Cleanup: enable signed/unsigned compare compiler warning
[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 <rculfhash-internal.h>
277 #include <stdio.h>
278 #include <pthread.h>
279 #include <signal.h>
280 #include "workqueue.h"
281 #include "urcu-die.h"
282 #include "urcu-utils.h"
283
284 /*
285 * Split-counters lazily update the global counter each 1024
286 * addition/removal. It automatically keeps track of resize required.
287 * We use the bucket length as indicator for need to expand for small
288 * tables and machines lacking per-cpu data support.
289 */
290 #define COUNT_COMMIT_ORDER 10
291 #define DEFAULT_SPLIT_COUNT_MASK 0xFUL
292 #define CHAIN_LEN_TARGET 1
293 #define CHAIN_LEN_RESIZE_THRESHOLD 3
294
295 /*
296 * Define the minimum table size.
297 */
298 #define MIN_TABLE_ORDER 0
299 #define MIN_TABLE_SIZE (1UL << MIN_TABLE_ORDER)
300
301 /*
302 * Minimum number of bucket nodes to touch per thread to parallelize grow/shrink.
303 */
304 #define MIN_PARTITION_PER_THREAD_ORDER 12
305 #define MIN_PARTITION_PER_THREAD (1UL << MIN_PARTITION_PER_THREAD_ORDER)
306
307 /*
308 * The removed flag needs to be updated atomically with the pointer.
309 * It indicates that no node must attach to the node scheduled for
310 * removal, and that node garbage collection must be performed.
311 * The bucket flag does not require to be updated atomically with the
312 * pointer, but it is added as a pointer low bit flag to save space.
313 * The "removal owner" flag is used to detect which of the "del"
314 * operation that has set the "removed flag" gets to return the removed
315 * node to its caller. Note that the replace operation does not need to
316 * iteract with the "removal owner" flag, because it validates that
317 * the "removed" flag is not set before performing its cmpxchg.
318 */
319 #define REMOVED_FLAG (1UL << 0)
320 #define BUCKET_FLAG (1UL << 1)
321 #define REMOVAL_OWNER_FLAG (1UL << 2)
322 #define FLAGS_MASK ((1UL << 3) - 1)
323
324 /* Value of the end pointer. Should not interact with flags. */
325 #define END_VALUE NULL
326
327 /*
328 * ht_items_count: Split-counters counting the number of node addition
329 * and removal in the table. Only used if the CDS_LFHT_ACCOUNTING flag
330 * is set at hash table creation.
331 *
332 * These are free-running counters, never reset to zero. They count the
333 * number of add/remove, and trigger every (1 << COUNT_COMMIT_ORDER)
334 * operations to update the global counter. We choose a power-of-2 value
335 * for the trigger to deal with 32 or 64-bit overflow of the counter.
336 */
337 struct ht_items_count {
338 unsigned long add, del;
339 } __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
340
341 /*
342 * resize_work: Contains arguments passed to worker thread
343 * responsible for performing lazy resize.
344 */
345 struct resize_work {
346 struct urcu_work work;
347 struct cds_lfht *ht;
348 };
349
350 /*
351 * partition_resize_work: Contains arguments passed to worker threads
352 * executing the hash table resize on partitions of the hash table
353 * assigned to each processor's worker thread.
354 */
355 struct partition_resize_work {
356 pthread_t thread_id;
357 struct cds_lfht *ht;
358 unsigned long i, start, len;
359 void (*fct)(struct cds_lfht *ht, unsigned long i,
360 unsigned long start, unsigned long len);
361 };
362
363 static struct urcu_workqueue *cds_lfht_workqueue;
364 static unsigned long cds_lfht_workqueue_user_count;
365
366 /*
367 * Mutex ensuring mutual exclusion between workqueue initialization and
368 * fork handlers. cds_lfht_fork_mutex nests inside call_rcu_mutex.
369 */
370 static pthread_mutex_t cds_lfht_fork_mutex = PTHREAD_MUTEX_INITIALIZER;
371
372 static struct urcu_atfork cds_lfht_atfork;
373
374 /*
375 * atfork handler nesting counters. Handle being registered to many urcu
376 * flavors, thus being possibly invoked more than once in the
377 * pthread_atfork list of callbacks.
378 */
379 static int cds_lfht_workqueue_atfork_nesting;
380
381 static void cds_lfht_init_worker(const struct rcu_flavor_struct *flavor);
382 static void cds_lfht_fini_worker(const struct rcu_flavor_struct *flavor);
383
384 #ifdef CONFIG_CDS_LFHT_ITER_DEBUG
385
386 static
387 void cds_lfht_iter_debug_set_ht(struct cds_lfht *ht, struct cds_lfht_iter *iter)
388 {
389 iter->lfht = ht;
390 }
391
392 #define cds_lfht_iter_debug_assert(...) assert(__VA_ARGS__)
393
394 #else
395
396 static
397 void cds_lfht_iter_debug_set_ht(struct cds_lfht *ht, struct cds_lfht_iter *iter)
398 {
399 }
400
401 #define cds_lfht_iter_debug_assert(...)
402
403 #endif
404
405 /*
406 * Algorithm to reverse bits in a word by lookup table, extended to
407 * 64-bit words.
408 * Source:
409 * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable
410 * Originally from Public Domain.
411 */
412
413 static const uint8_t BitReverseTable256[256] =
414 {
415 #define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64
416 #define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16)
417 #define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 )
418 R6(0), R6(2), R6(1), R6(3)
419 };
420 #undef R2
421 #undef R4
422 #undef R6
423
424 static
425 uint8_t bit_reverse_u8(uint8_t v)
426 {
427 return BitReverseTable256[v];
428 }
429
430 #if (CAA_BITS_PER_LONG == 32)
431 static
432 uint32_t bit_reverse_u32(uint32_t v)
433 {
434 return ((uint32_t) bit_reverse_u8(v) << 24) |
435 ((uint32_t) bit_reverse_u8(v >> 8) << 16) |
436 ((uint32_t) bit_reverse_u8(v >> 16) << 8) |
437 ((uint32_t) bit_reverse_u8(v >> 24));
438 }
439 #else
440 static
441 uint64_t bit_reverse_u64(uint64_t v)
442 {
443 return ((uint64_t) bit_reverse_u8(v) << 56) |
444 ((uint64_t) bit_reverse_u8(v >> 8) << 48) |
445 ((uint64_t) bit_reverse_u8(v >> 16) << 40) |
446 ((uint64_t) bit_reverse_u8(v >> 24) << 32) |
447 ((uint64_t) bit_reverse_u8(v >> 32) << 24) |
448 ((uint64_t) bit_reverse_u8(v >> 40) << 16) |
449 ((uint64_t) bit_reverse_u8(v >> 48) << 8) |
450 ((uint64_t) bit_reverse_u8(v >> 56));
451 }
452 #endif
453
454 static
455 unsigned long bit_reverse_ulong(unsigned long v)
456 {
457 #if (CAA_BITS_PER_LONG == 32)
458 return bit_reverse_u32(v);
459 #else
460 return bit_reverse_u64(v);
461 #endif
462 }
463
464 /*
465 * fls: returns the position of the most significant bit.
466 * Returns 0 if no bit is set, else returns the position of the most
467 * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit).
