Use workqueue in rculfhash
[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
283 /*
284 * Split-counters lazily update the global counter each 1024
285 * addition/removal. It automatically keeps track of resize required.
286 * We use the bucket length as indicator for need to expand for small
287 * tables and machines lacking per-cpu data support.
288 */
289 #define COUNT_COMMIT_ORDER 10
290 #define DEFAULT_SPLIT_COUNT_MASK 0xFUL
291 #define CHAIN_LEN_TARGET 1
292 #define CHAIN_LEN_RESIZE_THRESHOLD 3
293
294 /*
295 * Define the minimum table size.
296 */
297 #define MIN_TABLE_ORDER 0
298 #define MIN_TABLE_SIZE (1UL << MIN_TABLE_ORDER)
299
300 /*
301 * Minimum number of bucket nodes to touch per thread to parallelize grow/shrink.
302 */
303 #define MIN_PARTITION_PER_THREAD_ORDER 12
304 #define MIN_PARTITION_PER_THREAD (1UL << MIN_PARTITION_PER_THREAD_ORDER)
305
306 /*
307 * The removed flag needs to be updated atomically with the pointer.
308 * It indicates that no node must attach to the node scheduled for
309 * removal, and that node garbage collection must be performed.
310 * The bucket flag does not require to be updated atomically with the
311 * pointer, but it is added as a pointer low bit flag to save space.
312 * The "removal owner" flag is used to detect which of the "del"
313 * operation that has set the "removed flag" gets to return the removed
314 * node to its caller. Note that the replace operation does not need to
315 * iteract with the "removal owner" flag, because it validates that
316 * the "removed" flag is not set before performing its cmpxchg.
317 */
318 #define REMOVED_FLAG (1UL << 0)
319 #define BUCKET_FLAG (1UL << 1)
320 #define REMOVAL_OWNER_FLAG (1UL << 2)
321 #define FLAGS_MASK ((1UL << 3) - 1)
322
323 /* Value of the end pointer. Should not interact with flags. */
324 #define END_VALUE NULL
325
326 /*
327 * ht_items_count: Split-counters counting the number of node addition
328 * and removal in the table. Only used if the CDS_LFHT_ACCOUNTING flag
329 * is set at hash table creation.
330 *
331 * These are free-running counters, never reset to zero. They count the
332 * number of add/remove, and trigger every (1 << COUNT_COMMIT_ORDER)
333 * operations to update the global counter. We choose a power-of-2 value
334 * for the trigger to deal with 32 or 64-bit overflow of the counter.
335 */
336 struct ht_items_count {
337 unsigned long add, del;
338 } __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
339
340 /*
341 * resize_work: Contains arguments passed to worker thread
342 * responsible for performing lazy resize.
343 */
344 struct resize_work {
345 struct urcu_work work;
346 struct cds_lfht *ht;
347 };
348
349 /*
350 * partition_resize_work: Contains arguments passed to worker threads
351 * executing the hash table resize on partitions of the hash table
352 * assigned to each processor's worker thread.
353 */
354 struct partition_resize_work {
355 pthread_t thread_id;
356 struct cds_lfht *ht;
357 unsigned long i, start, len;
358 void (*fct)(struct cds_lfht *ht, unsigned long i,
359 unsigned long start, unsigned long len);
360 };
361
362 static struct urcu_workqueue *cds_lfht_workqueue;
363 static unsigned long cds_lfht_workqueue_user_count;
364
365 /*
366 * Mutex ensuring mutual exclusion between workqueue initialization and
367 * fork handlers. cds_lfht_fork_mutex nests inside call_rcu_mutex.
368 */
369 static pthread_mutex_t cds_lfht_fork_mutex = PTHREAD_MUTEX_INITIALIZER;
370
371 static struct urcu_atfork cds_lfht_atfork;
372
373 /*
374 * atfork handler nesting counters. Handle being registered to many urcu
375 * flavors, thus being possibly invoked more than once in the
376 * pthread_atfork list of callbacks.
377 */
378 static int cds_lfht_workqueue_atfork_nesting;
379
380 static void cds_lfht_init_worker(const struct rcu_flavor_struct *flavor);
381 static void cds_lfht_fini_worker(const struct rcu_flavor_struct *flavor);
382
383 /*
384 * Algorithm to reverse bits in a word by lookup table, extended to
385 * 64-bit words.
386 * Source:
387 * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable
388 * Originally from Public Domain.
389 */
390
391 static const uint8_t BitReverseTable256[256] =
392 {
393 #define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64
394 #define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16)
395 #define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 )
396 R6(0), R6(2), R6(1), R6(3)
397 };
398 #undef R2
399 #undef R4
400 #undef R6
401
402 static
403 uint8_t bit_reverse_u8(uint8_t v)
404 {
405 return BitReverseTable256[v];
406 }
407
408 #if (CAA_BITS_PER_LONG == 32)
409 static
410 uint32_t bit_reverse_u32(uint32_t v)
411 {
412 return ((uint32_t) bit_reverse_u8(v) << 24) |
413 ((uint32_t) bit_reverse_u8(v >> 8) << 16) |
414 ((uint32_t) bit_reverse_u8(v >> 16) << 8) |
415 ((uint32_t) bit_reverse_u8(v >> 24));
416 }
417 #else
418 static
419 uint64_t bit_reverse_u64(uint64_t v)
420 {
421 return ((uint64_t) bit_reverse_u8(v) << 56) |
422 ((uint64_t) bit_reverse_u8(v >> 8) << 48) |
423 ((uint64_t) bit_reverse_u8(v >> 16) << 40) |
424 ((uint64_t) bit_reverse_u8(v >> 24) << 32) |
425 ((uint64_t) bit_reverse_u8(v >> 32) << 24) |
426 ((uint64_t) bit_reverse_u8(v >> 40) << 16) |
427 ((uint64_t) bit_reverse_u8(v >> 48) << 8) |
428 ((uint64_t) bit_reverse_u8(v >> 56));
429 }
430 #endif
431
432 static
433 unsigned long bit_reverse_ulong(unsigned long v)
434 {
435 #if (CAA_BITS_PER_LONG == 32)
436 return bit_reverse_u32(v);
437 #else
438 return bit_reverse_u64(v);
439 #endif
440 }
441
442 /*
443 * fls: returns the position of the most significant bit.
444 * Returns 0 if no bit is set, else returns the position of the most
445 * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit).
446 */
447 #if defined(__i386) || defined(__x86_64)
448 static inline
449 unsigned int fls_u32(uint32_t x)
450 {
451 int r;
452
453 __asm__ ("bsrl %1,%0\n\t"
454 "jnz 1f\n\t"
455 "movl $-1,%0\n\t"
456 "1:\n\t"
457 : "=r" (r) : "rm" (x));
458 return r + 1;
459 }
460 #define HAS_FLS_U32
461 #endif
462
463 #if defined(__x86_64)
464 static inline
465 unsigned int fls_u64(uint64_t x)
466 {
467 long r;
468
469 __asm__ ("bsrq %1,%0\n\t"
470 "jnz 1f\n\t"
471 "movq $-1,%0\n\t"
472 "1:\n\t"
473 : "=r" (r) : "rm" (x));
474 return r + 1;
475 }
476 #define HAS_FLS_U64
477 #endif
478
479 #ifndef HAS_FLS_U64
480 static __attribute__((unused))
481 unsigned int fls_u64(uint64_t x)
482 {
483 unsigned int r = 64;
484
485 if (!x)
486 return 0;
487
488 if (!(x & 0xFFFFFFFF00000000ULL)) {
489 x <<= 32;
490 r -= 32;
491 }
492 if (!(x & 0xFFFF000000000000ULL)) {
493 x <<= 16;
494 r -= 16;
495 }
496 if (!(x & 0xFF00000000000000ULL)) {
497 x <<= 8;
498 r -= 8;
499 }
500 if (!(x & 0xF000000000000000ULL)) {
501 x <<= 4;
502 r -= 4;
503 }
504 if (!(x & 0xC000000000000000ULL)) {
505 x <<= 2;
506 r -= 2;
507 }
508 if (!(x & 0x8000000000000000ULL)) {
509 x <<= 1;
510 r -= 1;
511 }
512 return r;
513 }
514 #endif
515
516 #ifndef HAS_FLS_U32
517 static __attribute__((unused))
518 unsigned int fls_u32(uint32_t x)
519 {
520 unsigned int r = 32;
521
522 if (!x)
523 return 0;
524 if (!(x & 0xFFFF0000U)) {
525 x <<= 16;
526 r -= 16;
527 }
528 if (!(x & 0xFF000000U)) {
529 x <<= 8;
530 r -= 8;
531 }
532 if (!(x & 0xF0000000U)) {
533 x <<= 4;
534 r -= 4;
535 }
536 if (!(x & 0xC0000000U)) {
537 x <<= 2;
538 r -= 2;
539 }
540 if (!(x & 0x80000000U)) {
541 x <<= 1;
542 r -= 1;
543 }
544 return r;
545 }
546 #endif
547
548 unsigned int cds_lfht_fls_ulong(unsigned long x)
549 {
550 #if (CAA_BITS_PER_LONG == 32)
551 return fls_u32(x);
552 #else
553 return fls_u64(x);
554 #endif
555 }
556
557 /*
558 * Return the minimum order for which x <= (1UL << order).
