Fix: futex.h: include headers outside extern C
[urcu.git] / src / rculfhash.c
... / ...
CommitLineData
1/*
2 * rculfhash.c
3 *
4 * Userspace RCU library - Lock-Free Resizable RCU Hash Table
5 *
6 * Copyright 2010-2011 - Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
7 * Copyright 2011 - Lai Jiangshan <laijs@cn.fujitsu.com>
8 *
9 * This library is free software; you can redistribute it and/or
10 * modify it under the terms of the GNU Lesser General Public
11 * License as published by the Free Software Foundation; either
12 * version 2.1 of the License, or (at your option) any later version.
13 *
14 * This library is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
17 * Lesser General Public License for more details.
18 *
19 * You should have received a copy of the GNU Lesser General Public
20 * License along with this library; if not, write to the Free Software
21 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
22 */
23
24/*
25 * Based on the following articles:
26 * - Ori Shalev and Nir Shavit. Split-ordered lists: Lock-free
27 * extensible hash tables. J. ACM 53, 3 (May 2006), 379-405.
28 * - Michael, M. M. High performance dynamic lock-free hash tables
29 * and list-based sets. In Proceedings of the fourteenth annual ACM
30 * symposium on Parallel algorithms and architectures, ACM Press,
31 * (2002), 73-82.
32 *
33 * Some specificities of this Lock-Free Resizable RCU Hash Table
34 * implementation:
35 *
36 * - RCU read-side critical section allows readers to perform hash
37 * table lookups, as well as traversals, and use the returned objects
38 * safely by allowing memory reclaim to take place only after a grace
39 * period.
40 * - Add and remove operations are lock-free, and do not need to
41 * allocate memory. They need to be executed within RCU read-side
42 * critical section to ensure the objects they read are valid and to
43 * deal with the cmpxchg ABA problem.
44 * - add and add_unique operations are supported. add_unique checks if
45 * the node key already exists in the hash table. It ensures not to
46 * populate a duplicate key if the node key already exists in the hash
47 * table.
48 * - The resize operation executes concurrently with
49 * add/add_unique/add_replace/remove/lookup/traversal.
50 * - Hash table nodes are contained within a split-ordered list. This
51 * list is ordered by incrementing reversed-bits-hash value.
52 * - An index of bucket nodes is kept. These bucket nodes are the hash
53 * table "buckets". These buckets are internal nodes that allow to
54 * perform a fast hash lookup, similarly to a skip list. These
55 * buckets are chained together in the split-ordered list, which
56 * allows recursive expansion by inserting new buckets between the
57 * existing buckets. The split-ordered list allows adding new buckets
58 * between existing buckets as the table needs to grow.
59 * - The resize operation for small tables only allows expanding the
60 * hash table. It is triggered automatically by detecting long chains
61 * in the add operation.
62 * - The resize operation for larger tables (and available through an
63 * API) allows both expanding and shrinking the hash table.
64 * - Split-counters are used to keep track of the number of
65 * nodes within the hash table for automatic resize triggering.
66 * - Resize operation initiated by long chain detection is executed by a
67 * worker thread, which keeps lock-freedom of add and remove.
68 * - Resize operations are protected by a mutex.
69 * - The removal operation is split in two parts: first, a "removed"
70 * flag is set in the next pointer within the node to remove. Then,
71 * a "garbage collection" is performed in the bucket containing the
72 * removed node (from the start of the bucket up to the removed node).
73 * All encountered nodes with "removed" flag set in their next
74 * pointers are removed from the linked-list. If the cmpxchg used for
75 * removal fails (due to concurrent garbage-collection or concurrent
76 * add), we retry from the beginning of the bucket. This ensures that
77 * the node with "removed" flag set is removed from the hash table
78 * (not visible to lookups anymore) before the RCU read-side critical
79 * section held across removal ends. Furthermore, this ensures that
80 * the node with "removed" flag set is removed from the linked-list
81 * before its memory is reclaimed. After setting the "removal" flag,
82 * only the thread which removal is the first to set the "removal
83 * owner" flag (with an xchg) into a node's next pointer is considered
84 * to have succeeded its removal (and thus owns the node to reclaim).
85 * Because we garbage-collect starting from an invariant node (the
86 * start-of-bucket bucket node) up to the "removed" node (or find a
87 * reverse-hash that is higher), we are sure that a successful
88 * traversal of the chain leads to a chain that is present in the
89 * linked-list (the start node is never removed) and that it does not
90 * contain the "removed" node anymore, even if concurrent delete/add
91 * operations are changing the structure of the list concurrently.
92 * - The add operations perform garbage collection of buckets if they
93 * encounter nodes with removed flag set in the bucket where they want
94 * to add their new node. This ensures lock-freedom of add operation by
95 * helping the remover unlink nodes from the list rather than to wait
96 * for it do to so.
97 * - There are three memory backends for the hash table buckets: the
98 * "order table", the "chunks", and the "mmap".
99 * - These bucket containers contain a compact version of the hash table
100 * nodes.
101 * - The RCU "order table":
102 * - has a first level table indexed by log2(hash index) which is
103 * copied and expanded by the resize operation. This order table
104 * allows finding the "bucket node" tables.
105 * - There is one bucket node table per hash index order. The size of
106 * each bucket node table is half the number of hashes contained in
107 * this order (except for order 0).
108 * - The RCU "chunks" is best suited for close interaction with a page
109 * allocator. It uses a linear array as index to "chunks" containing
110 * each the same number of buckets.
111 * - The RCU "mmap" memory backend uses a single memory map to hold
112 * all buckets.
113 * - synchronize_rcu is used to garbage-collect the old bucket node table.
114 *
115 * Ordering Guarantees:
116 *
117 * To discuss these guarantees, we first define "read" operation as any
118 * of the the basic cds_lfht_lookup, cds_lfht_next_duplicate,
119 * cds_lfht_first, cds_lfht_next operation, as well as
120 * cds_lfht_add_unique (failure).
121 *
122 * We define "read traversal" operation as any of the following
123 * group of operations
124 * - cds_lfht_lookup followed by iteration with cds_lfht_next_duplicate
125 * (and/or cds_lfht_next, although less common).
126 * - cds_lfht_add_unique (failure) followed by iteration with
127 * cds_lfht_next_duplicate (and/or cds_lfht_next, although less
128 * common).
129 * - cds_lfht_first followed iteration with cds_lfht_next (and/or
130 * cds_lfht_next_duplicate, although less common).
131 *
132 * We define "write" operations as any of cds_lfht_add, cds_lfht_replace,
133 * cds_lfht_add_unique (success), cds_lfht_add_replace, cds_lfht_del.
134 *
135 * When cds_lfht_add_unique succeeds (returns the node passed as
136 * parameter), it acts as a "write" operation. When cds_lfht_add_unique
137 * fails (returns a node different from the one passed as parameter), it
138 * acts as a "read" operation. A cds_lfht_add_unique failure is a
139 * cds_lfht_lookup "read" operation, therefore, any ordering guarantee
140 * referring to "lookup" imply any of "lookup" or cds_lfht_add_unique
141 * (failure).
142 *
143 * We define "prior" and "later" node as nodes observable by reads and
144 * read traversals respectively before and after a write or sequence of
145 * write operations.
146 *
147 * Hash-table operations are often cascaded, for example, the pointer
148 * returned by a cds_lfht_lookup() might be passed to a cds_lfht_next(),
149 * whose return value might in turn be passed to another hash-table
150 * operation. This entire cascaded series of operations must be enclosed
151 * by a pair of matching rcu_read_lock() and rcu_read_unlock()
152 * operations.
153 *
154 * The following ordering guarantees are offered by this hash table:
155 *
156 * A.1) "read" after "write": if there is ordering between a write and a
157 * later read, then the read is guaranteed to see the write or some
158 * later write.
159 * A.2) "read traversal" after "write": given that there is dependency
160 * ordering between reads in a "read traversal", if there is
161 * ordering between a write and the first read of the traversal,
162 * then the "read traversal" is guaranteed to see the write or
163 * some later write.
164 * B.1) "write" after "read": if there is ordering between a read and a
165 * later write, then the read will never see the write.
166 * B.2) "write" after "read traversal": given that there is dependency
167 * ordering between reads in a "read traversal", if there is
168 * ordering between the last read of the traversal and a later
169 * write, then the "read traversal" will never see the write.
170 * C) "write" while "read traversal": if a write occurs during a "read
171 * traversal", the traversal may, or may not, see the write.
172 * D.1) "write" after "write": if there is ordering between a write and
173 * a later write, then the later write is guaranteed to see the
174 * effects of the first write.
175 * D.2) Concurrent "write" pairs: The system will assign an arbitrary
176 * order to any pair of concurrent conflicting writes.
177 * Non-conflicting writes (for example, to different keys) are
178 * unordered.
179 * E) If a grace period separates a "del" or "replace" operation
180 * and a subsequent operation, then that subsequent operation is
181 * guaranteed not to see the removed item.
182 * F) Uniqueness guarantee: given a hash table that does not contain
183 * duplicate items for a given key, there will only be one item in
184 * the hash table after an arbitrary sequence of add_unique and/or
185 * add_replace operations. Note, however, that a pair of
186 * concurrent read operations might well access two different items
187 * with that key.
188 * G.1) If a pair of lookups for a given key are ordered (e.g. by a
189 * memory barrier), then the second lookup will return the same
190 * node as the previous lookup, or some later node.
191 * G.2) A "read traversal" that starts after the end of a prior "read
192 * traversal" (ordered by memory barriers) is guaranteed to see the
193 * same nodes as the previous traversal, or some later nodes.
194 * G.3) Concurrent "read" pairs: concurrent reads are unordered. For
195 * example, if a pair of reads to the same key run concurrently
196 * with an insertion of that same key, the reads remain unordered
197 * regardless of their return values. In other words, you cannot
198 * rely on the values returned by the reads to deduce ordering.
199 *
200 * Progress guarantees:
201 *
202 * * Reads are wait-free. These operations always move forward in the
203 * hash table linked list, and this list has no loop.
204 * * Writes are lock-free. Any retry loop performed by a write operation
205 * is triggered by progress made within another update operation.
