872d6ff1dea4d1a8533f5bc0c3ba70fe0a438db3
[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 <stdio.h>
262 #include <stdint.h>
263 #include <string.h>
264 #include <sched.h>
265 #include <unistd.h>
266
267 #include "compat-getcpu.h"
268 #include <urcu/assert.h>
269 #include <urcu/pointer.h>
270 #include <urcu/call-rcu.h>
271 #include <urcu/flavor.h>
272 #include <urcu/arch.h>
273 #include <urcu/uatomic.h>
274 #include <urcu/compiler.h>
275 #include <urcu/rculfhash.h>
276 #include <urcu/static/urcu-signal-nr.h>
277 #include <rculfhash-internal.h>
278 #include <stdio.h>
279 #include <pthread.h>
280 #include <signal.h>
281 #include "workqueue.h"
282 #include "urcu-die.h"
283 #include "urcu-utils.h"
284
285 /*
286 * Split-counters lazily update the global counter each 1024
287 * addition/removal. It automatically keeps track of resize required.
288 * We use the bucket length as indicator for need to expand for small
289 * tables and machines lacking per-cpu data support.
290 */
291 #define COUNT_COMMIT_ORDER 10
292 #define DEFAULT_SPLIT_COUNT_MASK 0xFUL
293 #define CHAIN_LEN_TARGET 1
294 #define CHAIN_LEN_RESIZE_THRESHOLD 3
295
296 /*
297 * Define the minimum table size.
298 */
299 #define MIN_TABLE_ORDER 0
300 #define MIN_TABLE_SIZE (1UL << MIN_TABLE_ORDER)
301
302 /*
303 * Minimum number of bucket nodes to touch per thread to parallelize grow/shrink.
304 */
305 #define MIN_PARTITION_PER_THREAD_ORDER 12
306 #define MIN_PARTITION_PER_THREAD (1UL << MIN_PARTITION_PER_THREAD_ORDER)
307
308 /*
309 * The removed flag needs to be updated atomically with the pointer.
310 * It indicates that no node must attach to the node scheduled for
311 * removal, and that node garbage collection must be performed.
312 * The bucket flag does not require to be updated atomically with the
313 * pointer, but it is added as a pointer low bit flag to save space.
314 * The "removal owner" flag is used to detect which of the "del"
315 * operation that has set the "removed flag" gets to return the removed
316 * node to its caller. Note that the replace operation does not need to
317 * iteract with the "removal owner" flag, because it validates that
318 * the "removed" flag is not set before performing its cmpxchg.
319 */
320 #define REMOVED_FLAG (1UL << 0)
321 #define BUCKET_FLAG (1UL << 1)
322 #define REMOVAL_OWNER_FLAG (1UL << 2)
323 #define FLAGS_MASK ((1UL << 3) - 1)
324
325 /* Value of the end pointer. Should not interact with flags. */
326 #define END_VALUE NULL
327
328 /*
329 * ht_items_count: Split-counters counting the number of node addition
330 * and removal in the table. Only used if the CDS_LFHT_ACCOUNTING flag
331 * is set at hash table creation.
332 *
333 * These are free-running counters, never reset to zero. They count the
334 * number of add/remove, and trigger every (1 << COUNT_COMMIT_ORDER)
335 * operations to update the global counter. We choose a power-of-2 value
336 * for the trigger to deal with 32 or 64-bit overflow of the counter.
337 */
338 struct ht_items_count {
339 unsigned long add, del;
340 } __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
341
342 /*
343 * resize_work: Contains arguments passed to worker thread
344 * responsible for performing lazy resize.
345 */
346 struct resize_work {
347 struct urcu_work work;
348 struct cds_lfht *ht;
349 };
350
351 /*
352 * partition_resize_work: Contains arguments passed to worker threads
353 * executing the hash table resize on partitions of the hash table
354 * assigned to each processor's worker thread.
355 */
356 struct partition_resize_work {
357 pthread_t thread_id;
358 struct cds_lfht *ht;
359 unsigned long i, start, len;
360 void (*fct)(struct cds_lfht *ht, unsigned long i,
361 unsigned long start, unsigned long len);
362 };
363
364 static struct urcu_workqueue *cds_lfht_workqueue;
365 static unsigned long cds_lfht_workqueue_user_count;
366
367 /*
368 * Mutex ensuring mutual exclusion between workqueue initialization and
369 * fork handlers. cds_lfht_fork_mutex nests inside call_rcu_mutex.
370 */
371 static pthread_mutex_t cds_lfht_fork_mutex = PTHREAD_MUTEX_INITIALIZER;
372
373 static struct urcu_atfork cds_lfht_atfork;
374
375 /*
376 * atfork handler nesting counters. Handle being registered to many urcu
377 * flavors, thus being possibly invoked more than once in the
378 * pthread_atfork list of callbacks.
379 */
380 static int cds_lfht_workqueue_atfork_nesting;
381
382 static void cds_lfht_init_worker(const struct rcu_flavor_struct *flavor);
383 static void cds_lfht_fini_worker(const struct rcu_flavor_struct *flavor);
384
385 #ifdef CONFIG_CDS_LFHT_ITER_DEBUG
386
387 static
388 void cds_lfht_iter_debug_set_ht(struct cds_lfht *ht, struct cds_lfht_iter *iter)
389 {
390 iter->lfht = ht;
391 }
392
393 #define cds_lfht_iter_debug_assert(...) urcu_posix_assert(__VA_ARGS__)
394
395 #else
396
397 static
398 void cds_lfht_iter_debug_set_ht(struct cds_lfht *ht __attribute__((unused)),
399 struct cds_lfht_iter *iter __attribute__((unused)))
400 {
401 }
402
403 #define cds_lfht_iter_debug_assert(...)
404
405 #endif
406
407 /*
408 * Algorithm to reverse bits in a word by lookup table, extended to
409 * 64-bit words.
410 * Source:
411 * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable
412 * Originally from Public Domain.
413 */
414
415 static const uint8_t BitReverseTable256[256] =
416 {
417 #define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64
418 #define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16)
419 #define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 )
420 R6(0), R6(2), R6(1), R6(3)
421 };
422 #undef R2
423 #undef R4
424 #undef R6
425
426 static
427 uint8_t bit_reverse_u8(uint8_t v)
428 {
429 return BitReverseTable256[v];
430 }
431
432 #if (CAA_BITS_PER_LONG == 32)
433 static
434 uint32_t bit_reverse_u32(uint32_t v)
435 {
436 return ((uint32_t) bit_reverse_u8(v) << 24) |
437 ((uint32_t) bit_reverse_u8(v >> 8) << 16) |
438 ((uint32_t) bit_reverse_u8(v >> 16) << 8) |
439 ((uint32_t) bit_reverse_u8(v >> 24));
440 }
441 #else
442 static
443 uint64_t bit_reverse_u64(uint64_t v)
444 {
445 return ((uint64_t) bit_reverse_u8(v) << 56) |
446 ((uint64_t) bit_reverse_u8(v >> 8) << 48) |
447 ((uint64_t) bit_reverse_u8(v >> 16) << 40) |
448 ((uint64_t) bit_reverse_u8(v >> 24) << 32) |
449 ((uint64_t) bit_reverse_u8(v >> 32) << 24) |
450 ((uint64_t) bit_reverse_u8(v >> 40) << 16) |
451 ((uint64_t) bit_reverse_u8(v >> 48) << 8) |
452 ((uint64_t) bit_reverse_u8(v >> 56));
453 }
454 #endif
455
456 static
457 unsigned long bit_reverse_ulong(unsigned long v)
458 {
459 #if (CAA_BITS_PER_LONG == 32)
460 return bit_reverse_u32(v);
461 #else
462 return bit_reverse_u64(v);
463 #endif
464 }
465
466 /*
467 * fls: returns the position of the most significant bit.
468 * Returns 0 if no bit is set, else returns the position of the most
469 * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit).