559 * Return -1 if x is 0.
560 */
561 int cds_lfht_get_count_order_u32(uint32_t x)
562 {
563 if (!x)
564 return -1;
565
566 return fls_u32(x - 1);
567 }
568
569 /*
570 * Return the minimum order for which x <= (1UL << order).
571 * Return -1 if x is 0.
572 */
573 int cds_lfht_get_count_order_ulong(unsigned long x)
574 {
575 if (!x)
576 return -1;
577
578 return cds_lfht_fls_ulong(x - 1);
579 }
580
581 static
582 void cds_lfht_resize_lazy_grow(struct cds_lfht *ht, unsigned long size, int growth);
583
584 static
585 void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
586 unsigned long count);
587
588 static long nr_cpus_mask = -1;
589 static long split_count_mask = -1;
590 static int split_count_order = -1;
591
592 #if defined(HAVE_SYSCONF)
593 static void ht_init_nr_cpus_mask(void)
594 {
595 long maxcpus;
596
597 maxcpus = sysconf(_SC_NPROCESSORS_CONF);
598 if (maxcpus <= 0) {
599 nr_cpus_mask = -2;
600 return;
601 }
602 /*
603 * round up number of CPUs to next power of two, so we
604 * can use & for modulo.
605 */
606 maxcpus = 1UL << cds_lfht_get_count_order_ulong(maxcpus);
607 nr_cpus_mask = maxcpus - 1;
608 }
609 #else /* #if defined(HAVE_SYSCONF) */
610 static void ht_init_nr_cpus_mask(void)
611 {
612 nr_cpus_mask = -2;
613 }
614 #endif /* #else #if defined(HAVE_SYSCONF) */
615
616 static
617 void alloc_split_items_count(struct cds_lfht *ht)
618 {
619 if (nr_cpus_mask == -1) {
620 ht_init_nr_cpus_mask();
621 if (nr_cpus_mask < 0)
622 split_count_mask = DEFAULT_SPLIT_COUNT_MASK;
623 else
624 split_count_mask = nr_cpus_mask;
625 split_count_order =
626 cds_lfht_get_count_order_ulong(split_count_mask + 1);
627 }
628
629 assert(split_count_mask >= 0);
630
631 if (ht->flags & CDS_LFHT_ACCOUNTING) {
632 ht->split_count = calloc(split_count_mask + 1,
633 sizeof(struct ht_items_count));
634 assert(ht->split_count);
635 } else {
636 ht->split_count = NULL;
637 }
638 }
639
640 static
641 void free_split_items_count(struct cds_lfht *ht)
642 {
643 poison_free(ht->split_count);
644 }
645
646 static
647 int ht_get_split_count_index(unsigned long hash)
648 {
649 int cpu;
650
651 assert(split_count_mask >= 0);
652 cpu = urcu_sched_getcpu();
653 if (caa_unlikely(cpu < 0))
654 return hash & split_count_mask;
655 else
656 return cpu & split_count_mask;
657 }
658
659 static
660 void ht_count_add(struct cds_lfht *ht, unsigned long size, unsigned long hash)
661 {
662 unsigned long split_count;
663 int index;
664 long count;
665
666 if (caa_unlikely(!ht->split_count))
667 return;
668 index = ht_get_split_count_index(hash);
669 split_count = uatomic_add_return(&ht->split_count[index].add, 1);
670 if (caa_likely(split_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))
671 return;
672 /* Only if number of add multiple of 1UL << COUNT_COMMIT_ORDER */
673
674 dbg_printf("add split count %lu\n", split_count);
675 count = uatomic_add_return(&ht->count,
676 1UL << COUNT_COMMIT_ORDER);
677 if (caa_likely(count & (count - 1)))
678 return;
679 /* Only if global count is power of 2 */
680
681 if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) < size)
682 return;
683 dbg_printf("add set global %ld\n", count);
684 cds_lfht_resize_lazy_count(ht, size,
685 count >> (CHAIN_LEN_TARGET - 1));
686 }
687
688 static
689 void ht_count_del(struct cds_lfht *ht, unsigned long size, unsigned long hash)
690 {
691 unsigned long split_count;
692 int index;
693 long count;
694
695 if (caa_unlikely(!ht->split_count))
696 return;
697 index = ht_get_split_count_index(hash);
698 split_count = uatomic_add_return(&ht->split_count[index].del, 1);
699 if (caa_likely(split_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))
700 return;
701 /* Only if number of deletes multiple of 1UL << COUNT_COMMIT_ORDER */
702
703 dbg_printf("del split count %lu\n", split_count);
704 count = uatomic_add_return(&ht->count,
705 -(1UL << COUNT_COMMIT_ORDER));
706 if (caa_likely(count & (count - 1)))
707 return;
708 /* Only if global count is power of 2 */
709
710 if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) >= size)
711 return;
712 dbg_printf("del set global %ld\n", count);
713 /*
714 * Don't shrink table if the number of nodes is below a
715 * certain threshold.
716 */
717 if (count < (1UL << COUNT_COMMIT_ORDER) * (split_count_mask + 1))
718 return;
719 cds_lfht_resize_lazy_count(ht, size,
720 count >> (CHAIN_LEN_TARGET - 1));
721 }
722
723 static
724 void check_resize(struct cds_lfht *ht, unsigned long size, uint32_t chain_len)
725 {
726 unsigned long count;
727
728 if (!(ht->flags & CDS_LFHT_AUTO_RESIZE))
729 return;
730 count = uatomic_read(&ht->count);
731 /*
732 * Use bucket-local length for small table expand and for
733 * environments lacking per-cpu data support.
734 */
735 if (count >= (1UL << (COUNT_COMMIT_ORDER + split_count_order)))
736 return;
737 if (chain_len > 100)
738 dbg_printf("WARNING: large chain length: %u.\n",
739 chain_len);
740 if (chain_len >= CHAIN_LEN_RESIZE_THRESHOLD) {
741 int growth;
742
743 /*
744 * Ideal growth calculated based on chain length.
745 */
746 growth = cds_lfht_get_count_order_u32(chain_len
747 - (CHAIN_LEN_TARGET - 1));
748 if ((ht->flags & CDS_LFHT_ACCOUNTING)
749 && (size << growth)
750 >= (1UL << (COUNT_COMMIT_ORDER
751 + split_count_order))) {
752 /*
753 * If ideal growth expands the hash table size
754 * beyond the "small hash table" sizes, use the
755 * maximum small hash table size to attempt
756 * expanding the hash table. This only applies
757 * when node accounting is available, otherwise
758 * the chain length is used to expand the hash
759 * table in every case.