206 *
207 * Bucket node tables:
208 *
209 * hash table hash table the last all bucket node tables
210 * order size bucket node 0 1 2 3 4 5 6(index)
211 * table size
212 * 0 1 1 1
213 * 1 2 1 1 1
214 * 2 4 2 1 1 2
215 * 3 8 4 1 1 2 4
216 * 4 16 8 1 1 2 4 8
217 * 5 32 16 1 1 2 4 8 16
218 * 6 64 32 1 1 2 4 8 16 32
219 *
220 * When growing/shrinking, we only focus on the last bucket node table
221 * which size is (!order ? 1 : (1 << (order -1))).
222 *
223 * Example for growing/shrinking:
224 * grow hash table from order 5 to 6: init the index=6 bucket node table
225 * shrink hash table from order 6 to 5: fini the index=6 bucket node table
226 *
227 * A bit of ascii art explanation:
228 *
229 * The order index is the off-by-one compared to the actual power of 2
230 * because we use index 0 to deal with the 0 special-case.
231 *
232 * This shows the nodes for a small table ordered by reversed bits:
233 *
234 * bits reverse
235 * 0 000 000
236 * 4 100 001
237 * 2 010 010
238 * 6 110 011
239 * 1 001 100
240 * 5 101 101
241 * 3 011 110
242 * 7 111 111
243 *
244 * This shows the nodes in order of non-reversed bits, linked by
245 * reversed-bit order.
246 *
247 * order bits reverse
248 * 0 0 000 000
249 * 1 | 1 001 100 <-
250 * 2 | | 2 010 010 <- |
251 * | | | 3 011 110 | <- |
252 * 3 -> | | | 4 100 001 | |
253 * -> | | 5 101 101 |
254 * -> | 6 110 011
255 * -> 7 111 111
256 */
257
258#define _LGPL_SOURCE
259#include <stdlib.h>
260#include <errno.h>
261#include <stdio.h>
262#include <stdint.h>
263#include <string.h>
264#include <sched.h>
265#include <unistd.h>
266
267#include "compat-getcpu.h"
268#include <urcu/assert.h>
269#include <urcu/pointer.h>
270#include <urcu/call-rcu.h>
271#include <urcu/flavor.h>
272#include <urcu/arch.h>
273#include <urcu/uatomic.h>
274#include <urcu/compiler.h>
275#include <urcu/rculfhash.h>
276#include <urcu/static/urcu-signal-nr.h>
277#include <rculfhash-internal.h>
278#include <stdio.h>
279#include <pthread.h>
280#include <signal.h>
281#include "workqueue.h"
282#include "urcu-die.h"
283#include "urcu-utils.h"
284#include "compat-smp.h"
285
286/*
287 * Split-counters lazily update the global counter each 1024
288 * addition/removal. It automatically keeps track of resize required.
289 * We use the bucket length as indicator for need to expand for small
290 * tables and machines lacking per-cpu data support.
291 */
292#define COUNT_COMMIT_ORDER 10
293#define DEFAULT_SPLIT_COUNT_MASK 0xFUL
294#define CHAIN_LEN_TARGET 1
295#define CHAIN_LEN_RESIZE_THRESHOLD 3
296
297/*
298 * Define the minimum table size.
299 */
300#define MIN_TABLE_ORDER 0
301#define MIN_TABLE_SIZE (1UL << MIN_TABLE_ORDER)
302
303/*
304 * Minimum number of bucket nodes to touch per thread to parallelize grow/shrink.
305 */
306#define MIN_PARTITION_PER_THREAD_ORDER 12
307#define MIN_PARTITION_PER_THREAD (1UL << MIN_PARTITION_PER_THREAD_ORDER)
308
309/*
310 * The removed flag needs to be updated atomically with the pointer.
311 * It indicates that no node must attach to the node scheduled for
312 * removal, and that node garbage collection must be performed.
313 * The bucket flag does not require to be updated atomically with the
314 * pointer, but it is added as a pointer low bit flag to save space.
315 * The "removal owner" flag is used to detect which of the "del"
316 * operation that has set the "removed flag" gets to return the removed
317 * node to its caller. Note that the replace operation does not need to
318 * iteract with the "removal owner" flag, because it validates that
319 * the "removed" flag is not set before performing its cmpxchg.
320 */
321#define REMOVED_FLAG (1UL << 0)
322#define BUCKET_FLAG (1UL << 1)
323#define REMOVAL_OWNER_FLAG (1UL << 2)
324#define FLAGS_MASK ((1UL << 3) - 1)
325
326/* Value of the end pointer. Should not interact with flags. */
327#define END_VALUE NULL
328
329/*
330 * ht_items_count: Split-counters counting the number of node addition
331 * and removal in the table. Only used if the CDS_LFHT_ACCOUNTING flag
332 * is set at hash table creation.
333 *
334 * These are free-running counters, never reset to zero. They count the
335 * number of add/remove, and trigger every (1 << COUNT_COMMIT_ORDER)
336 * operations to update the global counter. We choose a power-of-2 value
337 * for the trigger to deal with 32 or 64-bit overflow of the counter.
338 */
339struct ht_items_count {
340 unsigned long add, del;
341} __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
342
343/*
344 * resize_work: Contains arguments passed to worker thread
345 * responsible for performing lazy resize.
346 */
347struct resize_work {
348 struct urcu_work work;
349 struct cds_lfht *ht;
350};
351
352/*
353 * partition_resize_work: Contains arguments passed to worker threads
354 * executing the hash table resize on partitions of the hash table
355 * assigned to each processor's worker thread.
356 */
357struct partition_resize_work {
358 pthread_t thread_id;
359 struct cds_lfht *ht;
360 unsigned long i, start, len;
361 void (*fct)(struct cds_lfht *ht, unsigned long i,
362 unsigned long start, unsigned long len);
363};
364
365static struct urcu_workqueue *cds_lfht_workqueue;
366static unsigned long cds_lfht_workqueue_user_count;
367
368/*
369 * Mutex ensuring mutual exclusion between workqueue initialization and
370 * fork handlers. cds_lfht_fork_mutex nests inside call_rcu_mutex.
371 */
372static pthread_mutex_t cds_lfht_fork_mutex = PTHREAD_MUTEX_INITIALIZER;
373
374static struct urcu_atfork cds_lfht_atfork;
375
376/*
377 * atfork handler nesting counters. Handle being registered to many urcu
378 * flavors, thus being possibly invoked more than once in the
379 * pthread_atfork list of callbacks.
380 */
381static int cds_lfht_workqueue_atfork_nesting;
382
383static void cds_lfht_init_worker(const struct rcu_flavor_struct *flavor);
384static void cds_lfht_fini_worker(const struct rcu_flavor_struct *flavor);
385
386#ifdef CONFIG_CDS_LFHT_ITER_DEBUG
387
388static
389void cds_lfht_iter_debug_set_ht(struct cds_lfht *ht, struct cds_lfht_iter *iter)
390{
391 iter->lfht = ht;
392}
393
394#define cds_lfht_iter_debug_assert(...) urcu_posix_assert(__VA_ARGS__)
395
396#else
397
398static
399void cds_lfht_iter_debug_set_ht(struct cds_lfht *ht __attribute__((unused)),
400 struct cds_lfht_iter *iter __attribute__((unused)))
401{
402}
403
404#define cds_lfht_iter_debug_assert(...)
405
406#endif
407
408/*
409 * Algorithm to reverse bits in a word by lookup table, extended to
410 * 64-bit words.
411 * Source:
412 * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable
413 * Originally from Public Domain.
414 */
415
416static const uint8_t BitReverseTable256[256] =
417{
418#define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64
419#define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16)
420#define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 )
421 R6(0), R6(2), R6(1), R6(3)
422};
423#undef R2
424#undef R4
425#undef R6
426
427static
428uint8_t bit_reverse_u8(uint8_t v)
429{
430 return BitReverseTable256[v];
431}
432
433#if (CAA_BITS_PER_LONG == 32)
434static
435uint32_t bit_reverse_u32(uint32_t v)
436{
437 return ((uint32_t) bit_reverse_u8(v) << 24) |
438 ((uint32_t) bit_reverse_u8(v >> 8) << 16) |
439 ((uint32_t) bit_reverse_u8(v >> 16) << 8) |
440 ((uint32_t) bit_reverse_u8(v >> 24));
441}
442#else
443static
444uint64_t bit_reverse_u64(uint64_t v)
445{
446 return ((uint64_t) bit_reverse_u8(v) << 56) |
447 ((uint64_t) bit_reverse_u8(v >> 8) << 48) |
448 ((uint64_t) bit_reverse_u8(v >> 16) << 40) |
449 ((uint64_t) bit_reverse_u8(v >> 24) << 32) |
450 ((uint64_t) bit_reverse_u8(v >> 32) << 24) |
451 ((uint64_t) bit_reverse_u8(v >> 40) << 16) |
452 ((uint64_t) bit_reverse_u8(v >> 48) << 8) |
453 ((uint64_t) bit_reverse_u8(v >> 56));
454}
455#endif
456
457static
458unsigned long bit_reverse_ulong(unsigned long v)
459{
460#if (CAA_BITS_PER_LONG == 32)
461 return bit_reverse_u32(v);
462#else
463 return bit_reverse_u64(v);
464#endif
465}
466
467/*
468 * fls: returns the position of the most significant bit.
469 * Returns 0 if no bit is set, else returns the position of the most
470 * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit).