470 */
471 #if defined(URCU_ARCH_X86)
472 static inline
473 unsigned int fls_u32(uint32_t x)
474 {
475 int r;
476
477 __asm__ ("bsrl %1,%0\n\t"
478 "jnz 1f\n\t"
479 "movl $-1,%0\n\t"
480 "1:\n\t"
481 : "=r" (r) : "rm" (x));
482 return r + 1;
483 }
484 #define HAS_FLS_U32
485 #endif
486
487 #if defined(URCU_ARCH_AMD64)
488 static inline
489 unsigned int fls_u64(uint64_t x)
490 {
491 long r;
492
493 __asm__ ("bsrq %1,%0\n\t"
494 "jnz 1f\n\t"
495 "movq $-1,%0\n\t"
496 "1:\n\t"
497 : "=r" (r) : "rm" (x));
498 return r + 1;
499 }
500 #define HAS_FLS_U64
501 #endif
502
503 #ifndef HAS_FLS_U64
504 static __attribute__((unused))
505 unsigned int fls_u64(uint64_t x)
506 {
507 unsigned int r = 64;
508
509 if (!x)
510 return 0;
511
512 if (!(x & 0xFFFFFFFF00000000ULL)) {
513 x <<= 32;
514 r -= 32;
515 }
516 if (!(x & 0xFFFF000000000000ULL)) {
517 x <<= 16;
518 r -= 16;
519 }
520 if (!(x & 0xFF00000000000000ULL)) {
521 x <<= 8;
522 r -= 8;
523 }
524 if (!(x & 0xF000000000000000ULL)) {
525 x <<= 4;
526 r -= 4;
527 }
528 if (!(x & 0xC000000000000000ULL)) {
529 x <<= 2;
530 r -= 2;
531 }
532 if (!(x & 0x8000000000000000ULL)) {
533 x <<= 1;
534 r -= 1;
535 }
536 return r;
537 }
538 #endif
539
540 #ifndef HAS_FLS_U32
541 static __attribute__((unused))
542 unsigned int fls_u32(uint32_t x)
543 {
544 unsigned int r = 32;
545
546 if (!x)
547 return 0;
548 if (!(x & 0xFFFF0000U)) {
549 x <<= 16;
550 r -= 16;
551 }
552 if (!(x & 0xFF000000U)) {
553 x <<= 8;
554 r -= 8;
555 }
556 if (!(x & 0xF0000000U)) {
557 x <<= 4;
558 r -= 4;
559 }
560 if (!(x & 0xC0000000U)) {
561 x <<= 2;
562 r -= 2;
563 }
564 if (!(x & 0x80000000U)) {
565 x <<= 1;
566 r -= 1;
567 }
568 return r;
569 }
570 #endif
571
572 unsigned int cds_lfht_fls_ulong(unsigned long x)
573 {
574 #if (CAA_BITS_PER_LONG == 32)
575 return fls_u32(x);
576 #else
577 return fls_u64(x);
578 #endif
579 }
580
581 /*
582 * Return the minimum order for which x <= (1UL << order).
583 * Return -1 if x is 0.
584 */
585 static
586 int cds_lfht_get_count_order_u32(uint32_t x)
587 {
588 if (!x)
589 return -1;
590
591 return fls_u32(x - 1);
592 }
593
594 /*
595 * Return the minimum order for which x <= (1UL << order).
596 * Return -1 if x is 0.
597 */
598 int cds_lfht_get_count_order_ulong(unsigned long x)
599 {
600 if (!x)
601 return -1;
602
603 return cds_lfht_fls_ulong(x - 1);
604 }
605
606 static
607 void cds_lfht_resize_lazy_grow(struct cds_lfht *ht, unsigned long size, int growth);
608
609 static
610 void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
611 unsigned long count);
612
613 static void mutex_lock(pthread_mutex_t *mutex)
614 {
615 int ret;
616
617 #ifndef DISTRUST_SIGNALS_EXTREME
618 ret = pthread_mutex_lock(mutex);
619 if (ret)
620 urcu_die(ret);
621 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
622 while ((ret = pthread_mutex_trylock(mutex)) != 0) {
623 if (ret != EBUSY && ret != EINTR)
624 urcu_die(ret);
625 if (CMM_LOAD_SHARED(URCU_TLS(rcu_reader).need_mb)) {
626 cmm_smp_mb();
627 _CMM_STORE_SHARED(URCU_TLS(rcu_reader).need_mb, 0);
628 cmm_smp_mb();
629 }
630 (void) poll(NULL, 0, 10);
631 }
632 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
633 }
634
635 static void mutex_unlock(pthread_mutex_t *mutex)
636 {
637 int ret;
638
639 ret = pthread_mutex_unlock(mutex);
640 if (ret)
641 urcu_die(ret);
642 }
643
644 static long nr_cpus_mask = -1;
645 static long split_count_mask = -1;
646 static int split_count_order = -1;
647
648 #if defined(HAVE_SYSCONF)
649 static void ht_init_nr_cpus_mask(void)
650 {
651 long maxcpus;
652
653 maxcpus = sysconf(_SC_NPROCESSORS_CONF);
654 if (maxcpus <= 0) {
655 nr_cpus_mask = -2;
656 return;
657 }
658 /*
659 * round up number of CPUs to next power of two, so we
660 * can use & for modulo.
661 */
662 maxcpus = 1UL << cds_lfht_get_count_order_ulong(maxcpus);
663 nr_cpus_mask = maxcpus - 1;
664 }
665 #else /* #if defined(HAVE_SYSCONF) */
666 static void ht_init_nr_cpus_mask(void)
667 {
668 nr_cpus_mask = -2;
669 }
670 #endif /* #else #if defined(HAVE_SYSCONF) */
671
672 static
673 void alloc_split_items_count(struct cds_lfht *ht)
674 {
675 if (nr_cpus_mask == -1) {
676 ht_init_nr_cpus_mask();
677 if (nr_cpus_mask < 0)
678 split_count_mask = DEFAULT_SPLIT_COUNT_MASK;
679 else
680 split_count_mask = nr_cpus_mask;
681 split_count_order =
682 cds_lfht_get_count_order_ulong(split_count_mask + 1);
683 }
684
685 urcu_posix_assert(split_count_mask >= 0);
686
687 if (ht->flags & CDS_LFHT_ACCOUNTING) {
688 ht->split_count = calloc(split_count_mask + 1,
689 sizeof(struct ht_items_count));
690 urcu_posix_assert(ht->split_count);
691 } else {
692 ht->split_count = NULL;
693 }
694 }
695
696 static
697 void free_split_items_count(struct cds_lfht *ht)
698 {
699 poison_free(ht->split_count);
700 }
701
702 static
703 int ht_get_split_count_index(unsigned long hash)
704 {
705 int cpu;
706
707 urcu_posix_assert(split_count_mask >= 0);
708 cpu = urcu_sched_getcpu();
709 if (caa_unlikely(cpu < 0))
710 return hash & split_count_mask;
711 else
712 return cpu & split_count_mask;
713 }
714
715 static
716 void ht_count_add(struct cds_lfht *ht, unsigned long size, unsigned long hash)
717 {
718 unsigned long split_count, count;
719 int index;
720
721 if (caa_unlikely(!ht->split_count))
722 return;
723 index = ht_get_split_count_index(hash);
724 split_count = uatomic_add_return(&ht->split_count[index].add, 1);
725 if (caa_likely(split_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))
726 return;
727 /* Only if number of add multiple of 1UL << COUNT_COMMIT_ORDER */
728
729 dbg_printf("add split count %lu\n", split_count);
730 count = uatomic_add_return(&ht->count,
731 1UL << COUNT_COMMIT_ORDER);
732 if (caa_likely(count & (count - 1)))
733 return;
734 /* Only if global count is power of 2 */
735
736 if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) < size)
737 return;
738 dbg_printf("add set global %lu\n", count);
739 cds_lfht_resize_lazy_count(ht, size,
740 count >> (CHAIN_LEN_TARGET - 1));
741 }
742
743 static
744 void ht_count_del(struct cds_lfht *ht, unsigned long size, unsigned long hash)
745 {
746 unsigned long split_count, count;
747 int index;
748
749 if (caa_unlikely(!ht->split_count))
750 return;
751 index = ht_get_split_count_index(hash);
752 split_count = uatomic_add_return(&ht->split_count[index].del, 1);
753 if (caa_likely(split_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))
754 return;
755 /* Only if number of deletes multiple of 1UL << COUNT_COMMIT_ORDER */
756
757 dbg_printf("del split count %lu\n", split_count);
758 count = uatomic_add_return(&ht->count,
759 -(1UL << COUNT_COMMIT_ORDER));
760 if (caa_likely(count & (count - 1)))
761 return;
762 /* Only if global count is power of 2 */
763
764 if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) >= size)
765 return;
766 dbg_printf("del set global %lu\n", count);
767 /*
768 * Don't shrink table if the number of nodes is below a
769 * certain threshold.
770 */
771 if (count < (1UL << COUNT_COMMIT_ORDER) * (split_count_mask + 1))
772 return;
773 cds_lfht_resize_lazy_count(ht, size,
774 count >> (CHAIN_LEN_TARGET - 1));
775 }
776
777 static
778 void check_resize(struct cds_lfht *ht, unsigned long size, uint32_t chain_len)
779 {
780 unsigned long count;
781
782 if (!(ht->flags & CDS_LFHT_AUTO_RESIZE))
783 return;
784 count = uatomic_read(&ht->count);
785 /*
786 * Use bucket-local length for small table expand and for
787 * environments lacking per-cpu data support.
788 */
789 if (count >= (1UL << (COUNT_COMMIT_ORDER + split_count_order)))
790 return;
791 if (chain_len > 100)
792 dbg_printf("WARNING: large chain length: %u.\n",
793 chain_len);
794 if (chain_len >= CHAIN_LEN_RESIZE_THRESHOLD) {
795 int growth;
796
797 /*
798 * Ideal growth calculated based on chain length.