468 */
469 #if defined(__i386) || defined(__x86_64)
470 static inline
471 unsigned int fls_u32(uint32_t x)
472 {
473 int r;
474
475 __asm__ ("bsrl %1,%0\n\t"
476 "jnz 1f\n\t"
477 "movl $-1,%0\n\t"
478 "1:\n\t"
479 : "=r" (r) : "rm" (x));
480 return r + 1;
481 }
482 #define HAS_FLS_U32
483 #endif
484
485 #if defined(__x86_64)
486 static inline
487 unsigned int fls_u64(uint64_t x)
488 {
489 long r;
490
491 __asm__ ("bsrq %1,%0\n\t"
492 "jnz 1f\n\t"
493 "movq $-1,%0\n\t"
494 "1:\n\t"
495 : "=r" (r) : "rm" (x));
496 return r + 1;
497 }
498 #define HAS_FLS_U64
499 #endif
500
501 #ifndef HAS_FLS_U64
502 static __attribute__((unused))
503 unsigned int fls_u64(uint64_t x)
504 {
505 unsigned int r = 64;
506
507 if (!x)
508 return 0;
509
510 if (!(x & 0xFFFFFFFF00000000ULL)) {
511 x <<= 32;
512 r -= 32;
513 }
514 if (!(x & 0xFFFF000000000000ULL)) {
515 x <<= 16;
516 r -= 16;
517 }
518 if (!(x & 0xFF00000000000000ULL)) {
519 x <<= 8;
520 r -= 8;
521 }
522 if (!(x & 0xF000000000000000ULL)) {
523 x <<= 4;
524 r -= 4;
525 }
526 if (!(x & 0xC000000000000000ULL)) {
527 x <<= 2;
528 r -= 2;
529 }
530 if (!(x & 0x8000000000000000ULL)) {
531 x <<= 1;
532 r -= 1;
533 }
534 return r;
535 }
536 #endif
537
538 #ifndef HAS_FLS_U32
539 static __attribute__((unused))
540 unsigned int fls_u32(uint32_t x)
541 {
542 unsigned int r = 32;
543
544 if (!x)
545 return 0;
546 if (!(x & 0xFFFF0000U)) {
547 x <<= 16;
548 r -= 16;
549 }
550 if (!(x & 0xFF000000U)) {
551 x <<= 8;
552 r -= 8;
553 }
554 if (!(x & 0xF0000000U)) {
555 x <<= 4;
556 r -= 4;
557 }
558 if (!(x & 0xC0000000U)) {
559 x <<= 2;
560 r -= 2;
561 }
562 if (!(x & 0x80000000U)) {
563 x <<= 1;
564 r -= 1;
565 }
566 return r;
567 }
568 #endif
569
570 unsigned int cds_lfht_fls_ulong(unsigned long x)
571 {
572 #if (CAA_BITS_PER_LONG == 32)
573 return fls_u32(x);
574 #else
575 return fls_u64(x);
576 #endif
577 }
578
579 /*
580 * Return the minimum order for which x <= (1UL << order).
581 * Return -1 if x is 0.
582 */
583 int cds_lfht_get_count_order_u32(uint32_t x)
584 {
585 if (!x)
586 return -1;
587
588 return fls_u32(x - 1);
589 }
590
591 /*
592 * Return the minimum order for which x <= (1UL << order).
593 * Return -1 if x is 0.
594 */
595 int cds_lfht_get_count_order_ulong(unsigned long x)
596 {
597 if (!x)
598 return -1;
599
600 return cds_lfht_fls_ulong(x - 1);
601 }
602
603 static
604 void cds_lfht_resize_lazy_grow(struct cds_lfht *ht, unsigned long size, int growth);
605
606 static
607 void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
608 unsigned long count);
609
610 static void mutex_lock(pthread_mutex_t *mutex)
611 {
612 int ret;
613
614 #ifndef DISTRUST_SIGNALS_EXTREME
615 ret = pthread_mutex_lock(mutex);
616 if (ret)
617 urcu_die(ret);
618 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
619 while ((ret = pthread_mutex_trylock(mutex)) != 0) {
620 if (ret != EBUSY && ret != EINTR)
621 urcu_die(ret);
622 if (CMM_LOAD_SHARED(URCU_TLS(rcu_reader).need_mb)) {
623 cmm_smp_mb();
624 _CMM_STORE_SHARED(URCU_TLS(rcu_reader).need_mb, 0);
625 cmm_smp_mb();
626 }
627 (void) poll(NULL, 0, 10);
628 }
629 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
630 }
631
632 static void mutex_unlock(pthread_mutex_t *mutex)
633 {
634 int ret;
635
636 ret = pthread_mutex_unlock(mutex);
637 if (ret)
638 urcu_die(ret);
639 }
640
641 static long nr_cpus_mask = -1;
642 static long split_count_mask = -1;
643 static int split_count_order = -1;
644
645 #if defined(HAVE_SYSCONF)
646 static void ht_init_nr_cpus_mask(void)
647 {
648 long maxcpus;
649
650 maxcpus = sysconf(_SC_NPROCESSORS_CONF);
651 if (maxcpus <= 0) {
652 nr_cpus_mask = -2;
653 return;
654 }
655 /*
656 * round up number of CPUs to next power of two, so we
657 * can use & for modulo.
658 */
659 maxcpus = 1UL << cds_lfht_get_count_order_ulong(maxcpus);
660 nr_cpus_mask = maxcpus - 1;
661 }
662 #else /* #if defined(HAVE_SYSCONF) */
663 static void ht_init_nr_cpus_mask(void)
664 {
665 nr_cpus_mask = -2;
666 }
667 #endif /* #else #if defined(HAVE_SYSCONF) */
668
669 static
670 void alloc_split_items_count(struct cds_lfht *ht)
671 {
672 if (nr_cpus_mask == -1) {
673 ht_init_nr_cpus_mask();
674 if (nr_cpus_mask < 0)
675 split_count_mask = DEFAULT_SPLIT_COUNT_MASK;
676 else
677 split_count_mask = nr_cpus_mask;
678 split_count_order =
679 cds_lfht_get_count_order_ulong(split_count_mask + 1);
680 }
681
682 assert(split_count_mask >= 0);
683
684 if (ht->flags & CDS_LFHT_ACCOUNTING) {
685 ht->split_count = calloc(split_count_mask + 1,
686 sizeof(struct ht_items_count));
687 assert(ht->split_count);
688 } else {
689 ht->split_count = NULL;
690 }
691 }
692
693 static
694 void free_split_items_count(struct cds_lfht *ht)
695 {
696 poison_free(ht->split_count);
697 }
698
699 static
700 int ht_get_split_count_index(unsigned long hash)
701 {
702 int cpu;
703
704 assert(split_count_mask >= 0);
705 cpu = urcu_sched_getcpu();
706 if (caa_unlikely(cpu < 0))
707 return hash & split_count_mask;
708 else
709 return cpu & split_count_mask;
710 }
711
712 static
713 void ht_count_add(struct cds_lfht *ht, unsigned long size, unsigned long hash)
714 {
715 unsigned long split_count, count;
716 int index;
717
718 if (caa_unlikely(!ht->split_count))
719 return;
720 index = ht_get_split_count_index(hash);
721 split_count = uatomic_add_return(&ht->split_count[index].add, 1);
722 if (caa_likely(split_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))
723 return;
724 /* Only if number of add multiple of 1UL << COUNT_COMMIT_ORDER */
725
726 dbg_printf("add split count %lu\n", split_count);
727 count = uatomic_add_return(&ht->count,
728 1UL << COUNT_COMMIT_ORDER);
729 if (caa_likely(count & (count - 1)))
730 return;
731 /* Only if global count is power of 2 */
732
733 if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) < size)
734 return;
735 dbg_printf("add set global %lu\n", count);
736 cds_lfht_resize_lazy_count(ht, size,
737 count >> (CHAIN_LEN_TARGET - 1));
738 }
739
740 static
741 void ht_count_del(struct cds_lfht *ht, unsigned long size, unsigned long hash)
742 {
743 unsigned long split_count, count;
744 int index;
745
746 if (caa_unlikely(!ht->split_count))
747 return;
748 index = ht_get_split_count_index(hash);
749 split_count = uatomic_add_return(&ht->split_count[index].del, 1);
750 if (caa_likely(split_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))
751 return;
752 /* Only if number of deletes multiple of 1UL << COUNT_COMMIT_ORDER */
753
754 dbg_printf("del split count %lu\n", split_count);
755 count = uatomic_add_return(&ht->count,
756 -(1UL << COUNT_COMMIT_ORDER));
757 if (caa_likely(count & (count - 1)))
758 return;
759 /* Only if global count is power of 2 */
760
761 if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) >= size)
762 return;
763 dbg_printf("del set global %ld\n", count);
764 /*
765 * Don't shrink table if the number of nodes is below a
766 * certain threshold.
767 */
768 if (count < (1UL << COUNT_COMMIT_ORDER) * (split_count_mask + 1))
769 return;
770 cds_lfht_resize_lazy_count(ht, size,
771 count >> (CHAIN_LEN_TARGET - 1));
772 }
773
774 static
775 void check_resize(struct cds_lfht *ht, unsigned long size, uint32_t chain_len)
776 {
777 unsigned long count;
778
779 if (!(ht->flags & CDS_LFHT_AUTO_RESIZE))
780 return;
781 count = uatomic_read(&ht->count);
782 /*
783 * Use bucket-local length for small table expand and for
784 * environments lacking per-cpu data support.
785 */
786 if (count >= (1UL << (COUNT_COMMIT_ORDER + split_count_order)))
787 return;
788 if (chain_len > 100)
789 dbg_printf("WARNING: large chain length: %u.\n",
790 chain_len);
791 if (chain_len >= CHAIN_LEN_RESIZE_THRESHOLD) {
792 int growth;
793
794 /*
795 * Ideal growth calculated based on chain length.