471 */
472#if defined(URCU_ARCH_X86)
473static inline
474unsigned int fls_u32(uint32_t x)
475{
476 int r;
477
478 __asm__ ("bsrl %1,%0\n\t"
479 "jnz 1f\n\t"
480 "movl $-1,%0\n\t"
481 "1:\n\t"
482 : "=r" (r) : "rm" (x));
483 return r + 1;
484}
485#define HAS_FLS_U32
486#endif
487
488#if defined(URCU_ARCH_AMD64)
489static inline
490unsigned int fls_u64(uint64_t x)
491{
492 long r;
493
494 __asm__ ("bsrq %1,%0\n\t"
495 "jnz 1f\n\t"
496 "movq $-1,%0\n\t"
497 "1:\n\t"
498 : "=r" (r) : "rm" (x));
499 return r + 1;
500}
501#define HAS_FLS_U64
502#endif
503
504#ifndef HAS_FLS_U64
505static __attribute__((unused))
506unsigned int fls_u64(uint64_t x)
507{
508 unsigned int r = 64;
509
510 if (!x)
511 return 0;
512
513 if (!(x & 0xFFFFFFFF00000000ULL)) {
514 x <<= 32;
515 r -= 32;
516 }
517 if (!(x & 0xFFFF000000000000ULL)) {
518 x <<= 16;
519 r -= 16;
520 }
521 if (!(x & 0xFF00000000000000ULL)) {
522 x <<= 8;
523 r -= 8;
524 }
525 if (!(x & 0xF000000000000000ULL)) {
526 x <<= 4;
527 r -= 4;
528 }
529 if (!(x & 0xC000000000000000ULL)) {
530 x <<= 2;
531 r -= 2;
532 }
533 if (!(x & 0x8000000000000000ULL)) {
534 x <<= 1;
535 r -= 1;
536 }
537 return r;
538}
539#endif
540
541#ifndef HAS_FLS_U32
542static __attribute__((unused))
543unsigned int fls_u32(uint32_t x)
544{
545 unsigned int r = 32;
546
547 if (!x)
548 return 0;
549 if (!(x & 0xFFFF0000U)) {
550 x <<= 16;
551 r -= 16;
552 }
553 if (!(x & 0xFF000000U)) {
554 x <<= 8;
555 r -= 8;
556 }
557 if (!(x & 0xF0000000U)) {
558 x <<= 4;
559 r -= 4;
560 }
561 if (!(x & 0xC0000000U)) {
562 x <<= 2;
563 r -= 2;
564 }
565 if (!(x & 0x80000000U)) {
566 x <<= 1;
567 r -= 1;
568 }
569 return r;
570}
571#endif
572
573unsigned int cds_lfht_fls_ulong(unsigned long x)
574{
575#if (CAA_BITS_PER_LONG == 32)
576 return fls_u32(x);
577#else
578 return fls_u64(x);
579#endif
580}
581
582/*
583 * Return the minimum order for which x <= (1UL << order).
584 * Return -1 if x is 0.
585 */
586static
587int cds_lfht_get_count_order_u32(uint32_t x)
588{
589 if (!x)
590 return -1;
591
592 return fls_u32(x - 1);
593}
594
595/*
596 * Return the minimum order for which x <= (1UL << order).
597 * Return -1 if x is 0.
598 */
599int cds_lfht_get_count_order_ulong(unsigned long x)
600{
601 if (!x)
602 return -1;
603
604 return cds_lfht_fls_ulong(x - 1);
605}
606
607static
608void cds_lfht_resize_lazy_grow(struct cds_lfht *ht, unsigned long size, int growth);
609
610static
611void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
612 unsigned long count);
613
614static void mutex_lock(pthread_mutex_t *mutex)
615{
616 int ret;
617
618#ifndef DISTRUST_SIGNALS_EXTREME
619 ret = pthread_mutex_lock(mutex);
620 if (ret)
621 urcu_die(ret);
622#else /* #ifndef DISTRUST_SIGNALS_EXTREME */
623 while ((ret = pthread_mutex_trylock(mutex)) != 0) {
624 if (ret != EBUSY && ret != EINTR)
625 urcu_die(ret);
626 if (CMM_LOAD_SHARED(URCU_TLS(rcu_reader).need_mb)) {
627 cmm_smp_mb();
628 _CMM_STORE_SHARED(URCU_TLS(rcu_reader).need_mb, 0);
629 cmm_smp_mb();
630 }
631 (void) poll(NULL, 0, 10);
632 }
633#endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
634}
635
636static void mutex_unlock(pthread_mutex_t *mutex)
637{
638 int ret;
639
640 ret = pthread_mutex_unlock(mutex);
641 if (ret)
642 urcu_die(ret);
643}
644
645static long nr_cpus_mask = -1;
646static long split_count_mask = -1;
647static int split_count_order = -1;
648
649static void ht_init_nr_cpus_mask(void)
650{
651 long maxcpus;
652
653 maxcpus = get_possible_cpus_array_len();
654 if (maxcpus <= 0) {
655 nr_cpus_mask = -2;
656 return;
657 }
658 /*
659 * round up number of CPUs to next power of two, so we
660 * can use & for modulo.
661 */
662 maxcpus = 1UL << cds_lfht_get_count_order_ulong(maxcpus);
663 nr_cpus_mask = maxcpus - 1;
664}
665
666static
667void alloc_split_items_count(struct cds_lfht *ht)
668{
669 if (nr_cpus_mask == -1) {
670 ht_init_nr_cpus_mask();
671 if (nr_cpus_mask < 0)
672 split_count_mask = DEFAULT_SPLIT_COUNT_MASK;
673 else
674 split_count_mask = nr_cpus_mask;
675 split_count_order =
676 cds_lfht_get_count_order_ulong(split_count_mask + 1);
677 }
678
679 urcu_posix_assert(split_count_mask >= 0);
680
681 if (ht->flags & CDS_LFHT_ACCOUNTING) {
682 ht->split_count = calloc(split_count_mask + 1,
683 sizeof(struct ht_items_count));
684 urcu_posix_assert(ht->split_count);
685 } else {
686 ht->split_count = NULL;
687 }
688}
689
690static
691void free_split_items_count(struct cds_lfht *ht)
692{
693 poison_free(ht->split_count);
694}
695
696static
697int ht_get_split_count_index(unsigned long hash)
698{
699 int cpu;
700
701 urcu_posix_assert(split_count_mask >= 0);
702 cpu = urcu_sched_getcpu();
703 if (caa_unlikely(cpu < 0))
704 return hash & split_count_mask;
705 else
706 return cpu & split_count_mask;
707}
708
709static
710void ht_count_add(struct cds_lfht *ht, unsigned long size, unsigned long hash)
711{
712 unsigned long split_count, count;
713 int index;
714
715 if (caa_unlikely(!ht->split_count))
716 return;
717 index = ht_get_split_count_index(hash);
718 split_count = uatomic_add_return(&ht->split_count[index].add, 1);
719 if (caa_likely(split_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))
720 return;
721 /* Only if number of add multiple of 1UL << COUNT_COMMIT_ORDER */
722
723 dbg_printf("add split count %lu\n", split_count);
724 count = uatomic_add_return(&ht->count,
725 1UL << COUNT_COMMIT_ORDER);
726 if (caa_likely(count & (count - 1)))
727 return;
728 /* Only if global count is power of 2 */
729
730 if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) < size)
731 return;
732 dbg_printf("add set global %lu\n", count);
733 cds_lfht_resize_lazy_count(ht, size,
734 count >> (CHAIN_LEN_TARGET - 1));
735}
736
737static
738void ht_count_del(struct cds_lfht *ht, unsigned long size, unsigned long hash)
739{
740 unsigned long split_count, count;
741 int index;
742
743 if (caa_unlikely(!ht->split_count))
744 return;
745 index = ht_get_split_count_index(hash);
746 split_count = uatomic_add_return(&ht->split_count[index].del, 1);
747 if (caa_likely(split_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))
748 return;
749 /* Only if number of deletes multiple of 1UL << COUNT_COMMIT_ORDER */
750
751 dbg_printf("del split count %lu\n", split_count);
752 count = uatomic_add_return(&ht->count,
753 -(1UL << COUNT_COMMIT_ORDER));
754 if (caa_likely(count & (count - 1)))
755 return;
756 /* Only if global count is power of 2 */
757
758 if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) >= size)
759 return;
760 dbg_printf("del set global %lu\n", count);
761 /*
762 * Don't shrink table if the number of nodes is below a
763 * certain threshold.
764 */
765 if (count < (1UL << COUNT_COMMIT_ORDER) * (split_count_mask + 1))
766 return;
767 cds_lfht_resize_lazy_count(ht, size,
768 count >> (CHAIN_LEN_TARGET - 1));
769}
770
771static
772void check_resize(struct cds_lfht *ht, unsigned long size, uint32_t chain_len)
773{
774 unsigned long count;
775
776 if (!(ht->flags & CDS_LFHT_AUTO_RESIZE))
777 return;
778 count = uatomic_read(&ht->count);
779 /*
780 * Use bucket-local length for small table expand and for
781 * environments lacking per-cpu data support.
782 */
783 if (count >= (1UL << (COUNT_COMMIT_ORDER + split_count_order)))
784 return;
785 if (chain_len > 100)
786 dbg_printf("WARNING: large chain length: %u.\n",
787 chain_len);
788 if (chain_len >= CHAIN_LEN_RESIZE_THRESHOLD) {
789 int growth;
790
791 /*
792 * Ideal growth calculated based on chain length.
793 */
794 growth = cds_lfht_get_count_order_u32(chain_len
795 - (CHAIN_LEN_TARGET - 1));
796 if ((ht->flags & CDS_LFHT_ACCOUNTING)
797 && (size << growth)
798 >= (1UL << (COUNT_COMMIT_ORDER
799 + split_count_order))) {
800 /*
801 * If ideal growth expands the hash table size
802 * beyond the "small hash table" sizes, use the
803 * maximum small hash table size to attempt
804 * expanding the hash table. This only applies
805 * when node accounting is available, otherwise
806 * the chain length is used to expand the hash
807 * table in every case.