760 */
761 growth = COUNT_COMMIT_ORDER + split_count_order
762 - cds_lfht_get_count_order_ulong(size);
763 if (growth <= 0)
764 return;
765 }
766 cds_lfht_resize_lazy_grow(ht, size, growth);
767 }
768 }
769
770 static
771 struct cds_lfht_node *clear_flag(struct cds_lfht_node *node)
772 {
773 return (struct cds_lfht_node *) (((unsigned long) node) & ~FLAGS_MASK);
774 }
775
776 static
777 int is_removed(struct cds_lfht_node *node)
778 {
779 return ((unsigned long) node) & REMOVED_FLAG;
780 }
781
782 static
783 int is_bucket(struct cds_lfht_node *node)
784 {
785 return ((unsigned long) node) & BUCKET_FLAG;
786 }
787
788 static
789 struct cds_lfht_node *flag_bucket(struct cds_lfht_node *node)
790 {
791 return (struct cds_lfht_node *) (((unsigned long) node) | BUCKET_FLAG);
792 }
793
794 static
795 int is_removal_owner(struct cds_lfht_node *node)
796 {
797 return ((unsigned long) node) & REMOVAL_OWNER_FLAG;
798 }
799
800 static
801 struct cds_lfht_node *flag_removal_owner(struct cds_lfht_node *node)
802 {
803 return (struct cds_lfht_node *) (((unsigned long) node) | REMOVAL_OWNER_FLAG);
804 }
805
806 static
807 struct cds_lfht_node *flag_removed_or_removal_owner(struct cds_lfht_node *node)
808 {
809 return (struct cds_lfht_node *) (((unsigned long) node) | REMOVED_FLAG | REMOVAL_OWNER_FLAG);
810 }
811
812 static
813 struct cds_lfht_node *get_end(void)
814 {
815 return (struct cds_lfht_node *) END_VALUE;
816 }
817
818 static
819 int is_end(struct cds_lfht_node *node)
820 {
821 return clear_flag(node) == (struct cds_lfht_node *) END_VALUE;
822 }
823
824 static
825 unsigned long _uatomic_xchg_monotonic_increase(unsigned long *ptr,
826 unsigned long v)
827 {
828 unsigned long old1, old2;
829
830 old1 = uatomic_read(ptr);
831 do {
832 old2 = old1;
833 if (old2 >= v)
834 return old2;
835 } while ((old1 = uatomic_cmpxchg(ptr, old2, v)) != old2);
836 return old2;
837 }
838
839 static
840 void cds_lfht_alloc_bucket_table(struct cds_lfht *ht, unsigned long order)
841 {
842 return ht->mm->alloc_bucket_table(ht, order);
843 }
844
845 /*
846 * cds_lfht_free_bucket_table() should be called with decreasing order.
847 * When cds_lfht_free_bucket_table(0) is called, it means the whole
848 * lfht is destroyed.
849 */
850 static
851 void cds_lfht_free_bucket_table(struct cds_lfht *ht, unsigned long order)
852 {
853 return ht->mm->free_bucket_table(ht, order);
854 }
855
856 static inline
857 struct cds_lfht_node *bucket_at(struct cds_lfht *ht, unsigned long index)
858 {
859 return ht->bucket_at(ht, index);
860 }
861
862 static inline
863 struct cds_lfht_node *lookup_bucket(struct cds_lfht *ht, unsigned long size,
864 unsigned long hash)
865 {
866 assert(size > 0);
867 return bucket_at(ht, hash & (size - 1));
868 }
869
870 /*
871 * Remove all logically deleted nodes from a bucket up to a certain node key.
872 */
873 static
874 void _cds_lfht_gc_bucket(struct cds_lfht_node *bucket, struct cds_lfht_node *node)
875 {
876 struct cds_lfht_node *iter_prev, *iter, *next, *new_next;
877
878 assert(!is_bucket(bucket));
879 assert(!is_removed(bucket));
880 assert(!is_removal_owner(bucket));
881 assert(!is_bucket(node));
882 assert(!is_removed(node));
883 assert(!is_removal_owner(node));
884 for (;;) {
885 iter_prev = bucket;
886 /* We can always skip the bucket node initially */
887 iter = rcu_dereference(iter_prev->next);
888 assert(!is_removed(iter));
889 assert(!is_removal_owner(iter));
890 assert(iter_prev->reverse_hash <= node->reverse_hash);
891 /*
892 * We should never be called with bucket (start of chain)
893 * and logically removed node (end of path compression
894 * marker) being the actual same node. This would be a
895 * bug in the algorithm implementation.
896 */
897 assert(bucket != node);
898 for (;;) {
899 if (caa_unlikely(is_end(iter)))
900 return;
901 if (caa_likely(clear_flag(iter)->reverse_hash > node->reverse_hash))
902 return;
903 next = rcu_dereference(clear_flag(iter)->next);
904 if (caa_likely(is_removed(next)))
905 break;
906 iter_prev = clear_flag(iter);
907 iter = next;
908 }
909 assert(!is_removed(iter));
910 assert(!is_removal_owner(iter));
911 if (is_bucket(iter))
912 new_next = flag_bucket(clear_flag(next));
913 else
914 new_next = clear_flag(next);
915 (void) uatomic_cmpxchg(&iter_prev->next, iter, new_next);
916 }
917 }
918
919 static
920 int _cds_lfht_replace(struct cds_lfht *ht, unsigned long size,
921 struct cds_lfht_node *old_node,
922 struct cds_lfht_node *old_next,
923 struct cds_lfht_node *new_node)
924 {
925 struct cds_lfht_node *bucket, *ret_next;
926
927 if (!old_node) /* Return -ENOENT if asked to replace NULL node */
928 return -ENOENT;
929
930 assert(!is_removed(old_node));
931 assert(!is_removal_owner(old_node));
932 assert(!is_bucket(old_node));
933 assert(!is_removed(new_node));
934 assert(!is_removal_owner(new_node));
935 assert(!is_bucket(new_node));
936 assert(new_node != old_node);
937 for (;;) {
938 /* Insert after node to be replaced */
939 if (is_removed(old_next)) {
940 /*
941 * Too late, the old node has been removed under us
942 * between lookup and replace. Fail.
943 */
944 return -ENOENT;
945 }
946 assert(old_next == clear_flag(old_next));
947 assert(new_node != old_next);
948 /*
949 * REMOVAL_OWNER flag is _NEVER_ set before the REMOVED
950 * flag. It is either set atomically at the same time
951 * (replace) or after (del).
952 */
953 assert(!is_removal_owner(old_next));
954 new_node->next = old_next;
955 /*
956 * Here is the whole trick for lock-free replace: we add
957 * the replacement node _after_ the node we want to
958 * replace by atomically setting its next pointer at the
959 * same time we set its removal flag. Given that
960 * the lookups/get next use an iterator aware of the
961 * next pointer, they will either skip the old node due
962 * to the removal flag and see the new node, or use
963 * the old node, but will not see the new one.
964 * This is a replacement of a node with another node
965 * that has the same value: we are therefore not
966 * removing a value from the hash table. We set both the
967 * REMOVED and REMOVAL_OWNER flags atomically so we own
968 * the node after successful cmpxchg.
969 */
970 ret_next = uatomic_cmpxchg(&old_node->next,
971 old_next, flag_removed_or_removal_owner(new_node));
972 if (ret_next == old_next)
973 break; /* We performed the replacement. */
974 old_next = ret_next;
975 }
976
977 /*
978 * Ensure that the old node is not visible to readers anymore:
979 * lookup for the node, and remove it (along with any other
980 * logically removed node) if found.
981 */
982 bucket = lookup_bucket(ht, size, bit_reverse_ulong(old_node->reverse_hash));
983 _cds_lfht_gc_bucket(bucket, new_node);
984
985 assert(is_removed(CMM_LOAD_SHARED(old_node->next)));
986 return 0;
987 }
988
989 /*
990 * A non-NULL unique_ret pointer uses the "add unique" (or uniquify) add
991 * mode. A NULL unique_ret allows creation of duplicate keys.
992 */
993 static
994 void _cds_lfht_add(struct cds_lfht *ht,
995 unsigned long hash,
996 cds_lfht_match_fct match,
997 const void *key,
998 unsigned long size,
999 struct cds_lfht_node *node,
1000 struct cds_lfht_iter *unique_ret,
1001 int bucket_flag)
1002 {
1003 struct cds_lfht_node *iter_prev, *iter, *next, *new_node, *new_next,
1004 *return_node;
1005 struct cds_lfht_node *bucket;
1006
1007 assert(!is_bucket(node));
1008 assert(!is_removed(node));
1009 assert(!is_removal_owner(node));
1010 bucket = lookup_bucket(ht, size, hash);
1011 for (;;) {
1012 uint32_t chain_len = 0;
1013
1014 /*
1015 * iter_prev points to the non-removed node prior to the
1016 * insert location.