796 */
797 growth = cds_lfht_get_count_order_u32(chain_len
798 - (CHAIN_LEN_TARGET - 1));
799 if ((ht->flags & CDS_LFHT_ACCOUNTING)
800 && (size << growth)
801 >= (1UL << (COUNT_COMMIT_ORDER
802 + split_count_order))) {
803 /*
804 * If ideal growth expands the hash table size
805 * beyond the "small hash table" sizes, use the
806 * maximum small hash table size to attempt
807 * expanding the hash table. This only applies
808 * when node accounting is available, otherwise
809 * the chain length is used to expand the hash
810 * table in every case.
811 */
812 growth = COUNT_COMMIT_ORDER + split_count_order
813 - cds_lfht_get_count_order_ulong(size);
814 if (growth <= 0)
815 return;
816 }
817 cds_lfht_resize_lazy_grow(ht, size, growth);
818 }
819 }
820
821 static
822 struct cds_lfht_node *clear_flag(struct cds_lfht_node *node)
823 {
824 return (struct cds_lfht_node *) (((unsigned long) node) & ~FLAGS_MASK);
825 }
826
827 static
828 int is_removed(struct cds_lfht_node *node)
829 {
830 return ((unsigned long) node) & REMOVED_FLAG;
831 }
832
833 static
834 int is_bucket(struct cds_lfht_node *node)
835 {
836 return ((unsigned long) node) & BUCKET_FLAG;
837 }
838
839 static
840 struct cds_lfht_node *flag_bucket(struct cds_lfht_node *node)
841 {
842 return (struct cds_lfht_node *) (((unsigned long) node) | BUCKET_FLAG);
843 }
844
845 static
846 int is_removal_owner(struct cds_lfht_node *node)
847 {
848 return ((unsigned long) node) & REMOVAL_OWNER_FLAG;
849 }
850
851 static
852 struct cds_lfht_node *flag_removal_owner(struct cds_lfht_node *node)
853 {
854 return (struct cds_lfht_node *) (((unsigned long) node) | REMOVAL_OWNER_FLAG);
855 }
856
857 static
858 struct cds_lfht_node *flag_removed_or_removal_owner(struct cds_lfht_node *node)
859 {
860 return (struct cds_lfht_node *) (((unsigned long) node) | REMOVED_FLAG | REMOVAL_OWNER_FLAG);
861 }
862
863 static
864 struct cds_lfht_node *get_end(void)
865 {
866 return (struct cds_lfht_node *) END_VALUE;
867 }
868
869 static
870 int is_end(struct cds_lfht_node *node)
871 {
872 return clear_flag(node) == (struct cds_lfht_node *) END_VALUE;
873 }
874
875 static
876 unsigned long _uatomic_xchg_monotonic_increase(unsigned long *ptr,
877 unsigned long v)
878 {
879 unsigned long old1, old2;
880
881 old1 = uatomic_read(ptr);
882 do {
883 old2 = old1;
884 if (old2 >= v)
885 return old2;
886 } while ((old1 = uatomic_cmpxchg(ptr, old2, v)) != old2);
887 return old2;
888 }
889
890 static
891 void cds_lfht_alloc_bucket_table(struct cds_lfht *ht, unsigned long order)
892 {
893 return ht->mm->alloc_bucket_table(ht, order);
894 }
895
896 /*
897 * cds_lfht_free_bucket_table() should be called with decreasing order.
898 * When cds_lfht_free_bucket_table(0) is called, it means the whole
899 * lfht is destroyed.
900 */
901 static
902 void cds_lfht_free_bucket_table(struct cds_lfht *ht, unsigned long order)
903 {
904 return ht->mm->free_bucket_table(ht, order);
905 }
906
907 static inline
908 struct cds_lfht_node *bucket_at(struct cds_lfht *ht, unsigned long index)
909 {
910 return ht->bucket_at(ht, index);
911 }
912
913 static inline
914 struct cds_lfht_node *lookup_bucket(struct cds_lfht *ht, unsigned long size,
915 unsigned long hash)
916 {
917 assert(size > 0);
918 return bucket_at(ht, hash & (size - 1));
919 }
920
921 /*
922 * Remove all logically deleted nodes from a bucket up to a certain node key.
923 */
924 static
925 void _cds_lfht_gc_bucket(struct cds_lfht_node *bucket, struct cds_lfht_node *node)
926 {
927 struct cds_lfht_node *iter_prev, *iter, *next, *new_next;
928
929 assert(!is_bucket(bucket));
930 assert(!is_removed(bucket));
931 assert(!is_removal_owner(bucket));
932 assert(!is_bucket(node));
933 assert(!is_removed(node));
934 assert(!is_removal_owner(node));
935 for (;;) {
936 iter_prev = bucket;
937 /* We can always skip the bucket node initially */
938 iter = rcu_dereference(iter_prev->next);
939 assert(!is_removed(iter));
940 assert(!is_removal_owner(iter));
941 assert(iter_prev->reverse_hash <= node->reverse_hash);
942 /*
943 * We should never be called with bucket (start of chain)
944 * and logically removed node (end of path compression
945 * marker) being the actual same node. This would be a
946 * bug in the algorithm implementation.
947 */
948 assert(bucket != node);
949 for (;;) {
950 if (caa_unlikely(is_end(iter)))
951 return;
952 if (caa_likely(clear_flag(iter)->reverse_hash > node->reverse_hash))
953 return;
954 next = rcu_dereference(clear_flag(iter)->next);
955 if (caa_likely(is_removed(next)))
956 break;
957 iter_prev = clear_flag(iter);
958 iter = next;
959 }
960 assert(!is_removed(iter));
961 assert(!is_removal_owner(iter));
962 if (is_bucket(iter))
963 new_next = flag_bucket(clear_flag(next));
964 else
965 new_next = clear_flag(next);
966 (void) uatomic_cmpxchg(&iter_prev->next, iter, new_next);
967 }
968 }
969
970 static
971 int _cds_lfht_replace(struct cds_lfht *ht, unsigned long size,
972 struct cds_lfht_node *old_node,
973 struct cds_lfht_node *old_next,
974 struct cds_lfht_node *new_node)
975 {
976 struct cds_lfht_node *bucket, *ret_next;
977
978 if (!old_node) /* Return -ENOENT if asked to replace NULL node */
979 return -ENOENT;
980
981 assert(!is_removed(old_node));
982 assert(!is_removal_owner(old_node));
983 assert(!is_bucket(old_node));
984 assert(!is_removed(new_node));
985 assert(!is_removal_owner(new_node));
986 assert(!is_bucket(new_node));
987 assert(new_node != old_node);
988 for (;;) {
989 /* Insert after node to be replaced */
990 if (is_removed(old_next)) {
991 /*
992 * Too late, the old node has been removed under us
993 * between lookup and replace. Fail.
994 */
995 return -ENOENT;
996 }
997 assert(old_next == clear_flag(old_next));
998 assert(new_node != old_next);
999 /*
1000 * REMOVAL_OWNER flag is _NEVER_ set before the REMOVED
1001 * flag. It is either set atomically at the same time
1002 * (replace) or after (del).
1003 */
1004 assert(!is_removal_owner(old_next));
1005 new_node->next = old_next;
1006 /*
1007 * Here is the whole trick for lock-free replace: we add
1008 * the replacement node _after_ the node we want to
1009 * replace by atomically setting its next pointer at the
1010 * same time we set its removal flag. Given that
1011 * the lookups/get next use an iterator aware of the
1012 * next pointer, they will either skip the old node due
1013 * to the removal flag and see the new node, or use
1014 * the old node, but will not see the new one.
1015 * This is a replacement of a node with another node
1016 * that has the same value: we are therefore not
1017 * removing a value from the hash table. We set both the
1018 * REMOVED and REMOVAL_OWNER flags atomically so we own
1019 * the node after successful cmpxchg.
1020 */
1021 ret_next = uatomic_cmpxchg(&old_node->next,
1022 old_next, flag_removed_or_removal_owner(new_node));
1023 if (ret_next == old_next)
1024 break; /* We performed the replacement. */
1025 old_next = ret_next;
1026 }
1027
1028 /*
1029 * Ensure that the old node is not visible to readers anymore:
1030 * lookup for the node, and remove it (along with any other
1031 * logically removed node) if found.
1032 */
1033 bucket = lookup_bucket(ht, size, bit_reverse_ulong(old_node->reverse_hash));
1034 _cds_lfht_gc_bucket(bucket, new_node);
1035
1036 assert(is_removed(CMM_LOAD_SHARED(old_node->next)));
1037 return 0;
1038 }
1039
1040 /*
1041 * A non-NULL unique_ret pointer uses the "add unique" (or uniquify) add
1042 * mode. A NULL unique_ret allows creation of duplicate keys.
