fix: rculfhash don't change qsbr online state
[urcu.git] / 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 * call_rcu 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 #define _GNU_SOURCE
260 #include <stdlib.h>
261 #include <errno.h>
262 #include <assert.h>
263 #include <stdio.h>
264 #include <stdint.h>
265 #include <string.h>
266 #include <sched.h>
267
268 #include "config.h"
269 #include <urcu.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
280 /*
281 * Split-counters lazily update the global counter each 1024
282 * addition/removal. It automatically keeps track of resize required.
283 * We use the bucket length as indicator for need to expand for small
284 * tables and machines lacking per-cpu data support.
285 */
286 #define COUNT_COMMIT_ORDER 10
287 #define DEFAULT_SPLIT_COUNT_MASK 0xFUL
288 #define CHAIN_LEN_TARGET 1
289 #define CHAIN_LEN_RESIZE_THRESHOLD 3
290
291 /*
292 * Define the minimum table size.
293 */
294 #define MIN_TABLE_ORDER 0
295 #define MIN_TABLE_SIZE (1UL << MIN_TABLE_ORDER)
296
297 /*
298 * Minimum number of bucket nodes to touch per thread to parallelize grow/shrink.
299 */
300 #define MIN_PARTITION_PER_THREAD_ORDER 12
301 #define MIN_PARTITION_PER_THREAD (1UL << MIN_PARTITION_PER_THREAD_ORDER)
302
303 /*
304 * The removed flag needs to be updated atomically with the pointer.
305 * It indicates that no node must attach to the node scheduled for
306 * removal, and that node garbage collection must be performed.
307 * The bucket flag does not require to be updated atomically with the
308 * pointer, but it is added as a pointer low bit flag to save space.
309 * The "removal owner" flag is used to detect which of the "del"
310 * operation that has set the "removed flag" gets to return the removed
311 * node to its caller. Note that the replace operation does not need to
312 * iteract with the "removal owner" flag, because it validates that
313 * the "removed" flag is not set before performing its cmpxchg.
314 */
315 #define REMOVED_FLAG (1UL << 0)
316 #define BUCKET_FLAG (1UL << 1)
317 #define REMOVAL_OWNER_FLAG (1UL << 2)
318 #define FLAGS_MASK ((1UL << 3) - 1)
319
320 /* Value of the end pointer. Should not interact with flags. */
321 #define END_VALUE NULL
322
323 /*
324 * ht_items_count: Split-counters counting the number of node addition
325 * and removal in the table. Only used if the CDS_LFHT_ACCOUNTING flag
326 * is set at hash table creation.
327 *
328 * These are free-running counters, never reset to zero. They count the
329 * number of add/remove, and trigger every (1 << COUNT_COMMIT_ORDER)
330 * operations to update the global counter. We choose a power-of-2 value
331 * for the trigger to deal with 32 or 64-bit overflow of the counter.
332 */
333 struct ht_items_count {
334 unsigned long add, del;
335 } __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
336
337 /*
338 * rcu_resize_work: Contains arguments passed to RCU worker thread
339 * responsible for performing lazy resize.
340 */
341 struct rcu_resize_work {
342 struct rcu_head head;
343 struct cds_lfht *ht;
344 };
345
346 /*
347 * partition_resize_work: Contains arguments passed to worker threads
348 * executing the hash table resize on partitions of the hash table
349 * assigned to each processor's worker thread.
350 */
351 struct partition_resize_work {
352 pthread_t thread_id;
353 struct cds_lfht *ht;
354 unsigned long i, start, len;
355 void (*fct)(struct cds_lfht *ht, unsigned long i,
356 unsigned long start, unsigned long len);
357 };
358
359 /*
360 * Algorithm to reverse bits in a word by lookup table, extended to
361 * 64-bit words.
362 * Source:
363 * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable
364 * Originally from Public Domain.