1017 */
1018 iter_prev = bucket;
1019 /* We can always skip the bucket node initially */
1020 iter = rcu_dereference(iter_prev->next);
1021 assert(iter_prev->reverse_hash <= node->reverse_hash);
1022 for (;;) {
1023 if (caa_unlikely(is_end(iter)))
1024 goto insert;
1025 if (caa_likely(clear_flag(iter)->reverse_hash > node->reverse_hash))
1026 goto insert;
1027
1028 /* bucket node is the first node of the identical-hash-value chain */
1029 if (bucket_flag && clear_flag(iter)->reverse_hash == node->reverse_hash)
1030 goto insert;
1031
1032 next = rcu_dereference(clear_flag(iter)->next);
1033 if (caa_unlikely(is_removed(next)))
1034 goto gc_node;
1035
1036 /* uniquely add */
1037 if (unique_ret
1038 && !is_bucket(next)
1039 && clear_flag(iter)->reverse_hash == node->reverse_hash) {
1040 struct cds_lfht_iter d_iter = { .node = node, .next = iter, };
1041
1042 /*
1043 * uniquely adding inserts the node as the first
1044 * node of the identical-hash-value node chain.
1045 *
1046 * This semantic ensures no duplicated keys
1047 * should ever be observable in the table
1048 * (including traversing the table node by
1049 * node by forward iterations)
1050 */
1051 cds_lfht_next_duplicate(ht, match, key, &d_iter);
1052 if (!d_iter.node)
1053 goto insert;
1054
1055 *unique_ret = d_iter;
1056 return;
1057 }
1058
1059 /* Only account for identical reverse hash once */
1060 if (iter_prev->reverse_hash != clear_flag(iter)->reverse_hash
1061 && !is_bucket(next))
1062 check_resize(ht, size, ++chain_len);
1063 iter_prev = clear_flag(iter);
1064 iter = next;
1065 }
1066
1067 insert:
1068 assert(node != clear_flag(iter));
1069 assert(!is_removed(iter_prev));
1070 assert(!is_removal_owner(iter_prev));
1071 assert(!is_removed(iter));
1072 assert(!is_removal_owner(iter));
1073 assert(iter_prev != node);
1074 if (!bucket_flag)
1075 node->next = clear_flag(iter);
1076 else
1077 node->next = flag_bucket(clear_flag(iter));
1078 if (is_bucket(iter))
1079 new_node = flag_bucket(node);
1080 else
1081 new_node = node;
1082 if (uatomic_cmpxchg(&iter_prev->next, iter,
1083 new_node) != iter) {
1084 continue; /* retry */
1085 } else {
1086 return_node = node;
1087 goto end;
1088 }
1089
1090 gc_node:
1091 assert(!is_removed(iter));
1092 assert(!is_removal_owner(iter));
1093 if (is_bucket(iter))
1094 new_next = flag_bucket(clear_flag(next));
1095 else
1096 new_next = clear_flag(next);
1097 (void) uatomic_cmpxchg(&iter_prev->next, iter, new_next);
1098 /* retry */
1099 }
1100 end:
1101 if (unique_ret) {
1102 unique_ret->node = return_node;
1103 /* unique_ret->next left unset, never used. */
1104 }
1105 }
1106
1107 static
1108 int _cds_lfht_del(struct cds_lfht *ht, unsigned long size,
1109 struct cds_lfht_node *node)
1110 {
1111 struct cds_lfht_node *bucket, *next;
1112
1113 if (!node) /* Return -ENOENT if asked to delete NULL node */
1114 return -ENOENT;
1115
1116 /* logically delete the node */
1117 assert(!is_bucket(node));
1118 assert(!is_removed(node));
1119 assert(!is_removal_owner(node));
1120
1121 /*
1122 * We are first checking if the node had previously been
1123 * logically removed (this check is not atomic with setting the
1124 * logical removal flag). Return -ENOENT if the node had
1125 * previously been removed.
1126 */
1127 next = CMM_LOAD_SHARED(node->next); /* next is not dereferenced */
1128 if (caa_unlikely(is_removed(next)))
1129 return -ENOENT;
1130 assert(!is_bucket(next));
1131 /*
1132 * The del operation semantic guarantees a full memory barrier
1133 * before the uatomic_or atomic commit of the deletion flag.
1134 */
1135 cmm_smp_mb__before_uatomic_or();
1136 /*
1137 * We set the REMOVED_FLAG unconditionally. Note that there may
1138 * be more than one concurrent thread setting this flag.
1139 * Knowing which wins the race will be known after the garbage
1140 * collection phase, stay tuned!
1141 */
1142 uatomic_or(&node->next, REMOVED_FLAG);
1143 /* We performed the (logical) deletion. */
1144
1145 /*
1146 * Ensure that the node is not visible to readers anymore: lookup for
1147 * the node, and remove it (along with any other logically removed node)
1148 * if found.
1149 */
1150 bucket = lookup_bucket(ht, size, bit_reverse_ulong(node->reverse_hash));
1151 _cds_lfht_gc_bucket(bucket, node);
1152
1153 assert(is_removed(CMM_LOAD_SHARED(node->next)));
1154 /*
1155 * Last phase: atomically exchange node->next with a version
1156 * having "REMOVAL_OWNER_FLAG" set. If the returned node->next
1157 * pointer did _not_ have "REMOVAL_OWNER_FLAG" set, we now own
1158 * the node and win the removal race.
1159 * It is interesting to note that all "add" paths are forbidden
1160 * to change the next pointer starting from the point where the
1161 * REMOVED_FLAG is set, so here using a read, followed by a
1162 * xchg() suffice to guarantee that the xchg() will ever only
1163 * set the "REMOVAL_OWNER_FLAG" (or change nothing if the flag
1164 * was already set).
1165 */
1166 if (!is_removal_owner(uatomic_xchg(&node->next,
1167 flag_removal_owner(node->next))))
1168 return 0;
1169 else
1170 return -ENOENT;
1171 }
1172
1173 static
1174 void *partition_resize_thread(void *arg)
1175 {
1176 struct partition_resize_work *work = arg;
1177
1178 work->ht->flavor->register_thread();
1179 work->fct(work->ht, work->i, work->start, work->len);
1180 work->ht->flavor->unregister_thread();
1181 return NULL;
1182 }
1183
1184 static
1185 void partition_resize_helper(struct cds_lfht *ht, unsigned long i,
1186 unsigned long len,
1187 void (*fct)(struct cds_lfht *ht, unsigned long i,
1188 unsigned long start, unsigned long len))
1189 {
1190 unsigned long partition_len, start = 0;
1191 struct partition_resize_work *work;
1192 int thread, ret;
1193 unsigned long nr_threads;
1194
1195 assert(nr_cpus_mask != -1);
1196 if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD)
1197 goto fallback;
1198
1199 /*
1200 * Note: nr_cpus_mask + 1 is always power of 2.
1201 * We spawn just the number of threads we need to satisfy the minimum
1202 * partition size, up to the number of CPUs in the system.
1203 */
1204 if (nr_cpus_mask > 0) {
1205 nr_threads = min(nr_cpus_mask + 1,
1206 len >> MIN_PARTITION_PER_THREAD_ORDER);
1207 } else {
1208 nr_threads = 1;
1209 }
1210 partition_len = len >> cds_lfht_get_count_order_ulong(nr_threads);
1211 work = calloc(nr_threads, sizeof(*work));
1212 if (!work) {
1213 dbg_printf("error allocating for resize, single-threading\n");
1214 goto fallback;
1215 }
1216 for (thread = 0; thread < nr_threads; thread++) {
1217 work[thread].ht = ht;
1218 work[thread].i = i;
1219 work[thread].len = partition_len;
1220 work[thread].start = thread * partition_len;
1221 work[thread].fct = fct;
1222 ret = pthread_create(&(work[thread].thread_id), ht->resize_attr,
1223 partition_resize_thread, &work[thread]);
1224 if (ret == EAGAIN) {
1225 /*
1226 * Out of resources: wait and join the threads
1227 * we've created, then handle leftovers.
1228 */
1229 dbg_printf("error spawning for resize, single-threading\n");
1230 start = work[thread].start;
1231 len -= start;
1232 nr_threads = thread;
1233 break;
1234 }
1235 assert(!ret);
1236 }
1237 for (thread = 0; thread < nr_threads; thread++) {
1238 ret = pthread_join(work[thread].thread_id, NULL);
1239 assert(!ret);
1240 }
1241 free(work);
1242
1243 /*
1244 * A pthread_create failure above will either lead in us having
1245 * no threads to join or starting at a non-zero offset,
1246 * fallback to single thread processing of leftovers.
1247 */
1248 if (start == 0 && nr_threads > 0)
1249 return;
1250 fallback:
1251 fct(ht, i, start, len);
1252 }
1253
1254 /*
1255 * Holding RCU read lock to protect _cds_lfht_add against memory
1256 * reclaim that could be performed by other worker threads (ABA
1257 * problem).
