95849 1579 for (;;) {
8ed51e04 1580 if (caa_unlikely(is_end(node))) {
96ad1112 1581 node = next = NULL;
abc490a1 1582 break;
bb7b2f26 1583 }
04db56f8 1584 if (caa_unlikely(node->reverse_hash > reverse_hash)) {
96ad1112 1585 node = next = NULL;
abc490a1 1586 break;
2ed95849 1587 }
04db56f8 1588 next = rcu_dereference(node->next);
7f52427b 1589 assert(node == clear_flag(node));
8ed51e04 1590 if (caa_likely(!is_removed(next))
1ee8f000 1591 && !is_bucket(next)
04db56f8 1592 && node->reverse_hash == reverse_hash
0422d92c 1593 && caa_likely(match(node, key))) {
273399de 1594 break;
2ed95849 1595 }
1b81fe1a 1596 node = clear_flag(next);
2ed95849 1597 }
a85eff52 1598 assert(!node || !is_bucket(CMM_LOAD_SHARED(node->next)));
adc0de68
MD
1599 iter->node = node;
1600 iter->next = next;
abc490a1 1601}
e0ba718a 1602
0422d92c 1603void cds_lfht_next_duplicate(struct cds_lfht *ht, cds_lfht_match_fct match,
996ff57c 1604 const void *key, struct cds_lfht_iter *iter)
a481e5ff 1605{
adc0de68 1606 struct cds_lfht_node *node, *next;
a481e5ff 1607 unsigned long reverse_hash;
a481e5ff 1608
adc0de68 1609 node = iter->node;
04db56f8 1610 reverse_hash = node->reverse_hash;
adc0de68 1611 next = iter->next;
a481e5ff
MD
1612 node = clear_flag(next);
1613
1614 for (;;) {
8ed51e04 1615 if (caa_unlikely(is_end(node))) {
96ad1112 1616 node = next = NULL;
a481e5ff 1617 break;
bb7b2f26 1618 }
04db56f8 1619 if (caa_unlikely(node->reverse_hash > reverse_hash)) {
96ad1112 1620 node = next = NULL;
a481e5ff
MD
1621 break;
1622 }
04db56f8 1623 next = rcu_dereference(node->next);
8ed51e04 1624 if (caa_likely(!is_removed(next))
1ee8f000 1625 && !is_bucket(next)
04db56f8 1626 && caa_likely(match(node, key))) {
a481e5ff
MD
1627 break;
1628 }
1629 node = clear_flag(next);
1630 }
a85eff52 1631 assert(!node || !is_bucket(CMM_LOAD_SHARED(node->next)));
adc0de68
MD
1632 iter->node = node;
1633 iter->next = next;
a481e5ff
MD
1634}
1635
4e9b9fbf
MD
1636void cds_lfht_next(struct cds_lfht *ht, struct cds_lfht_iter *iter)
1637{
1638 struct cds_lfht_node *node, *next;
1639
853395e1 1640 node = clear_flag(iter->next);
4e9b9fbf 1641 for (;;) {
8ed51e04 1642 if (caa_unlikely(is_end(node))) {
4e9b9fbf
MD
1643 node = next = NULL;
1644 break;
1645 }
04db56f8 1646 next = rcu_dereference(node->next);
8ed51e04 1647 if (caa_likely(!is_removed(next))
1ee8f000 1648 && !is_bucket(next)) {
4e9b9fbf
MD
1649 break;
1650 }
1651 node = clear_flag(next);
1652 }
a85eff52 1653 assert(!node || !is_bucket(CMM_LOAD_SHARED(node->next)));
4e9b9fbf
MD
1654 iter->node = node;
1655 iter->next = next;
1656}
1657
1658void cds_lfht_first(struct cds_lfht *ht, struct cds_lfht_iter *iter)
1659{
4e9b9fbf 1660 /*
1ee8f000 1661 * Get next after first bucket node. The first bucket node is the
4e9b9fbf
MD
1662 * first node of the linked list.