799 */
800 growth = cds_lfht_get_count_order_u32(chain_len
801 - (CHAIN_LEN_TARGET - 1));
802 if ((ht->flags & CDS_LFHT_ACCOUNTING)
803 && (size << growth)
804 >= (1UL << (COUNT_COMMIT_ORDER
805 + split_count_order))) {
806 /*
807 * If ideal growth expands the hash table size
808 * beyond the "small hash table" sizes, use the
809 * maximum small hash table size to attempt
810 * expanding the hash table. This only applies
811 * when node accounting is available, otherwise
812 * the chain length is used to expand the hash
813 * table in every case.
814 */
815 growth = COUNT_COMMIT_ORDER + split_count_order
816 - cds_lfht_get_count_order_ulong(size);
817 if (growth <= 0)
818 return;
819 }
820 cds_lfht_resize_lazy_grow(ht, size, growth);
821 }
822 }
823
824 static
825 struct cds_lfht_node *clear_flag(struct cds_lfht_node *node)
826 {
827 return (struct cds_lfht_node *) (((unsigned long) node) & ~FLAGS_MASK);
828 }
829
830 static
831 int is_removed(const struct cds_lfht_node *node)
832 {
833 return ((unsigned long) node) & REMOVED_FLAG;
834 }
835
836 static
837 int is_bucket(struct cds_lfht_node *node)
838 {
839 return ((unsigned long) node) & BUCKET_FLAG;
840 }
841
842 static
843 struct cds_lfht_node *flag_bucket(struct cds_lfht_node *node)
844 {
845 return (struct cds_lfht_node *) (((unsigned long) node) | BUCKET_FLAG);
846 }
847
848 static
849 int is_removal_owner(struct cds_lfht_node *node)
850 {
851 return ((unsigned long) node) & REMOVAL_OWNER_FLAG;
852 }
853
854 static
855 struct cds_lfht_node *flag_removed(struct cds_lfht_node *node)
856 {
857 return (struct cds_lfht_node *) (((unsigned long) node) | REMOVED_FLAG);
858 }
859
860 static
861 struct cds_lfht_node *flag_removal_owner(struct cds_lfht_node *node)
862 {
863 return (struct cds_lfht_node *) (((unsigned long) node) | REMOVAL_OWNER_FLAG);
864 }
865
866 static
867 struct cds_lfht_node *flag_removed_or_removal_owner(struct cds_lfht_node *node)
868 {
869 return (struct cds_lfht_node *) (((unsigned long) node) | REMOVED_FLAG | REMOVAL_OWNER_FLAG);
870 }
871
872 static
873 struct cds_lfht_node *get_end(void)
874 {
875 return (struct cds_lfht_node *) END_VALUE;
876 }
877
878 static
879 int is_end(struct cds_lfht_node *node)
880 {
881 return clear_flag(node) == (struct cds_lfht_node *) END_VALUE;
882 }
883
884 static
885 unsigned long _uatomic_xchg_monotonic_increase(unsigned long *ptr,
886 unsigned long v)
887 {
888 unsigned long old1, old2;
889
890 old1 = uatomic_read(ptr);
891 do {
892 old2 = old1;
893 if (old2 >= v)
894 return old2;
895 } while ((old1 = uatomic_cmpxchg(ptr, old2, v)) != old2);
896 return old2;
897 }
898
899 static
900 void cds_lfht_alloc_bucket_table(struct cds_lfht *ht, unsigned long order)
901 {
902 return ht->mm->alloc_bucket_table(ht, order);
903 }
904
905 /*
906 * cds_lfht_free_bucket_table() should be called with decreasing order.
907 * When cds_lfht_free_bucket_table(0) is called, it means the whole
908 * lfht is destroyed.
909 */
910 static
911 void cds_lfht_free_bucket_table(struct cds_lfht *ht, unsigned long order)
912 {
913 return ht->mm->free_bucket_table(ht, order);
914 }
915
916 static inline
917 struct cds_lfht_node *bucket_at(struct cds_lfht *ht, unsigned long index)
918 {
919 return ht->bucket_at(ht, index);
920 }
921
922 static inline
923 struct cds_lfht_node *lookup_bucket(struct cds_lfht *ht, unsigned long size,
924 unsigned long hash)
925 {
926 urcu_posix_assert(size > 0);
927 return bucket_at(ht, hash & (size - 1));
928 }
929
930 /*
931 * Remove all logically deleted nodes from a bucket up to a certain node key.
932 */
933 static
934 void _cds_lfht_gc_bucket(struct cds_lfht_node *bucket, struct cds_lfht_node *node)
935 {
936 struct cds_lfht_node *iter_prev, *iter, *next, *new_next;
937
938 urcu_posix_assert(!is_bucket(bucket));
939 urcu_posix_assert(!is_removed(bucket));
940 urcu_posix_assert(!is_removal_owner(bucket));
941 urcu_posix_assert(!is_bucket(node));
942 urcu_posix_assert(!is_removed(node));
943 urcu_posix_assert(!is_removal_owner(node));
944 for (;;) {
945 iter_prev = bucket;
946 /* We can always skip the bucket node initially */
947 iter = rcu_dereference(iter_prev->next);
948 urcu_posix_assert(!is_removed(iter));
949 urcu_posix_assert(!is_removal_owner(iter));
950 urcu_posix_assert(iter_prev->reverse_hash <= node->reverse_hash);
951 /*
952 * We should never be called with bucket (start of chain)
953 * and logically removed node (end of path compression
954 * marker) being the actual same node. This would be a
955 * bug in the algorithm implementation.
956 */
957 urcu_posix_assert(bucket != node);
958 for (;;) {
959 if (caa_unlikely(is_end(iter)))
960 return;
961 if (caa_likely(clear_flag(iter)->reverse_hash > node->reverse_hash))
962 return;
963 next = rcu_dereference(clear_flag(iter)->next);
964 if (caa_likely(is_removed(next)))
965 break;
966 iter_prev = clear_flag(iter);
967 iter = next;
968 }
969 urcu_posix_assert(!is_removed(iter));
970 urcu_posix_assert(!is_removal_owner(iter));
971 if (is_bucket(iter))
972 new_next = flag_bucket(clear_flag(next));
973 else
974 new_next = clear_flag(next);
975 (void) uatomic_cmpxchg(&iter_prev->next, iter, new_next);
976 }
977 }
978
979 static
980 int _cds_lfht_replace(struct cds_lfht *ht, unsigned long size,
981 struct cds_lfht_node *old_node,
982 struct cds_lfht_node *old_next,
983 struct cds_lfht_node *new_node)
984 {
985 struct cds_lfht_node *bucket, *ret_next;
986
987 if (!old_node) /* Return -ENOENT if asked to replace NULL node */
988 return -ENOENT;
989
990 urcu_posix_assert(!is_removed(old_node));
991 urcu_posix_assert(!is_removal_owner(old_node));
992 urcu_posix_assert(!is_bucket(old_node));
993 urcu_posix_assert(!is_removed(new_node));
994 urcu_posix_assert(!is_removal_owner(new_node));
995 urcu_posix_assert(!is_bucket(new_node));
996 urcu_posix_assert(new_node != old_node);
997 for (;;) {
998 /* Insert after node to be replaced */
999 if (is_removed(old_next)) {
1000 /*
1001 * Too late, the old node has been removed under us
1002 * between lookup and replace. Fail.
1003 */
1004 return -ENOENT;
1005 }
1006 urcu_posix_assert(old_next == clear_flag(old_next));
1007 urcu_posix_assert(new_node != old_next);
1008 /*
1009 * REMOVAL_OWNER flag is _NEVER_ set before the REMOVED
1010 * flag. It is either set atomically at the same time
1011 * (replace) or after (del).
1012 */
1013 urcu_posix_assert(!is_removal_owner(old_next));
1014 new_node->next = old_next;
1015 /*
1016 * Here is the whole trick for lock-free replace: we add
1017 * the replacement node _after_ the node we want to
1018 * replace by atomically setting its next pointer at the
1019 * same time we set its removal flag. Given that
1020 * the lookups/get next use an iterator aware of the
1021 * next pointer, they will either skip the old node due
1022 * to the removal flag and see the new node, or use
1023 * the old node, but will not see the new one.
1024 * This is a replacement of a node with another node
1025 * that has the same value: we are therefore not
1026 * removing a value from the hash table. We set both the
1027 * REMOVED and REMOVAL_OWNER flags atomically so we own
1028 * the node after successful cmpxchg.
1029 */
1030 ret_next = uatomic_cmpxchg(&old_node->next,
1031 old_next, flag_removed_or_removal_owner(new_node));
1032 if (ret_next == old_next)
1033 break; /* We performed the replacement. */
1034 old_next = ret_next;
1035 }
1036
1037 /*
1038 * Ensure that the old node is not visible to readers anymore:
1039 * lookup for the node, and remove it (along with any other
1040 * logically removed node) if found.