365 */
366
367 static const uint8_t BitReverseTable256[256] =
368 {
369 #define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64
370 #define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16)
371 #define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 )
372 R6(0), R6(2), R6(1), R6(3)
373 };
374 #undef R2
375 #undef R4
376 #undef R6
377
378 static
379 uint8_t bit_reverse_u8(uint8_t v)
380 {
381 return BitReverseTable256[v];
382 }
383
384 #if (CAA_BITS_PER_LONG == 32)
385 static
386 uint32_t bit_reverse_u32(uint32_t v)
387 {
388 return ((uint32_t) bit_reverse_u8(v) << 24) |
389 ((uint32_t) bit_reverse_u8(v >> 8) << 16) |
390 ((uint32_t) bit_reverse_u8(v >> 16) << 8) |
391 ((uint32_t) bit_reverse_u8(v >> 24));
392 }
393 #else
394 static
395 uint64_t bit_reverse_u64(uint64_t v)
396 {
397 return ((uint64_t) bit_reverse_u8(v) << 56) |
398 ((uint64_t) bit_reverse_u8(v >> 8) << 48) |
399 ((uint64_t) bit_reverse_u8(v >> 16) << 40) |
400 ((uint64_t) bit_reverse_u8(v >> 24) << 32) |
401 ((uint64_t) bit_reverse_u8(v >> 32) << 24) |
402 ((uint64_t) bit_reverse_u8(v >> 40) << 16) |
403 ((uint64_t) bit_reverse_u8(v >> 48) << 8) |
404 ((uint64_t) bit_reverse_u8(v >> 56));
405 }
406 #endif
407
408 static
409 unsigned long bit_reverse_ulong(unsigned long v)
410 {
411 #if (CAA_BITS_PER_LONG == 32)
412 return bit_reverse_u32(v);
413 #else
414 return bit_reverse_u64(v);
415 #endif
416 }
417
418 /*
419 * fls: returns the position of the most significant bit.
420 * Returns 0 if no bit is set, else returns the position of the most
421 * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit).
422 */
423 #if defined(__i386) || defined(__x86_64)
424 static inline
425 unsigned int fls_u32(uint32_t x)
426 {
427 int r;
428
429 asm("bsrl %1,%0\n\t"
430 "jnz 1f\n\t"
431 "movl $-1,%0\n\t"
432 "1:\n\t"
433 : "=r" (r) : "rm" (x));
434 return r + 1;
435 }
436 #define HAS_FLS_U32
437 #endif
438
439 #if defined(__x86_64)
440 static inline
441 unsigned int fls_u64(uint64_t x)
442 {
443 long r;
444
445 asm("bsrq %1,%0\n\t"
446 "jnz 1f\n\t"
447 "movq $-1,%0\n\t"
448 "1:\n\t"
449 : "=r" (r) : "rm" (x));
450 return r + 1;
451 }
452 #define HAS_FLS_U64
453 #endif
454
455 #ifndef HAS_FLS_U64
456 static __attribute__((unused))
457 unsigned int fls_u64(uint64_t x)
458 {
459 unsigned int r = 64;
460
461 if (!x)
462 return 0;
463
464 if (!(x & 0xFFFFFFFF00000000ULL)) {
465 x <<= 32;
466 r -= 32;
467 }
468 if (!(x & 0xFFFF000000000000ULL)) {
469 x <<= 16;
470 r -= 16;
471 }
472 if (!(x & 0xFF00000000000000ULL)) {
473 x <<= 8;
474 r -= 8;
475 }
476 if (!(x & 0xF000000000000000ULL)) {
477 x <<= 4;
478 r -= 4;
479 }
480 if (!(x & 0xC000000000000000ULL)) {
481 x <<= 2;
482 r -= 2;
483 }
484 if (!(x & 0x8000000000000000ULL)) {
485 x <<= 1;
486 r -= 1;
487 }
488 return r;
489 }
490 #endif
491
492 #ifndef HAS_FLS_U32
493 static __attribute__((unused))
494 unsigned int fls_u32(uint32_t x)
495 {
496 unsigned int r = 32;
497
498 if (!x)
499 return 0;
500 if (!(x & 0xFFFF0000U)) {
501 x <<= 16;
502 r -= 16;
503 }
504 if (!(x & 0xFF000000U)) {
505 x <<= 8;
506 r -= 8;
507 }
508 if (!(x & 0xF0000000U)) {
509 x <<= 4;
510 r -= 4;
511 }
512 if (!(x & 0xC0000000U)) {
513 x <<= 2;
514 r -= 2;
515 }
516 if (!(x & 0x80000000U)) {
517 x <<= 1;
518 r -= 1;
519 }
520 return r;
521 }
522 #endif
523
524 unsigned int cds_lfht_fls_ulong(unsigned long x)
525 {
526 #if (CAA_BITS_PER_LONG == 32)
527 return fls_u32(x);
528 #else
529 return fls_u64(x);
530 #endif
531 }
532
533 /*
534 * Return the minimum order for which x <= (1UL << order).