808 */
809 growth = COUNT_COMMIT_ORDER + split_count_order
810 - cds_lfht_get_count_order_ulong(size);
811 if (growth <= 0)
812 return;
813 }
814 cds_lfht_resize_lazy_grow(ht, size, growth);
815 }
816}
817
818static
819struct cds_lfht_node *clear_flag(struct cds_lfht_node *node)
820{
821 return (struct cds_lfht_node *) (((unsigned long) node) & ~FLAGS_MASK);
822}
823
824static
825int is_removed(const struct cds_lfht_node *node)
826{
827 return ((unsigned long) node) & REMOVED_FLAG;
828}
829
830static
831int is_bucket(struct cds_lfht_node *node)
832{
833 return ((unsigned long) node) & BUCKET_FLAG;
834}
835
836static
837struct cds_lfht_node *flag_bucket(struct cds_lfht_node *node)
838{
839 return (struct cds_lfht_node *) (((unsigned long) node) | BUCKET_FLAG);
840}
841
842static
843int is_removal_owner(struct cds_lfht_node *node)
844{
845 return ((unsigned long) node) & REMOVAL_OWNER_FLAG;
846}
847
848static
849struct cds_lfht_node *flag_removed(struct cds_lfht_node *node)
850{
851 return (struct cds_lfht_node *) (((unsigned long) node) | REMOVED_FLAG);
852}
853
854static
855struct cds_lfht_node *flag_removal_owner(struct cds_lfht_node *node)
856{
857 return (struct cds_lfht_node *) (((unsigned long) node) | REMOVAL_OWNER_FLAG);
858}
859
860static
861struct cds_lfht_node *flag_removed_or_removal_owner(struct cds_lfht_node *node)
862{
863 return (struct cds_lfht_node *) (((unsigned long) node) | REMOVED_FLAG | REMOVAL_OWNER_FLAG);
864}
865
866static
867struct cds_lfht_node *get_end(void)
868{
869 return (struct cds_lfht_node *) END_VALUE;
870}
871
872static
873int is_end(struct cds_lfht_node *node)
874{
875 return clear_flag(node) == (struct cds_lfht_node *) END_VALUE;
876}
877
878static
879unsigned long _uatomic_xchg_monotonic_increase(unsigned long *ptr,
880 unsigned long v)
881{
882 unsigned long old1, old2;
883
884 old1 = uatomic_read(ptr);
885 do {
886 old2 = old1;
887 if (old2 >= v)
888 return old2;
889 } while ((old1 = uatomic_cmpxchg(ptr, old2, v)) != old2);
890 return old2;
891}
892
893static
894void cds_lfht_alloc_bucket_table(struct cds_lfht *ht, unsigned long order)
895{
896 return ht->mm->alloc_bucket_table(ht, order);
897}
898
899/*
900 * cds_lfht_free_bucket_table() should be called with decreasing order.
901 * When cds_lfht_free_bucket_table(0) is called, it means the whole
902 * lfht is destroyed.
903 */
904static
905void cds_lfht_free_bucket_table(struct cds_lfht *ht, unsigned long order)
906{
907 return ht->mm->free_bucket_table(ht, order);
908}
909
910static inline
911struct cds_lfht_node *bucket_at(struct cds_lfht *ht, unsigned long index)
912{
913 return ht->bucket_at(ht, index);
914}
915
916static inline
917struct cds_lfht_node *lookup_bucket(struct cds_lfht *ht, unsigned long size,
918 unsigned long hash)
919{
920 urcu_posix_assert(size > 0);
921 return bucket_at(ht, hash & (size - 1));
922}
923
924/*
925 * Remove all logically deleted nodes from a bucket up to a certain node key.
926 */
927static
928void _cds_lfht_gc_bucket(struct cds_lfht_node *bucket, struct cds_lfht_node *node)
929{
930 struct cds_lfht_node *iter_prev, *iter, *next, *new_next;
931
932 urcu_posix_assert(!is_bucket(bucket));
933 urcu_posix_assert(!is_removed(bucket));
934 urcu_posix_assert(!is_removal_owner(bucket));
935 urcu_posix_assert(!is_bucket(node));
936 urcu_posix_assert(!is_removed(node));
937 urcu_posix_assert(!is_removal_owner(node));
938 for (;;) {
939 iter_prev = bucket;
940 /* We can always skip the bucket node initially */
941 iter = rcu_dereference(iter_prev->next);
942 urcu_posix_assert(!is_removed(iter));
943 urcu_posix_assert(!is_removal_owner(iter));
944 urcu_posix_assert(iter_prev->reverse_hash <= node->reverse_hash);
945 /*
946 * We should never be called with bucket (start of chain)
947 * and logically removed node (end of path compression
948 * marker) being the actual same node. This would be a
949 * bug in the algorithm implementation.
950 */
951 urcu_posix_assert(bucket != node);
952 for (;;) {
953 if (caa_unlikely(is_end(iter)))
954 return;
955 if (caa_likely(clear_flag(iter)->reverse_hash > node->reverse_hash))
956 return;
957 next = rcu_dereference(clear_flag(iter)->next);
958 if (caa_likely(is_removed(next)))
959 break;
960 iter_prev = clear_flag(iter);
961 iter = next;
962 }
963 urcu_posix_assert(!is_removed(iter));
964 urcu_posix_assert(!is_removal_owner(iter));
965 if (is_bucket(iter))
966 new_next = flag_bucket(clear_flag(next));
967 else
968 new_next = clear_flag(next);
969 (void) uatomic_cmpxchg(&iter_prev->next, iter, new_next);
970 }
971}
972
973static
974int _cds_lfht_replace(struct cds_lfht *ht, unsigned long size,
975 struct cds_lfht_node *old_node,
976 struct cds_lfht_node *old_next,
977 struct cds_lfht_node *new_node)
978{
979 struct cds_lfht_node *bucket, *ret_next;
980
981 if (!old_node) /* Return -ENOENT if asked to replace NULL node */
982 return -ENOENT;
983
984 urcu_posix_assert(!is_removed(old_node));
985 urcu_posix_assert(!is_removal_owner(old_node));
986 urcu_posix_assert(!is_bucket(old_node));
987 urcu_posix_assert(!is_removed(new_node));
988 urcu_posix_assert(!is_removal_owner(new_node));
989 urcu_posix_assert(!is_bucket(new_node));
990 urcu_posix_assert(new_node != old_node);
991 for (;;) {
992 /* Insert after node to be replaced */
993 if (is_removed(old_next)) {
994 /*
995 * Too late, the old node has been removed under us
996 * between lookup and replace. Fail.
997 */
998 return -ENOENT;
999 }
1000 urcu_posix_assert(old_next == clear_flag(old_next));
1001 urcu_posix_assert(new_node != old_next);
1002 /*
1003 * REMOVAL_OWNER flag is _NEVER_ set before the REMOVED
1004 * flag. It is either set atomically at the same time
1005 * (replace) or after (del).
1006 */
1007 urcu_posix_assert(!is_removal_owner(old_next));
1008 new_node->next = old_next;
1009 /*
1010 * Here is the whole trick for lock-free replace: we add
1011 * the replacement node _after_ the node we want to
1012 * replace by atomically setting its next pointer at the
1013 * same time we set its removal flag. Given that
1014 * the lookups/get next use an iterator aware of the
1015 * next pointer, they will either skip the old node due
1016 * to the removal flag and see the new node, or use
1017 * the old node, but will not see the new one.
1018 * This is a replacement of a node with another node
1019 * that has the same value: we are therefore not
1020 * removing a value from the hash table. We set both the
1021 * REMOVED and REMOVAL_OWNER flags atomically so we own
1022 * the node after successful cmpxchg.
1023 */
1024 ret_next = uatomic_cmpxchg(&old_node->next,
1025 old_next, flag_removed_or_removal_owner(new_node));
1026 if (ret_next == old_next)
1027 break; /* We performed the replacement. */
1028 old_next = ret_next;
1029 }
1030
1031 /*
1032 * Ensure that the old node is not visible to readers anymore:
1033 * lookup for the node, and remove it (along with any other
1034 * logically removed node) if found.
1035 */
1036 bucket = lookup_bucket(ht, size, bit_reverse_ulong(old_node->reverse_hash));
1037 _cds_lfht_gc_bucket(bucket, new_node);
1038
1039 urcu_posix_assert(is_removed(CMM_LOAD_SHARED(old_node->next)));
1040 return 0;
1041}
1042
1043/*
1044 * A non-NULL unique_ret pointer uses the "add unique" (or uniquify) add
1045 * mode. A NULL unique_ret allows creation of duplicate keys.
1046 */
1047static
1048void _cds_lfht_add(struct cds_lfht *ht,
1049 unsigned long hash,
1050 cds_lfht_match_fct match,
1051 const void *key,
1052 unsigned long size,
1053 struct cds_lfht_node *node,
1054 struct cds_lfht_iter *unique_ret,
1055 int bucket_flag)
1056{
1057 struct cds_lfht_node *iter_prev, *iter, *next, *new_node, *new_next,
1058 *return_node;
1059 struct cds_lfht_node *bucket;
1060
1061 urcu_posix_assert(!is_bucket(node));
1062 urcu_posix_assert(!is_removed(node));
1063 urcu_posix_assert(!is_removal_owner(node));
1064 bucket = lookup_bucket(ht, size, hash);
1065 for (;;) {
1066 uint32_t chain_len = 0;
1067
1068 /*
1069 * iter_prev points to the non-removed node prior to the
1070 * insert location.
1071 */
1072 iter_prev = bucket;
1073 /* We can always skip the bucket node initially */
1074 iter = rcu_dereference(iter_prev->next);
1075 urcu_posix_assert(iter_prev->reverse_hash <= node->reverse_hash);
1076 for (;;) {
1077 if (caa_unlikely(is_end(iter)))
1078 goto insert;
1079 if (caa_likely(clear_flag(iter)->reverse_hash > node->reverse_hash))
1080 goto insert;
1081
1082 /* bucket node is the first node of the identical-hash-value chain */
1083 if (bucket_flag && clear_flag(iter)->reverse_hash == node->reverse_hash)
1084 goto insert;
1085
1086 next = rcu_dereference(clear_flag(iter)->next);
1087 if (caa_unlikely(is_removed(next)))
1088 goto gc_node;
1089
1090 /* uniquely add */
1091 if (unique_ret
1092 && !is_bucket(next)
1093 && clear_flag(iter)->reverse_hash == node->reverse_hash) {
1094 struct cds_lfht_iter d_iter = {
1095 .node = node,
1096 .next = iter,
1097#ifdef CONFIG_CDS_LFHT_ITER_DEBUG
1098 .lfht = ht,
1099#endif
1100 };
1101
1102 /*
1103 * uniquely adding inserts the node as the first
1104 * node of the identical-hash-value node chain.