1041 */
1042 bucket = lookup_bucket(ht, size, bit_reverse_ulong(old_node->reverse_hash));
1043 _cds_lfht_gc_bucket(bucket, new_node);
1044
1045 urcu_posix_assert(is_removed(CMM_LOAD_SHARED(old_node->next)));
1046 return 0;
1047 }
1048
1049 /*
1050 * A non-NULL unique_ret pointer uses the "add unique" (or uniquify) add
1051 * mode. A NULL unique_ret allows creation of duplicate keys.
1052 */
1053 static
1054 void _cds_lfht_add(struct cds_lfht *ht,
1055 unsigned long hash,
1056 cds_lfht_match_fct match,
1057 const void *key,
1058 unsigned long size,
1059 struct cds_lfht_node *node,
1060 struct cds_lfht_iter *unique_ret,
1061 int bucket_flag)
1062 {
1063 struct cds_lfht_node *iter_prev, *iter, *next, *new_node, *new_next,
1064 *return_node;
1065 struct cds_lfht_node *bucket;
1066
1067 urcu_posix_assert(!is_bucket(node));
1068 urcu_posix_assert(!is_removed(node));
1069 urcu_posix_assert(!is_removal_owner(node));
1070 bucket = lookup_bucket(ht, size, hash);
1071 for (;;) {
1072 uint32_t chain_len = 0;
1073
1074 /*
1075 * iter_prev points to the non-removed node prior to the
1076 * insert location.
1077 */
1078 iter_prev = bucket;
1079 /* We can always skip the bucket node initially */
1080 iter = rcu_dereference(iter_prev->next);
1081 urcu_posix_assert(iter_prev->reverse_hash <= node->reverse_hash);
1082 for (;;) {
1083 if (caa_unlikely(is_end(iter)))
1084 goto insert;
1085 if (caa_likely(clear_flag(iter)->reverse_hash > node->reverse_hash))
1086 goto insert;
1087
1088 /* bucket node is the first node of the identical-hash-value chain */
1089 if (bucket_flag && clear_flag(iter)->reverse_hash == node->reverse_hash)
1090 goto insert;
1091
1092 next = rcu_dereference(clear_flag(iter)->next);
1093 if (caa_unlikely(is_removed(next)))
1094 goto gc_node;
1095
1096 /* uniquely add */
1097 if (unique_ret
1098 && !is_bucket(next)
1099 && clear_flag(iter)->reverse_hash == node->reverse_hash) {
1100 struct cds_lfht_iter d_iter = {
1101 .node = node,
1102 .next = iter,
1103 #ifdef CONFIG_CDS_LFHT_ITER_DEBUG
1104 .lfht = ht,
1105 #endif
1106 };
1107
1108 /*
1109 * uniquely adding inserts the node as the first
1110 * node of the identical-hash-value node chain.
1111 *
1112 * This semantic ensures no duplicated keys
1113 * should ever be observable in the table
1114 * (including traversing the table node by
1115 * node by forward iterations)
1116 */
1117 cds_lfht_next_duplicate(ht, match, key, &d_iter);
1118 if (!d_iter.node)
1119 goto insert;
1120
1121 *unique_ret = d_iter;
1122 return;
1123 }
1124
1125 /* Only account for identical reverse hash once */
1126 if (iter_prev->reverse_hash != clear_flag(iter)->reverse_hash
1127 && !is_bucket(next))
1128 check_resize(ht, size, ++chain_len);
1129 iter_prev = clear_flag(iter);
1130 iter = next;
1131 }
1132
1133 insert:
1134 urcu_posix_assert(node != clear_flag(iter));
1135 urcu_posix_assert(!is_removed(iter_prev));
1136 urcu_posix_assert(!is_removal_owner(iter_prev));
1137 urcu_posix_assert(!is_removed(iter));
1138 urcu_posix_assert(!is_removal_owner(iter));
1139 urcu_posix_assert(iter_prev != node);
1140 if (!bucket_flag)
1141 node->next = clear_flag(iter);
1142 else
1143 node->next = flag_bucket(clear_flag(iter));
1144 if (is_bucket(iter))
1145 new_node = flag_bucket(node);
1146 else
1147 new_node = node;
1148 if (uatomic_cmpxchg(&iter_prev->next, iter,
1149 new_node) != iter) {
1150 continue; /* retry */
1151 } else {
1152 return_node = node;
1153 goto end;
1154 }
1155
1156 gc_node:
1157 urcu_posix_assert(!is_removed(iter));
1158 urcu_posix_assert(!is_removal_owner(iter));
1159 if (is_bucket(iter))
1160 new_next = flag_bucket(clear_flag(next));
1161 else
1162 new_next = clear_flag(next);
1163 (void) uatomic_cmpxchg(&iter_prev->next, iter, new_next);
1164 /* retry */
1165 }
1166 end:
1167 if (unique_ret) {
1168 unique_ret->node = return_node;
1169 /* unique_ret->next left unset, never used. */
1170 }
1171 }
1172
1173 static
1174 int _cds_lfht_del(struct cds_lfht *ht, unsigned long size,
1175 struct cds_lfht_node *node)
1176 {
1177 struct cds_lfht_node *bucket, *next;
1178
1179 if (!node) /* Return -ENOENT if asked to delete NULL node */
1180 return -ENOENT;
1181
1182 /* logically delete the node */
1183 urcu_posix_assert(!is_bucket(node));
1184 urcu_posix_assert(!is_removed(node));
1185 urcu_posix_assert(!is_removal_owner(node));
1186
1187 /*
1188 * We are first checking if the node had previously been
1189 * logically removed (this check is not atomic with setting the
1190 * logical removal flag). Return -ENOENT if the node had
1191 * previously been removed.
1192 */
1193 next = CMM_LOAD_SHARED(node->next); /* next is not dereferenced */
1194 if (caa_unlikely(is_removed(next)))
1195 return -ENOENT;
1196 urcu_posix_assert(!is_bucket(next));
1197 /*
1198 * The del operation semantic guarantees a full memory barrier
1199 * before the uatomic_or atomic commit of the deletion flag.
1200 */
1201 cmm_smp_mb__before_uatomic_or();
1202 /*
1203 * We set the REMOVED_FLAG unconditionally. Note that there may
1204 * be more than one concurrent thread setting this flag.
1205 * Knowing which wins the race will be known after the garbage
1206 * collection phase, stay tuned!
1207 */
1208 uatomic_or(&node->next, REMOVED_FLAG);
1209 /* We performed the (logical) deletion. */
1210
1211 /*
1212 * Ensure that the node is not visible to readers anymore: lookup for
1213 * the node, and remove it (along with any other logically removed node)
1214 * if found.
1215 */
1216 bucket = lookup_bucket(ht, size, bit_reverse_ulong(node->reverse_hash));
1217 _cds_lfht_gc_bucket(bucket, node);
1218
1219 urcu_posix_assert(is_removed(CMM_LOAD_SHARED(node->next)));
1220 /*
1221 * Last phase: atomically exchange node->next with a version
1222 * having "REMOVAL_OWNER_FLAG" set. If the returned node->next
1223 * pointer did _not_ have "REMOVAL_OWNER_FLAG" set, we now own
1224 * the node and win the removal race.
1225 * It is interesting to note that all "add" paths are forbidden
1226 * to change the next pointer starting from the point where the
1227 * REMOVED_FLAG is set, so here using a read, followed by a
1228 * xchg() suffice to guarantee that the xchg() will ever only
1229 * set the "REMOVAL_OWNER_FLAG" (or change nothing if the flag
1230 * was already set).
1231 */
1232 if (!is_removal_owner(uatomic_xchg(&node->next,
1233 flag_removal_owner(node->next))))
1234 return 0;
1235 else
1236 return -ENOENT;
1237 }
1238
1239 static
1240 void *partition_resize_thread(void *arg)
1241 {
1242 struct partition_resize_work *work = arg;
1243
1244 work->ht->flavor->register_thread();
1245 work->fct(work->ht, work->i, work->start, work->len);
1246 work->ht->flavor->unregister_thread();
1247 return NULL;
1248 }
1249
1250 static
1251 void partition_resize_helper(struct cds_lfht *ht, unsigned long i,
1252 unsigned long len,
1253 void (*fct)(struct cds_lfht *ht, unsigned long i,
1254 unsigned long start, unsigned long len))
1255 {
1256 unsigned long partition_len, start = 0;
1257 struct partition_resize_work *work;
1258 int ret;
1259 unsigned long thread, nr_threads;
1260
1261 urcu_posix_assert(nr_cpus_mask != -1);
1262 if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD)
1263 goto fallback;
1264
1265 /*
1266 * Note: nr_cpus_mask + 1 is always power of 2.
1267 * We spawn just the number of threads we need to satisfy the minimum
1268 * partition size, up to the number of CPUs in the system.