1258 *
1259 * When we reach a certain length, we can split this population phase over
1260 * many worker threads, based on the number of CPUs available in the system.
1261 * This should therefore take care of not having the expand lagging behind too
1262 * many concurrent insertion threads by using the scheduler's ability to
1263 * schedule bucket node population fairly with insertions.
1264 */
1265 static
1266 void init_table_populate_partition(struct cds_lfht *ht, unsigned long i,
1267 unsigned long start, unsigned long len)
1268 {
1269 unsigned long j, size = 1UL << (i - 1);
1270
1271 assert(i > MIN_TABLE_ORDER);
1272 ht->flavor->read_lock();
1273 for (j = size + start; j < size + start + len; j++) {
1274 struct cds_lfht_node *new_node = bucket_at(ht, j);
1275
1276 assert(j >= size && j < (size << 1));
1277 dbg_printf("init populate: order %lu index %lu hash %lu\n",
1278 i, j, j);
1279 new_node->reverse_hash = bit_reverse_ulong(j);
1280 _cds_lfht_add(ht, j, NULL, NULL, size, new_node, NULL, 1);
1281 }
1282 ht->flavor->read_unlock();
1283 }
1284
1285 static
1286 void init_table_populate(struct cds_lfht *ht, unsigned long i,
1287 unsigned long len)
1288 {
1289 partition_resize_helper(ht, i, len, init_table_populate_partition);
1290 }
1291
1292 static
1293 void init_table(struct cds_lfht *ht,
1294 unsigned long first_order, unsigned long last_order)
1295 {
1296 unsigned long i;
1297
1298 dbg_printf("init table: first_order %lu last_order %lu\n",
1299 first_order, last_order);
1300 assert(first_order > MIN_TABLE_ORDER);
1301 for (i = first_order; i <= last_order; i++) {
1302 unsigned long len;
1303
1304 len = 1UL << (i - 1);
1305 dbg_printf("init order %lu len: %lu\n", i, len);
1306
1307 /* Stop expand if the resize target changes under us */
1308 if (CMM_LOAD_SHARED(ht->resize_target) < (1UL << i))
1309 break;
1310
1311 cds_lfht_alloc_bucket_table(ht, i);
1312
1313 /*
1314 * Set all bucket nodes reverse hash values for a level and
1315 * link all bucket nodes into the table.
1316 */
1317 init_table_populate(ht, i, len);
1318
1319 /*
1320 * Update table size.
1321 */
1322 cmm_smp_wmb(); /* populate data before RCU size */
1323 CMM_STORE_SHARED(ht->size, 1UL << i);
1324
1325 dbg_printf("init new size: %lu\n", 1UL << i);
1326 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1327 break;
1328 }
1329 }
1330
1331 /*
1332 * Holding RCU read lock to protect _cds_lfht_remove against memory
1333 * reclaim that could be performed by other worker threads (ABA
1334 * problem).
1335 * For a single level, we logically remove and garbage collect each node.
1336 *
1337 * As a design choice, we perform logical removal and garbage collection on a
1338 * node-per-node basis to simplify this algorithm. We also assume keeping good
1339 * cache locality of the operation would overweight possible performance gain
1340 * that could be achieved by batching garbage collection for multiple levels.
1341 * However, this would have to be justified by benchmarks.
1342 *
1343 * Concurrent removal and add operations are helping us perform garbage
1344 * collection of logically removed nodes. We guarantee that all logically
1345 * removed nodes have been garbage-collected (unlinked) before work
1346 * enqueue is invoked to free a hole level of bucket nodes (after a
1347 * grace period).
1348 *
1349 * Logical removal and garbage collection can therefore be done in batch
1350 * or on a node-per-node basis, as long as the guarantee above holds.
1351 *
1352 * When we reach a certain length, we can split this removal over many worker
1353 * threads, based on the number of CPUs available in the system. This should
1354 * take care of not letting resize process lag behind too many concurrent
1355 * updater threads actively inserting into the hash table.
1356 */
1357 static
1358 void remove_table_partition(struct cds_lfht *ht, unsigned long i,
1359 unsigned long start, unsigned long len)
1360 {
1361 unsigned long j, size = 1UL << (i - 1);
1362
1363 assert(i > MIN_TABLE_ORDER);
1364 ht->flavor->read_lock();
1365 for (j = size + start; j < size + start + len; j++) {
1366 struct cds_lfht_node *fini_bucket = bucket_at(ht, j);
1367 struct cds_lfht_node *parent_bucket = bucket_at(ht, j - size);
1368
1369 assert(j >= size && j < (size << 1));
1370 dbg_printf("remove entry: order %lu index %lu hash %lu\n",
1371 i, j, j);
1372 /* Set the REMOVED_FLAG to freeze the ->next for gc */
1373 uatomic_or(&fini_bucket->next, REMOVED_FLAG);
1374 _cds_lfht_gc_bucket(parent_bucket, fini_bucket);
1375 }
1376 ht->flavor->read_unlock();
1377 }
1378
1379 static
1380 void remove_table(struct cds_lfht *ht, unsigned long i, unsigned long len)
1381 {
1382 partition_resize_helper(ht, i, len, remove_table_partition);
1383 }
1384
1385 /*
1386 * fini_table() is never called for first_order == 0, which is why
1387 * free_by_rcu_order == 0 can be used as criterion to know if free must
1388 * be called.
1389 */
1390 static
1391 void fini_table(struct cds_lfht *ht,
1392 unsigned long first_order, unsigned long last_order)
1393 {
1394 long i;
1395 unsigned long free_by_rcu_order = 0;
1396
1397 dbg_printf("fini table: first_order %lu last_order %lu\n",
1398 first_order, last_order);
1399 assert(first_order > MIN_TABLE_ORDER);
1400 for (i = last_order; i >= first_order; i--) {
1401 unsigned long len;
1402
1403 len = 1UL << (i - 1);
1404 dbg_printf("fini order %ld len: %lu\n", i, len);
1405
1406 /* Stop shrink if the resize target changes under us */
1407 if (CMM_LOAD_SHARED(ht->resize_target) > (1UL << (i - 1)))
1408 break;
1409
1410 cmm_smp_wmb(); /* populate data before RCU size */
1411 CMM_STORE_SHARED(ht->size, 1UL << (i - 1));
1412
1413 /*
1414 * We need to wait for all add operations to reach Q.S. (and
1415 * thus use the new table for lookups) before we can start
1416 * releasing the old bucket nodes. Otherwise their lookup will
1417 * return a logically removed node as insert position.
1418 */
1419 ht->flavor->update_synchronize_rcu();
1420 if (free_by_rcu_order)
1421 cds_lfht_free_bucket_table(ht, free_by_rcu_order);
1422
1423 /*
1424 * Set "removed" flag in bucket nodes about to be removed.
1425 * Unlink all now-logically-removed bucket node pointers.
1426 * Concurrent add/remove operation are helping us doing
1427 * the gc.
1428 */
1429 remove_table(ht, i, len);
1430
1431 free_by_rcu_order = i;
1432
1433 dbg_printf("fini new size: %lu\n", 1UL << i);
1434 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1435 break;
1436 }
1437
1438 if (free_by_rcu_order) {
1439 ht->flavor->update_synchronize_rcu();
1440 cds_lfht_free_bucket_table(ht, free_by_rcu_order);
1441 }
1442 }
1443
1444 static
1445 void cds_lfht_create_bucket(struct cds_lfht *ht, unsigned long size)
1446 {
1447 struct cds_lfht_node *prev, *node;
1448 unsigned long order, len, i;
1449
1450 cds_lfht_alloc_bucket_table(ht, 0);
1451
1452 dbg_printf("create bucket: order 0 index 0 hash 0\n");
1453 node = bucket_at(ht, 0);
1454 node->next = flag_bucket(get_end());
1455 node->reverse_hash = 0;
1456
1457 for (order = 1; order < cds_lfht_get_count_order_ulong(size) + 1; order++) {
1458 len = 1UL << (order - 1);
1459 cds_lfht_alloc_bucket_table(ht, order);
1460
1461 for (i = 0; i < len; i++) {
1462 /*
1463 * Now, we are trying to init the node with the
1464 * hash=(len+i) (which is also a bucket with the
1465 * index=(len+i)) and insert it into the hash table,
1466 * so this node has to be inserted after the bucket
1467 * with the index=(len+i)&(len-1)=i. And because there
1468 * is no other non-bucket node nor bucket node with
1469 * larger index/hash inserted, so the bucket node
1470 * being inserted should be inserted directly linked
1471 * after the bucket node with index=i.