1663 */
9d72a73f 1664 iter->next = bucket_at(ht, 0)->next;
4e9b9fbf
MD
1665 cds_lfht_next(ht, iter);
1666}
1667
0422d92c
MD
1668void cds_lfht_add(struct cds_lfht *ht, unsigned long hash,
1669 struct cds_lfht_node *node)
abc490a1 1670{
0422d92c 1671 unsigned long size;
ab7d5fc6 1672
709bacf9 1673 node->reverse_hash = bit_reverse_ulong(hash);
7b3893e4 1674 size = rcu_dereference(ht->size);
91a75cc5 1675 _cds_lfht_add(ht, hash, NULL, NULL, size, node, NULL, 0);
14360f1c 1676 ht_count_add(ht, size, hash);
3eca1b8c
MD
1677}
1678
14044b37 1679struct cds_lfht_node *cds_lfht_add_unique(struct cds_lfht *ht,
6f554439 1680 unsigned long hash,
0422d92c 1681 cds_lfht_match_fct match,
996ff57c 1682 const void *key,
48ed1c18 1683 struct cds_lfht_node *node)
3eca1b8c 1684{
0422d92c 1685 unsigned long size;
83beee94 1686 struct cds_lfht_iter iter;
3eca1b8c 1687
709bacf9 1688 node->reverse_hash = bit_reverse_ulong(hash);
7b3893e4 1689 size = rcu_dereference(ht->size);
91a75cc5 1690 _cds_lfht_add(ht, hash, match, key, size, node, &iter, 0);
83beee94 1691 if (iter.node == node)
14360f1c 1692 ht_count_add(ht, size, hash);
83beee94 1693 return iter.node;
2ed95849
MD
1694}
1695
9357c415 1696struct cds_lfht_node *cds_lfht_add_replace(struct cds_lfht *ht,
6f554439 1697 unsigned long hash,
0422d92c 1698 cds_lfht_match_fct match,
996ff57c 1699 const void *key,
48ed1c18
MD
1700 struct cds_lfht_node *node)
1701{
0422d92c 1702 unsigned long size;
83beee94 1703 struct cds_lfht_iter iter;
48ed1c18 1704
709bacf9 1705 node->reverse_hash = bit_reverse_ulong(hash);
7b3893e4 1706 size = rcu_dereference(ht->size);
83beee94 1707 for (;;) {
91a75cc5 1708 _cds_lfht_add(ht, hash, match, key, size, node, &iter, 0);
83beee94 1709 if (iter.node == node) {
14360f1c 1710 ht_count_add(ht, size, hash);
83beee94
MD
1711 return NULL;
1712 }
1713
1714 if (!_cds_lfht_replace(ht, size, iter.node, iter.next, node))
1715 return iter.node;
1716 }
48ed1c18
MD
1717}
1718
2e79c445
MD
1719int cds_lfht_replace(struct cds_lfht *ht,
1720 struct cds_lfht_iter *old_iter,
1721 unsigned long hash,
1722 cds_lfht_match_fct match,
1723 const void *key,
9357c415
MD
1724 struct cds_lfht_node *new_node)
1725{
1726 unsigned long size;
1727
709bacf9 1728 new_node->reverse_hash = bit_reverse_ulong(hash);
2e79c445
MD
1729 if (!old_iter->node)
1730 return -ENOENT;
1731 if (caa_unlikely(old_iter->node->reverse_hash != new_node->reverse_hash))
1732 return -EINVAL;
1733 if (caa_unlikely(!match(old_iter->node, key)))
1734 return -EINVAL;
7b3893e4 1735 size = rcu_dereference(ht->size);
9357c415
MD
1736 return _cds_lfht_replace(ht, size, old_iter->node, old_iter->next,
1737 new_node);
1738}
1739
bc8c3c74 1740int cds_lfht_del(struct cds_lfht *ht, struct cds_lfht_node *node)
2ed95849 1741{
95bc7fb9 1742 unsigned long size;
df44348d 1743 int ret;
abc490a1 1744
7b3893e4 1745 size = rcu_dereference(ht->size);
bc8c3c74 1746 ret = _cds_lfht_del(ht, size, node);
14360f1c 1747 if (!ret) {
