9106a745ff29a24dcefef9a1e671ee28fd68b7e1
[urcu.git] / src / rculfhash.c
1 /*
2 * rculfhash.c
3 *
4 * Userspace RCU library - Lock-Free Resizable RCU Hash Table
5 *
6 * Copyright 2010-2011 - Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
7 * Copyright 2011 - Lai Jiangshan <laijs@cn.fujitsu.com>
8 *
9 * This library is free software; you can redistribute it and/or
10 * modify it under the terms of the GNU Lesser General Public
11 * License as published by the Free Software Foundation; either
12 * version 2.1 of the License, or (at your option) any later version.
13 *
14 * This library is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
17 * Lesser General Public License for more details.
18 *
19 * You should have received a copy of the GNU Lesser General Public
20 * License along with this library; if not, write to the Free Software
21 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
22 */
23
24 /*
25 * Based on the following articles:
26 * - Ori Shalev and Nir Shavit. Split-ordered lists: Lock-free
27 * extensible hash tables. J. ACM 53, 3 (May 2006), 379-405.
28 * - Michael, M. M. High performance dynamic lock-free hash tables
29 * and list-based sets. In Proceedings of the fourteenth annual ACM
30 * symposium on Parallel algorithms and architectures, ACM Press,
31 * (2002), 73-82.
32 *
33 * Some specificities of this Lock-Free Resizable RCU Hash Table
34 * implementation:
35 *
36 * - RCU read-side critical section allows readers to perform hash
37 * table lookups, as well as traversals, and use the returned objects
38 * safely by allowing memory reclaim to take place only after a grace
39 * period.
40 * - Add and remove operations are lock-free, and do not need to
41 * allocate memory. They need to be executed within RCU read-side
42 * critical section to ensure the objects they read are valid and to
43 * deal with the cmpxchg ABA problem.
44 * - add and add_unique operations are supported. add_unique checks if
45 * the node key already exists in the hash table. It ensures not to
46 * populate a duplicate key if the node key already exists in the hash
47 * table.
48 * - The resize operation executes concurrently with
49 * add/add_unique/add_replace/remove/lookup/traversal.
50 * - Hash table nodes are contained within a split-ordered list. This
51 * list is ordered by incrementing reversed-bits-hash value.
52 * - An index of bucket nodes is kept. These bucket nodes are the hash
53 * table "buckets". These buckets are internal nodes that allow to
54 * perform a fast hash lookup, similarly to a skip list. These
55 * buckets are chained together in the split-ordered list, which
56 * allows recursive expansion by inserting new buckets between the
57 * existing buckets. The split-ordered list allows adding new buckets
58 * between existing buckets as the table needs to grow.
59 * - The resize operation for small tables only allows expanding the
60 * hash table. It is triggered automatically by detecting long chains
61 * in the add operation.
62 * - The resize operation for larger tables (and available through an
63 * API) allows both expanding and shrinking the hash table.
64 * - Split-counters are used to keep track of the number of
65 * nodes within the hash table for automatic resize triggering.
66 * - Resize operation initiated by long chain detection is executed by a
67 * 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 <assert.h>
262 #include <stdio.h>
263 #include <stdint.h>
264 #include <string.h>
265 #include <sched.h>
266 #include <unistd.h>
267
268 #include "compat-getcpu.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
285 /*
286 * Split-counters lazily update the global counter each 1024
287 * addition/removal. It automatically keeps track of resize required.
288 * We use the bucket length as indicator for need to expand for small
289 * tables and machines lacking per-cpu data support.
290 */
291 #define COUNT_COMMIT_ORDER 10
292 #define DEFAULT_SPLIT_COUNT_MASK 0xFUL
293 #define CHAIN_LEN_TARGET 1
294 #define CHAIN_LEN_RESIZE_THRESHOLD 3
295
296 /*
297 * Define the minimum table size.
298 */
299 #define MIN_TABLE_ORDER 0
300 #define MIN_TABLE_SIZE (1UL << MIN_TABLE_ORDER)
301
302 /*
303 * Minimum number of bucket nodes to touch per thread to parallelize grow/shrink.
304 */
305 #define MIN_PARTITION_PER_THREAD_ORDER 12
306 #define MIN_PARTITION_PER_THREAD (1UL << MIN_PARTITION_PER_THREAD_ORDER)
307
308 /*
309 * The removed flag needs to be updated atomically with the pointer.
310 * It indicates that no node must attach to the node scheduled for
311 * removal, and that node garbage collection must be performed.
312 * The bucket flag does not require to be updated atomically with the
313 * pointer, but it is added as a pointer low bit flag to save space.
314 * The "removal owner" flag is used to detect which of the "del"
315 * operation that has set the "removed flag" gets to return the removed
316 * node to its caller. Note that the replace operation does not need to
317 * iteract with the "removal owner" flag, because it validates that
318 * the "removed" flag is not set before performing its cmpxchg.
319 */
320 #define REMOVED_FLAG (1UL << 0)
321 #define BUCKET_FLAG (1UL << 1)
322 #define REMOVAL_OWNER_FLAG (1UL << 2)
323 #define FLAGS_MASK ((1UL << 3) - 1)
324
325 /* Value of the end pointer. Should not interact with flags. */
326 #define END_VALUE NULL
327
328 /*
329 * ht_items_count: Split-counters counting the number of node addition
330 * and removal in the table. Only used if the CDS_LFHT_ACCOUNTING flag
331 * is set at hash table creation.
332 *
333 * These are free-running counters, never reset to zero. They count the
334 * number of add/remove, and trigger every (1 << COUNT_COMMIT_ORDER)
335 * operations to update the global counter. We choose a power-of-2 value
336 * for the trigger to deal with 32 or 64-bit overflow of the counter.
337 */
338 struct ht_items_count {
339 unsigned long add, del;
340 } __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
341
342 /*
343 * resize_work: Contains arguments passed to worker thread
344 * responsible for performing lazy resize.
345 */
346 struct resize_work {
347 struct urcu_work work;
348 struct cds_lfht *ht;
349 };
350
351 /*
352 * partition_resize_work: Contains arguments passed to worker threads
353 * executing the hash table resize on partitions of the hash table
354 * assigned to each processor's worker thread.