1472 */
1473 prev = bucket_at(ht, i);
1474 node = bucket_at(ht, len + i);
1475
1476 dbg_printf("create bucket: order %lu index %lu hash %lu\n",
1477 order, len + i, len + i);
1478 node->reverse_hash = bit_reverse_ulong(len + i);
1479
1480 /* insert after prev */
1481 assert(is_bucket(prev->next));
1482 node->next = prev->next;
1483 prev->next = flag_bucket(node);
1484 }
1485 }
1486 }
1487
1488 struct cds_lfht *_cds_lfht_new(unsigned long init_size,
1489 unsigned long min_nr_alloc_buckets,
1490 unsigned long max_nr_buckets,
1491 int flags,
1492 const struct cds_lfht_mm_type *mm,
1493 const struct rcu_flavor_struct *flavor,
1494 pthread_attr_t *attr)
1495 {
1496 struct cds_lfht *ht;
1497 unsigned long order;
1498
1499 /* min_nr_alloc_buckets must be power of two */
1500 if (!min_nr_alloc_buckets || (min_nr_alloc_buckets & (min_nr_alloc_buckets - 1)))
1501 return NULL;
1502
1503 /* init_size must be power of two */
1504 if (!init_size || (init_size & (init_size - 1)))
1505 return NULL;
1506
1507 /*
1508 * Memory management plugin default.
1509 */
1510 if (!mm) {
1511 if (CAA_BITS_PER_LONG > 32
1512 && max_nr_buckets
1513 && max_nr_buckets <= (1ULL << 32)) {
1514 /*
1515 * For 64-bit architectures, with max number of
1516 * buckets small enough not to use the entire
1517 * 64-bit memory mapping space (and allowing a
1518 * fair number of hash table instances), use the
1519 * mmap allocator, which is faster than the
1520 * order allocator.
1521 */
1522 mm = &cds_lfht_mm_mmap;
1523 } else {
1524 /*
1525 * The fallback is to use the order allocator.
1526 */
1527 mm = &cds_lfht_mm_order;
1528 }
1529 }
1530
1531 /* max_nr_buckets == 0 for order based mm means infinite */
1532 if (mm == &cds_lfht_mm_order && !max_nr_buckets)
1533 max_nr_buckets = 1UL << (MAX_TABLE_ORDER - 1);
1534
1535 /* max_nr_buckets must be power of two */
1536 if (!max_nr_buckets || (max_nr_buckets & (max_nr_buckets - 1)))
1537 return NULL;
1538
1539 if (flags & CDS_LFHT_AUTO_RESIZE)
1540 cds_lfht_init_worker(flavor);
1541
1542 min_nr_alloc_buckets = max(min_nr_alloc_buckets, MIN_TABLE_SIZE);
1543 init_size = max(init_size, MIN_TABLE_SIZE);
1544 max_nr_buckets = max(max_nr_buckets, min_nr_alloc_buckets);
1545 init_size = min(init_size, max_nr_buckets);
1546
1547 ht = mm->alloc_cds_lfht(min_nr_alloc_buckets, max_nr_buckets);
1548 assert(ht);
1549 assert(ht->mm == mm);
1550 assert(ht->bucket_at == mm->bucket_at);
1551
1552 ht->flags = flags;
1553 ht->flavor = flavor;
1554 ht->resize_attr = attr;
1555 alloc_split_items_count(ht);
1556 /* this mutex should not nest in read-side C.S. */
1557 pthread_mutex_init(&ht->resize_mutex, NULL);
1558 order = cds_lfht_get_count_order_ulong(init_size);
1559 ht->resize_target = 1UL << order;
1560 cds_lfht_create_bucket(ht, 1UL << order);
1561 ht->size = 1UL << order;
1562 return ht;
1563 }
1564
1565 void cds_lfht_lookup(struct cds_lfht *ht, unsigned long hash,
1566 cds_lfht_match_fct match, const void *key,
1567 struct cds_lfht_iter *iter)
1568 {
1569 struct cds_lfht_node *node, *next, *bucket;
1570 unsigned long reverse_hash, size;
1571
1572 reverse_hash = bit_reverse_ulong(hash);
1573
1574 size = rcu_dereference(ht->size);
1575 bucket = lookup_bucket(ht, size, hash);
1576 /* We can always skip the bucket node initially */
1577 node = rcu_dereference(bucket->next);
1578 node = clear_flag(node);
1579 for (;;) {
1580 if (caa_unlikely(is_end(node))) {
1581 node = next = NULL;
1582 break;
1583 }
1584 if (caa_unlikely(node->reverse_hash > reverse_hash)) {
1585 node = next = NULL;
1586 break;
1587 }
1588 next = rcu_dereference(node->next);
1589 assert(node == clear_flag(node));
1590 if (caa_likely(!is_removed(next))
1591 && !is_bucket(next)
1592 && node->reverse_hash == reverse_hash
1593 && caa_likely(match(node, key))) {
1594 break;
1595 }
1596 node = clear_flag(next);
1597 }
1598 assert(!node || !is_bucket(CMM_LOAD_SHARED(node->next)));
1599 iter->node = node;
1600 iter->next = next;
1601 }
1602
1603 void cds_lfht_next_duplicate(struct cds_lfht *ht, cds_lfht_match_fct match,
1604 const void *key, struct cds_lfht_iter *iter)
1605 {
1606 struct cds_lfht_node *node, *next;
1607 unsigned long reverse_hash;
1608
1609 node = iter->node;
1610 reverse_hash = node->reverse_hash;
1611 next = iter->next;
1612 node = clear_flag(next);
1613
1614 for (;;) {
1615 if (caa_unlikely(is_end(node))) {
1616 node = next = NULL;
1617 break;
1618 }
1619 if (caa_unlikely(node->reverse_hash > reverse_hash)) {
1620 node = next = NULL;
1621 break;
1622 }
1623 next = rcu_dereference(node->next);
1624 if (caa_likely(!is_removed(next))
1625 && !is_bucket(next)
1626 && caa_likely(match(node, key))) {
1627 break;
1628 }
1629 node = clear_flag(next);
1630 }
1631 assert(!node || !is_bucket(CMM_LOAD_SHARED(node->next)));
1632 iter->node = node;
1633 iter->next = next;
1634 }
1635
1636 void cds_lfht_next(struct cds_lfht *ht, struct cds_lfht_iter *iter)
1637 {
1638 struct cds_lfht_node *node, *next;
1639
1640 node = clear_flag(iter->next);
1641 for (;;) {
1642 if (caa_unlikely(is_end(node))) {
1643 node = next = NULL;
1644 break;
1645 }
1646 next = rcu_dereference(node->next);
1647 if (caa_likely(!is_removed(next))
1648 && !is_bucket(next)) {
1649 break;
1650 }
1651 node = clear_flag(next);
1652 }
1653 assert(!node || !is_bucket(CMM_LOAD_SHARED(node->next)));
1654 iter->node = node;
1655 iter->next = next;
1656 }
1657
1658 void cds_lfht_first(struct cds_lfht *ht, struct cds_lfht_iter *iter)
1659 {
1660 /*
1661 * Get next after first bucket node. The first bucket node is the
1662 * first node of the linked list.