1269 */
1270 if (nr_cpus_mask > 0) {
1271 nr_threads = min_t(unsigned long, nr_cpus_mask + 1,
1272 len >> MIN_PARTITION_PER_THREAD_ORDER);
1273 } else {
1274 nr_threads = 1;
1275 }
1276 partition_len = len >> cds_lfht_get_count_order_ulong(nr_threads);
1277 work = calloc(nr_threads, sizeof(*work));
1278 if (!work) {
1279 dbg_printf("error allocating for resize, single-threading\n");
1280 goto fallback;
1281 }
1282 for (thread = 0; thread < nr_threads; thread++) {
1283 work[thread].ht = ht;
1284 work[thread].i = i;
1285 work[thread].len = partition_len;
1286 work[thread].start = thread * partition_len;
1287 work[thread].fct = fct;
1288 ret = pthread_create(&(work[thread].thread_id), ht->resize_attr,
1289 partition_resize_thread, &work[thread]);
1290 if (ret == EAGAIN) {
1291 /*
1292 * Out of resources: wait and join the threads
1293 * we've created, then handle leftovers.
1294 */
1295 dbg_printf("error spawning for resize, single-threading\n");
1296 start = work[thread].start;
1297 len -= start;
1298 nr_threads = thread;
1299 break;
1300 }
1301 urcu_posix_assert(!ret);
1302 }
1303 for (thread = 0; thread < nr_threads; thread++) {
1304 ret = pthread_join(work[thread].thread_id, NULL);
1305 urcu_posix_assert(!ret);
1306 }
1307 free(work);
1308
1309 /*
1310 * A pthread_create failure above will either lead in us having
1311 * no threads to join or starting at a non-zero offset,
1312 * fallback to single thread processing of leftovers.
1313 */
1314 if (start == 0 && nr_threads > 0)
1315 return;
1316 fallback:
1317 fct(ht, i, start, len);
1318 }
1319
1320 /*
1321 * Holding RCU read lock to protect _cds_lfht_add against memory
1322 * reclaim that could be performed by other worker threads (ABA
1323 * problem).
1324 *
1325 * When we reach a certain length, we can split this population phase over
1326 * many worker threads, based on the number of CPUs available in the system.
1327 * This should therefore take care of not having the expand lagging behind too
1328 * many concurrent insertion threads by using the scheduler's ability to
1329 * schedule bucket node population fairly with insertions.
1330 */
1331 static
1332 void init_table_populate_partition(struct cds_lfht *ht, unsigned long i,
1333 unsigned long start, unsigned long len)
1334 {
1335 unsigned long j, size = 1UL << (i - 1);
1336
1337 urcu_posix_assert(i > MIN_TABLE_ORDER);
1338 ht->flavor->read_lock();
1339 for (j = size + start; j < size + start + len; j++) {
1340 struct cds_lfht_node *new_node = bucket_at(ht, j);
1341
1342 urcu_posix_assert(j >= size && j < (size << 1));
1343 dbg_printf("init populate: order %lu index %lu hash %lu\n",
1344 i, j, j);
1345 new_node->reverse_hash = bit_reverse_ulong(j);
1346 _cds_lfht_add(ht, j, NULL, NULL, size, new_node, NULL, 1);
1347 }
1348 ht->flavor->read_unlock();
1349 }
1350
1351 static
1352 void init_table_populate(struct cds_lfht *ht, unsigned long i,
1353 unsigned long len)
1354 {
1355 partition_resize_helper(ht, i, len, init_table_populate_partition);
1356 }
1357
1358 static
1359 void init_table(struct cds_lfht *ht,
1360 unsigned long first_order, unsigned long last_order)
1361 {
1362 unsigned long i;
1363
1364 dbg_printf("init table: first_order %lu last_order %lu\n",
1365 first_order, last_order);
1366 urcu_posix_assert(first_order > MIN_TABLE_ORDER);
1367 for (i = first_order; i <= last_order; i++) {
1368 unsigned long len;
1369
1370 len = 1UL << (i - 1);
1371 dbg_printf("init order %lu len: %lu\n", i, len);
1372
1373 /* Stop expand if the resize target changes under us */
1374 if (CMM_LOAD_SHARED(ht->resize_target) < (1UL << i))
1375 break;
1376
1377 cds_lfht_alloc_bucket_table(ht, i);
1378
1379 /*
1380 * Set all bucket nodes reverse hash values for a level and
1381 * link all bucket nodes into the table.
1382 */
1383 init_table_populate(ht, i, len);
1384
1385 /*
1386 * Update table size.
1387 */
1388 cmm_smp_wmb(); /* populate data before RCU size */
1389 CMM_STORE_SHARED(ht->size, 1UL << i);
1390
1391 dbg_printf("init new size: %lu\n", 1UL << i);
1392 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1393 break;
1394 }
1395 }
1396
1397 /*
1398 * Holding RCU read lock to protect _cds_lfht_remove against memory
1399 * reclaim that could be performed by other worker threads (ABA
1400 * problem).
1401 * For a single level, we logically remove and garbage collect each node.
1402 *
1403 * As a design choice, we perform logical removal and garbage collection on a
1404 * node-per-node basis to simplify this algorithm. We also assume keeping good
1405 * cache locality of the operation would overweight possible performance gain
1406 * that could be achieved by batching garbage collection for multiple levels.
1407 * However, this would have to be justified by benchmarks.
1408 *
1409 * Concurrent removal and add operations are helping us perform garbage
1410 * collection of logically removed nodes. We guarantee that all logically
1411 * removed nodes have been garbage-collected (unlinked) before work
1412 * enqueue is invoked to free a hole level of bucket nodes (after a
1413 * grace period).
1414 *
1415 * Logical removal and garbage collection can therefore be done in batch
1416 * or on a node-per-node basis, as long as the guarantee above holds.
1417 *
1418 * When we reach a certain length, we can split this removal over many worker
1419 * threads, based on the number of CPUs available in the system. This should
1420 * take care of not letting resize process lag behind too many concurrent
1421 * updater threads actively inserting into the hash table.
1422 */
1423 static
1424 void remove_table_partition(struct cds_lfht *ht, unsigned long i,
1425 unsigned long start, unsigned long len)
1426 {
1427 unsigned long j, size = 1UL << (i - 1);
1428
1429 urcu_posix_assert(i > MIN_TABLE_ORDER);
1430 ht->flavor->read_lock();
1431 for (j = size + start; j < size + start + len; j++) {
1432 struct cds_lfht_node *fini_bucket = bucket_at(ht, j);
1433 struct cds_lfht_node *parent_bucket = bucket_at(ht, j - size);
1434
1435 urcu_posix_assert(j >= size && j < (size << 1));
1436 dbg_printf("remove entry: order %lu index %lu hash %lu\n",
1437 i, j, j);
1438 /* Set the REMOVED_FLAG to freeze the ->next for gc */
1439 uatomic_or(&fini_bucket->next, REMOVED_FLAG);
1440 _cds_lfht_gc_bucket(parent_bucket, fini_bucket);
1441 }
1442 ht->flavor->read_unlock();
1443 }
1444
1445 static
1446 void remove_table(struct cds_lfht *ht, unsigned long i, unsigned long len)
1447 {
1448 partition_resize_helper(ht, i, len, remove_table_partition);
1449 }
1450
1451 /*
1452 * fini_table() is never called for first_order == 0, which is why
1453 * free_by_rcu_order == 0 can be used as criterion to know if free must
1454 * be called.
1455 */
1456 static
1457 void fini_table(struct cds_lfht *ht,
1458 unsigned long first_order, unsigned long last_order)
1459 {
1460 unsigned long free_by_rcu_order = 0, i;
1461
1462 dbg_printf("fini table: first_order %lu last_order %lu\n",
1463 first_order, last_order);
1464 urcu_posix_assert(first_order > MIN_TABLE_ORDER);
1465 for (i = last_order; i >= first_order; i--) {
1466 unsigned long len;
1467
1468 len = 1UL << (i - 1);
1469 dbg_printf("fini order %ld len: %lu\n", i, len);
1470
1471 /* Stop shrink if the resize target changes under us */
1472 if (CMM_LOAD_SHARED(ht->resize_target) > (1UL << (i - 1)))
1473 break;
1474
1475 cmm_smp_wmb(); /* populate data before RCU size */
1476 CMM_STORE_SHARED(ht->size, 1UL << (i - 1));
1477
1478 /*
1479 * We need to wait for all add operations to reach Q.S. (and
1480 * thus use the new table for lookups) before we can start
1481 * releasing the old bucket nodes. Otherwise their lookup will
1482 * return a logically removed node as insert position.
1483 */
1484 ht->flavor->update_synchronize_rcu();
1485 if (free_by_rcu_order)
1486 cds_lfht_free_bucket_table(ht, free_by_rcu_order);
1487
1488 /*
1489 * Set "removed" flag in bucket nodes about to be removed.
1490 * Unlink all now-logically-removed bucket node pointers.
1491 * Concurrent add/remove operation are helping us doing
1492 * the gc.
1493 */
1494 remove_table(ht, i, len);
1495
1496 free_by_rcu_order = i;
1497
1498 dbg_printf("fini new size: %lu\n", 1UL << i);
1499 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1500 break;
1501 }
1502
1503 if (free_by_rcu_order) {
1504 ht->flavor->update_synchronize_rcu();
1505 cds_lfht_free_bucket_table(ht, free_by_rcu_order);
1506 }
1507 }
1508
1509 /*
1510 * Never called with size < 1.