355 */
356 struct partition_resize_work {
357 pthread_t thread_id;
358 struct cds_lfht *ht;
359 unsigned long i, start, len;
360 void (*fct)(struct cds_lfht *ht, unsigned long i,
361 unsigned long start, unsigned long len);
362 };
363
364 static struct urcu_workqueue *cds_lfht_workqueue;
365 static unsigned long cds_lfht_workqueue_user_count;
366
367 /*
368 * Mutex ensuring mutual exclusion between workqueue initialization and
369 * fork handlers. cds_lfht_fork_mutex nests inside call_rcu_mutex.
370 */
371 static pthread_mutex_t cds_lfht_fork_mutex = PTHREAD_MUTEX_INITIALIZER;
372
373 static struct urcu_atfork cds_lfht_atfork;
374
375 /*
376 * atfork handler nesting counters. Handle being registered to many urcu
377 * flavors, thus being possibly invoked more than once in the
378 * pthread_atfork list of callbacks.
379 */
380 static int cds_lfht_workqueue_atfork_nesting;
381
382 static void cds_lfht_init_worker(const struct rcu_flavor_struct *flavor);
383 static void cds_lfht_fini_worker(const struct rcu_flavor_struct *flavor);
384
385 #ifdef CONFIG_CDS_LFHT_ITER_DEBUG
386
387 static
388 void cds_lfht_iter_debug_set_ht(struct cds_lfht *ht, struct cds_lfht_iter *iter)
389 {
390 iter->lfht = ht;
391 }
392
393 #define cds_lfht_iter_debug_assert(...) assert(__VA_ARGS__)
394
395 #else
396
397 static
398 void cds_lfht_iter_debug_set_ht(struct cds_lfht *ht __attribute__((unused)),
399 struct cds_lfht_iter *iter __attribute__((unused)))
400 {
401 }
402
403 #define cds_lfht_iter_debug_assert(...)
404
405 #endif
406
407 /*
408 * Algorithm to reverse bits in a word by lookup table, extended to
409 * 64-bit words.
410 * Source:
411 * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable
412 * Originally from Public Domain.
413 */
414
415 static const uint8_t BitReverseTable256[256] =
416 {
417 #define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64
418 #define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16)
419 #define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 )
420 R6(0), R6(2), R6(1), R6(3)
421 };
422 #undef R2
423 #undef R4
424 #undef R6
425
426 static
427 uint8_t bit_reverse_u8(uint8_t v)
428 {
429 return BitReverseTable256[v];
430 }
431
432 #if (CAA_BITS_PER_LONG == 32)
433 static
434 uint32_t bit_reverse_u32(uint32_t v)
435 {
436 return ((uint32_t) bit_reverse_u8(v) << 24) |
437 ((uint32_t) bit_reverse_u8(v >> 8) << 16) |
438 ((uint32_t) bit_reverse_u8(v >> 16) << 8) |
439 ((uint32_t) bit_reverse_u8(v >> 24));
440 }
441 #else
442 static
443 uint64_t bit_reverse_u64(uint64_t v)
444 {
445 return ((uint64_t) bit_reverse_u8(v) << 56) |
446 ((uint64_t) bit_reverse_u8(v >> 8) << 48) |
447 ((uint64_t) bit_reverse_u8(v >> 16) << 40) |
448 ((uint64_t) bit_reverse_u8(v >> 24) << 32) |
449 ((uint64_t) bit_reverse_u8(v >> 32) << 24) |
450 ((uint64_t) bit_reverse_u8(v >> 40) << 16) |
451 ((uint64_t) bit_reverse_u8(v >> 48) << 8) |
452 ((uint64_t) bit_reverse_u8(v >> 56));
453 }
454 #endif
455
456 static
457 unsigned long bit_reverse_ulong(unsigned long v)
458 {
459 #if (CAA_BITS_PER_LONG == 32)
460 return bit_reverse_u32(v);
461 #else
462 return bit_reverse_u64(v);
463 #endif
464 }
465
466 /*
467 * fls: returns the position of the most significant bit.
468 * Returns 0 if no bit is set, else returns the position of the most
469 * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit).
470 */
471 #if defined(URCU_ARCH_X86)
472 static inline
473 unsigned int fls_u32(uint32_t x)
474 {
475 int r;
476
477 __asm__ ("bsrl %1,%0\n\t"
478 "jnz 1f\n\t"
479 "movl $-1,%0\n\t"
480 "1:\n\t"
481 : "=r" (r) : "rm" (x));
482 return r + 1;
483 }
484 #define HAS_FLS_U32
485 #endif
486
487 #if defined(URCU_ARCH_AMD64)
488 static inline
489 unsigned int fls_u64(uint64_t x)
490 {
491 long r;
492
493 __asm__ ("bsrq %1,%0\n\t"
494 "jnz 1f\n\t"
495 "movq $-1,%0\n\t"
496 "1:\n\t"
497 : "=r" (r) : "rm" (x));
498 return r + 1;
499 }
500 #define HAS_FLS_U64
501 #endif
502
503 #ifndef HAS_FLS_U64
504 static __attribute__((unused))
505 unsigned int fls_u64(uint64_t x)
506 {
507 unsigned int r = 64;
508
509 if (!x)
510 return 0;
511
512 if (!(x & 0xFFFFFFFF00000000ULL)) {
513 x <<= 32;
514 r -= 32;
515 }
516 if (!(x & 0xFFFF000000000000ULL)) {
517 x <<= 16;
518 r -= 16;
519 }
520 if (!(x & 0xFF00000000000000ULL)) {
521 x <<= 8;
522 r -= 8;
523 }
524 if (!(x & 0xF000000000000000ULL)) {
525 x <<= 4;
526 r -= 4;
527 }
528 if (!(x & 0xC000000000000000ULL)) {
529 x <<= 2;
530 r -= 2;
531 }
532 if (!(x & 0x8000000000000000ULL)) {
533 x <<= 1;
534 r -= 1;
535 }
536 return r;
537 }
538 #endif
539
540 #ifndef HAS_FLS_U32
541 static __attribute__((unused))
542 unsigned int fls_u32(uint32_t x)
543 {
544 unsigned int r = 32;
545
546 if (!x)
547 return 0;
548 if (!(x & 0xFFFF0000U)) {
549 x <<= 16;
550 r -= 16;
551 }
552 if (!(x & 0xFF000000U)) {
553 x <<= 8;
554 r -= 8;
555 }
556 if (!(x & 0xF0000000U)) {
557 x <<= 4;
558 r -= 4;
559 }
560 if (!(x & 0xC0000000U)) {
561 x <<= 2;
562 r -= 2;
563 }
564 if (!(x & 0x80000000U)) {
565 x <<= 1;
566 r -= 1;
567 }
568 return r;
569 }
570 #endif
571
572 unsigned int cds_lfht_fls_ulong(unsigned long x)
573 {
574 #if (CAA_BITS_PER_LONG == 32)
575 return fls_u32(x);
576 #else
577 return fls_u64(x);
578 #endif
579 }
580
581 /*
582 * Return the minimum order for which x <= (1UL << order).