95bc7fb9
MD
1748 unsigned long hash;
1749
bc8c3c74 1750 hash = bit_reverse_ulong(node->reverse_hash);
14360f1c
LJ
1751 ht_count_del(ht, size, hash);
1752 }
df44348d 1753 return ret;
2ed95849 1754}
ab7d5fc6 1755
df55172a
MD
1756int cds_lfht_is_node_deleted(struct cds_lfht_node *node)
1757{
a85eff52 1758 return is_removed(CMM_LOAD_SHARED(node->next));
df55172a
MD
1759}
1760
abc490a1 1761static
1ee8f000 1762int cds_lfht_delete_bucket(struct cds_lfht *ht)
674f7a69 1763{
14044b37 1764 struct cds_lfht_node *node;
4105056a 1765 unsigned long order, i, size;
674f7a69 1766
abc490a1 1767 /* Check that the table is empty */
9d72a73f 1768 node = bucket_at(ht, 0);
abc490a1 1769 do {
04db56f8 1770 node = clear_flag(node)->next;
1ee8f000 1771 if (!is_bucket(node))
abc490a1 1772 return -EPERM;
273399de 1773 assert(!is_removed(node));
2f943cd7 1774 assert(!is_removal_owner(node));
bb7b2f26 1775 } while (!is_end(node));
4105056a
MD
1776 /*
1777 * size accessed without rcu_dereference because hash table is
1778 * being destroyed.
1779 */
7b3893e4 1780 size = ht->size;
1f67ba50 1781 /* Internal sanity check: all nodes left should be buckets */
48f1b16d
LJ
1782 for (i = 0; i < size; i++) {
1783 node = bucket_at(ht, i);
1784 dbg_printf("delete bucket: index %lu expected hash %lu hash %lu\n",
1785 i, i, bit_reverse_ulong(node->reverse_hash));
1786 assert(is_bucket(node->next));
1787 }
24365af7 1788
5bc6b66f 1789 for (order = cds_lfht_get_count_order_ulong(size); (long)order >= 0; order--)
48f1b16d 1790 cds_lfht_free_bucket_table(ht, order);
5488222b 1791
abc490a1 1792 return 0;
674f7a69
MD
1793}
1794
1795/*
1796 * Should only be called when no more concurrent readers nor writers can
1797 * possibly access the table.
1798 */
b7d619b0 1799int cds_lfht_destroy(struct cds_lfht *ht, pthread_attr_t **attr)
674f7a69 1800{
d0ec0ed2
MD
1801 int ret;
1802
1803 if (ht->flags & CDS_LFHT_AUTO_RESIZE) {
1804 /* Cancel ongoing resize operations. */
1805 _CMM_STORE_SHARED(ht->in_progress_destroy, 1);
1806 /* Wait for in-flight resize operations to complete */
1807 urcu_workqueue_flush_queued_work(cds_lfht_workqueue);
10e68472 1808 }
1ee8f000 1809 ret = cds_lfht_delete_bucket(ht);
abc490a1
MD
1810 if (ret)
1811 return ret;
5afadd12 1812 free_split_items_count(ht);
b7d619b0
MD
1813 if (attr)
1814 *attr = ht->resize_attr;
59629f09
MD
1815 ret = pthread_mutex_destroy(&ht->resize_mutex);
1816 if (ret)
1817 ret = -EBUSY;
d0ec0ed2
MD
1818 if (ht->flags & CDS_LFHT_AUTO_RESIZE)
1819 cds_lfht_fini_worker(ht->flavor);
98808fb1 1820 poison_free(ht);
5e28c532 1821 return ret;
674f7a69
MD
1822}
1823
14044b37 1824void cds_lfht_count_nodes(struct cds_lfht *ht,
d933dd0e 1825 long *approx_before,
273399de 1826 unsigned long *count,
d933dd0e 1827 long *approx_after)
273399de 1828{
14044b37 1829 struct cds_lfht_node *node, *next;
caf3653d 1830 unsigned long nr_bucket = 0, nr_removed = 0;
273399de 1831
7ed7682f 1832 *approx_before = 0;