1043 */
1044 static
1045 void _cds_lfht_add(struct cds_lfht *ht,
1046 unsigned long hash,
1047 cds_lfht_match_fct match,
1048 const void *key,
1049 unsigned long size,
1050 struct cds_lfht_node *node,
1051 struct cds_lfht_iter *unique_ret,
1052 int bucket_flag)
1053 {
1054 struct cds_lfht_node *iter_prev, *iter, *next, *new_node, *new_next,
1055 *return_node;
1056 struct cds_lfht_node *bucket;
1057
1058 assert(!is_bucket(node));
1059 assert(!is_removed(node));
1060 assert(!is_removal_owner(node));
1061 bucket = lookup_bucket(ht, size, hash);
1062 for (;;) {
1063 uint32_t chain_len = 0;
1064
1065 /*
1066 * iter_prev points to the non-removed node prior to the
1067 * insert location.
1068 */
1069 iter_prev = bucket;
1070 /* We can always skip the bucket node initially */
1071 iter = rcu_dereference(iter_prev->next);
1072 assert(iter_prev->reverse_hash <= node->reverse_hash);
1073 for (;;) {
1074 if (caa_unlikely(is_end(iter)))
1075 goto insert;
1076 if (caa_likely(clear_flag(iter)->reverse_hash > node->reverse_hash))
1077 goto insert;
1078
1079 /* bucket node is the first node of the identical-hash-value chain */
1080 if (bucket_flag && clear_flag(iter)->reverse_hash == node->reverse_hash)
1081 goto insert;
1082
1083 next = rcu_dereference(clear_flag(iter)->next);
1084 if (caa_unlikely(is_removed(next)))
1085 goto gc_node;
1086
1087 /* uniquely add */
1088 if (unique_ret
1089 && !is_bucket(next)
1090 && clear_flag(iter)->reverse_hash == node->reverse_hash) {
1091 struct cds_lfht_iter d_iter = {
1092 .node = node,
1093 .next = iter,
1094 #ifdef CONFIG_CDS_LFHT_ITER_DEBUG
1095 .lfht = ht,
1096 #endif
1097 };
1098
1099 /*
1100 * uniquely adding inserts the node as the first
1101 * node of the identical-hash-value node chain.
1102 *
1103 * This semantic ensures no duplicated keys
1104 * should ever be observable in the table
1105 * (including traversing the table node by
1106 * node by forward iterations)
1107 */
1108 cds_lfht_next_duplicate(ht, match, key, &d_iter);
1109 if (!d_iter.node)
1110 goto insert;
1111
1112 *unique_ret = d_iter;
1113 return;
1114 }
1115
1116 /* Only account for identical reverse hash once */
1117 if (iter_prev->reverse_hash != clear_flag(iter)->reverse_hash
1118 && !is_bucket(next))
1119 check_resize(ht, size, ++chain_len);
1120 iter_prev = clear_flag(iter);
1121 iter = next;
1122 }
1123
1124 insert:
1125 assert(node != clear_flag(iter));
1126 assert(!is_removed(iter_prev));
1127 assert(!is_removal_owner(iter_prev));
1128 assert(!is_removed(iter));
1129 assert(!is_removal_owner(iter));
1130 assert(iter_prev != node);
1131 if (!bucket_flag)
1132 node->next = clear_flag(iter);
1133 else
1134 node->next = flag_bucket(clear_flag(iter));
1135 if (is_bucket(iter))
1136 new_node = flag_bucket(node);
1137 else
1138 new_node = node;
1139 if (uatomic_cmpxchg(&iter_prev->next, iter,
1140 new_node) != iter) {
1141 continue; /* retry */
1142 } else {
1143 return_node = node;
1144 goto end;
1145 }
1146
1147 gc_node:
1148 assert(!is_removed(iter));
1149 assert(!is_removal_owner(iter));
1150 if (is_bucket(iter))
1151 new_next = flag_bucket(clear_flag(next));
1152 else
1153 new_next = clear_flag(next);
1154 (void) uatomic_cmpxchg(&iter_prev->next, iter, new_next);
1155 /* retry */
1156 }
1157 end:
1158 if (unique_ret) {
1159 unique_ret->node = return_node;
1160 /* unique_ret->next left unset, never used. */
1161 }
1162 }
1163
1164 static
1165 int _cds_lfht_del(struct cds_lfht *ht, unsigned long size,
1166 struct cds_lfht_node *node)
1167 {
1168 struct cds_lfht_node *bucket, *next;
1169
1170 if (!node) /* Return -ENOENT if asked to delete NULL node */
1171 return -ENOENT;
1172
1173 /* logically delete the node */
1174 assert(!is_bucket(node));
1175 assert(!is_removed(node));
1176 assert(!is_removal_owner(node));
1177
1178 /*
1179 * We are first checking if the node had previously been
1180 * logically removed (this check is not atomic with setting the
1181 * logical removal flag). Return -ENOENT if the node had
1182 * previously been removed.
1183 */
1184 next = CMM_LOAD_SHARED(node->next); /* next is not dereferenced */
1185 if (caa_unlikely(is_removed(next)))
1186 return -ENOENT;
1187 assert(!is_bucket(next));
1188 /*
1189 * The del operation semantic guarantees a full memory barrier
1190 * before the uatomic_or atomic commit of the deletion flag.
1191 */
1192 cmm_smp_mb__before_uatomic_or();
1193 /*
1194 * We set the REMOVED_FLAG unconditionally. Note that there may
1195 * be more than one concurrent thread setting this flag.
1196 * Knowing which wins the race will be known after the garbage
1197 * collection phase, stay tuned!
1198 */
1199 uatomic_or(&node->next, REMOVED_FLAG);
1200 /* We performed the (logical) deletion. */
1201
1202 /*
1203 * Ensure that the node is not visible to readers anymore: lookup for
1204 * the node, and remove it (along with any other logically removed node)
1205 * if found.
1206 */
1207 bucket = lookup_bucket(ht, size, bit_reverse_ulong(node->reverse_hash));
1208 _cds_lfht_gc_bucket(bucket, node);
1209
1210 assert(is_removed(CMM_LOAD_SHARED(node->next)));
1211 /*
1212 * Last phase: atomically exchange node->next with a version
1213 * having "REMOVAL_OWNER_FLAG" set. If the returned node->next
1214 * pointer did _not_ have "REMOVAL_OWNER_FLAG" set, we now own
1215 * the node and win the removal race.
1216 * It is interesting to note that all "add" paths are forbidden
1217 * to change the next pointer starting from the point where the
1218 * REMOVED_FLAG is set, so here using a read, followed by a
1219 * xchg() suffice to guarantee that the xchg() will ever only
1220 * set the "REMOVAL_OWNER_FLAG" (or change nothing if the flag
1221 * was already set).
1222 */
1223 if (!is_removal_owner(uatomic_xchg(&node->next,
1224 flag_removal_owner(node->next))))
1225 return 0;
1226 else
1227 return -ENOENT;
1228 }
1229
1230 static
1231 void *partition_resize_thread(void *arg)
1232 {
1233 struct partition_resize_work *work = arg;
1234
1235 work->ht->flavor->register_thread();
1236 work->fct(work->ht, work->i, work->start, work->len);
1237 work->ht->flavor->unregister_thread();
1238 return NULL;
1239 }
1240
1241 static
1242 void partition_resize_helper(struct cds_lfht *ht, unsigned long i,
1243 unsigned long len,
1244 void (*fct)(struct cds_lfht *ht, unsigned long i,
1245 unsigned long start, unsigned long len))
1246 {
1247 unsigned long partition_len, start = 0;
1248 struct partition_resize_work *work;
1249 int ret;
1250 unsigned long thread, nr_threads;
1251
1252 assert(nr_cpus_mask != -1);
1253 if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD)
1254 goto fallback;
1255
1256 /*
1257 * Note: nr_cpus_mask + 1 is always power of 2.
1258 * We spawn just the number of threads we need to satisfy the minimum
1259 * partition size, up to the number of CPUs in the system.
1260 */
1261 if (nr_cpus_mask > 0) {
1262 nr_threads = min_t(unsigned long, nr_cpus_mask + 1,
1263 len >> MIN_PARTITION_PER_THREAD_ORDER);
1264 } else {
1265 nr_threads = 1;
1266 }
1267 partition_len = len >> cds_lfht_get_count_order_ulong(nr_threads);
1268 work = calloc(nr_threads, sizeof(*work));
1269 if (!work) {
1270 dbg_printf("error allocating for resize, single-threading\n");
1271 goto fallback;
1272 }
1273 for (thread = 0; thread < nr_threads; thread++) {
1274 work[thread].ht = ht;
1275 work[thread].i = i;
1276 work[thread].len = partition_len;
1277 work[thread].start = thread * partition_len;
1278 work[thread].fct = fct;
1279 ret = pthread_create(&(work[thread].thread_id), ht->resize_attr,
1280 partition_resize_thread, &work[thread]);
1281 if (ret == EAGAIN) {
1282 /*
1283 * Out of resources: wait and join the threads
1284 * we've created, then handle leftovers.