583 * Return -1 if x is 0.
584 */
585 static
586 int cds_lfht_get_count_order_u32(uint32_t x)
587 {
588 if (!x)
589 return -1;
590
591 return fls_u32(x - 1);
592 }
593
594 /*
595 * Return the minimum order for which x <= (1UL << order).
596 * Return -1 if x is 0.
597 */
598 int cds_lfht_get_count_order_ulong(unsigned long x)
599 {
600 if (!x)
601 return -1;
602
603 return cds_lfht_fls_ulong(x - 1);
604 }
605
606 static
607 void cds_lfht_resize_lazy_grow(struct cds_lfht *ht, unsigned long size, int growth);
608
609 static
610 void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
611 unsigned long count);
612
613 static void mutex_lock(pthread_mutex_t *mutex)
614 {
615 int ret;
616
617 #ifndef DISTRUST_SIGNALS_EXTREME
618 ret = pthread_mutex_lock(mutex);
619 if (ret)
620 urcu_die(ret);
621 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
622 while ((ret = pthread_mutex_trylock(mutex)) != 0) {
623 if (ret != EBUSY && ret != EINTR)
624 urcu_die(ret);
625 if (CMM_LOAD_SHARED(URCU_TLS(rcu_reader).need_mb)) {
626 cmm_smp_mb();
627 _CMM_STORE_SHARED(URCU_TLS(rcu_reader).need_mb, 0);
628 cmm_smp_mb();
629 }
630 (void) poll(NULL, 0, 10);
631 }
632 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
633 }
634
635 static void mutex_unlock(pthread_mutex_t *mutex)
636 {
637 int ret;
638
639 ret = pthread_mutex_unlock(mutex);
640 if (ret)
641 urcu_die(ret);
642 }
643
644 static long nr_cpus_mask = -1;
645 static long split_count_mask = -1;
646 static int split_count_order = -1;
647
648 #if defined(HAVE_SYSCONF)
649 static void ht_init_nr_cpus_mask(void)
650 {
651 long maxcpus;
652
653 maxcpus = sysconf(_SC_NPROCESSORS_CONF);
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 #else /* #if defined(HAVE_SYSCONF) */
666 static void ht_init_nr_cpus_mask(void)
667 {
668 nr_cpus_mask = -2;
669 }
670 #endif /* #else #if defined(HAVE_SYSCONF) */
671
672 static
673 void alloc_split_items_count(struct cds_lfht *ht)
674 {
675 if (nr_cpus_mask == -1) {
676 ht_init_nr_cpus_mask();
677 if (nr_cpus_mask < 0)
678 split_count_mask = DEFAULT_SPLIT_COUNT_MASK;
679 else
680 split_count_mask = nr_cpus_mask;
681 split_count_order =
682 cds_lfht_get_count_order_ulong(split_count_mask + 1);
683 }
684
685 assert(split_count_mask >= 0);
686
687 if (ht->flags & CDS_LFHT_ACCOUNTING) {
688 ht->split_count = calloc(split_count_mask + 1,
689 sizeof(struct ht_items_count));
690 assert(ht->split_count);
691 } else {
692 ht->split_count = NULL;
693 }
694 }
695
696 static
697 void free_split_items_count(struct cds_lfht *ht)
698 {
699 poison_free(ht->split_count);
700 }
701
702 static
703 int ht_get_split_count_index(unsigned long hash)
704 {
705 int cpu;
706
707 assert(split_count_mask >= 0);
708 cpu = urcu_sched_getcpu();
709 if (caa_unlikely(cpu < 0))
710 return hash & split_count_mask;
711 else
712 return cpu & split_count_mask;
713 }
714
715 static
716 void ht_count_add(struct cds_lfht *ht, unsigned long size, unsigned long hash)
717 {
718 unsigned long split_count, count;
719 int index;
720
721 if (caa_unlikely(!ht->split_count))
722 return;
723 index = ht_get_split_count_index(hash);
724 split_count = uatomic_add_return(&ht->split_count[index].add, 1);
725 if (caa_likely(split_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))
726 return;
727 /* Only if number of add multiple of 1UL << COUNT_COMMIT_ORDER */
728
729 dbg_printf("add split count %lu\n", split_count);
730 count = uatomic_add_return(&ht->count,
731 1UL << COUNT_COMMIT_ORDER);
732 if (caa_likely(count & (count - 1)))
733 return;
734 /* Only if global count is power of 2 */
735
736 if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) < size)
737 return;
738 dbg_printf("add set global %lu\n", count);
739 cds_lfht_resize_lazy_count(ht, size,
740 count >> (CHAIN_LEN_TARGET - 1));
741 }
742
743 static
744 void ht_count_del(struct cds_lfht *ht, unsigned long size, unsigned long hash)
745 {
746 unsigned long split_count, count;
747 int index;
748
749 if (caa_unlikely(!ht->split_count))
750 return;
751 index = ht_get_split_count_index(hash);
752 split_count = uatomic_add_return(&ht->split_count[index].del, 1);
753 if (caa_likely(split_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))
754 return;
755 /* Only if number of deletes multiple of 1UL << COUNT_COMMIT_ORDER */
756
757 dbg_printf("del split count %lu\n", split_count);
758 count = uatomic_add_return(&ht->count,
759 -(1UL << COUNT_COMMIT_ORDER));
760 if (caa_likely(count & (count - 1)))
761 return;
762 /* Only if global count is power of 2 */
763
764 if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) >= size)
765 return;
766 dbg_printf("del set global %ld\n", count);
767 /*
768 * Don't shrink table if the number of nodes is below a
769 * certain threshold.