5afadd12 1833 if (ht->split_count) {
973e5e1b
MD
1834 int i;
1835
4c42f1b8
LJ
1836 for (i = 0; i < split_count_mask + 1; i++) {
1837 *approx_before += uatomic_read(&ht->split_count[i].add);
1838 *approx_before -= uatomic_read(&ht->split_count[i].del);
973e5e1b
MD
1839 }
1840 }
1841
273399de 1842 *count = 0;
273399de 1843
1ee8f000 1844 /* Count non-bucket nodes in the table */
9d72a73f 1845 node = bucket_at(ht, 0);
273399de 1846 do {
04db56f8 1847 next = rcu_dereference(node->next);
b198f0fd 1848 if (is_removed(next)) {
1ee8f000 1849 if (!is_bucket(next))
caf3653d 1850 (nr_removed)++;
973e5e1b 1851 else
1ee8f000
LJ
1852 (nr_bucket)++;
1853 } else if (!is_bucket(next))
273399de 1854 (*count)++;
24365af7 1855 else
1ee8f000 1856 (nr_bucket)++;
273399de 1857 node = clear_flag(next);
bb7b2f26 1858 } while (!is_end(node));
caf3653d 1859 dbg_printf("number of logically removed nodes: %lu\n", nr_removed);
1ee8f000 1860 dbg_printf("number of bucket nodes: %lu\n", nr_bucket);
7ed7682f 1861 *approx_after = 0;
5afadd12 1862 if (ht->split_count) {
973e5e1b
MD
1863 int i;
1864
4c42f1b8
LJ
1865 for (i = 0; i < split_count_mask + 1; i++) {
1866 *approx_after += uatomic_read(&ht->split_count[i].add);
1867 *approx_after -= uatomic_read(&ht->split_count[i].del);
973e5e1b
MD
1868 }
1869 }
273399de
MD
1870}
1871
1475579c 1872/* called with resize mutex held */
abc490a1 1873static
4105056a 1874void _do_cds_lfht_grow(struct cds_lfht *ht,
1475579c 1875 unsigned long old_size, unsigned long new_size)
abc490a1 1876{
1475579c 1877 unsigned long old_order, new_order;
1475579c 1878
5bc6b66f
MD
1879 old_order = cds_lfht_get_count_order_ulong(old_size);
1880 new_order = cds_lfht_get_count_order_ulong(new_size);
1a401918
LJ
1881 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1882 old_size, old_order, new_size, new_order);
1475579c 1883 assert(new_size > old_size);
93d46c39 1884 init_table(ht, old_order + 1, new_order);
abc490a1
MD
1885}
1886
1887/* called with resize mutex held */
1888static
4105056a 1889void _do_cds_lfht_shrink(struct cds_lfht *ht,
1475579c 1890 unsigned long old_size, unsigned long new_size)
464a1ec9 1891{
1475579c 1892 unsigned long old_order, new_order;
464a1ec9 1893
d0d8f9aa 1894 new_size = max(new_size, MIN_TABLE_SIZE);
5bc6b66f
MD
1895 old_order = cds_lfht_get_count_order_ulong(old_size);
1896 new_order = cds_lfht_get_count_order_ulong(new_size);
1a401918
LJ
1897 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1898 old_size, old_order, new_size, new_order);
1475579c 1899 assert(new_size < old_size);
1475579c 1900
1ee8f000 1901 /* Remove and unlink all bucket nodes to remove. */
93d46c39 1902 fini_table(ht, new_order + 1, old_order);
464a1ec9
MD
1903}
1904
1475579c
MD
1905
1906/* called with resize mutex held */
1907static
1908void _do_cds_lfht_resize(struct cds_lfht *ht)
1909{
1910 unsigned long new_size, old_size;
4105056a
MD
1911
1912 /*
1913 * Resize table, re-do if the target size has changed under us.