1285 */
1286 dbg_printf("error spawning for resize, single-threading\n");
1287 start = work[thread].start;
1288 len -= start;
1289 nr_threads = thread;
1290 break;
1291 }
1292 assert(!ret);
1293 }
1294 for (thread = 0; thread < nr_threads; thread++) {
1295 ret = pthread_join(work[thread].thread_id, NULL);
1296 assert(!ret);
1297 }
1298 free(work);
1299
1300 /*
1301 * A pthread_create failure above will either lead in us having
1302 * no threads to join or starting at a non-zero offset,
1303 * fallback to single thread processing of leftovers.
1304 */
1305 if (start == 0 && nr_threads > 0)
1306 return;
1307 fallback:
1308 fct(ht, i, start, len);
1309 }
1310
1311 /*
1312 * Holding RCU read lock to protect _cds_lfht_add against memory
1313 * reclaim that could be performed by other worker threads (ABA
1314 * problem).
1315 *
1316 * When we reach a certain length, we can split this population phase over
1317 * many worker threads, based on the number of CPUs available in the system.
1318 * This should therefore take care of not having the expand lagging behind too
1319 * many concurrent insertion threads by using the scheduler's ability to
1320 * schedule bucket node population fairly with insertions.
1321 */
1322 static
1323 void init_table_populate_partition(struct cds_lfht *ht, unsigned long i,
1324 unsigned long start, unsigned long len)
1325 {
1326 unsigned long j, size = 1UL << (i - 1);
1327
1328 assert(i > MIN_TABLE_ORDER);
1329 ht->flavor->read_lock();
1330 for (j = size + start; j < size + start + len; j++) {
1331 struct cds_lfht_node *new_node = bucket_at(ht, j);
1332
1333 assert(j >= size && j < (size << 1));
1334 dbg_printf("init populate: order %lu index %lu hash %lu\n",
1335 i, j, j);
1336 new_node->reverse_hash = bit_reverse_ulong(j);
1337 _cds_lfht_add(ht, j, NULL, NULL, size, new_node, NULL, 1);
1338 }
1339 ht->flavor->read_unlock();
1340 }
1341
1342 static
1343 void init_table_populate(struct cds_lfht *ht, unsigned long i,
1344 unsigned long len)
1345 {
1346 partition_resize_helper(ht, i, len, init_table_populate_partition);
1347 }
1348
1349 static
1350 void init_table(struct cds_lfht *ht,
1351 unsigned long first_order, unsigned long last_order)
1352 {
1353 unsigned long i;
1354
1355 dbg_printf("init table: first_order %lu last_order %lu\n",
1356 first_order, last_order);
1357 assert(first_order > MIN_TABLE_ORDER);
1358 for (i = first_order; i <= last_order; i++) {
1359 unsigned long len;
1360
1361 len = 1UL << (i - 1);
1362 dbg_printf("init order %lu len: %lu\n", i, len);
1363
1364 /* Stop expand if the resize target changes under us */
1365 if (CMM_LOAD_SHARED(ht->resize_target) < (1UL << i))
1366 break;
1367
1368 cds_lfht_alloc_bucket_table(ht, i);
1369
1370 /*
1371 * Set all bucket nodes reverse hash values for a level and
1372 * link all bucket nodes into the table.
1373 */
1374 init_table_populate(ht, i, len);
1375
1376 /*
1377 * Update table size.
1378 */
1379 cmm_smp_wmb(); /* populate data before RCU size */
1380 CMM_STORE_SHARED(ht->size, 1UL << i);
1381
1382 dbg_printf("init new size: %lu\n", 1UL << i);
1383 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1384 break;
1385 }
1386 }
1387
1388 /*
1389 * Holding RCU read lock to protect _cds_lfht_remove against memory
1390 * reclaim that could be performed by other worker threads (ABA
1391 * problem).
1392 * For a single level, we logically remove and garbage collect each node.
1393 *
1394 * As a design choice, we perform logical removal and garbage collection on a
1395 * node-per-node basis to simplify this algorithm. We also assume keeping good
1396 * cache locality of the operation would overweight possible performance gain
1397 * that could be achieved by batching garbage collection for multiple levels.
1398 * However, this would have to be justified by benchmarks.
1399 *
1400 * Concurrent removal and add operations are helping us perform garbage
1401 * collection of logically removed nodes. We guarantee that all logically
1402 * removed nodes have been garbage-collected (unlinked) before work
1403 * enqueue is invoked to free a hole level of bucket nodes (after a
1404 * grace period).
1405 *
1406 * Logical removal and garbage collection can therefore be done in batch
1407 * or on a node-per-node basis, as long as the guarantee above holds.
1408 *
1409 * When we reach a certain length, we can split this removal over many worker
1410 * threads, based on the number of CPUs available in the system. This should
1411 * take care of not letting resize process lag behind too many concurrent
1412 * updater threads actively inserting into the hash table.
1413 */
1414 static
1415 void remove_table_partition(struct cds_lfht *ht, unsigned long i,
1416 unsigned long start, unsigned long len)
1417 {
1418 unsigned long j, size = 1UL << (i - 1);
1419
1420 assert(i > MIN_TABLE_ORDER);
1421 ht->flavor->read_lock();
1422 for (j = size + start; j < size + start + len; j++) {
1423 struct cds_lfht_node *fini_bucket = bucket_at(ht, j);
1424 struct cds_lfht_node *parent_bucket = bucket_at(ht, j - size);
1425
1426 assert(j >= size && j < (size << 1));
1427 dbg_printf("remove entry: order %lu index %lu hash %lu\n",
1428 i, j, j);
1429 /* Set the REMOVED_FLAG to freeze the ->next for gc */
1430 uatomic_or(&fini_bucket->next, REMOVED_FLAG);
1431 _cds_lfht_gc_bucket(parent_bucket, fini_bucket);
1432 }
1433 ht->flavor->read_unlock();
1434 }
1435
1436 static
1437 void remove_table(struct cds_lfht *ht, unsigned long i, unsigned long len)
1438 {
1439 partition_resize_helper(ht, i, len, remove_table_partition);
1440 }
1441
1442 /*
1443 * fini_table() is never called for first_order == 0, which is why
1444 * free_by_rcu_order == 0 can be used as criterion to know if free must
1445 * be called.
1446 */
1447 static
1448 void fini_table(struct cds_lfht *ht,
1449 unsigned long first_order, unsigned long last_order)
1450 {
1451 unsigned long free_by_rcu_order = 0, i;
1452
1453 dbg_printf("fini table: first_order %lu last_order %lu\n",
1454 first_order, last_order);
1455 assert(first_order > MIN_TABLE_ORDER);
1456 for (i = last_order; i >= first_order; i--) {
1457 unsigned long len;
1458
1459 len = 1UL << (i - 1);
1460 dbg_printf("fini order %ld len: %lu\n", i, len);
1461
1462 /* Stop shrink if the resize target changes under us */
1463 if (CMM_LOAD_SHARED(ht->resize_target) > (1UL << (i - 1)))
1464 break;
1465
1466 cmm_smp_wmb(); /* populate data before RCU size */
1467 CMM_STORE_SHARED(ht->size, 1UL << (i - 1));
1468
1469 /*
1470 * We need to wait for all add operations to reach Q.S. (and
1471 * thus use the new table for lookups) before we can start
1472 * releasing the old bucket nodes. Otherwise their lookup will
1473 * return a logically removed node as insert position.
1474 */
1475 ht->flavor->update_synchronize_rcu();
1476 if (free_by_rcu_order)
1477 cds_lfht_free_bucket_table(ht, free_by_rcu_order);
1478
1479 /*
1480 * Set "removed" flag in bucket nodes about to be removed.
1481 * Unlink all now-logically-removed bucket node pointers.
1482 * Concurrent add/remove operation are helping us doing
1483 * the gc.
1484 */
1485 remove_table(ht, i, len);
1486
1487 free_by_rcu_order = i;
1488
1489 dbg_printf("fini new size: %lu\n", 1UL << i);
1490 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1491 break;
1492 }
1493
1494 if (free_by_rcu_order) {
1495 ht->flavor->update_synchronize_rcu();
1496 cds_lfht_free_bucket_table(ht, free_by_rcu_order);
1497 }
1498 }
1499
1500 /*
1501 * Never called with size < 1.