770 */
771 if (count < (1UL << COUNT_COMMIT_ORDER) * (split_count_mask + 1))
772 return;
773 cds_lfht_resize_lazy_count(ht, size,
774 count >> (CHAIN_LEN_TARGET - 1));
775 }
776
777 static
778 void check_resize(struct cds_lfht *ht, unsigned long size, uint32_t chain_len)
779 {
780 unsigned long count;
781
782 if (!(ht->flags & CDS_LFHT_AUTO_RESIZE))
783 return;
784 count = uatomic_read(&ht->count);
785 /*
786 * Use bucket-local length for small table expand and for
787 * environments lacking per-cpu data support.
788 */
789 if (count >= (1UL << (COUNT_COMMIT_ORDER + split_count_order)))
790 return;
791 if (chain_len > 100)
792 dbg_printf("WARNING: large chain length: %u.\n",
793 chain_len);
794 if (chain_len >= CHAIN_LEN_RESIZE_THRESHOLD) {
795 int growth;
796
797 /*
798 * Ideal growth calculated based on chain length.
799 */
800 growth = cds_lfht_get_count_order_u32(chain_len
801 - (CHAIN_LEN_TARGET - 1));
802 if ((ht->flags & CDS_LFHT_ACCOUNTING)
803 && (size << growth)
804 >= (1UL << (COUNT_COMMIT_ORDER
805 + split_count_order))) {
806 /*
807 * If ideal growth expands the hash table size
808 * beyond the "small hash table" sizes, use the
809 * maximum small hash table size to attempt
810 * expanding the hash table. This only applies
811 * when node accounting is available, otherwise
812 * the chain length is used to expand the hash
813 * table in every case.
814 */
815 growth = COUNT_COMMIT_ORDER + split_count_order
816 - cds_lfht_get_count_order_ulong(size);
817 if (growth <= 0)
818 return;
819 }
820 cds_lfht_resize_lazy_grow(ht, size, growth);
821 }
822 }
823
824 static
825 struct cds_lfht_node *clear_flag(struct cds_lfht_node *node)
826 {
827 return (struct cds_lfht_node *) (((unsigned long) node) & ~FLAGS_MASK);
828 }
829
830 static
831 int is_removed(const struct cds_lfht_node *node)
832 {
833 return ((unsigned long) node) & REMOVED_FLAG;
834 }
835
836 static
837 int is_bucket(struct cds_lfht_node *node)
838 {
839 return ((unsigned long) node) & BUCKET_FLAG;
840 }
841
842 static
843 struct cds_lfht_node *flag_bucket(struct cds_lfht_node *node)
844 {
845 return (struct cds_lfht_node *) (((unsigned long) node) | BUCKET_FLAG);
846 }
847
848 static
849 int is_removal_owner(struct cds_lfht_node *node)
850 {
851 return ((unsigned long) node) & REMOVAL_OWNER_FLAG;
852 }
853
854 static
855 struct 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
860 static
861 struct 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
866 static
867 struct cds_lfht_node *get_end(void)
868 {
869 return (struct cds_lfht_node *) END_VALUE;
870 }
871
872 static
873 int is_end(struct cds_lfht_node *node)
874 {
875 return clear_flag(node) == (struct cds_lfht_node *) END_VALUE;
876 }
877
878 static
879 unsigned 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
893 static
894 void 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 */
904 static
905 void cds_lfht_free_bucket_table(struct cds_lfht *ht, unsigned long order)
906 {
907 return ht->mm->free_bucket_table(ht, order);
908 }
909
910 static inline
911 struct cds_lfht_node *bucket_at(struct cds_lfht *ht, unsigned long index)
912 {
913 return ht->bucket_at(ht, index);
914 }
915
916 static inline
917 struct cds_lfht_node *lookup_bucket(struct cds_lfht *ht, unsigned long size,
918 unsigned long hash)
919 {
920 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 */
927 static
928 void _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 assert(!is_bucket(bucket));
933 assert(!is_removed(bucket));
934 assert(!is_removal_owner(bucket));
935 assert(!is_bucket(node));
936 assert(!is_removed(node));
937 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 assert(!is_removed(iter));
943 assert(!is_removal_owner(iter));
944 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 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 assert(!is_removed(iter));
964 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
973 static
974 int _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 assert(!is_removed(old_node));
985 assert(!is_removal_owner(old_node));
986 assert(!is_bucket(old_node));
987 assert(!is_removed(new_node));
988 assert(!is_removal_owner(new_node));
989 assert(!is_bucket(new_node));
990 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 assert(old_next == clear_flag(old_next));
1001 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 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 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 */
1047 static
1048 void _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 assert(!is_bucket(node));
1062 assert(!is_removed(node));
1063 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 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 assert(node != clear_flag(iter));
1129 assert(!is_removed(iter_prev));
1130 assert(!is_removal_owner(iter_prev));
1131 assert(!is_removed(iter));
1132 assert(!is_removal_owner(iter));
1133 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 assert(!is_removed(iter));
1152 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 }
1160 end:
1161 if (unique_ret) {
1162 unique_ret->node = return_node;
1163 /* unique_ret->next left unset, never used. */
1164 }
1165 }
1166
1167 static
1168 int _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 assert(!is_bucket(node));
1178 assert(!is_removed(node));
1179 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 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 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
1233 static
1234 void *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
1244 static
1245 void 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 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 assert(!ret);
1296 }
1297 for (thread = 0; thread < nr_threads; thread++) {
1298 ret = pthread_join(work[thread].thread_id, NULL);
1299 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;
1310 fallback:
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 */
1325 static
1326 void 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 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 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
1345 static
1346 void 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
1352 static
1353 void 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 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 */
1417 static
1418 void 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 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 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
1439 static
1440 void 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 */
1450 static
1451 void 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 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 */
1506 static
1507 void 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 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 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 */
1561 static
1562 const 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 */
1573 static
1574 const 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
1581 struct cds_lfht *_cds_lfht_new(unsigned long init_size,
1582 unsigned long min_nr_alloc_buckets,
1583 unsigned long max_nr_buckets,
1584 int flags,
1585 const struct cds_lfht_mm_type *mm,
1586 const struct rcu_flavor_struct *flavor,
1587 pthread_attr_t *attr)
1588 {
1589 struct cds_lfht *ht;
1590 unsigned long order;
1591
1592 /* min_nr_alloc_buckets must be power of two */
1593 if (!min_nr_alloc_buckets || (min_nr_alloc_buckets & (min_nr_alloc_buckets - 1)))
1594 return NULL;
1595
1596 /* init_size must be power of two */
1597 if (!init_size || (init_size & (init_size - 1)))
1598 return NULL;
1599
1600 /*
1601 * Memory management plugin default.