1511 */
1512 static
1513 void cds_lfht_create_bucket(struct cds_lfht *ht, unsigned long size)
1514 {
1515 struct cds_lfht_node *prev, *node;
1516 unsigned long order, len, i;
1517 int bucket_order;
1518
1519 cds_lfht_alloc_bucket_table(ht, 0);
1520
1521 dbg_printf("create bucket: order 0 index 0 hash 0\n");
1522 node = bucket_at(ht, 0);
1523 node->next = flag_bucket(get_end());
1524 node->reverse_hash = 0;
1525
1526 bucket_order = cds_lfht_get_count_order_ulong(size);
1527 urcu_posix_assert(bucket_order >= 0);
1528
1529 for (order = 1; order < (unsigned long) bucket_order + 1; order++) {
1530 len = 1UL << (order - 1);
1531 cds_lfht_alloc_bucket_table(ht, order);
1532
1533 for (i = 0; i < len; i++) {
1534 /*
1535 * Now, we are trying to init the node with the
1536 * hash=(len+i) (which is also a bucket with the
1537 * index=(len+i)) and insert it into the hash table,
1538 * so this node has to be inserted after the bucket
1539 * with the index=(len+i)&(len-1)=i. And because there
1540 * is no other non-bucket node nor bucket node with
1541 * larger index/hash inserted, so the bucket node
1542 * being inserted should be inserted directly linked
1543 * after the bucket node with index=i.
1544 */
1545 prev = bucket_at(ht, i);
1546 node = bucket_at(ht, len + i);
1547
1548 dbg_printf("create bucket: order %lu index %lu hash %lu\n",
1549 order, len + i, len + i);
1550 node->reverse_hash = bit_reverse_ulong(len + i);
1551
1552 /* insert after prev */
1553 urcu_posix_assert(is_bucket(prev->next));
1554 node->next = prev->next;
1555 prev->next = flag_bucket(node);
1556 }
1557 }
1558 }
1559
1560 #if (CAA_BITS_PER_LONG > 32)
1561 /*
1562 * For 64-bit architectures, with max number of buckets small enough not to
1563 * use the entire 64-bit memory mapping space (and allowing a fair number of
1564 * hash table instances), use the mmap allocator, which is faster. Otherwise,
1565 * fallback to the order allocator.
1566 */
1567 static
1568 const struct cds_lfht_mm_type *get_mm_type(unsigned long max_nr_buckets)
1569 {
1570 if (max_nr_buckets && max_nr_buckets <= (1ULL << 32))
1571 return &cds_lfht_mm_mmap;
1572 else
1573 return &cds_lfht_mm_order;
1574 }
1575 #else
1576 /*
1577 * For 32-bit architectures, use the order allocator.
1578 */
1579 static
1580 const struct cds_lfht_mm_type *get_mm_type(
1581 unsigned long max_nr_buckets __attribute__((unused)))
1582 {
1583 return &cds_lfht_mm_order;
1584 }
1585 #endif
1586
1587 void cds_lfht_node_init_deleted(struct cds_lfht_node *node)
1588 {
1589 cds_lfht_node_init(node);
1590 node->next = flag_removed(NULL);
1591 }
1592
1593 struct cds_lfht *_cds_lfht_new(unsigned long init_size,
1594 unsigned long min_nr_alloc_buckets,
1595 unsigned long max_nr_buckets,
1596 int flags,
1597 const struct cds_lfht_mm_type *mm,
1598 const struct rcu_flavor_struct *flavor,
1599 pthread_attr_t *attr)
1600 {
1601 struct cds_lfht *ht;
1602 unsigned long order;
1603
1604 /* min_nr_alloc_buckets must be power of two */
1605 if (!min_nr_alloc_buckets || (min_nr_alloc_buckets & (min_nr_alloc_buckets - 1)))
1606 return NULL;
1607
1608 /* init_size must be power of two */
1609 if (!init_size || (init_size & (init_size - 1)))
1610 return NULL;
1611
1612 /*
1613 * Memory management plugin default.
1614 */
1615 if (!mm)
1616 mm = get_mm_type(max_nr_buckets);
1617
1618 /* max_nr_buckets == 0 for order based mm means infinite */
1619 if (mm == &cds_lfht_mm_order && !max_nr_buckets)
1620 max_nr_buckets = 1UL << (MAX_TABLE_ORDER - 1);
1621
1622 /* max_nr_buckets must be power of two */
1623 if (!max_nr_buckets || (max_nr_buckets & (max_nr_buckets - 1)))
1624 return NULL;
1625
1626 if (flags & CDS_LFHT_AUTO_RESIZE)
1627 cds_lfht_init_worker(flavor);
1628
1629 min_nr_alloc_buckets = max(min_nr_alloc_buckets, MIN_TABLE_SIZE);
1630 init_size = max(init_size, MIN_TABLE_SIZE);
1631 max_nr_buckets = max(max_nr_buckets, min_nr_alloc_buckets);
1632 init_size = min(init_size, max_nr_buckets);
1633
1634 ht = mm->alloc_cds_lfht(min_nr_alloc_buckets, max_nr_buckets);
1635 urcu_posix_assert(ht);
1636 urcu_posix_assert(ht->mm == mm);
1637 urcu_posix_assert(ht->bucket_at == mm->bucket_at);
1638
1639 ht->flags = flags;
1640 ht->flavor = flavor;
1641 ht->resize_attr = attr;
1642 alloc_split_items_count(ht);
1643 /* this mutex should not nest in read-side C.S. */
1644 pthread_mutex_init(&ht->resize_mutex, NULL);
1645 order = cds_lfht_get_count_order_ulong(init_size);
1646 ht->resize_target = 1UL << order;
1647 cds_lfht_create_bucket(ht, 1UL << order);
1648 ht->size = 1UL << order;
1649 return ht;
1650 }
1651
1652 void cds_lfht_lookup(struct cds_lfht *ht, unsigned long hash,
1653 cds_lfht_match_fct match, const void *key,
1654 struct cds_lfht_iter *iter)
1655 {
1656 struct cds_lfht_node *node, *next, *bucket;
1657 unsigned long reverse_hash, size;
1658
1659 cds_lfht_iter_debug_set_ht(ht, iter);
1660
1661 reverse_hash = bit_reverse_ulong(hash);
1662
1663 size = rcu_dereference(ht->size);
1664 bucket = lookup_bucket(ht, size, hash);
1665 /* We can always skip the bucket node initially */
1666 node = rcu_dereference(bucket->next);
1667 node = clear_flag(node);
1668 for (;;) {
1669 if (caa_unlikely(is_end(node))) {
1670 node = next = NULL;
1671 break;
1672 }
1673 if (caa_unlikely(node->reverse_hash > reverse_hash)) {
1674 node = next = NULL;
1675 break;
1676 }
1677 next = rcu_dereference(node->next);
1678 urcu_posix_assert(node == clear_flag(node));
1679 if (caa_likely(!is_removed(next))
1680 && !is_bucket(next)
1681 && node->reverse_hash == reverse_hash
1682 && caa_likely(match(node, key))) {
1683 break;
1684 }
1685 node = clear_flag(next);
1686 }
1687 urcu_posix_assert(!node || !is_bucket(CMM_LOAD_SHARED(node->next)));
1688 iter->node = node;
1689 iter->next = next;
1690 }
1691
1692 void cds_lfht_next_duplicate(struct cds_lfht *ht __attribute__((unused)),
1693 cds_lfht_match_fct match,
1694 const void *key, struct cds_lfht_iter *iter)
1695 {
1696 struct cds_lfht_node *node, *next;
1697 unsigned long reverse_hash;
1698
1699 cds_lfht_iter_debug_assert(ht == iter->lfht);
1700 node = iter->node;
1701 reverse_hash = node->reverse_hash;
1702 next = iter->next;
1703 node = clear_flag(next);
1704
1705 for (;;) {
1706 if (caa_unlikely(is_end(node))) {
1707 node = next = NULL;
1708 break;
1709 }
1710 if (caa_unlikely(node->reverse_hash > reverse_hash)) {
1711 node = next = NULL;
1712 break;
1713 }
1714 next = rcu_dereference(node->next);
1715 if (caa_likely(!is_removed(next))
1716 && !is_bucket(next)
1717 && caa_likely(match(node, key))) {
1718 break;
1719 }
1720 node = clear_flag(next);
1721 }
1722 urcu_posix_assert(!node || !is_bucket(CMM_LOAD_SHARED(node->next)));
1723 iter->node = node;
1724 iter->next = next;
1725 }
1726
1727 void cds_lfht_next(struct cds_lfht *ht __attribute__((unused)),
1728 struct cds_lfht_iter *iter)
1729 {
1730 struct cds_lfht_node *node, *next;
1731
1732 cds_lfht_iter_debug_assert(ht == iter->lfht);
1733 node = clear_flag(iter->next);
1734 for (;;) {
1735 if (caa_unlikely(is_end(node))) {
1736 node = next = NULL;
1737 break;
1738 }
1739 next = rcu_dereference(node->next);
1740 if (caa_likely(!is_removed(next))
1741 && !is_bucket(next)) {
1742 break;
1743 }
1744 node = clear_flag(next);
1745 }
1746 urcu_posix_assert(!node || !is_bucket(CMM_LOAD_SHARED(node->next)));
1747 iter->node = node;
1748 iter->next = next;
1749 }
1750
1751 void cds_lfht_first(struct cds_lfht *ht, struct cds_lfht_iter *iter)
1752 {
1753 cds_lfht_iter_debug_set_ht(ht, iter);
1754 /*
1755 * Get next after first bucket node. The first bucket node is the
1756 * first node of the linked list.