1914 */
1915 do {
d2be3620
MD
1916 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1917 break;
7b3893e4
LJ
1918 ht->resize_initiated = 1;
1919 old_size = ht->size;
1920 new_size = CMM_LOAD_SHARED(ht->resize_target);
4105056a
MD
1921 if (old_size < new_size)
1922 _do_cds_lfht_grow(ht, old_size, new_size);
1923 else if (old_size > new_size)
1924 _do_cds_lfht_shrink(ht, old_size, new_size);
7b3893e4 1925 ht->resize_initiated = 0;
4105056a
MD
1926 /* write resize_initiated before read resize_target */
1927 cmm_smp_mb();
7b3893e4 1928 } while (ht->size != CMM_LOAD_SHARED(ht->resize_target));
1475579c
MD
1929}
1930
abc490a1 1931static
ab65b890 1932unsigned long resize_target_grow(struct cds_lfht *ht, unsigned long new_size)
464a1ec9 1933{
7b3893e4 1934 return _uatomic_xchg_monotonic_increase(&ht->resize_target, new_size);
464a1ec9
MD
1935}
1936
1475579c 1937static
4105056a 1938void resize_target_update_count(struct cds_lfht *ht,
b8af5011 1939 unsigned long count)
1475579c 1940{
d0d8f9aa 1941 count = max(count, MIN_TABLE_SIZE);
747d725c 1942 count = min(count, ht->max_nr_buckets);
7b3893e4 1943 uatomic_set(&ht->resize_target, count);
1475579c
MD
1944}
1945
1946void cds_lfht_resize(struct cds_lfht *ht, unsigned long new_size)
464a1ec9 1947{
10e68472
MD
1948 resize_target_update_count(ht, new_size);
1949 CMM_STORE_SHARED(ht->resize_initiated, 1);
1475579c
MD
1950 pthread_mutex_lock(&ht->resize_mutex);
1951 _do_cds_lfht_resize(ht);
1952 pthread_mutex_unlock(&ht->resize_mutex);
abc490a1 1953}
464a1ec9 1954
abc490a1 1955static
d0ec0ed2 1956void do_resize_cb(struct urcu_work *work)
abc490a1 1957{
d0ec0ed2
MD
1958 struct resize_work *resize_work =
1959 caa_container_of(work, struct resize_work, work);
1960 struct cds_lfht *ht = resize_work->ht;
abc490a1 1961
d0ec0ed2 1962 ht->flavor->register_thread();
abc490a1 1963 pthread_mutex_lock(&ht->resize_mutex);
14044b37 1964 _do_cds_lfht_resize(ht);
abc490a1 1965 pthread_mutex_unlock(&ht->resize_mutex);
d0ec0ed2 1966 ht->flavor->unregister_thread();
98808fb1 1967 poison_free(work);
464a1ec9
MD
1968}
1969
abc490a1 1970static
f1f119ee 1971void __cds_lfht_resize_lazy_launch(struct cds_lfht *ht)
ab7d5fc6 1972{
d0ec0ed2 1973 struct resize_work *work;
abc490a1 1974
4105056a
MD
1975 /* Store resize_target before read resize_initiated */
1976 cmm_smp_mb();
7b3893e4 1977 if (!CMM_LOAD_SHARED(ht->resize_initiated)) {
ed35e6d8 1978 if (CMM_LOAD_SHARED(ht->in_progress_destroy)) {
59290e9d 1979 return;
ed35e6d8 1980 }
f9830efd 1981 work = malloc(sizeof(*work));
741f378e
MD
1982 if (work == NULL) {
1983 dbg_printf("error allocating resize work, bailing out\n");
741f378e
MD
1984 return;
1985 }