1602 */
1603 if (!mm)
1604 mm = get_mm_type(max_nr_buckets);
1605
1606 /* max_nr_buckets == 0 for order based mm means infinite */
1607 if (mm == &cds_lfht_mm_order && !max_nr_buckets)
1608 max_nr_buckets = 1UL << (MAX_TABLE_ORDER - 1);
1609
1610 /* max_nr_buckets must be power of two */
1611 if (!max_nr_buckets || (max_nr_buckets & (max_nr_buckets - 1)))
1612 return NULL;
1613
1614 if (flags & CDS_LFHT_AUTO_RESIZE)
1615 cds_lfht_init_worker(flavor);
1616
1617 min_nr_alloc_buckets = max(min_nr_alloc_buckets, MIN_TABLE_SIZE);
1618 init_size = max(init_size, MIN_TABLE_SIZE);
1619 max_nr_buckets = max(max_nr_buckets, min_nr_alloc_buckets);
1620 init_size = min(init_size, max_nr_buckets);
1621
1622 ht = mm->alloc_cds_lfht(min_nr_alloc_buckets, max_nr_buckets);
1623 assert(ht);
1624 assert(ht->mm == mm);
1625 assert(ht->bucket_at == mm->bucket_at);
1626
1627 ht->flags = flags;
1628 ht->flavor = flavor;
1629 ht->resize_attr = attr;
1630 alloc_split_items_count(ht);
1631 /* this mutex should not nest in read-side C.S. */
1632 pthread_mutex_init(&ht->resize_mutex, NULL);
1633 order = cds_lfht_get_count_order_ulong(init_size);
1634 ht->resize_target = 1UL << order;
1635 cds_lfht_create_bucket(ht, 1UL << order);
1636 ht->size = 1UL << order;
1637 return ht;
1638 }
1639
1640 void cds_lfht_lookup(struct cds_lfht *ht, unsigned long hash,
1641 cds_lfht_match_fct match, const void *key,
1642 struct cds_lfht_iter *iter)
1643 {
1644 struct cds_lfht_node *node, *next, *bucket;
1645 unsigned long reverse_hash, size;
1646
1647 cds_lfht_iter_debug_set_ht(ht, iter);
1648
1649 reverse_hash = bit_reverse_ulong(hash);
1650
1651 size = rcu_dereference(ht->size);
1652 bucket = lookup_bucket(ht, size, hash);
1653 /* We can always skip the bucket node initially */
1654 node = rcu_dereference(bucket->next);
1655 node = clear_flag(node);
1656 for (;;) {
1657 if (caa_unlikely(is_end(node))) {
1658 node = next = NULL;
1659 break;
1660 }
1661 if (caa_unlikely(node->reverse_hash > reverse_hash)) {
1662 node = next = NULL;
1663 break;
1664 }
1665 next = rcu_dereference(node->next);
1666 assert(node == clear_flag(node));
1667 if (caa_likely(!is_removed(next))
1668 && !is_bucket(next)
1669 && node->reverse_hash == reverse_hash
1670 && caa_likely(match(node, key))) {
1671 break;
1672 }
1673 node = clear_flag(next);
1674 }
1675 assert(!node || !is_bucket(CMM_LOAD_SHARED(node->next)));
1676 iter->node = node;
1677 iter->next = next;
1678 }
1679
1680 void cds_lfht_next_duplicate(struct cds_lfht *ht __attribute__((unused)),
1681 cds_lfht_match_fct match,
1682 const void *key, struct cds_lfht_iter *iter)
1683 {
1684 struct cds_lfht_node *node, *next;
1685 unsigned long reverse_hash;
1686
1687 cds_lfht_iter_debug_assert(ht == iter->lfht);
1688 node = iter->node;
1689 reverse_hash = node->reverse_hash;
1690 next = iter->next;
1691 node = clear_flag(next);
1692
1693 for (;;) {
1694 if (caa_unlikely(is_end(node))) {
1695 node = next = NULL;
1696 break;
1697 }
1698 if (caa_unlikely(node->reverse_hash > reverse_hash)) {
1699 node = next = NULL;
1700 break;
1701 }
1702 next = rcu_dereference(node->next);
1703 if (caa_likely(!is_removed(next))
1704 && !is_bucket(next)
1705 && caa_likely(match(node, key))) {
1706 break;
1707 }
1708 node = clear_flag(next);
1709 }
1710 assert(!node || !is_bucket(CMM_LOAD_SHARED(node->next)));
1711 iter->node = node;
1712 iter->next = next;
1713 }
1714
1715 void cds_lfht_next(struct cds_lfht *ht __attribute__((unused)),
1716 struct cds_lfht_iter *iter)
1717 {
1718 struct cds_lfht_node *node, *next;
1719
1720 cds_lfht_iter_debug_assert(ht == iter->lfht);
1721 node = clear_flag(iter->next);
1722 for (;;) {
1723 if (caa_unlikely(is_end(node))) {
1724 node = next = NULL;
1725 break;
1726 }
1727 next = rcu_dereference(node->next);
1728 if (caa_likely(!is_removed(next))
1729 && !is_bucket(next)) {
1730 break;
1731 }
1732 node = clear_flag(next);
1733 }
1734 assert(!node || !is_bucket(CMM_LOAD_SHARED(node->next)));
1735 iter->node = node;
1736 iter->next = next;
1737 }
1738
1739 void cds_lfht_first(struct cds_lfht *ht, struct cds_lfht_iter *iter)
1740 {
1741 cds_lfht_iter_debug_set_ht(ht, iter);
1742 /*
1743 * Get next after first bucket node. The first bucket node is the
1744 * first node of the linked list.