1105 *
1106 * This semantic ensures no duplicated keys
1107 * should ever be observable in the table
1108 * (including traversing the table node by
1109 * node by forward iterations)
1110 */
1111 cds_lfht_next_duplicate(ht, match, key, &d_iter);
1112 if (!d_iter.node)
1113 goto insert;
1114
1115 *unique_ret = d_iter;
1116 return;
1117 }
1118
1119 /* Only account for identical reverse hash once */
1120 if (iter_prev->reverse_hash != clear_flag(iter)->reverse_hash
1121 && !is_bucket(next))
1122 check_resize(ht, size, ++chain_len);
1123 iter_prev = clear_flag(iter);
1124 iter = next;
1125 }
1126
1127 insert:
1128 urcu_posix_assert(node != clear_flag(iter));
1129 urcu_posix_assert(!is_removed(iter_prev));
1130 urcu_posix_assert(!is_removal_owner(iter_prev));
1131 urcu_posix_assert(!is_removed(iter));
1132 urcu_posix_assert(!is_removal_owner(iter));
1133 urcu_posix_assert(iter_prev != node);
1134 if (!bucket_flag)
1135 node->next = clear_flag(iter);
1136 else
1137 node->next = flag_bucket(clear_flag(iter));
1138 if (is_bucket(iter))
1139 new_node = flag_bucket(node);
1140 else
1141 new_node = node;
1142 if (uatomic_cmpxchg(&iter_prev->next, iter,
1143 new_node) != iter) {
1144 continue; /* retry */
1145 } else {
1146 return_node = node;
1147 goto end;
1148 }
1149
1150 gc_node:
1151 urcu_posix_assert(!is_removed(iter));
1152 urcu_posix_assert(!is_removal_owner(iter));
1153 if (is_bucket(iter))
1154 new_next = flag_bucket(clear_flag(next));
1155 else
1156 new_next = clear_flag(next);
1157 (void) uatomic_cmpxchg(&iter_prev->next, iter, new_next);
1158 /* retry */
1159 }
1160end:
1161 if (unique_ret) {
1162 unique_ret->node = return_node;
1163 /* unique_ret->next left unset, never used. */
1164 }
1165}
1166
1167static
1168int _cds_lfht_del(struct cds_lfht *ht, unsigned long size,
1169 struct cds_lfht_node *node)
1170{
1171 struct cds_lfht_node *bucket, *next;
1172
1173 if (!node) /* Return -ENOENT if asked to delete NULL node */
1174 return -ENOENT;
1175
1176 /* logically delete the node */
1177 urcu_posix_assert(!is_bucket(node));
1178 urcu_posix_assert(!is_removed(node));
1179 urcu_posix_assert(!is_removal_owner(node));
1180
1181 /*
1182 * We are first checking if the node had previously been
1183 * logically removed (this check is not atomic with setting the
1184 * logical removal flag). Return -ENOENT if the node had
1185 * previously been removed.
1186 */
1187 next = CMM_LOAD_SHARED(node->next); /* next is not dereferenced */
1188 if (caa_unlikely(is_removed(next)))
1189 return -ENOENT;
1190 urcu_posix_assert(!is_bucket(next));
1191 /*
1192 * The del operation semantic guarantees a full memory barrier
1193 * before the uatomic_or atomic commit of the deletion flag.
1194 */
1195 cmm_smp_mb__before_uatomic_or();
1196 /*
1197 * We set the REMOVED_FLAG unconditionally. Note that there may
1198 * be more than one concurrent thread setting this flag.
1199 * Knowing which wins the race will be known after the garbage
1200 * collection phase, stay tuned!
1201 */
1202 uatomic_or(&node->next, REMOVED_FLAG);
1203 /* We performed the (logical) deletion. */
1204
1205 /*
1206 * Ensure that the node is not visible to readers anymore: lookup for
1207 * the node, and remove it (along with any other logically removed node)
1208 * if found.
1209 */
1210 bucket = lookup_bucket(ht, size, bit_reverse_ulong(node->reverse_hash));
1211 _cds_lfht_gc_bucket(bucket, node);
1212
1213 urcu_posix_assert(is_removed(CMM_LOAD_SHARED(node->next)));
1214 /*
1215 * Last phase: atomically exchange node->next with a version
1216 * having "REMOVAL_OWNER_FLAG" set. If the returned node->next
1217 * pointer did _not_ have "REMOVAL_OWNER_FLAG" set, we now own
1218 * the node and win the removal race.
1219 * It is interesting to note that all "add" paths are forbidden
1220 * to change the next pointer starting from the point where the
1221 * REMOVED_FLAG is set, so here using a read, followed by a
1222 * xchg() suffice to guarantee that the xchg() will ever only
1223 * set the "REMOVAL_OWNER_FLAG" (or change nothing if the flag
1224 * was already set).
1225 */
1226 if (!is_removal_owner(uatomic_xchg(&node->next,
1227 flag_removal_owner(node->next))))
1228 return 0;
1229 else
1230 return -ENOENT;
1231}
1232
1233static
1234void *partition_resize_thread(void *arg)
1235{
1236 struct partition_resize_work *work = arg;
1237
1238 work->ht->flavor->register_thread();
1239 work->fct(work->ht, work->i, work->start, work->len);
1240 work->ht->flavor->unregister_thread();
1241 return NULL;
1242}
1243
1244static
1245void partition_resize_helper(struct cds_lfht *ht, unsigned long i,
1246 unsigned long len,
1247 void (*fct)(struct cds_lfht *ht, unsigned long i,
1248 unsigned long start, unsigned long len))
1249{
1250 unsigned long partition_len, start = 0;
1251 struct partition_resize_work *work;
1252 int ret;
1253 unsigned long thread, nr_threads;
1254
1255 urcu_posix_assert(nr_cpus_mask != -1);
1256 if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD)
1257 goto fallback;
1258
1259 /*
1260 * Note: nr_cpus_mask + 1 is always power of 2.
1261 * We spawn just the number of threads we need to satisfy the minimum
1262 * partition size, up to the number of CPUs in the system.
1263 */
1264 if (nr_cpus_mask > 0) {
1265 nr_threads = min_t(unsigned long, nr_cpus_mask + 1,
1266 len >> MIN_PARTITION_PER_THREAD_ORDER);
1267 } else {
1268 nr_threads = 1;
1269 }
1270 partition_len = len >> cds_lfht_get_count_order_ulong(nr_threads);
1271 work = calloc(nr_threads, sizeof(*work));
1272 if (!work) {
1273 dbg_printf("error allocating for resize, single-threading\n");
1274 goto fallback;
1275 }
1276 for (thread = 0; thread < nr_threads; thread++) {
1277 work[thread].ht = ht;
1278 work[thread].i = i;
1279 work[thread].len = partition_len;
1280 work[thread].start = thread * partition_len;
1281 work[thread].fct = fct;
1282 ret = pthread_create(&(work[thread].thread_id), ht->resize_attr,
1283 partition_resize_thread, &work[thread]);
1284 if (ret == EAGAIN) {
1285 /*
1286 * Out of resources: wait and join the threads
1287 * we've created, then handle leftovers.
1288 */
1289 dbg_printf("error spawning for resize, single-threading\n");
1290 start = work[thread].start;
1291 len -= start;
1292 nr_threads = thread;
1293 break;
1294 }
1295 urcu_posix_assert(!ret);
1296 }
1297 for (thread = 0; thread < nr_threads; thread++) {
1298 ret = pthread_join(work[thread].thread_id, NULL);
1299 urcu_posix_assert(!ret);
1300 }
1301 free(work);
1302
1303 /*
1304 * A pthread_create failure above will either lead in us having
1305 * no threads to join or starting at a non-zero offset,
1306 * fallback to single thread processing of leftovers.
1307 */
1308 if (start == 0 && nr_threads > 0)
1309 return;
1310fallback:
1311 fct(ht, i, start, len);
1312}
1313
1314/*
1315 * Holding RCU read lock to protect _cds_lfht_add against memory
1316 * reclaim that could be performed by other worker threads (ABA
1317 * problem).
1318 *
1319 * When we reach a certain length, we can split this population phase over
1320 * many worker threads, based on the number of CPUs available in the system.
1321 * This should therefore take care of not having the expand lagging behind too
1322 * many concurrent insertion threads by using the scheduler's ability to
1323 * schedule bucket node population fairly with insertions.
1324 */
1325static
1326void init_table_populate_partition(struct cds_lfht *ht, unsigned long i,
1327 unsigned long start, unsigned long len)
1328{
1329 unsigned long j, size = 1UL << (i - 1);
1330
1331 urcu_posix_assert(i > MIN_TABLE_ORDER);
1332 ht->flavor->read_lock();
1333 for (j = size + start; j < size + start + len; j++) {
1334 struct cds_lfht_node *new_node = bucket_at(ht, j);
1335
1336 urcu_posix_assert(j >= size && j < (size << 1));
1337 dbg_printf("init populate: order %lu index %lu hash %lu\n",
1338 i, j, j);
1339 new_node->reverse_hash = bit_reverse_ulong(j);
1340 _cds_lfht_add(ht, j, NULL, NULL, size, new_node, NULL, 1);
1341 }
1342 ht->flavor->read_unlock();
1343}
1344
1345static
1346void init_table_populate(struct cds_lfht *ht, unsigned long i,
1347 unsigned long len)
1348{
1349 partition_resize_helper(ht, i, len, init_table_populate_partition);
1350}
1351
1352static
1353void init_table(struct cds_lfht *ht,
1354 unsigned long first_order, unsigned long last_order)
1355{
1356 unsigned long i;
1357
1358 dbg_printf("init table: first_order %lu last_order %lu\n",
1359 first_order, last_order);
1360 urcu_posix_assert(first_order > MIN_TABLE_ORDER);
1361 for (i = first_order; i <= last_order; i++) {
1362 unsigned long len;
1363
1364 len = 1UL << (i - 1);
1365 dbg_printf("init order %lu len: %lu\n", i, len);
1366
1367 /* Stop expand if the resize target changes under us */
1368 if (CMM_LOAD_SHARED(ht->resize_target) < (1UL << i))
1369 break;
1370
1371 cds_lfht_alloc_bucket_table(ht, i);
1372
1373 /*
1374 * Set all bucket nodes reverse hash values for a level and
1375 * link all bucket nodes into the table.
1376 */
1377 init_table_populate(ht, i, len);
1378
1379 /*
1380 * Update table size.
1381 */
1382 cmm_smp_wmb(); /* populate data before RCU size */
1383 CMM_STORE_SHARED(ht->size, 1UL << i);
1384
1385 dbg_printf("init new size: %lu\n", 1UL << i);
1386 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1387 break;
1388 }
1389}
1390
1391/*
1392 * Holding RCU read lock to protect _cds_lfht_remove against memory
1393 * reclaim that could be performed by other worker threads (ABA
1394 * problem).
1395 * For a single level, we logically remove and garbage collect each node.
1396 *
1397 * As a design choice, we perform logical removal and garbage collection on a
1398 * node-per-node basis to simplify this algorithm. We also assume keeping good
1399 * cache locality of the operation would overweight possible performance gain
1400 * that could be achieved by batching garbage collection for multiple levels.