1663 */
1664 iter->next = bucket_at(ht, 0)->next;
1665 cds_lfht_next(ht, iter);
1666 }
1667
1668 void cds_lfht_add(struct cds_lfht *ht, unsigned long hash,
1669 struct cds_lfht_node *node)
1670 {
1671 unsigned long size;
1672
1673 node->reverse_hash = bit_reverse_ulong(hash);
1674 size = rcu_dereference(ht->size);
1675 _cds_lfht_add(ht, hash, NULL, NULL, size, node, NULL, 0);
1676 ht_count_add(ht, size, hash);
1677 }
1678
1679 struct cds_lfht_node *cds_lfht_add_unique(struct cds_lfht *ht,
1680 unsigned long hash,
1681 cds_lfht_match_fct match,
1682 const void *key,
1683 struct cds_lfht_node *node)
1684 {
1685 unsigned long size;
1686 struct cds_lfht_iter iter;
1687
1688 node->reverse_hash = bit_reverse_ulong(hash);
1689 size = rcu_dereference(ht->size);
1690 _cds_lfht_add(ht, hash, match, key, size, node, &iter, 0);
1691 if (iter.node == node)
1692 ht_count_add(ht, size, hash);
1693 return iter.node;
1694 }
1695
1696 struct cds_lfht_node *cds_lfht_add_replace(struct cds_lfht *ht,
1697 unsigned long hash,
1698 cds_lfht_match_fct match,
1699 const void *key,
1700 struct cds_lfht_node *node)
1701 {
1702 unsigned long size;
1703 struct cds_lfht_iter iter;
1704
1705 node->reverse_hash = bit_reverse_ulong(hash);
1706 size = rcu_dereference(ht->size);
1707 for (;;) {
1708 _cds_lfht_add(ht, hash, match, key, size, node, &iter, 0);
1709 if (iter.node == node) {
1710 ht_count_add(ht, size, hash);
1711 return NULL;
1712 }
1713
1714 if (!_cds_lfht_replace(ht, size, iter.node, iter.next, node))
1715 return iter.node;
1716 }
1717 }
1718
1719 int cds_lfht_replace(struct cds_lfht *ht,
1720 struct cds_lfht_iter *old_iter,
1721 unsigned long hash,
1722 cds_lfht_match_fct match,
1723 const void *key,
1724 struct cds_lfht_node *new_node)
1725 {
1726 unsigned long size;
1727
1728 new_node->reverse_hash = bit_reverse_ulong(hash);
1729 if (!old_iter->node)
1730 return -ENOENT;
1731 if (caa_unlikely(old_iter->node->reverse_hash != new_node->reverse_hash))
1732 return -EINVAL;
1733 if (caa_unlikely(!match(old_iter->node, key)))
1734 return -EINVAL;
1735 size = rcu_dereference(ht->size);
1736 return _cds_lfht_replace(ht, size, old_iter->node, old_iter->next,
1737 new_node);
1738 }
1739
1740 int cds_lfht_del(struct cds_lfht *ht, struct cds_lfht_node *node)
1741 {
1742 unsigned long size;
1743 int ret;
1744
1745 size = rcu_dereference(ht->size);
1746 ret = _cds_lfht_del(ht, size, node);
1747 if (!ret) {
1748 unsigned long hash;
1749
1750 hash = bit_reverse_ulong(node->reverse_hash);
1751 ht_count_del(ht, size, hash);
1752 }
1753 return ret;
1754 }
1755
1756 int cds_lfht_is_node_deleted(struct cds_lfht_node *node)
1757 {
1758 return is_removed(CMM_LOAD_SHARED(node->next));
1759 }
1760
1761 static
1762 int cds_lfht_delete_bucket(struct cds_lfht *ht)
1763 {
1764 struct cds_lfht_node *node;
1765 unsigned long order, i, size;
1766
1767 /* Check that the table is empty */
1768 node = bucket_at(ht, 0);
1769 do {
1770 node = clear_flag(node)->next;
1771 if (!is_bucket(node))
1772 return -EPERM;
1773 assert(!is_removed(node));
1774 assert(!is_removal_owner(node));
1775 } while (!is_end(node));
1776 /*
1777 * size accessed without rcu_dereference because hash table is
1778 * being destroyed.
1779 */
1780 size = ht->size;
1781 /* Internal sanity check: all nodes left should be buckets */
1782 for (i = 0; i < size; i++) {
1783 node = bucket_at(ht, i);
1784 dbg_printf("delete bucket: index %lu expected hash %lu hash %lu\n",
1785 i, i, bit_reverse_ulong(node->reverse_hash));
1786 assert(is_bucket(node->next));
1787 }
1788
1789 for (order = cds_lfht_get_count_order_ulong(size); (long)order >= 0; order--)
1790 cds_lfht_free_bucket_table(ht, order);
1791
1792 return 0;
1793 }
1794
1795 /*
1796 * Should only be called when no more concurrent readers nor writers can
1797 * possibly access the table.
1798 */
1799 int cds_lfht_destroy(struct cds_lfht *ht, pthread_attr_t **attr)
1800 {
1801 int ret;
1802
1803 if (ht->flags & CDS_LFHT_AUTO_RESIZE) {
1804 /* Cancel ongoing resize operations. */
1805 _CMM_STORE_SHARED(ht->in_progress_destroy, 1);
1806 /* Wait for in-flight resize operations to complete */
1807 urcu_workqueue_flush_queued_work(cds_lfht_workqueue);
1808 }
1809 ret = cds_lfht_delete_bucket(ht);
1810 if (ret)
1811 return ret;
1812 free_split_items_count(ht);
1813 if (attr)
1814 *attr = ht->resize_attr;
1815 ret = pthread_mutex_destroy(&ht->resize_mutex);
1816 if (ret)
1817 ret = -EBUSY;
1818 if (ht->flags & CDS_LFHT_AUTO_RESIZE)
1819 cds_lfht_fini_worker(ht->flavor);
1820 poison_free(ht);
1821 return ret;
1822 }
1823
1824 void cds_lfht_count_nodes(struct cds_lfht *ht,
1825 long *approx_before,
1826 unsigned long *count,
1827 long *approx_after)
1828 {
1829 struct cds_lfht_node *node, *next;
1830 unsigned long nr_bucket = 0, nr_removed = 0;
1831
1832 *approx_before = 0;
1833 if (ht->split_count) {
1834 int i;
1835
1836 for (i = 0; i < split_count_mask + 1; i++) {
1837 *approx_before += uatomic_read(&ht->split_count[i].add);
1838 *approx_before -= uatomic_read(&ht->split_count[i].del);
1839 }
1840 }
1841
1842 *count = 0;
1843
1844 /* Count non-bucket nodes in the table */
1845 node = bucket_at(ht, 0);
1846 do {
1847 next = rcu_dereference(node->next);
1848 if (is_removed(next)) {
1849 if (!is_bucket(next))
1850 (nr_removed)++;
1851 else
1852 (nr_bucket)++;
1853 } else if (!is_bucket(next))
1854 (*count)++;
1855 else
1856 (nr_bucket)++;
1857 node = clear_flag(next);
1858 } while (!is_end(node));
1859 dbg_printf("number of logically removed nodes: %lu\n", nr_removed);
1860 dbg_printf("number of bucket nodes: %lu\n", nr_bucket);
1861 *approx_after = 0;
1862 if (ht->split_count) {
1863 int i;
1864
1865 for (i = 0; i < split_count_mask + 1; i++) {
1866 *approx_after += uatomic_read(&ht->split_count[i].add);
1867 *approx_after -= uatomic_read(&ht->split_count[i].del);
1868 }
1869 }
1870 }
1871
1872 /* called with resize mutex held */
1873 static
1874 void _do_cds_lfht_grow(struct cds_lfht *ht,
1875 unsigned long old_size, unsigned long new_size)
1876 {
1877 unsigned long old_order, new_order;
1878
1879 old_order = cds_lfht_get_count_order_ulong(old_size);
1880 new_order = cds_lfht_get_count_order_ulong(new_size);
1881 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1882 old_size, old_order, new_size, new_order);
1883 assert(new_size > old_size);
1884 init_table(ht, old_order + 1, new_order);
1885 }
1886
1887 /* called with resize mutex held */
1888 static
1889 void _do_cds_lfht_shrink(struct cds_lfht *ht,
1890 unsigned long old_size, unsigned long new_size)
1891 {
1892 unsigned long old_order, new_order;
1893
1894 new_size = max(new_size, MIN_TABLE_SIZE);
1895 old_order = cds_lfht_get_count_order_ulong(old_size);
1896 new_order = cds_lfht_get_count_order_ulong(new_size);
1897 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1898 old_size, old_order, new_size, new_order);
1899 assert(new_size < old_size);
1900
1901 /* Remove and unlink all bucket nodes to remove. */
1902 fini_table(ht, new_order + 1, old_order);
1903 }
1904
1905
1906 /* called with resize mutex held */
1907 static
1908 void _do_cds_lfht_resize(struct cds_lfht *ht)
1909 {
1910 unsigned long new_size, old_size;
1911
1912 /*
1913 * Resize table, re-do if the target size has changed under us.