1745 */
1746 iter->next = bucket_at(ht, 0)->next;
1747 cds_lfht_next(ht, iter);
1748 }
1749
1750 void cds_lfht_add(struct cds_lfht *ht, unsigned long hash,
1751 struct cds_lfht_node *node)
1752 {
1753 unsigned long size;
1754
1755 node->reverse_hash = bit_reverse_ulong(hash);
1756 size = rcu_dereference(ht->size);
1757 _cds_lfht_add(ht, hash, NULL, NULL, size, node, NULL, 0);
1758 ht_count_add(ht, size, hash);
1759 }
1760
1761 struct cds_lfht_node *cds_lfht_add_unique(struct cds_lfht *ht,
1762 unsigned long hash,
1763 cds_lfht_match_fct match,
1764 const void *key,
1765 struct cds_lfht_node *node)
1766 {
1767 unsigned long size;
1768 struct cds_lfht_iter iter;
1769
1770 node->reverse_hash = bit_reverse_ulong(hash);
1771 size = rcu_dereference(ht->size);
1772 _cds_lfht_add(ht, hash, match, key, size, node, &iter, 0);
1773 if (iter.node == node)
1774 ht_count_add(ht, size, hash);
1775 return iter.node;
1776 }
1777
1778 struct cds_lfht_node *cds_lfht_add_replace(struct cds_lfht *ht,
1779 unsigned long hash,
1780 cds_lfht_match_fct match,
1781 const void *key,
1782 struct cds_lfht_node *node)
1783 {
1784 unsigned long size;
1785 struct cds_lfht_iter iter;
1786
1787 node->reverse_hash = bit_reverse_ulong(hash);
1788 size = rcu_dereference(ht->size);
1789 for (;;) {
1790 _cds_lfht_add(ht, hash, match, key, size, node, &iter, 0);
1791 if (iter.node == node) {
1792 ht_count_add(ht, size, hash);
1793 return NULL;
1794 }
1795
1796 if (!_cds_lfht_replace(ht, size, iter.node, iter.next, node))
1797 return iter.node;
1798 }
1799 }
1800
1801 int cds_lfht_replace(struct cds_lfht *ht,
1802 struct cds_lfht_iter *old_iter,
1803 unsigned long hash,
1804 cds_lfht_match_fct match,
1805 const void *key,
1806 struct cds_lfht_node *new_node)
1807 {
1808 unsigned long size;
1809
1810 new_node->reverse_hash = bit_reverse_ulong(hash);
1811 if (!old_iter->node)
1812 return -ENOENT;
1813 if (caa_unlikely(old_iter->node->reverse_hash != new_node->reverse_hash))
1814 return -EINVAL;
1815 if (caa_unlikely(!match(old_iter->node, key)))
1816 return -EINVAL;
1817 size = rcu_dereference(ht->size);
1818 return _cds_lfht_replace(ht, size, old_iter->node, old_iter->next,
1819 new_node);
1820 }
1821
1822 int cds_lfht_del(struct cds_lfht *ht, struct cds_lfht_node *node)
1823 {
1824 unsigned long size;
1825 int ret;
1826
1827 size = rcu_dereference(ht->size);
1828 ret = _cds_lfht_del(ht, size, node);
1829 if (!ret) {
1830 unsigned long hash;
1831
1832 hash = bit_reverse_ulong(node->reverse_hash);
1833 ht_count_del(ht, size, hash);
1834 }
1835 return ret;
1836 }
1837
1838 int cds_lfht_is_node_deleted(const struct cds_lfht_node *node)
1839 {
1840 return is_removed(CMM_LOAD_SHARED(node->next));
1841 }
1842
1843 static
1844 int cds_lfht_delete_bucket(struct cds_lfht *ht)
1845 {
1846 struct cds_lfht_node *node;
1847 unsigned long order, i, size;
1848
1849 /* Check that the table is empty */
1850 node = bucket_at(ht, 0);
1851 do {
1852 node = clear_flag(node)->next;
1853 if (!is_bucket(node))
1854 return -EPERM;
1855 assert(!is_removed(node));
1856 assert(!is_removal_owner(node));
1857 } while (!is_end(node));
1858 /*
1859 * size accessed without rcu_dereference because hash table is
1860 * being destroyed.
1861 */
1862 size = ht->size;
1863 /* Internal sanity check: all nodes left should be buckets */
1864 for (i = 0; i < size; i++) {
1865 node = bucket_at(ht, i);
1866 dbg_printf("delete bucket: index %lu expected hash %lu hash %lu\n",
1867 i, i, bit_reverse_ulong(node->reverse_hash));
1868 assert(is_bucket(node->next));
1869 }
1870
1871 for (order = cds_lfht_get_count_order_ulong(size); (long)order >= 0; order--)
1872 cds_lfht_free_bucket_table(ht, order);
1873
1874 return 0;
1875 }
1876
1877 /*
1878 * Should only be called when no more concurrent readers nor writers can
1879 * possibly access the table.