1502 */
1503 static
1504 void cds_lfht_create_bucket(struct cds_lfht *ht, unsigned long size)
1505 {
1506 struct cds_lfht_node *prev, *node;
1507 unsigned long order, len, i;
1508 int bucket_order;
1509
1510 cds_lfht_alloc_bucket_table(ht, 0);
1511
1512 dbg_printf("create bucket: order 0 index 0 hash 0\n");
1513 node = bucket_at(ht, 0);
1514 node->next = flag_bucket(get_end());
1515 node->reverse_hash = 0;
1516
1517 bucket_order = cds_lfht_get_count_order_ulong(size);
1518 assert(bucket_order >= 0);
1519
1520 for (order = 1; order < (unsigned long) bucket_order + 1; order++) {
1521 len = 1UL << (order - 1);
1522 cds_lfht_alloc_bucket_table(ht, order);
1523
1524 for (i = 0; i < len; i++) {
1525 /*
1526 * Now, we are trying to init the node with the
1527 * hash=(len+i) (which is also a bucket with the
1528 * index=(len+i)) and insert it into the hash table,
1529 * so this node has to be inserted after the bucket
1530 * with the index=(len+i)&(len-1)=i. And because there
1531 * is no other non-bucket node nor bucket node with
1532 * larger index/hash inserted, so the bucket node
1533 * being inserted should be inserted directly linked
1534 * after the bucket node with index=i.
1535 */
1536 prev = bucket_at(ht, i);
1537 node = bucket_at(ht, len + i);
1538
1539 dbg_printf("create bucket: order %lu index %lu hash %lu\n",
1540 order, len + i, len + i);
1541 node->reverse_hash = bit_reverse_ulong(len + i);
1542
1543 /* insert after prev */
1544 assert(is_bucket(prev->next));
1545 node->next = prev->next;
1546 prev->next = flag_bucket(node);
1547 }
1548 }
1549 }
1550
1551 #if (CAA_BITS_PER_LONG > 32)
1552 /*
1553 * For 64-bit architectures, with max number of buckets small enough not to
1554 * use the entire 64-bit memory mapping space (and allowing a fair number of
1555 * hash table instances), use the mmap allocator, which is faster. Otherwise,
1556 * fallback to the order allocator.
1557 */
1558 static
1559 const struct cds_lfht_mm_type *get_mm_type(unsigned long max_nr_buckets)
1560 {
1561 if (max_nr_buckets && max_nr_buckets <= (1ULL << 32))
1562 return &cds_lfht_mm_mmap;
1563 else
1564 return &cds_lfht_mm_order;
1565 }
1566 #else
1567 /*
1568 * For 32-bit architectures, use the order allocator.
1569 */
1570 static
1571 const struct cds_lfht_mm_type *get_mm_type(unsigned long max_nr_buckets)
1572 {
1573 return &cds_lfht_mm_order;
1574 }
1575 #endif
1576
1577 struct cds_lfht *_cds_lfht_new(unsigned long init_size,
1578 unsigned long min_nr_alloc_buckets,
1579 unsigned long max_nr_buckets,
1580 int flags,
1581 const struct cds_lfht_mm_type *mm,
1582 const struct rcu_flavor_struct *flavor,
1583 pthread_attr_t *attr)
1584 {
1585 struct cds_lfht *ht;
1586 unsigned long order;
1587
1588 /* min_nr_alloc_buckets must be power of two */
1589 if (!min_nr_alloc_buckets || (min_nr_alloc_buckets & (min_nr_alloc_buckets - 1)))
1590 return NULL;
1591
1592 /* init_size must be power of two */
1593 if (!init_size || (init_size & (init_size - 1)))
1594 return NULL;
1595
1596 /*
1597 * Memory management plugin default.
1598 */
1599 if (!mm)
1600 mm = get_mm_type(max_nr_buckets);
1601
1602 /* max_nr_buckets == 0 for order based mm means infinite */
1603 if (mm == &cds_lfht_mm_order && !max_nr_buckets)
1604 max_nr_buckets = 1UL << (MAX_TABLE_ORDER - 1);
1605
1606 /* max_nr_buckets must be power of two */
1607 if (!max_nr_buckets || (max_nr_buckets & (max_nr_buckets - 1)))
1608 return NULL;
1609
1610 if (flags & CDS_LFHT_AUTO_RESIZE)
1611 cds_lfht_init_worker(flavor);
1612
1613 min_nr_alloc_buckets = max(min_nr_alloc_buckets, MIN_TABLE_SIZE);
1614 init_size = max(init_size, MIN_TABLE_SIZE);
1615 max_nr_buckets = max(max_nr_buckets, min_nr_alloc_buckets);
1616 init_size = min(init_size, max_nr_buckets);
1617
1618 ht = mm->alloc_cds_lfht(min_nr_alloc_buckets, max_nr_buckets);
1619 assert(ht);
1620 assert(ht->mm == mm);
1621 assert(ht->bucket_at == mm->bucket_at);
1622
1623 ht->flags = flags;
1624 ht->flavor = flavor;
1625 ht->resize_attr = attr;
1626 alloc_split_items_count(ht);
1627 /* this mutex should not nest in read-side C.S. */
1628 pthread_mutex_init(&ht->resize_mutex, NULL);
1629 order = cds_lfht_get_count_order_ulong(init_size);
1630 ht->resize_target = 1UL << order;
1631 cds_lfht_create_bucket(ht, 1UL << order);
1632 ht->size = 1UL << order;
1633 return ht;
1634 }
1635
1636 void cds_lfht_lookup(struct cds_lfht *ht, unsigned long hash,
1637 cds_lfht_match_fct match, const void *key,
1638 struct cds_lfht_iter *iter)
1639 {
1640 struct cds_lfht_node *node, *next, *bucket;
1641 unsigned long reverse_hash, size;
1642
1643 cds_lfht_iter_debug_set_ht(ht, iter);
1644
1645 reverse_hash = bit_reverse_ulong(hash);
1646
1647 size = rcu_dereference(ht->size);
1648 bucket = lookup_bucket(ht, size, hash);
1649 /* We can always skip the bucket node initially */
1650 node = rcu_dereference(bucket->next);
1651 node = clear_flag(node);
1652 for (;;) {
1653 if (caa_unlikely(is_end(node))) {
1654 node = next = NULL;
1655 break;
1656 }
1657 if (caa_unlikely(node->reverse_hash > reverse_hash)) {
1658 node = next = NULL;
1659 break;
1660 }
1661 next = rcu_dereference(node->next);
1662 assert(node == clear_flag(node));
1663 if (caa_likely(!is_removed(next))
1664 && !is_bucket(next)
1665 && node->reverse_hash == reverse_hash
1666 && caa_likely(match(node, key))) {
1667 break;
1668 }
1669 node = clear_flag(next);
1670 }
1671 assert(!node || !is_bucket(CMM_LOAD_SHARED(node->next)));
1672 iter->node = node;
1673 iter->next = next;
1674 }
1675
1676 void cds_lfht_next_duplicate(struct cds_lfht *ht, cds_lfht_match_fct match,
1677 const void *key, struct cds_lfht_iter *iter)
1678 {
1679 struct cds_lfht_node *node, *next;
1680 unsigned long reverse_hash;
1681
1682 cds_lfht_iter_debug_assert(ht == iter->lfht);
1683 node = iter->node;
1684 reverse_hash = node->reverse_hash;
1685 next = iter->next;
1686 node = clear_flag(next);
1687
1688 for (;;) {
1689 if (caa_unlikely(is_end(node))) {
1690 node = next = NULL;
1691 break;
1692 }
1693 if (caa_unlikely(node->reverse_hash > reverse_hash)) {
1694 node = next = NULL;
1695 break;
1696 }
1697 next = rcu_dereference(node->next);
1698 if (caa_likely(!is_removed(next))
1699 && !is_bucket(next)
1700 && caa_likely(match(node, key))) {
1701 break;
1702 }
1703 node = clear_flag(next);
1704 }
1705 assert(!node || !is_bucket(CMM_LOAD_SHARED(node->next)));
1706 iter->node = node;
1707 iter->next = next;
1708 }
1709
1710 void cds_lfht_next(struct cds_lfht *ht, struct cds_lfht_iter *iter)
1711 {
1712 struct cds_lfht_node *node, *next;
1713
1714 cds_lfht_iter_debug_assert(ht == iter->lfht);
1715 node = clear_flag(iter->next);
1716 for (;;) {
1717 if (caa_unlikely(is_end(node))) {
1718 node = next = NULL;
1719 break;
1720 }
1721 next = rcu_dereference(node->next);
1722 if (caa_likely(!is_removed(next))
1723 && !is_bucket(next)) {
1724 break;
1725 }
1726 node = clear_flag(next);
1727 }
1728 assert(!node || !is_bucket(CMM_LOAD_SHARED(node->next)));
1729 iter->node = node;
1730 iter->next = next;
1731 }
1732
1733 void cds_lfht_first(struct cds_lfht *ht, struct cds_lfht_iter *iter)
1734 {
1735 cds_lfht_iter_debug_set_ht(ht, iter);
1736 /*
1737 * Get next after first bucket node. The first bucket node is the
1738 * first node of the linked list.