1914 */
1915 do {
1916 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1917 break;
1918 ht->resize_initiated = 1;
1919 old_size = ht->size;
1920 new_size = CMM_LOAD_SHARED(ht->resize_target);
1921 if (old_size < new_size)
1922 _do_cds_lfht_grow(ht, old_size, new_size);
1923 else if (old_size > new_size)
1924 _do_cds_lfht_shrink(ht, old_size, new_size);
1925 ht->resize_initiated = 0;
1926 /* write resize_initiated before read resize_target */
1927 cmm_smp_mb();
1928 } while (ht->size != CMM_LOAD_SHARED(ht->resize_target));
1929 }
1930
1931 static
1932 unsigned long resize_target_grow(struct cds_lfht *ht, unsigned long new_size)
1933 {
1934 return _uatomic_xchg_monotonic_increase(&ht->resize_target, new_size);
1935 }
1936
1937 static
1938 void resize_target_update_count(struct cds_lfht *ht,
1939 unsigned long count)
1940 {
1941 count = max(count, MIN_TABLE_SIZE);
1942 count = min(count, ht->max_nr_buckets);
1943 uatomic_set(&ht->resize_target, count);
1944 }
1945
1946 void cds_lfht_resize(struct cds_lfht *ht, unsigned long new_size)
1947 {
1948 resize_target_update_count(ht, new_size);
1949 CMM_STORE_SHARED(ht->resize_initiated, 1);
1950 pthread_mutex_lock(&ht->resize_mutex);
1951 _do_cds_lfht_resize(ht);
1952 pthread_mutex_unlock(&ht->resize_mutex);
1953 }
1954
1955 static
1956 void do_resize_cb(struct urcu_work *work)
1957 {
1958 struct resize_work *resize_work =
1959 caa_container_of(work, struct resize_work, work);
1960 struct cds_lfht *ht = resize_work->ht;
1961
1962 ht->flavor->register_thread();
1963 pthread_mutex_lock(&ht->resize_mutex);
1964 _do_cds_lfht_resize(ht);
1965 pthread_mutex_unlock(&ht->resize_mutex);
1966 ht->flavor->unregister_thread();
1967 poison_free(work);
1968 }
1969
1970 static
1971 void __cds_lfht_resize_lazy_launch(struct cds_lfht *ht)
1972 {
1973 struct resize_work *work;
1974
1975 /* Store resize_target before read resize_initiated */
1976 cmm_smp_mb();
1977 if (!CMM_LOAD_SHARED(ht->resize_initiated)) {
1978 if (CMM_LOAD_SHARED(ht->in_progress_destroy)) {
1979 return;
1980 }
1981 work = malloc(sizeof(*work));
1982 if (work == NULL) {
1983 dbg_printf("error allocating resize work, bailing out\n");
1984 return;
1985 }
1986 work->ht = ht;
1987 urcu_workqueue_queue_work(cds_lfht_workqueue,
1988 &work->work, do_resize_cb);
1989 CMM_STORE_SHARED(ht->resize_initiated, 1);
1990 }
1991 }
1992
1993 static
1994 void cds_lfht_resize_lazy_grow(struct cds_lfht *ht, unsigned long size, int growth)
1995 {
1996 unsigned long target_size = size << growth;
1997
1998 target_size = min(target_size, ht->max_nr_buckets);
1999 if (resize_target_grow(ht, target_size) >= target_size)
2000 return;
2001
2002 __cds_lfht_resize_lazy_launch(ht);
2003 }
2004
2005 /*
2006 * We favor grow operations over shrink. A shrink operation never occurs
2007 * if a grow operation is queued for lazy execution. A grow operation
2008 * cancels any pending shrink lazy execution.
2009 */
2010 static
2011 void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
2012 unsigned long count)
2013 {
2014 if (!(ht->flags & CDS_LFHT_AUTO_RESIZE))
2015 return;
2016 count = max(count, MIN_TABLE_SIZE);
2017 count = min(count, ht->max_nr_buckets);
2018 if (count == size)
2019 return; /* Already the right size, no resize needed */
2020 if (count > size) { /* lazy grow */
2021 if (resize_target_grow(ht, count) >= count)
2022 return;
2023 } else { /* lazy shrink */
2024 for (;;) {
2025 unsigned long s;
2026
2027 s = uatomic_cmpxchg(&ht->resize_target, size, count);
2028 if (s == size)
2029 break; /* no resize needed */
2030 if (s > size)
2031 return; /* growing is/(was just) in progress */
2032 if (s <= count)
2033 return; /* some other thread do shrink */
2034 size = s;
2035 }
2036 }
2037 __cds_lfht_resize_lazy_launch(ht);
2038 }
2039
2040 static void mutex_lock(pthread_mutex_t *mutex)
2041 {
2042 int ret;
2043
2044 #ifndef DISTRUST_SIGNALS_EXTREME
2045 ret = pthread_mutex_lock(mutex);
2046 if (ret)
2047 urcu_die(ret);
2048 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
2049 while ((ret = pthread_mutex_trylock(mutex)) != 0) {
2050 if (ret != EBUSY && ret != EINTR)
2051 urcu_die(ret);
2052 if (CMM_LOAD_SHARED(URCU_TLS(rcu_reader).need_mb)) {
2053 cmm_smp_mb();
2054 _CMM_STORE_SHARED(URCU_TLS(rcu_reader).need_mb, 0);
2055 cmm_smp_mb();
2056 }
2057 (void) poll(NULL, 0, 10);
2058 }
2059 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
2060 }
2061
2062 static void mutex_unlock(pthread_mutex_t *mutex)
2063 {
2064 int ret;
2065
2066 ret = pthread_mutex_unlock(mutex);
2067 if (ret)
2068 urcu_die(ret);
2069 }
2070
2071 static void cds_lfht_before_fork(void *priv)
2072 {
2073 if (cds_lfht_workqueue_atfork_nesting++)
2074 return;
2075 mutex_lock(&cds_lfht_fork_mutex);
2076 if (!cds_lfht_workqueue)
2077 return;
2078 urcu_workqueue_pause_worker(cds_lfht_workqueue);
2079 }
2080
2081 static void cds_lfht_after_fork_parent(void *priv)
2082 {
2083 if (--cds_lfht_workqueue_atfork_nesting)
2084 return;
2085 if (!cds_lfht_workqueue)
2086 goto end;
2087 urcu_workqueue_resume_worker(cds_lfht_workqueue);
2088 end:
2089 mutex_unlock(&cds_lfht_fork_mutex);
2090 }
2091
2092 static void cds_lfht_after_fork_child(void *priv)
2093 {
2094 if (--cds_lfht_workqueue_atfork_nesting)
2095 return;
2096 if (!cds_lfht_workqueue)
2097 goto end;
2098 urcu_workqueue_create_worker(cds_lfht_workqueue);
2099 end:
2100 mutex_unlock(&cds_lfht_fork_mutex);
2101 }
2102
2103 static struct urcu_atfork cds_lfht_atfork = {
2104 .before_fork = cds_lfht_before_fork,
2105 .after_fork_parent = cds_lfht_after_fork_parent,
2106 .after_fork_child = cds_lfht_after_fork_child,
2107 };
2108
2109 /* Block all signals to ensure we don't disturb the application. */
2110 static void cds_lfht_worker_init(struct urcu_workqueue *workqueue,
2111 void *priv)
2112 {
2113 int ret;
2114 sigset_t mask;
2115
2116 /* Block signal for entire process, so only our thread processes it. */
2117 ret = sigfillset(&mask);
2118 if (ret)
2119 urcu_die(errno);
2120 ret = pthread_sigmask(SIG_BLOCK, &mask, NULL);
2121 if (ret)
2122 urcu_die(ret);
2123 }
2124
2125 static void cds_lfht_init_worker(const struct rcu_flavor_struct *flavor)
2126 {
2127 flavor->register_rculfhash_atfork(&cds_lfht_atfork);
2128
2129 mutex_lock(&cds_lfht_fork_mutex);
2130 if (cds_lfht_workqueue_user_count++)
2131 goto end;
2132 cds_lfht_workqueue = urcu_workqueue_create(0, -1, NULL,
2133 NULL, cds_lfht_worker_init, NULL, NULL, NULL, NULL, NULL);
2134 end:
2135 mutex_unlock(&cds_lfht_fork_mutex);
2136 }
2137
2138 static void cds_lfht_fini_worker(const struct rcu_flavor_struct *flavor)
2139 {
2140 mutex_lock(&cds_lfht_fork_mutex);
2141 if (--cds_lfht_workqueue_user_count)
2142 goto end;
2143 urcu_workqueue_destroy(cds_lfht_workqueue);
2144 cds_lfht_workqueue = NULL;
2145 end:
2146 mutex_unlock(&cds_lfht_fork_mutex);
2147
2148 flavor->unregister_rculfhash_atfork(&cds_lfht_atfork);
2149 }
This page took 0.126207 seconds and 4 git commands to generate.