1880 */
1881 int cds_lfht_destroy(struct cds_lfht *ht, pthread_attr_t **attr)
1882 {
1883 int ret;
1884
1885 if (ht->flags & CDS_LFHT_AUTO_RESIZE) {
1886 /* Cancel ongoing resize operations. */
1887 _CMM_STORE_SHARED(ht->in_progress_destroy, 1);
1888 /* Wait for in-flight resize operations to complete */
1889 urcu_workqueue_flush_queued_work(cds_lfht_workqueue);
1890 }
1891 ret = cds_lfht_delete_bucket(ht);
1892 if (ret)
1893 return ret;
1894 free_split_items_count(ht);
1895 if (attr)
1896 *attr = ht->resize_attr;
1897 ret = pthread_mutex_destroy(&ht->resize_mutex);
1898 if (ret)
1899 ret = -EBUSY;
1900 if (ht->flags & CDS_LFHT_AUTO_RESIZE)
1901 cds_lfht_fini_worker(ht->flavor);
1902 poison_free(ht);
1903 return ret;
1904 }
1905
1906 void cds_lfht_count_nodes(struct cds_lfht *ht,
1907 long *approx_before,
1908 unsigned long *count,
1909 long *approx_after)
1910 {
1911 struct cds_lfht_node *node, *next;
1912 unsigned long nr_bucket = 0, nr_removed = 0;
1913
1914 *approx_before = 0;
1915 if (ht->split_count) {
1916 int i;
1917
1918 for (i = 0; i < split_count_mask + 1; i++) {
1919 *approx_before += uatomic_read(&ht->split_count[i].add);
1920 *approx_before -= uatomic_read(&ht->split_count[i].del);
1921 }
1922 }
1923
1924 *count = 0;
1925
1926 /* Count non-bucket nodes in the table */
1927 node = bucket_at(ht, 0);
1928 do {
1929 next = rcu_dereference(node->next);
1930 if (is_removed(next)) {
1931 if (!is_bucket(next))
1932 (nr_removed)++;
1933 else
1934 (nr_bucket)++;
1935 } else if (!is_bucket(next))
1936 (*count)++;
1937 else
1938 (nr_bucket)++;
1939 node = clear_flag(next);
1940 } while (!is_end(node));
1941 dbg_printf("number of logically removed nodes: %lu\n", nr_removed);
1942 dbg_printf("number of bucket nodes: %lu\n", nr_bucket);
1943 *approx_after = 0;
1944 if (ht->split_count) {
1945 int i;
1946
1947 for (i = 0; i < split_count_mask + 1; i++) {
1948 *approx_after += uatomic_read(&ht->split_count[i].add);
1949 *approx_after -= uatomic_read(&ht->split_count[i].del);
1950 }
1951 }
1952 }
1953
1954 /* called with resize mutex held */
1955 static
1956 void _do_cds_lfht_grow(struct cds_lfht *ht,
1957 unsigned long old_size, unsigned long new_size)
1958 {
1959 unsigned long old_order, new_order;
1960
1961 old_order = cds_lfht_get_count_order_ulong(old_size);
1962 new_order = cds_lfht_get_count_order_ulong(new_size);
1963 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1964 old_size, old_order, new_size, new_order);
1965 assert(new_size > old_size);
1966 init_table(ht, old_order + 1, new_order);
1967 }
1968
1969 /* called with resize mutex held */
1970 static
1971 void _do_cds_lfht_shrink(struct cds_lfht *ht,
1972 unsigned long old_size, unsigned long new_size)
1973 {
1974 unsigned long old_order, new_order;
1975
1976 new_size = max(new_size, MIN_TABLE_SIZE);
1977 old_order = cds_lfht_get_count_order_ulong(old_size);
1978 new_order = cds_lfht_get_count_order_ulong(new_size);
1979 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1980 old_size, old_order, new_size, new_order);
1981 assert(new_size < old_size);
1982
1983 /* Remove and unlink all bucket nodes to remove. */
1984 fini_table(ht, new_order + 1, old_order);
1985 }
1986
1987
1988 /* called with resize mutex held */
1989 static
1990 void _do_cds_lfht_resize(struct cds_lfht *ht)
1991 {
1992 unsigned long new_size, old_size;
1993
1994 /*
1995 * Resize table, re-do if the target size has changed under us.
1996 */
1997 do {
1998 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1999 break;
2000 ht->resize_initiated = 1;
2001 old_size = ht->size;
2002 new_size = CMM_LOAD_SHARED(ht->resize_target);
2003 if (old_size < new_size)
2004 _do_cds_lfht_grow(ht, old_size, new_size);
2005 else if (old_size > new_size)
2006 _do_cds_lfht_shrink(ht, old_size, new_size);
2007 ht->resize_initiated = 0;
2008 /* write resize_initiated before read resize_target */
2009 cmm_smp_mb();
2010 } while (ht->size != CMM_LOAD_SHARED(ht->resize_target));
2011 }
2012
2013 static
2014 unsigned long resize_target_grow(struct cds_lfht *ht, unsigned long new_size)
2015 {
2016 return _uatomic_xchg_monotonic_increase(&ht->resize_target, new_size);
2017 }
2018
2019 static
2020 void resize_target_update_count(struct cds_lfht *ht,
2021 unsigned long count)
2022 {
2023 count = max(count, MIN_TABLE_SIZE);
2024 count = min(count, ht->max_nr_buckets);
2025 uatomic_set(&ht->resize_target, count);
2026 }
2027
2028 void cds_lfht_resize(struct cds_lfht *ht, unsigned long new_size)
2029 {
2030 resize_target_update_count(ht, new_size);
2031 CMM_STORE_SHARED(ht->resize_initiated, 1);
2032 mutex_lock(&ht->resize_mutex);
2033 _do_cds_lfht_resize(ht);
2034 mutex_unlock(&ht->resize_mutex);
2035 }
2036
2037 static
2038 void do_resize_cb(struct urcu_work *work)
2039 {
2040 struct resize_work *resize_work =
2041 caa_container_of(work, struct resize_work, work);
2042 struct cds_lfht *ht = resize_work->ht;
2043
2044 ht->flavor->register_thread();
2045 mutex_lock(&ht->resize_mutex);
2046 _do_cds_lfht_resize(ht);
2047 mutex_unlock(&ht->resize_mutex);
2048 ht->flavor->unregister_thread();
2049 poison_free(work);
2050 }
2051
2052 static
2053 void __cds_lfht_resize_lazy_launch(struct cds_lfht *ht)
2054 {
2055 struct resize_work *work;
2056
2057 /* Store resize_target before read resize_initiated */
2058 cmm_smp_mb();
2059 if (!CMM_LOAD_SHARED(ht->resize_initiated)) {
2060 if (CMM_LOAD_SHARED(ht->in_progress_destroy)) {
2061 return;
2062 }
2063 work = malloc(sizeof(*work));
2064 if (work == NULL) {
2065 dbg_printf("error allocating resize work, bailing out\n");
2066 return;
2067 }
2068 work->ht = ht;
2069 urcu_workqueue_queue_work(cds_lfht_workqueue,
2070 &work->work, do_resize_cb);
2071 CMM_STORE_SHARED(ht->resize_initiated, 1);
2072 }
2073 }
2074
2075 static
2076 void cds_lfht_resize_lazy_grow(struct cds_lfht *ht, unsigned long size, int growth)
2077 {
2078 unsigned long target_size = size << growth;
2079
2080 target_size = min(target_size, ht->max_nr_buckets);
2081 if (resize_target_grow(ht, target_size) >= target_size)
2082 return;
2083
2084 __cds_lfht_resize_lazy_launch(ht);
2085 }
2086
2087 /*
2088 * We favor grow operations over shrink. A shrink operation never occurs
2089 * if a grow operation is queued for lazy execution. A grow operation
2090 * cancels any pending shrink lazy execution.