1401 * However, this would have to be justified by benchmarks.
1402 *
1403 * Concurrent removal and add operations are helping us perform garbage
1404 * collection of logically removed nodes. We guarantee that all logically
1405 * removed nodes have been garbage-collected (unlinked) before work
1406 * enqueue is invoked to free a hole level of bucket nodes (after a
1407 * grace period).
1408 *
1409 * Logical removal and garbage collection can therefore be done in batch
1410 * or on a node-per-node basis, as long as the guarantee above holds.
1411 *
1412 * When we reach a certain length, we can split this removal over many worker
1413 * threads, based on the number of CPUs available in the system. This should
1414 * take care of not letting resize process lag behind too many concurrent
1415 * updater threads actively inserting into the hash table.
1416 */
1417static
1418void remove_table_partition(struct cds_lfht *ht, unsigned long i,
1419 unsigned long start, unsigned long len)
1420{
1421 unsigned long j, size = 1UL << (i - 1);
1422
1423 urcu_posix_assert(i > MIN_TABLE_ORDER);
1424 ht->flavor->read_lock();
1425 for (j = size + start; j < size + start + len; j++) {
1426 struct cds_lfht_node *fini_bucket = bucket_at(ht, j);
1427 struct cds_lfht_node *parent_bucket = bucket_at(ht, j - size);
1428
1429 urcu_posix_assert(j >= size && j < (size << 1));
1430 dbg_printf("remove entry: order %lu index %lu hash %lu\n",
1431 i, j, j);
1432 /* Set the REMOVED_FLAG to freeze the ->next for gc */
1433 uatomic_or(&fini_bucket->next, REMOVED_FLAG);
1434 _cds_lfht_gc_bucket(parent_bucket, fini_bucket);
1435 }
1436 ht->flavor->read_unlock();
1437}
1438
1439static
1440void remove_table(struct cds_lfht *ht, unsigned long i, unsigned long len)
1441{
1442 partition_resize_helper(ht, i, len, remove_table_partition);
1443}
1444
1445/*
1446 * fini_table() is never called for first_order == 0, which is why
1447 * free_by_rcu_order == 0 can be used as criterion to know if free must
1448 * be called.
1449 */
1450static
1451void fini_table(struct cds_lfht *ht,
1452 unsigned long first_order, unsigned long last_order)
1453{
1454 unsigned long free_by_rcu_order = 0, i;
1455
1456 dbg_printf("fini table: first_order %lu last_order %lu\n",
1457 first_order, last_order);
1458 urcu_posix_assert(first_order > MIN_TABLE_ORDER);
1459 for (i = last_order; i >= first_order; i--) {
1460 unsigned long len;
1461
1462 len = 1UL << (i - 1);
1463 dbg_printf("fini order %ld len: %lu\n", i, len);
1464
1465 /* Stop shrink if the resize target changes under us */
1466 if (CMM_LOAD_SHARED(ht->resize_target) > (1UL << (i - 1)))
1467 break;
1468
1469 cmm_smp_wmb(); /* populate data before RCU size */
1470 CMM_STORE_SHARED(ht->size, 1UL << (i - 1));
1471
1472 /*
1473 * We need to wait for all add operations to reach Q.S. (and
1474 * thus use the new table for lookups) before we can start
1475 * releasing the old bucket nodes. Otherwise their lookup will
1476 * return a logically removed node as insert position.
1477 */
1478 ht->flavor->update_synchronize_rcu();
1479 if (free_by_rcu_order)
1480 cds_lfht_free_bucket_table(ht, free_by_rcu_order);
1481
1482 /*
1483 * Set "removed" flag in bucket nodes about to be removed.
1484 * Unlink all now-logically-removed bucket node pointers.
1485 * Concurrent add/remove operation are helping us doing
1486 * the gc.
1487 */
1488 remove_table(ht, i, len);
1489
1490 free_by_rcu_order = i;
1491
1492 dbg_printf("fini new size: %lu\n", 1UL << i);
1493 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1494 break;
1495 }
1496
1497 if (free_by_rcu_order) {
1498 ht->flavor->update_synchronize_rcu();
1499 cds_lfht_free_bucket_table(ht, free_by_rcu_order);
1500 }
1501}
1502
1503/*
1504 * Never called with size < 1.
1505 */
1506static
1507void cds_lfht_create_bucket(struct cds_lfht *ht, unsigned long size)
1508{
1509 struct cds_lfht_node *prev, *node;
1510 unsigned long order, len, i;
1511 int bucket_order;
1512
1513 cds_lfht_alloc_bucket_table(ht, 0);
1514
1515 dbg_printf("create bucket: order 0 index 0 hash 0\n");
1516 node = bucket_at(ht, 0);
1517 node->next = flag_bucket(get_end());
1518 node->reverse_hash = 0;
1519
1520 bucket_order = cds_lfht_get_count_order_ulong(size);
1521 urcu_posix_assert(bucket_order >= 0);
1522
1523 for (order = 1; order < (unsigned long) bucket_order + 1; order++) {
1524 len = 1UL << (order - 1);
1525 cds_lfht_alloc_bucket_table(ht, order);
1526
1527 for (i = 0; i < len; i++) {
1528 /*
1529 * Now, we are trying to init the node with the
1530 * hash=(len+i) (which is also a bucket with the
1531 * index=(len+i)) and insert it into the hash table,
1532 * so this node has to be inserted after the bucket
1533 * with the index=(len+i)&(len-1)=i. And because there
1534 * is no other non-bucket node nor bucket node with
1535 * larger index/hash inserted, so the bucket node
1536 * being inserted should be inserted directly linked
1537 * after the bucket node with index=i.
1538 */
1539 prev = bucket_at(ht, i);
1540 node = bucket_at(ht, len + i);
1541
1542 dbg_printf("create bucket: order %lu index %lu hash %lu\n",
1543 order, len + i, len + i);
1544 node->reverse_hash = bit_reverse_ulong(len + i);
1545
1546 /* insert after prev */
1547 urcu_posix_assert(is_bucket(prev->next));
1548 node->next = prev->next;
1549 prev->next = flag_bucket(node);
1550 }
1551 }
1552}
1553
1554#if (CAA_BITS_PER_LONG > 32)
1555/*
1556 * For 64-bit architectures, with max number of buckets small enough not to
1557 * use the entire 64-bit memory mapping space (and allowing a fair number of
1558 * hash table instances), use the mmap allocator, which is faster. Otherwise,
1559 * fallback to the order allocator.
1560 */
1561static
1562const struct cds_lfht_mm_type *get_mm_type(unsigned long max_nr_buckets)
1563{
1564 if (max_nr_buckets && max_nr_buckets <= (1ULL << 32))
1565 return &cds_lfht_mm_mmap;
1566 else
1567 return &cds_lfht_mm_order;
1568}
1569#else
1570/*
1571 * For 32-bit architectures, use the order allocator.
1572 */
1573static
1574const struct cds_lfht_mm_type *get_mm_type(
1575 unsigned long max_nr_buckets __attribute__((unused)))
1576{
1577 return &cds_lfht_mm_order;
1578}
1579#endif
1580
1581void cds_lfht_node_init_deleted(struct cds_lfht_node *node)
1582{
1583 cds_lfht_node_init(node);
1584 node->next = flag_removed(NULL);
1585}
1586
1587struct cds_lfht *_cds_lfht_new(unsigned long init_size,
1588 unsigned long min_nr_alloc_buckets,
1589 unsigned long max_nr_buckets,
1590 int flags,
1591 const struct cds_lfht_mm_type *mm,
1592 const struct rcu_flavor_struct *flavor,
1593 pthread_attr_t *attr)
1594{
1595 struct cds_lfht *ht;
1596 unsigned long order;
1597
1598 /* min_nr_alloc_buckets must be power of two */
1599 if (!min_nr_alloc_buckets || (min_nr_alloc_buckets & (min_nr_alloc_buckets - 1)))
1600 return NULL;
1601
1602 /* init_size must be power of two */
1603 if (!init_size || (init_size & (init_size - 1)))
1604 return NULL;
1605
1606 /*
1607 * Memory management plugin default.