1757 */
1758 iter->next = bucket_at(ht, 0)->next;
1759 cds_lfht_next(ht, iter);
1760 }
1761
1762 void cds_lfht_add(struct cds_lfht *ht, unsigned long hash,
1763 struct cds_lfht_node *node)
1764 {
1765 unsigned long size;
1766
1767 node->reverse_hash = bit_reverse_ulong(hash);
1768 size = rcu_dereference(ht->size);
1769 _cds_lfht_add(ht, hash, NULL, NULL, size, node, NULL, 0);
1770 ht_count_add(ht, size, hash);
1771 }
1772
1773 struct cds_lfht_node *cds_lfht_add_unique(struct cds_lfht *ht,
1774 unsigned long hash,
1775 cds_lfht_match_fct match,
1776 const void *key,
1777 struct cds_lfht_node *node)
1778 {
1779 unsigned long size;
1780 struct cds_lfht_iter iter;
1781
1782 node->reverse_hash = bit_reverse_ulong(hash);
1783 size = rcu_dereference(ht->size);
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 iter.node;
1788 }
1789
1790 struct cds_lfht_node *cds_lfht_add_replace(struct cds_lfht *ht,
1791 unsigned long hash,
1792 cds_lfht_match_fct match,
1793 const void *key,
1794 struct cds_lfht_node *node)
1795 {
1796 unsigned long size;
1797 struct cds_lfht_iter iter;
1798
1799 node->reverse_hash = bit_reverse_ulong(hash);
1800 size = rcu_dereference(ht->size);
1801 for (;;) {
1802 _cds_lfht_add(ht, hash, match, key, size, node, &iter, 0);
1803 if (iter.node == node) {
1804 ht_count_add(ht, size, hash);
1805 return NULL;
1806 }
1807
1808 if (!_cds_lfht_replace(ht, size, iter.node, iter.next, node))
1809 return iter.node;
1810 }
1811 }
1812
1813 int cds_lfht_replace(struct cds_lfht *ht,
1814 struct cds_lfht_iter *old_iter,
1815 unsigned long hash,
1816 cds_lfht_match_fct match,
1817 const void *key,
1818 struct cds_lfht_node *new_node)
1819 {
1820 unsigned long size;
1821
1822 new_node->reverse_hash = bit_reverse_ulong(hash);
1823 if (!old_iter->node)
1824 return -ENOENT;
1825 if (caa_unlikely(old_iter->node->reverse_hash != new_node->reverse_hash))
1826 return -EINVAL;
1827 if (caa_unlikely(!match(old_iter->node, key)))
1828 return -EINVAL;
1829 size = rcu_dereference(ht->size);
1830 return _cds_lfht_replace(ht, size, old_iter->node, old_iter->next,
1831 new_node);
1832 }
1833
1834 int cds_lfht_del(struct cds_lfht *ht, struct cds_lfht_node *node)
1835 {
1836 unsigned long size;
1837 int ret;
1838
1839 size = rcu_dereference(ht->size);
1840 ret = _cds_lfht_del(ht, size, node);
1841 if (!ret) {
1842 unsigned long hash;
1843
1844 hash = bit_reverse_ulong(node->reverse_hash);
1845 ht_count_del(ht, size, hash);
1846 }
1847 return ret;
1848 }
1849
1850 int cds_lfht_is_node_deleted(const struct cds_lfht_node *node)
1851 {
1852 return is_removed(CMM_LOAD_SHARED(node->next));
1853 }
1854
1855 static
1856 int cds_lfht_delete_bucket(struct cds_lfht *ht)
1857 {
1858 struct cds_lfht_node *node;
1859 unsigned long order, i, size;
1860
1861 /* Check that the table is empty */
1862 node = bucket_at(ht, 0);
1863 do {
1864 node = clear_flag(node)->next;
1865 if (!is_bucket(node))
1866 return -EPERM;
1867 urcu_posix_assert(!is_removed(node));
1868 urcu_posix_assert(!is_removal_owner(node));
1869 } while (!is_end(node));
1870 /*
1871 * size accessed without rcu_dereference because hash table is
1872 * being destroyed.
1873 */
1874 size = ht->size;
1875 /* Internal sanity check: all nodes left should be buckets */
1876 for (i = 0; i < size; i++) {
1877 node = bucket_at(ht, i);
1878 dbg_printf("delete bucket: index %lu expected hash %lu hash %lu\n",
1879 i, i, bit_reverse_ulong(node->reverse_hash));
1880 urcu_posix_assert(is_bucket(node->next));
1881 }
1882
1883 for (order = cds_lfht_get_count_order_ulong(size); (long)order >= 0; order--)
1884 cds_lfht_free_bucket_table(ht, order);
1885
1886 return 0;
1887 }
1888
1889 /*
1890 * Should only be called when no more concurrent readers nor writers can
1891 * possibly access the table.
1892 */
1893 int cds_lfht_destroy(struct cds_lfht *ht, pthread_attr_t **attr)
1894 {
1895 int ret;
1896
1897 if (ht->flags & CDS_LFHT_AUTO_RESIZE) {
1898 /* Cancel ongoing resize operations. */
1899 _CMM_STORE_SHARED(ht->in_progress_destroy, 1);
1900 /* Wait for in-flight resize operations to complete */
1901 urcu_workqueue_flush_queued_work(cds_lfht_workqueue);
1902 }
1903 ret = cds_lfht_delete_bucket(ht);
1904 if (ret)
1905 return ret;
1906 free_split_items_count(ht);
1907 if (attr)
1908 *attr = ht->resize_attr;
1909 ret = pthread_mutex_destroy(&ht->resize_mutex);
1910 if (ret)
1911 ret = -EBUSY;
1912 if (ht->flags & CDS_LFHT_AUTO_RESIZE)
1913 cds_lfht_fini_worker(ht->flavor);
1914 poison_free(ht);
1915 return ret;
1916 }
1917
1918 void cds_lfht_count_nodes(struct cds_lfht *ht,
1919 long *approx_before,
1920 unsigned long *count,
1921 long *approx_after)
1922 {
1923 struct cds_lfht_node *node, *next;
1924 unsigned long nr_bucket = 0, nr_removed = 0;
1925
1926 *approx_before = 0;
1927 if (ht->split_count) {
1928 int i;
1929
1930 for (i = 0; i < split_count_mask + 1; i++) {
1931 *approx_before += uatomic_read(&ht->split_count[i].add);
1932 *approx_before -= uatomic_read(&ht->split_count[i].del);
1933 }
1934 }
1935
1936 *count = 0;
1937
1938 /* Count non-bucket nodes in the table */
1939 node = bucket_at(ht, 0);
1940 do {
1941 next = rcu_dereference(node->next);
1942 if (is_removed(next)) {
1943 if (!is_bucket(next))
1944 (nr_removed)++;
1945 else
1946 (nr_bucket)++;
1947 } else if (!is_bucket(next))
1948 (*count)++;
1949 else
1950 (nr_bucket)++;
1951 node = clear_flag(next);
1952 } while (!is_end(node));
1953 dbg_printf("number of logically removed nodes: %lu\n", nr_removed);
1954 dbg_printf("number of bucket nodes: %lu\n", nr_bucket);
1955 *approx_after = 0;
1956 if (ht->split_count) {
1957 int i;
1958
1959 for (i = 0; i < split_count_mask + 1; i++) {
1960 *approx_after += uatomic_read(&ht->split_count[i].add);
1961 *approx_after -= uatomic_read(&ht->split_count[i].del);
1962 }
1963 }
1964 }
1965
1966 /* called with resize mutex held */
1967 static
1968 void _do_cds_lfht_grow(struct cds_lfht *ht,
1969 unsigned long old_size, unsigned long new_size)
1970 {
1971 unsigned long old_order, new_order;
1972
1973 old_order = cds_lfht_get_count_order_ulong(old_size);
1974 new_order = cds_lfht_get_count_order_ulong(new_size);
1975 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1976 old_size, old_order, new_size, new_order);
1977 urcu_posix_assert(new_size > old_size);
1978 init_table(ht, old_order + 1, new_order);
1979 }
1980
1981 /* called with resize mutex held */
1982 static
1983 void _do_cds_lfht_shrink(struct cds_lfht *ht,
1984 unsigned long old_size, unsigned long new_size)
1985 {
1986 unsigned long old_order, new_order;
1987
1988 new_size = max(new_size, MIN_TABLE_SIZE);
1989 old_order = cds_lfht_get_count_order_ulong(old_size);
1990 new_order = cds_lfht_get_count_order_ulong(new_size);
1991 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1992 old_size, old_order, new_size, new_order);
1993 urcu_posix_assert(new_size < old_size);
1994
1995 /* Remove and unlink all bucket nodes to remove. */
1996 fini_table(ht, new_order + 1, old_order);
1997 }
1998
1999
2000 /* called with resize mutex held */
2001 static
2002 void _do_cds_lfht_resize(struct cds_lfht *ht)
2003 {
2004 unsigned long new_size, old_size;
2005
2006 /*
2007 * Resize table, re-do if the target size has changed under us.