2091 */
2092 static
2093 void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
2094 unsigned long count)
2095 {
2096 if (!(ht->flags & CDS_LFHT_AUTO_RESIZE))
2097 return;
2098 count = max(count, MIN_TABLE_SIZE);
2099 count = min(count, ht->max_nr_buckets);
2100 if (count == size)
2101 return; /* Already the right size, no resize needed */
2102 if (count > size) { /* lazy grow */
2103 if (resize_target_grow(ht, count) >= count)
2104 return;
2105 } else { /* lazy shrink */
2106 for (;;) {
2107 unsigned long s;
2108
2109 s = uatomic_cmpxchg(&ht->resize_target, size, count);
2110 if (s == size)
2111 break; /* no resize needed */
2112 if (s > size)
2113 return; /* growing is/(was just) in progress */
2114 if (s <= count)
2115 return; /* some other thread do shrink */
2116 size = s;
2117 }
2118 }
2119 __cds_lfht_resize_lazy_launch(ht);
2120 }
2121
2122 static void cds_lfht_before_fork(void *priv __attribute__((unused)))
2123 {
2124 if (cds_lfht_workqueue_atfork_nesting++)
2125 return;
2126 mutex_lock(&cds_lfht_fork_mutex);
2127 if (!cds_lfht_workqueue)
2128 return;
2129 urcu_workqueue_pause_worker(cds_lfht_workqueue);
2130 }
2131
2132 static void cds_lfht_after_fork_parent(void *priv __attribute__((unused)))
2133 {
2134 if (--cds_lfht_workqueue_atfork_nesting)
2135 return;
2136 if (!cds_lfht_workqueue)
2137 goto end;
2138 urcu_workqueue_resume_worker(cds_lfht_workqueue);
2139 end:
2140 mutex_unlock(&cds_lfht_fork_mutex);
2141 }
2142
2143 static void cds_lfht_after_fork_child(void *priv __attribute__((unused)))
2144 {
2145 if (--cds_lfht_workqueue_atfork_nesting)
2146 return;
2147 if (!cds_lfht_workqueue)
2148 goto end;
2149 urcu_workqueue_create_worker(cds_lfht_workqueue);
2150 end:
2151 mutex_unlock(&cds_lfht_fork_mutex);
2152 }
2153
2154 static struct urcu_atfork cds_lfht_atfork = {
2155 .before_fork = cds_lfht_before_fork,
2156 .after_fork_parent = cds_lfht_after_fork_parent,
2157 .after_fork_child = cds_lfht_after_fork_child,
2158 };
2159
2160 /*
2161 * Block all signals for the workqueue worker thread to ensure we don't
2162 * disturb the application. The SIGRCU signal needs to be unblocked for
2163 * the urcu-signal flavor.
2164 */
2165 static void cds_lfht_worker_init(
2166 struct urcu_workqueue *workqueue __attribute__((unused)),
2167 void *priv __attribute__((unused)))
2168 {
2169 int ret;
2170 sigset_t mask;
2171
2172 ret = sigfillset(&mask);
2173 if (ret)
2174 urcu_die(errno);
2175 ret = sigdelset(&mask, SIGRCU);
2176 if (ret)
2177 urcu_die(errno);
2178 ret = pthread_sigmask(SIG_SETMASK, &mask, NULL);
2179 if (ret)
2180 urcu_die(ret);
2181 }
2182
2183 static void cds_lfht_init_worker(const struct rcu_flavor_struct *flavor)
2184 {
2185 flavor->register_rculfhash_atfork(&cds_lfht_atfork);
2186
2187 mutex_lock(&cds_lfht_fork_mutex);
2188 if (cds_lfht_workqueue_user_count++)
2189 goto end;
2190 cds_lfht_workqueue = urcu_workqueue_create(0, -1, NULL,
2191 NULL, cds_lfht_worker_init, NULL, NULL, NULL, NULL, NULL);
2192 end:
2193 mutex_unlock(&cds_lfht_fork_mutex);
2194 }
2195
2196 static void cds_lfht_fini_worker(const struct rcu_flavor_struct *flavor)
2197 {
2198 mutex_lock(&cds_lfht_fork_mutex);
2199 if (--cds_lfht_workqueue_user_count)
2200 goto end;
2201 urcu_workqueue_destroy(cds_lfht_workqueue);
2202 cds_lfht_workqueue = NULL;
2203 end:
2204 mutex_unlock(&cds_lfht_fork_mutex);
2205
2206 flavor->unregister_rculfhash_atfork(&cds_lfht_atfork);
2207 }
This page took 0.127089 seconds and 3 git commands to generate.