1739 */
1740 iter->next = bucket_at(ht, 0)->next;
1741 cds_lfht_next(ht, iter);
1742 }
1743
1744 void cds_lfht_add(struct cds_lfht *ht, unsigned long hash,
1745 struct cds_lfht_node *node)
1746 {
1747 unsigned long size;
1748
1749 node->reverse_hash = bit_reverse_ulong(hash);
1750 size = rcu_dereference(ht->size);
1751 _cds_lfht_add(ht, hash, NULL, NULL, size, node, NULL, 0);
1752 ht_count_add(ht, size, hash);
1753 }
1754
1755 struct cds_lfht_node *cds_lfht_add_unique(struct cds_lfht *ht,
1756 unsigned long hash,
1757 cds_lfht_match_fct match,
1758 const void *key,
1759 struct cds_lfht_node *node)
1760 {
1761 unsigned long size;
1762 struct cds_lfht_iter iter;
1763
1764 node->reverse_hash = bit_reverse_ulong(hash);
1765 size = rcu_dereference(ht->size);
1766 _cds_lfht_add(ht, hash, match, key, size, node, &iter, 0);
1767 if (iter.node == node)
1768 ht_count_add(ht, size, hash);
1769 return iter.node;
1770 }
1771
1772 struct cds_lfht_node *cds_lfht_add_replace(struct cds_lfht *ht,
1773 unsigned long hash,
1774 cds_lfht_match_fct match,
1775 const void *key,
1776 struct cds_lfht_node *node)
1777 {
1778 unsigned long size;
1779 struct cds_lfht_iter iter;
1780
1781 node->reverse_hash = bit_reverse_ulong(hash);
1782 size = rcu_dereference(ht->size);
1783 for (;;) {
1784 _cds_lfht_add(ht, hash, match, key, size, node, &iter, 0);
1785 if (iter.node == node) {
1786 ht_count_add(ht, size, hash);
1787 return NULL;
1788 }
1789
1790 if (!_cds_lfht_replace(ht, size, iter.node, iter.next, node))
1791 return iter.node;
1792 }
1793 }
1794
1795 int cds_lfht_replace(struct cds_lfht *ht,
1796 struct cds_lfht_iter *old_iter,
1797 unsigned long hash,
1798 cds_lfht_match_fct match,
1799 const void *key,
1800 struct cds_lfht_node *new_node)
1801 {
1802 unsigned long size;
1803
1804 new_node->reverse_hash = bit_reverse_ulong(hash);
1805 if (!old_iter->node)
1806 return -ENOENT;
1807 if (caa_unlikely(old_iter->node->reverse_hash != new_node->reverse_hash))
1808 return -EINVAL;
1809 if (caa_unlikely(!match(old_iter->node, key)))
1810 return -EINVAL;
1811 size = rcu_dereference(ht->size);
1812 return _cds_lfht_replace(ht, size, old_iter->node, old_iter->next,
1813 new_node);
1814 }
1815
1816 int cds_lfht_del(struct cds_lfht *ht, struct cds_lfht_node *node)
1817 {
1818 unsigned long size;
1819 int ret;
1820
1821 size = rcu_dereference(ht->size);
1822 ret = _cds_lfht_del(ht, size, node);
1823 if (!ret) {
1824 unsigned long hash;
1825
1826 hash = bit_reverse_ulong(node->reverse_hash);
1827 ht_count_del(ht, size, hash);
1828 }
1829 return ret;
1830 }
1831
1832 int cds_lfht_is_node_deleted(struct cds_lfht_node *node)
1833 {
1834 return is_removed(CMM_LOAD_SHARED(node->next));
1835 }
1836
1837 static
1838 int cds_lfht_delete_bucket(struct cds_lfht *ht)
1839 {
1840 struct cds_lfht_node *node;
1841 unsigned long order, i, size;
1842
1843 /* Check that the table is empty */
1844 node = bucket_at(ht, 0);
1845 do {
1846 node = clear_flag(node)->next;
1847 if (!is_bucket(node))
1848 return -EPERM;
1849 assert(!is_removed(node));
1850 assert(!is_removal_owner(node));
1851 } while (!is_end(node));
1852 /*
1853 * size accessed without rcu_dereference because hash table is
1854 * being destroyed.
1855 */
1856 size = ht->size;
1857 /* Internal sanity check: all nodes left should be buckets */
1858 for (i = 0; i < size; i++) {
1859 node = bucket_at(ht, i);
1860 dbg_printf("delete bucket: index %lu expected hash %lu hash %lu\n",
1861 i, i, bit_reverse_ulong(node->reverse_hash));
1862 assert(is_bucket(node->next));
1863 }
1864
1865 for (order = cds_lfht_get_count_order_ulong(size); (long)order >= 0; order--)
1866 cds_lfht_free_bucket_table(ht, order);
1867
1868 return 0;
1869 }
1870
1871 /*
1872 * Should only be called when no more concurrent readers nor writers can
1873 * possibly access the table.
1874 */
1875 int cds_lfht_destroy(struct cds_lfht *ht, pthread_attr_t **attr)
1876 {
1877 int ret;
1878
1879 if (ht->flags & CDS_LFHT_AUTO_RESIZE) {
1880 /* Cancel ongoing resize operations. */
1881 _CMM_STORE_SHARED(ht->in_progress_destroy, 1);
1882 /* Wait for in-flight resize operations to complete */
1883 urcu_workqueue_flush_queued_work(cds_lfht_workqueue);
1884 }
1885 ret = cds_lfht_delete_bucket(ht);
1886 if (ret)
1887 return ret;
1888 free_split_items_count(ht);
1889 if (attr)
1890 *attr = ht->resize_attr;
1891 ret = pthread_mutex_destroy(&ht->resize_mutex);
1892 if (ret)
1893 ret = -EBUSY;
1894 if (ht->flags & CDS_LFHT_AUTO_RESIZE)
1895 cds_lfht_fini_worker(ht->flavor);
1896 poison_free(ht);
1897 return ret;
1898 }
1899
1900 void cds_lfht_count_nodes(struct cds_lfht *ht,
1901 long *approx_before,
1902 unsigned long *count,
1903 long *approx_after)
1904 {
1905 struct cds_lfht_node *node, *next;
1906 unsigned long nr_bucket = 0, nr_removed = 0;
1907
1908 *approx_before = 0;
1909 if (ht->split_count) {
1910 int i;
1911
1912 for (i = 0; i < split_count_mask + 1; i++) {
1913 *approx_before += uatomic_read(&ht->split_count[i].add);
1914 *approx_before -= uatomic_read(&ht->split_count[i].del);
1915 }
1916 }
1917
1918 *count = 0;
1919
1920 /* Count non-bucket nodes in the table */
1921 node = bucket_at(ht, 0);
1922 do {
1923 next = rcu_dereference(node->next);
1924 if (is_removed(next)) {
1925 if (!is_bucket(next))
1926 (nr_removed)++;
1927 else
1928 (nr_bucket)++;
1929 } else if (!is_bucket(next))
1930 (*count)++;
1931 else
1932 (nr_bucket)++;
1933 node = clear_flag(next);
1934 } while (!is_end(node));
1935 dbg_printf("number of logically removed nodes: %lu\n", nr_removed);
1936 dbg_printf("number of bucket nodes: %lu\n", nr_bucket);
1937 *approx_after = 0;
1938 if (ht->split_count) {
1939 int i;
1940
1941 for (i = 0; i < split_count_mask + 1; i++) {
1942 *approx_after += uatomic_read(&ht->split_count[i].add);
1943 *approx_after -= uatomic_read(&ht->split_count[i].del);
1944 }
1945 }
1946 }
1947
1948 /* called with resize mutex held */
1949 static
1950 void _do_cds_lfht_grow(struct cds_lfht *ht,
1951 unsigned long old_size, unsigned long new_size)
1952 {
1953 unsigned long old_order, new_order;
1954
1955 old_order = cds_lfht_get_count_order_ulong(old_size);
1956 new_order = cds_lfht_get_count_order_ulong(new_size);
1957 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1958 old_size, old_order, new_size, new_order);
1959 assert(new_size > old_size);
1960 init_table(ht, old_order + 1, new_order);
1961 }
1962
1963 /* called with resize mutex held */
1964 static
1965 void _do_cds_lfht_shrink(struct cds_lfht *ht,
1966 unsigned long old_size, unsigned long new_size)
1967 {
1968 unsigned long old_order, new_order;
1969
1970 new_size = max(new_size, MIN_TABLE_SIZE);
1971 old_order = cds_lfht_get_count_order_ulong(old_size);
1972 new_order = cds_lfht_get_count_order_ulong(new_size);
1973 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1974 old_size, old_order, new_size, new_order);
1975 assert(new_size < old_size);
1976
1977 /* Remove and unlink all bucket nodes to remove. */
1978 fini_table(ht, new_order + 1, old_order);
1979 }
1980
1981
1982 /* called with resize mutex held */
1983 static
1984 void _do_cds_lfht_resize(struct cds_lfht *ht)
1985 {
1986 unsigned long new_size, old_size;
1987
1988 /*
1989 * Resize table, re-do if the target size has changed under us.