1608 */
1609 if (!mm)
1610 mm = get_mm_type(max_nr_buckets);
1611
1612 /* max_nr_buckets == 0 for order based mm means infinite */
1613 if (mm == &cds_lfht_mm_order && !max_nr_buckets)
1614 max_nr_buckets = 1UL << (MAX_TABLE_ORDER - 1);
1615
1616 /* max_nr_buckets must be power of two */
1617 if (!max_nr_buckets || (max_nr_buckets & (max_nr_buckets - 1)))
1618 return NULL;
1619
1620 if (flags & CDS_LFHT_AUTO_RESIZE)
1621 cds_lfht_init_worker(flavor);
1622
1623 min_nr_alloc_buckets = max(min_nr_alloc_buckets, MIN_TABLE_SIZE);
1624 init_size = max(init_size, MIN_TABLE_SIZE);
1625 max_nr_buckets = max(max_nr_buckets, min_nr_alloc_buckets);
1626 init_size = min(init_size, max_nr_buckets);
1627
1628 ht = mm->alloc_cds_lfht(min_nr_alloc_buckets, max_nr_buckets);
1629 urcu_posix_assert(ht);
1630 urcu_posix_assert(ht->mm == mm);
1631 urcu_posix_assert(ht->bucket_at == mm->bucket_at);
1632
1633 ht->flags = flags;
1634 ht->flavor = flavor;
1635 ht->resize_attr = attr;
1636 alloc_split_items_count(ht);
1637 /* this mutex should not nest in read-side C.S. */
1638 pthread_mutex_init(&ht->resize_mutex, NULL);
1639 order = cds_lfht_get_count_order_ulong(init_size);
1640 ht->resize_target = 1UL << order;
1641 cds_lfht_create_bucket(ht, 1UL << order);
1642 ht->size = 1UL << order;
1643 return ht;
1644}
1645
1646void cds_lfht_lookup(struct cds_lfht *ht, unsigned long hash,
1647 cds_lfht_match_fct match, const void *key,
1648 struct cds_lfht_iter *iter)
1649{
1650 struct cds_lfht_node *node, *next, *bucket;
1651 unsigned long reverse_hash, size;
1652
1653 cds_lfht_iter_debug_set_ht(ht, iter);
1654
1655 reverse_hash = bit_reverse_ulong(hash);
1656
1657 size = rcu_dereference(ht->size);
1658 bucket = lookup_bucket(ht, size, hash);
1659 /* We can always skip the bucket node initially */
1660 node = rcu_dereference(bucket->next);
1661 node = clear_flag(node);
1662 for (;;) {
1663 if (caa_unlikely(is_end(node))) {
1664 node = next = NULL;
1665 break;
1666 }
1667 if (caa_unlikely(node->reverse_hash > reverse_hash)) {
1668 node = next = NULL;
1669 break;
1670 }
1671 next = rcu_dereference(node->next);
1672 urcu_posix_assert(node == clear_flag(node));
1673 if (caa_likely(!is_removed(next))
1674 && !is_bucket(next)
1675 && node->reverse_hash == reverse_hash
1676 && caa_likely(match(node, key))) {
1677 break;
1678 }
1679 node = clear_flag(next);
1680 }
1681 urcu_posix_assert(!node || !is_bucket(CMM_LOAD_SHARED(node->next)));
1682 iter->node = node;
1683 iter->next = next;
1684}
1685
1686void cds_lfht_next_duplicate(struct cds_lfht *ht __attribute__((unused)),
1687 cds_lfht_match_fct match,
1688 const void *key, struct cds_lfht_iter *iter)
1689{
1690 struct cds_lfht_node *node, *next;
1691 unsigned long reverse_hash;
1692
1693 cds_lfht_iter_debug_assert(ht == iter->lfht);
1694 node = iter->node;
1695 reverse_hash = node->reverse_hash;
1696 next = iter->next;
1697 node = clear_flag(next);
1698
1699 for (;;) {
1700 if (caa_unlikely(is_end(node))) {
1701 node = next = NULL;
1702 break;
1703 }
1704 if (caa_unlikely(node->reverse_hash > reverse_hash)) {
1705 node = next = NULL;
1706 break;
1707 }
1708 next = rcu_dereference(node->next);
1709 if (caa_likely(!is_removed(next))
1710 && !is_bucket(next)
1711 && caa_likely(match(node, key))) {
1712 break;
1713 }
1714 node = clear_flag(next);
1715 }
1716 urcu_posix_assert(!node || !is_bucket(CMM_LOAD_SHARED(node->next)));
1717 iter->node = node;
1718 iter->next = next;
1719}
1720
1721void cds_lfht_next(struct cds_lfht *ht __attribute__((unused)),
1722 struct cds_lfht_iter *iter)
1723{
1724 struct cds_lfht_node *node, *next;
1725
1726 cds_lfht_iter_debug_assert(ht == iter->lfht);
1727 node = clear_flag(iter->next);
1728 for (;;) {
1729 if (caa_unlikely(is_end(node))) {
1730 node = next = NULL;
1731 break;
1732 }
1733 next = rcu_dereference(node->next);
1734 if (caa_likely(!is_removed(next))
1735 && !is_bucket(next)) {
1736 break;
1737 }
1738 node = clear_flag(next);
1739 }
1740 urcu_posix_assert(!node || !is_bucket(CMM_LOAD_SHARED(node->next)));
1741 iter->node = node;
1742 iter->next = next;
1743}
1744
1745void cds_lfht_first(struct cds_lfht *ht, struct cds_lfht_iter *iter)
1746{
1747 cds_lfht_iter_debug_set_ht(ht, iter);
1748 /*
1749 * Get next after first bucket node. The first bucket node is the
1750 * first node of the linked list.
1751 */
1752 iter->next = bucket_at(ht, 0)->next;
1753 cds_lfht_next(ht, iter);
1754}
1755
1756void cds_lfht_add(struct cds_lfht *ht, unsigned long hash,
1757 struct cds_lfht_node *node)
1758{
1759 unsigned long size;
1760
1761 node->reverse_hash = bit_reverse_ulong(hash);
1762 size = rcu_dereference(ht->size);
1763 _cds_lfht_add(ht, hash, NULL, NULL, size, node, NULL, 0);
1764 ht_count_add(ht, size, hash);
1765}
1766
1767struct cds_lfht_node *cds_lfht_add_unique(struct cds_lfht *ht,
1768 unsigned long hash,
1769 cds_lfht_match_fct match,
1770 const void *key,
1771 struct cds_lfht_node *node)
1772{
1773 unsigned long size;
1774 struct cds_lfht_iter iter;
1775
1776 node->reverse_hash = bit_reverse_ulong(hash);
1777 size = rcu_dereference(ht->size);
1778 _cds_lfht_add(ht, hash, match, key, size, node, &iter, 0);
1779 if (iter.node == node)
1780 ht_count_add(ht, size, hash);
1781 return iter.node;
1782}
1783
1784struct cds_lfht_node *cds_lfht_add_replace(struct cds_lfht *ht,
1785 unsigned long hash,
1786 cds_lfht_match_fct match,
1787 const void *key,
1788 struct cds_lfht_node *node)
1789{
1790 unsigned long size;
1791 struct cds_lfht_iter iter;
1792
1793 node->reverse_hash = bit_reverse_ulong(hash);
1794 size = rcu_dereference(ht->size);
1795 for (;;) {
1796 _cds_lfht_add(ht, hash, match, key, size, node, &iter, 0);
1797 if (iter.node == node) {
1798 ht_count_add(ht, size, hash);
1799 return NULL;
1800 }
1801
1802 if (!_cds_lfht_replace(ht, size, iter.node, iter.next, node))
1803 return iter.node;
1804 }
1805}
1806
1807int cds_lfht_replace(struct cds_lfht *ht,
1808 struct cds_lfht_iter *old_iter,
1809 unsigned long hash,
1810 cds_lfht_match_fct match,
1811 const void *key,
1812 struct cds_lfht_node *new_node)
1813{
1814 unsigned long size;
1815
1816 new_node->reverse_hash = bit_reverse_ulong(hash);
1817 if (!old_iter->node)
1818 return -ENOENT;
1819 if (caa_unlikely(old_iter->node->reverse_hash != new_node->reverse_hash))
1820 return -EINVAL;
1821 if (caa_unlikely(!match(old_iter->node, key)))
1822 return -EINVAL;
1823 size = rcu_dereference(ht->size);
1824 return _cds_lfht_replace(ht, size, old_iter->node, old_iter->next,
1825 new_node);
1826}
1827
1828int cds_lfht_del(struct cds_lfht *ht, struct cds_lfht_node *node)
1829{
1830 unsigned long size;
1831 int ret;
1832
1833 size = rcu_dereference(ht->size);
1834 ret = _cds_lfht_del(ht, size, node);
1835 if (!ret) {
1836 unsigned long hash;
1837
1838 hash = bit_reverse_ulong(node->reverse_hash);
1839 ht_count_del(ht, size, hash);
1840 }
1841 return ret;
1842}
1843
1844int cds_lfht_is_node_deleted(const struct cds_lfht_node *node)
1845{
1846 return is_removed(CMM_LOAD_SHARED(node->next));
1847}
1848
1849static
1850int cds_lfht_delete_bucket(struct cds_lfht *ht)
1851{
1852 struct cds_lfht_node *node;
1853 unsigned long order, i, size;
1854
1855 /* Check that the table is empty */
1856 node = bucket_at(ht, 0);
1857 do {
1858 node = clear_flag(node)->next;
1859 if (!is_bucket(node))
1860 return -EPERM;
1861 urcu_posix_assert(!is_removed(node));
1862 urcu_posix_assert(!is_removal_owner(node));
1863 } while (!is_end(node));
1864 /*
1865 * size accessed without rcu_dereference because hash table is
1866 * being destroyed.
1867 */
1868 size = ht->size;
1869 /* Internal sanity check: all nodes left should be buckets */
1870 for (i = 0; i < size; i++) {
1871 node = bucket_at(ht, i);
1872 dbg_printf("delete bucket: index %lu expected hash %lu hash %lu\n",
1873 i, i, bit_reverse_ulong(node->reverse_hash));
1874 urcu_posix_assert(is_bucket(node->next));
1875 }
1876
1877 for (order = cds_lfht_get_count_order_ulong(size); (long)order >= 0; order--)
1878 cds_lfht_free_bucket_table(ht, order);
1879
1880 return 0;
1881}
1882
1883/*
1884 * Should only be called when no more concurrent readers nor writers can
1885 * possibly access the table.