2008 */
2009 do {
2010 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
2011 break;
2012 ht->resize_initiated = 1;
2013 old_size = ht->size;
2014 new_size = CMM_LOAD_SHARED(ht->resize_target);
2015 if (old_size < new_size)
2016 _do_cds_lfht_grow(ht, old_size, new_size);
2017 else if (old_size > new_size)
2018 _do_cds_lfht_shrink(ht, old_size, new_size);
2019 ht->resize_initiated = 0;
2020 /* write resize_initiated before read resize_target */
2021 cmm_smp_mb();
2022 } while (ht->size != CMM_LOAD_SHARED(ht->resize_target));
2023 }
2024
2025 static
2026 unsigned long resize_target_grow(struct cds_lfht *ht, unsigned long new_size)
2027 {
2028 return _uatomic_xchg_monotonic_increase(&ht->resize_target, new_size);
2029 }
2030
2031 static
2032 void resize_target_update_count(struct cds_lfht *ht,
2033 unsigned long count)
2034 {
2035 count = max(count, MIN_TABLE_SIZE);
2036 count = min(count, ht->max_nr_buckets);
2037 uatomic_set(&ht->resize_target, count);
2038 }
2039
2040 void cds_lfht_resize(struct cds_lfht *ht, unsigned long new_size)
2041 {
2042 resize_target_update_count(ht, new_size);
2043 CMM_STORE_SHARED(ht->resize_initiated, 1);
2044 mutex_lock(&ht->resize_mutex);
2045 _do_cds_lfht_resize(ht);
2046 mutex_unlock(&ht->resize_mutex);
2047 }
2048
2049 static
2050 void do_resize_cb(struct urcu_work *work)
2051 {
2052 struct resize_work *resize_work =
2053 caa_container_of(work, struct resize_work, work);
2054 struct cds_lfht *ht = resize_work->ht;
2055
2056 ht->flavor->register_thread();
2057 mutex_lock(&ht->resize_mutex);
2058 _do_cds_lfht_resize(ht);
2059 mutex_unlock(&ht->resize_mutex);
2060 ht->flavor->unregister_thread();
2061 poison_free(work);
2062 }
2063
2064 static
2065 void __cds_lfht_resize_lazy_launch(struct cds_lfht *ht)
2066 {
2067 struct resize_work *work;
2068
2069 /* Store resize_target before read resize_initiated */
2070 cmm_smp_mb();
2071 if (!CMM_LOAD_SHARED(ht->resize_initiated)) {
2072 if (CMM_LOAD_SHARED(ht->in_progress_destroy)) {
2073 return;
2074 }
2075 work = malloc(sizeof(*work));
2076 if (work == NULL) {
2077 dbg_printf("error allocating resize work, bailing out\n");
2078 return;
2079 }
2080 work->ht = ht;
2081 urcu_workqueue_queue_work(cds_lfht_workqueue,
2082 &work->work, do_resize_cb);
2083 CMM_STORE_SHARED(ht->resize_initiated, 1);
2084 }
2085 }
2086
2087 static
2088 void cds_lfht_resize_lazy_grow(struct cds_lfht *ht, unsigned long size, int growth)
2089 {
2090 unsigned long target_size = size << growth;
2091
2092 target_size = min(target_size, ht->max_nr_buckets);
2093 if (resize_target_grow(ht, target_size) >= target_size)
2094 return;
2095
2096 __cds_lfht_resize_lazy_launch(ht);
2097 }
2098
2099 /*
2100 * We favor grow operations over shrink. A shrink operation never occurs
2101 * if a grow operation is queued for lazy execution. A grow operation
2102 * cancels any pending shrink lazy execution.
2103 */
2104 static
2105 void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
2106 unsigned long count)
2107 {
2108 if (!(ht->flags & CDS_LFHT_AUTO_RESIZE))
2109 return;
2110 count = max(count, MIN_TABLE_SIZE);
2111 count = min(count, ht->max_nr_buckets);
2112 if (count == size)
2113 return; /* Already the right size, no resize needed */
2114 if (count > size) { /* lazy grow */
2115 if (resize_target_grow(ht, count) >= count)
2116 return;
2117 } else { /* lazy shrink */
2118 for (;;) {
2119 unsigned long s;
2120
2121 s = uatomic_cmpxchg(&ht->resize_target, size, count);
2122 if (s == size)
2123 break; /* no resize needed */
2124 if (s > size)
2125 return; /* growing is/(was just) in progress */
2126 if (s <= count)
2127 return; /* some other thread do shrink */
2128 size = s;
2129 }
2130 }
2131 __cds_lfht_resize_lazy_launch(ht);
2132 }
2133
2134 static void cds_lfht_before_fork(void *priv __attribute__((unused)))
2135 {
2136 if (cds_lfht_workqueue_atfork_nesting++)
2137 return;
2138 mutex_lock(&cds_lfht_fork_mutex);
2139 if (!cds_lfht_workqueue)
2140 return;
2141 urcu_workqueue_pause_worker(cds_lfht_workqueue);
2142 }
2143
2144 static void cds_lfht_after_fork_parent(void *priv __attribute__((unused)))
2145 {
2146 if (--cds_lfht_workqueue_atfork_nesting)
2147 return;
2148 if (!cds_lfht_workqueue)
2149 goto end;
2150 urcu_workqueue_resume_worker(cds_lfht_workqueue);
2151 end:
2152 mutex_unlock(&cds_lfht_fork_mutex);
2153 }
2154
2155 static void cds_lfht_after_fork_child(void *priv __attribute__((unused)))
2156 {
2157 if (--cds_lfht_workqueue_atfork_nesting)
2158 return;
2159 if (!cds_lfht_workqueue)
2160 goto end;
2161 urcu_workqueue_create_worker(cds_lfht_workqueue);
2162 end:
2163 mutex_unlock(&cds_lfht_fork_mutex);
2164 }
2165
2166 static struct urcu_atfork cds_lfht_atfork = {
2167 .before_fork = cds_lfht_before_fork,
2168 .after_fork_parent = cds_lfht_after_fork_parent,
2169 .after_fork_child = cds_lfht_after_fork_child,
2170 };
2171
2172 /*
2173 * Block all signals for the workqueue worker thread to ensure we don't
2174 * disturb the application. The SIGRCU signal needs to be unblocked for
2175 * the urcu-signal flavor.
2176 */
2177 static void cds_lfht_worker_init(
2178 struct urcu_workqueue *workqueue __attribute__((unused)),
2179 void *priv __attribute__((unused)))
2180 {
2181 int ret;
2182 sigset_t mask;
2183
2184 ret = sigfillset(&mask);
2185 if (ret)
2186 urcu_die(errno);
2187 ret = sigdelset(&mask, SIGRCU);
2188 if (ret)
2189 urcu_die(errno);
2190 ret = pthread_sigmask(SIG_SETMASK, &mask, NULL);
2191 if (ret)
2192 urcu_die(ret);
2193 }
2194
2195 static void cds_lfht_init_worker(const struct rcu_flavor_struct *flavor)
2196 {
2197 flavor->register_rculfhash_atfork(&cds_lfht_atfork);
2198
2199 mutex_lock(&cds_lfht_fork_mutex);
2200 if (cds_lfht_workqueue_user_count++)
2201 goto end;
2202 cds_lfht_workqueue = urcu_workqueue_create(0, -1, NULL,
2203 NULL, cds_lfht_worker_init, NULL, NULL, NULL, NULL, NULL);
2204 end:
2205 mutex_unlock(&cds_lfht_fork_mutex);
2206 }
2207
2208 static void cds_lfht_fini_worker(const struct rcu_flavor_struct *flavor)
2209 {
2210 mutex_lock(&cds_lfht_fork_mutex);
2211 if (--cds_lfht_workqueue_user_count)
2212 goto end;
2213 urcu_workqueue_destroy(cds_lfht_workqueue);
2214 cds_lfht_workqueue = NULL;
2215 end:
2216 mutex_unlock(&cds_lfht_fork_mutex);
2217
2218 flavor->unregister_rculfhash_atfork(&cds_lfht_atfork);
2219 }
This page took 0.13132 seconds and 4 git commands to generate.