1990 */
1991 do {
1992 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1993 break;
1994 ht->resize_initiated = 1;
1995 old_size = ht->size;
1996 new_size = CMM_LOAD_SHARED(ht->resize_target);
1997 if (old_size < new_size)
1998 _do_cds_lfht_grow(ht, old_size, new_size);
1999 else if (old_size > new_size)
2000 _do_cds_lfht_shrink(ht, old_size, new_size);
2001 ht->resize_initiated = 0;
2002 /* write resize_initiated before read resize_target */
2003 cmm_smp_mb();
2004 } while (ht->size != CMM_LOAD_SHARED(ht->resize_target));
2005 }
2006
2007 static
2008 unsigned long resize_target_grow(struct cds_lfht *ht, unsigned long new_size)
2009 {
2010 return _uatomic_xchg_monotonic_increase(&ht->resize_target, new_size);
2011 }
2012
2013 static
2014 void resize_target_update_count(struct cds_lfht *ht,
2015 unsigned long count)
2016 {
2017 count = max(count, MIN_TABLE_SIZE);
2018 count = min(count, ht->max_nr_buckets);
2019 uatomic_set(&ht->resize_target, count);
2020 }
2021
2022 void cds_lfht_resize(struct cds_lfht *ht, unsigned long new_size)
2023 {
2024 resize_target_update_count(ht, new_size);
2025 CMM_STORE_SHARED(ht->resize_initiated, 1);
2026 mutex_lock(&ht->resize_mutex);
2027 _do_cds_lfht_resize(ht);
2028 mutex_unlock(&ht->resize_mutex);
2029 }
2030
2031 static
2032 void do_resize_cb(struct urcu_work *work)
2033 {
2034 struct resize_work *resize_work =
2035 caa_container_of(work, struct resize_work, work);
2036 struct cds_lfht *ht = resize_work->ht;
2037
2038 ht->flavor->register_thread();
2039 mutex_lock(&ht->resize_mutex);
2040 _do_cds_lfht_resize(ht);
2041 mutex_unlock(&ht->resize_mutex);
2042 ht->flavor->unregister_thread();
2043 poison_free(work);
2044 }
2045
2046 static
2047 void __cds_lfht_resize_lazy_launch(struct cds_lfht *ht)
2048 {
2049 struct resize_work *work;
2050
2051 /* Store resize_target before read resize_initiated */
2052 cmm_smp_mb();
2053 if (!CMM_LOAD_SHARED(ht->resize_initiated)) {
2054 if (CMM_LOAD_SHARED(ht->in_progress_destroy)) {
2055 return;
2056 }
2057 work = malloc(sizeof(*work));
2058 if (work == NULL) {
2059 dbg_printf("error allocating resize work, bailing out\n");
2060 return;
2061 }
2062 work->ht = ht;
2063 urcu_workqueue_queue_work(cds_lfht_workqueue,
2064 &work->work, do_resize_cb);
2065 CMM_STORE_SHARED(ht->resize_initiated, 1);
2066 }
2067 }
2068
2069 static
2070 void cds_lfht_resize_lazy_grow(struct cds_lfht *ht, unsigned long size, int growth)
2071 {
2072 unsigned long target_size = size << growth;
2073
2074 target_size = min(target_size, ht->max_nr_buckets);
2075 if (resize_target_grow(ht, target_size) >= target_size)
2076 return;
2077
2078 __cds_lfht_resize_lazy_launch(ht);
2079 }
2080
2081 /*
2082 * We favor grow operations over shrink. A shrink operation never occurs
2083 * if a grow operation is queued for lazy execution. A grow operation
2084 * cancels any pending shrink lazy execution.
2085 */
2086 static
2087 void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
2088 unsigned long count)
2089 {
2090 if (!(ht->flags & CDS_LFHT_AUTO_RESIZE))
2091 return;
2092 count = max(count, MIN_TABLE_SIZE);
2093 count = min(count, ht->max_nr_buckets);
2094 if (count == size)
2095 return; /* Already the right size, no resize needed */
2096 if (count > size) { /* lazy grow */
2097 if (resize_target_grow(ht, count) >= count)
2098 return;
2099 } else { /* lazy shrink */
2100 for (;;) {
2101 unsigned long s;
2102
2103 s = uatomic_cmpxchg(&ht->resize_target, size, count);
2104 if (s == size)
2105 break; /* no resize needed */
2106 if (s > size)
2107 return; /* growing is/(was just) in progress */
2108 if (s <= count)
2109 return; /* some other thread do shrink */
2110 size = s;
2111 }
2112 }
2113 __cds_lfht_resize_lazy_launch(ht);
2114 }
2115
2116 static void cds_lfht_before_fork(void *priv)
2117 {
2118 if (cds_lfht_workqueue_atfork_nesting++)
2119 return;
2120 mutex_lock(&cds_lfht_fork_mutex);
2121 if (!cds_lfht_workqueue)
2122 return;
2123 urcu_workqueue_pause_worker(cds_lfht_workqueue);
2124 }
2125
2126 static void cds_lfht_after_fork_parent(void *priv)
2127 {
2128 if (--cds_lfht_workqueue_atfork_nesting)
2129 return;
2130 if (!cds_lfht_workqueue)
2131 goto end;
2132 urcu_workqueue_resume_worker(cds_lfht_workqueue);
2133 end:
2134 mutex_unlock(&cds_lfht_fork_mutex);
2135 }
2136
2137 static void cds_lfht_after_fork_child(void *priv)
2138 {
2139 if (--cds_lfht_workqueue_atfork_nesting)
2140 return;
2141 if (!cds_lfht_workqueue)
2142 goto end;
2143 urcu_workqueue_create_worker(cds_lfht_workqueue);
2144 end:
2145 mutex_unlock(&cds_lfht_fork_mutex);
2146 }
2147
2148 static struct urcu_atfork cds_lfht_atfork = {
2149 .before_fork = cds_lfht_before_fork,
2150 .after_fork_parent = cds_lfht_after_fork_parent,
2151 .after_fork_child = cds_lfht_after_fork_child,
2152 };
2153
2154 /* Block all signals to ensure we don't disturb the application. */
2155 static void cds_lfht_worker_init(struct urcu_workqueue *workqueue,
2156 void *priv)
2157 {
2158 int ret;
2159 sigset_t mask;
2160
2161 /* Block signal for entire process, so only our thread processes it. */
2162 ret = sigfillset(&mask);
2163 if (ret)
2164 urcu_die(errno);
2165 ret = pthread_sigmask(SIG_BLOCK, &mask, NULL);
2166 if (ret)
2167 urcu_die(ret);
2168 }
2169
2170 static void cds_lfht_init_worker(const struct rcu_flavor_struct *flavor)
2171 {
2172 flavor->register_rculfhash_atfork(&cds_lfht_atfork);
2173
2174 mutex_lock(&cds_lfht_fork_mutex);
2175 if (cds_lfht_workqueue_user_count++)
2176 goto end;
2177 cds_lfht_workqueue = urcu_workqueue_create(0, -1, NULL,
2178 NULL, cds_lfht_worker_init, NULL, NULL, NULL, NULL, NULL);
2179 end:
2180 mutex_unlock(&cds_lfht_fork_mutex);
2181 }
2182
2183 static void cds_lfht_fini_worker(const struct rcu_flavor_struct *flavor)
2184 {
2185 mutex_lock(&cds_lfht_fork_mutex);
2186 if (--cds_lfht_workqueue_user_count)
2187 goto end;
2188 urcu_workqueue_destroy(cds_lfht_workqueue);
2189 cds_lfht_workqueue = NULL;
2190 end:
2191 mutex_unlock(&cds_lfht_fork_mutex);
2192
2193 flavor->unregister_rculfhash_atfork(&cds_lfht_atfork);
2194 }
This page took 0.130234 seconds and 4 git commands to generate.