1886 */
1887int cds_lfht_destroy(struct cds_lfht *ht, pthread_attr_t **attr)
1888{
1889 int ret;
1890
1891 if (ht->flags & CDS_LFHT_AUTO_RESIZE) {
1892 /* Cancel ongoing resize operations. */
1893 _CMM_STORE_SHARED(ht->in_progress_destroy, 1);
1894 /* Wait for in-flight resize operations to complete */
1895 urcu_workqueue_flush_queued_work(cds_lfht_workqueue);
1896 }
1897 ret = cds_lfht_delete_bucket(ht);
1898 if (ret)
1899 return ret;
1900 free_split_items_count(ht);
1901 if (attr)
1902 *attr = ht->resize_attr;
1903 ret = pthread_mutex_destroy(&ht->resize_mutex);
1904 if (ret)
1905 ret = -EBUSY;
1906 if (ht->flags & CDS_LFHT_AUTO_RESIZE)
1907 cds_lfht_fini_worker(ht->flavor);
1908 poison_free(ht);
1909 return ret;
1910}
1911
1912void cds_lfht_count_nodes(struct cds_lfht *ht,
1913 long *approx_before,
1914 unsigned long *count,
1915 long *approx_after)
1916{
1917 struct cds_lfht_node *node, *next;
1918 unsigned long nr_bucket = 0, nr_removed = 0;
1919
1920 *approx_before = 0;
1921 if (ht->split_count) {
1922 int i;
1923
1924 for (i = 0; i < split_count_mask + 1; i++) {
1925 *approx_before += uatomic_read(&ht->split_count[i].add);
1926 *approx_before -= uatomic_read(&ht->split_count[i].del);
1927 }
1928 }
1929
1930 *count = 0;
1931
1932 /* Count non-bucket nodes in the table */
1933 node = bucket_at(ht, 0);
1934 do {
1935 next = rcu_dereference(node->next);
1936 if (is_removed(next)) {
1937 if (!is_bucket(next))
1938 (nr_removed)++;
1939 else
1940 (nr_bucket)++;
1941 } else if (!is_bucket(next))
1942 (*count)++;
1943 else
1944 (nr_bucket)++;
1945 node = clear_flag(next);
1946 } while (!is_end(node));
1947 dbg_printf("number of logically removed nodes: %lu\n", nr_removed);
1948 dbg_printf("number of bucket nodes: %lu\n", nr_bucket);
1949 *approx_after = 0;
1950 if (ht->split_count) {
1951 int i;
1952
1953 for (i = 0; i < split_count_mask + 1; i++) {
1954 *approx_after += uatomic_read(&ht->split_count[i].add);
1955 *approx_after -= uatomic_read(&ht->split_count[i].del);
1956 }
1957 }
1958}
1959
1960/* called with resize mutex held */
1961static
1962void _do_cds_lfht_grow(struct cds_lfht *ht,
1963 unsigned long old_size, unsigned long new_size)
1964{
1965 unsigned long old_order, new_order;
1966
1967 old_order = cds_lfht_get_count_order_ulong(old_size);
1968 new_order = cds_lfht_get_count_order_ulong(new_size);
1969 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1970 old_size, old_order, new_size, new_order);
1971 urcu_posix_assert(new_size > old_size);
1972 init_table(ht, old_order + 1, new_order);
1973}
1974
1975/* called with resize mutex held */
1976static
1977void _do_cds_lfht_shrink(struct cds_lfht *ht,
1978 unsigned long old_size, unsigned long new_size)
1979{
1980 unsigned long old_order, new_order;
1981
1982 new_size = max(new_size, MIN_TABLE_SIZE);
1983 old_order = cds_lfht_get_count_order_ulong(old_size);
1984 new_order = cds_lfht_get_count_order_ulong(new_size);
1985 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1986 old_size, old_order, new_size, new_order);
1987 urcu_posix_assert(new_size < old_size);
1988
1989 /* Remove and unlink all bucket nodes to remove. */
1990 fini_table(ht, new_order + 1, old_order);
1991}
1992
1993
1994/* called with resize mutex held */
1995static
1996void _do_cds_lfht_resize(struct cds_lfht *ht)
1997{
1998 unsigned long new_size, old_size;
1999
2000 /*
2001 * Resize table, re-do if the target size has changed under us.
2002 */
2003 do {
2004 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
2005 break;
2006 ht->resize_initiated = 1;
2007 old_size = ht->size;
2008 new_size = CMM_LOAD_SHARED(ht->resize_target);
2009 if (old_size < new_size)
2010 _do_cds_lfht_grow(ht, old_size, new_size);
2011 else if (old_size > new_size)
2012 _do_cds_lfht_shrink(ht, old_size, new_size);
2013 ht->resize_initiated = 0;
2014 /* write resize_initiated before read resize_target */
2015 cmm_smp_mb();
2016 } while (ht->size != CMM_LOAD_SHARED(ht->resize_target));
2017}
2018
2019static
2020unsigned long resize_target_grow(struct cds_lfht *ht, unsigned long new_size)
2021{
2022 return _uatomic_xchg_monotonic_increase(&ht->resize_target, new_size);
2023}
2024
2025static
2026void resize_target_update_count(struct cds_lfht *ht,
2027 unsigned long count)
2028{
2029 count = max(count, MIN_TABLE_SIZE);
2030 count = min(count, ht->max_nr_buckets);
2031 uatomic_set(&ht->resize_target, count);
2032}
2033
2034void cds_lfht_resize(struct cds_lfht *ht, unsigned long new_size)
2035{
2036 resize_target_update_count(ht, new_size);
2037 CMM_STORE_SHARED(ht->resize_initiated, 1);
2038 mutex_lock(&ht->resize_mutex);
2039 _do_cds_lfht_resize(ht);
2040 mutex_unlock(&ht->resize_mutex);
2041}
2042
2043static
2044void do_resize_cb(struct urcu_work *work)
2045{
2046 struct resize_work *resize_work =
2047 caa_container_of(work, struct resize_work, work);
2048 struct cds_lfht *ht = resize_work->ht;
2049
2050 ht->flavor->register_thread();
2051 mutex_lock(&ht->resize_mutex);
2052 _do_cds_lfht_resize(ht);
2053 mutex_unlock(&ht->resize_mutex);
2054 ht->flavor->unregister_thread();
2055 poison_free(work);
2056}
2057
2058static
2059void __cds_lfht_resize_lazy_launch(struct cds_lfht *ht)
2060{
2061 struct resize_work *work;
2062
2063 /* Store resize_target before read resize_initiated */
2064 cmm_smp_mb();
2065 if (!CMM_LOAD_SHARED(ht->resize_initiated)) {
2066 if (CMM_LOAD_SHARED(ht->in_progress_destroy)) {
2067 return;
2068 }
2069 work = malloc(sizeof(*work));
2070 if (work == NULL) {
2071 dbg_printf("error allocating resize work, bailing out\n");
2072 return;
2073 }
2074 work->ht = ht;
2075 urcu_workqueue_queue_work(cds_lfht_workqueue,
2076 &work->work, do_resize_cb);
2077 CMM_STORE_SHARED(ht->resize_initiated, 1);
2078 }
2079}
2080
2081static
2082void cds_lfht_resize_lazy_grow(struct cds_lfht *ht, unsigned long size, int growth)
2083{
2084 unsigned long target_size = size << growth;
2085
2086 target_size = min(target_size, ht->max_nr_buckets);
2087 if (resize_target_grow(ht, target_size) >= target_size)
2088 return;
2089
2090 __cds_lfht_resize_lazy_launch(ht);
2091}
2092
2093/*
2094 * We favor grow operations over shrink. A shrink operation never occurs
2095 * if a grow operation is queued for lazy execution. A grow operation
2096 * cancels any pending shrink lazy execution.
2097 */
2098static
2099void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
2100 unsigned long count)
2101{
2102 if (!(ht->flags & CDS_LFHT_AUTO_RESIZE))
2103 return;
2104 count = max(count, MIN_TABLE_SIZE);
2105 count = min(count, ht->max_nr_buckets);
2106 if (count == size)
2107 return; /* Already the right size, no resize needed */
2108 if (count > size) { /* lazy grow */
2109 if (resize_target_grow(ht, count) >= count)
2110 return;
2111 } else { /* lazy shrink */
2112 for (;;) {
2113 unsigned long s;
2114
2115 s = uatomic_cmpxchg(&ht->resize_target, size, count);
2116 if (s == size)
2117 break; /* no resize needed */
2118 if (s > size)
2119 return; /* growing is/(was just) in progress */
2120 if (s <= count)
2121 return; /* some other thread do shrink */
2122 size = s;
2123 }
2124 }
2125 __cds_lfht_resize_lazy_launch(ht);
2126}
2127
2128static void cds_lfht_before_fork(void *priv __attribute__((unused)))
2129{
2130 if (cds_lfht_workqueue_atfork_nesting++)
2131 return;
2132 mutex_lock(&cds_lfht_fork_mutex);
2133 if (!cds_lfht_workqueue)
2134 return;
2135 urcu_workqueue_pause_worker(cds_lfht_workqueue);
2136}
2137
2138static void cds_lfht_after_fork_parent(void *priv __attribute__((unused)))
2139{
2140 if (--cds_lfht_workqueue_atfork_nesting)
2141 return;
2142 if (!cds_lfht_workqueue)
2143 goto end;
2144 urcu_workqueue_resume_worker(cds_lfht_workqueue);
2145end:
2146 mutex_unlock(&cds_lfht_fork_mutex);
2147}
2148
2149static void cds_lfht_after_fork_child(void *priv __attribute__((unused)))
2150{
2151 if (--cds_lfht_workqueue_atfork_nesting)
2152 return;
2153 if (!cds_lfht_workqueue)
2154 goto end;
2155 urcu_workqueue_create_worker(cds_lfht_workqueue);
2156end:
2157 mutex_unlock(&cds_lfht_fork_mutex);
2158}
2159
2160static struct urcu_atfork cds_lfht_atfork = {
2161 .before_fork = cds_lfht_before_fork,
2162 .after_fork_parent = cds_lfht_after_fork_parent,
2163 .after_fork_child = cds_lfht_after_fork_child,
2164};
2165
2166/*
2167 * Block all signals for the workqueue worker thread to ensure we don't
2168 * disturb the application. The SIGRCU signal needs to be unblocked for
2169 * the urcu-signal flavor.
2170 */
2171static void cds_lfht_worker_init(
2172 struct urcu_workqueue *workqueue __attribute__((unused)),
2173 void *priv __attribute__((unused)))
2174{
2175 int ret;
2176 sigset_t mask;
2177
2178 ret = sigfillset(&mask);
2179 if (ret)
2180 urcu_die(errno);
2181 ret = sigdelset(&mask, SIGRCU);
2182 if (ret)
2183 urcu_die(errno);
2184 ret = pthread_sigmask(SIG_SETMASK, &mask, NULL);
2185 if (ret)
2186 urcu_die(ret);
2187}
2188
2189static void cds_lfht_init_worker(const struct rcu_flavor_struct *flavor)
2190{
2191 flavor->register_rculfhash_atfork(&cds_lfht_atfork);
2192
2193 mutex_lock(&cds_lfht_fork_mutex);
2194 if (cds_lfht_workqueue_user_count++)
2195 goto end;
2196 cds_lfht_workqueue = urcu_workqueue_create(0, -1, NULL,
2197 NULL, cds_lfht_worker_init, NULL, NULL, NULL, NULL, NULL);
2198end:
2199 mutex_unlock(&cds_lfht_fork_mutex);
2200}
2201
2202static void cds_lfht_fini_worker(const struct rcu_flavor_struct *flavor)
2203{
2204 mutex_lock(&cds_lfht_fork_mutex);
2205 if (--cds_lfht_workqueue_user_count)
2206 goto end;
2207 urcu_workqueue_destroy(cds_lfht_workqueue);
2208 cds_lfht_workqueue = NULL;
2209end:
2210 mutex_unlock(&cds_lfht_fork_mutex);
2211
2212 flavor->unregister_rculfhash_atfork(&cds_lfht_atfork);
2213}
This page took 0.030085 seconds and 4 git commands to generate.