f9830efd 1986 work->ht = ht;
d0ec0ed2
MD
1987 urcu_workqueue_queue_work(cds_lfht_workqueue,
1988 &work->work, do_resize_cb);
7b3893e4 1989 CMM_STORE_SHARED(ht->resize_initiated, 1);
f9830efd 1990 }
ab7d5fc6 1991}
3171717f 1992
f1f119ee
LJ
1993static
1994void cds_lfht_resize_lazy_grow(struct cds_lfht *ht, unsigned long size, int growth)
1995{
1996 unsigned long target_size = size << growth;
1997
747d725c 1998 target_size = min(target_size, ht->max_nr_buckets);
f1f119ee
LJ
1999 if (resize_target_grow(ht, target_size) >= target_size)
2000 return;
2001
2002 __cds_lfht_resize_lazy_launch(ht);
2003}
2004
89bb121d
LJ
2005/*
2006 * We favor grow operations over shrink. A shrink operation never occurs
2007 * if a grow operation is queued for lazy execution. A grow operation
2008 * cancels any pending shrink lazy execution.
2009 */
3171717f 2010static
4105056a 2011void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
3171717f
MD
2012 unsigned long count)
2013{
b8af5011
MD
2014 if (!(ht->flags & CDS_LFHT_AUTO_RESIZE))
2015 return;
d0d8f9aa 2016 count = max(count, MIN_TABLE_SIZE);
747d725c 2017 count = min(count, ht->max_nr_buckets);
89bb121d
LJ
2018 if (count == size)
2019 return; /* Already the right size, no resize needed */
2020 if (count > size) { /* lazy grow */
2021 if (resize_target_grow(ht, count) >= count)
2022 return;
2023 } else { /* lazy shrink */
2024 for (;;) {
2025 unsigned long s;
2026
7b3893e4 2027 s = uatomic_cmpxchg(&ht->resize_target, size, count);
89bb121d
LJ
2028 if (s == size)
2029 break; /* no resize needed */
2030 if (s > size)
2031 return; /* growing is/(was just) in progress */
2032 if (s <= count)
2033 return; /* some other thread do shrink */
2034 size = s;
2035 }
2036 }
f1f119ee 2037 __cds_lfht_resize_lazy_launch(ht);
3171717f 2038}
d0ec0ed2
MD
2039
2040static void mutex_lock(pthread_mutex_t *mutex)
2041{
2042 int ret;
2043
2044#ifndef DISTRUST_SIGNALS_EXTREME
2045 ret = pthread_mutex_lock(mutex);
2046 if (ret)
2047 urcu_die(ret);
2048#else /* #ifndef DISTRUST_SIGNALS_EXTREME */
2049 while ((ret = pthread_mutex_trylock(mutex)) != 0) {
2050 if (ret != EBUSY && ret != EINTR)
2051 urcu_die(ret);
2052 if (CMM_LOAD_SHARED(URCU_TLS(rcu_reader).need_mb)) {
2053 cmm_smp_mb();
2054 _CMM_STORE_SHARED(URCU_TLS(rcu_reader).need_mb, 0);
2055 cmm_smp_mb();
2056 }
2057 (void) poll(NULL, 0, 10);
2058 }
2059#endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
2060}
2061
2062static void mutex_unlock(pthread_mutex_t *mutex)
2063{
2064 int ret;
2065
2066 ret = pthread_mutex_unlock(mutex);
2067 if (ret)
2068 urcu_die(ret);
2069}
2070
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 0.268748 seconds and 4 git commands to generate.