535 * Return -1 if x is 0.
536 */
537 int cds_lfht_get_count_order_u32(uint32_t x)
538 {
539 if (!x)
540 return -1;
541
542 return fls_u32(x - 1);
543 }
544
545 /*
546 * Return the minimum order for which x <= (1UL << order).
547 * Return -1 if x is 0.
548 */
549 int cds_lfht_get_count_order_ulong(unsigned long x)
550 {
551 if (!x)
552 return -1;
553
554 return cds_lfht_fls_ulong(x - 1);
555 }
556
557 static
558 void cds_lfht_resize_lazy_grow(struct cds_lfht *ht, unsigned long size, int growth);
559
560 static
561 void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
562 unsigned long count);
563
564 static long nr_cpus_mask = -1;
565 static long split_count_mask = -1;
566
567 #if defined(HAVE_SYSCONF)
568 static void ht_init_nr_cpus_mask(void)
569 {
570 long maxcpus;
571
572 maxcpus = sysconf(_SC_NPROCESSORS_CONF);
573 if (maxcpus <= 0) {
574 nr_cpus_mask = -2;
575 return;
576 }
577 /*
578 * round up number of CPUs to next power of two, so we
579 * can use & for modulo.
580 */
581 maxcpus = 1UL << cds_lfht_get_count_order_ulong(maxcpus);
582 nr_cpus_mask = maxcpus - 1;
583 }
584 #else /* #if defined(HAVE_SYSCONF) */
585 static void ht_init_nr_cpus_mask(void)
586 {
587 nr_cpus_mask = -2;
588 }
589 #endif /* #else #if defined(HAVE_SYSCONF) */
590
591 static
592 void alloc_split_items_count(struct cds_lfht *ht)
593 {
594 if (nr_cpus_mask == -1) {
595 ht_init_nr_cpus_mask();
596 if (nr_cpus_mask < 0)
597 split_count_mask = DEFAULT_SPLIT_COUNT_MASK;
598 else
599 split_count_mask = nr_cpus_mask;
600 }
601
602 assert(split_count_mask >= 0);
603
604 if (ht->flags & CDS_LFHT_ACCOUNTING) {
605 ht->split_count = calloc(split_count_mask + 1,
606 sizeof(struct ht_items_count));
607 assert(ht->split_count);
608 } else {
609 ht->split_count = NULL;
610 }
611 }
612
613 static
614 void free_split_items_count(struct cds_lfht *ht)
615 {
616 poison_free(ht->split_count);
617 }
618
619 #if defined(HAVE_SCHED_GETCPU)
620 static
621 int ht_get_split_count_index(unsigned long hash)
622 {
623 int cpu;
624
625 assert(split_count_mask >= 0);
626 cpu = sched_getcpu();
627 if (caa_unlikely(cpu < 0))
628 return hash & split_count_mask;
629 else
630 return cpu & split_count_mask;
631 }
632 #else /* #if defined(HAVE_SCHED_GETCPU) */
633 static
634 int ht_get_split_count_index(unsigned long hash)
635 {
636 return hash & split_count_mask;
637 }
638 #endif /* #else #if defined(HAVE_SCHED_GETCPU) */
639
640 static
641 void ht_count_add(struct cds_lfht *ht, unsigned long size, unsigned long hash)
642 {
643 unsigned long split_count;
644 int index;
645 long count;
646
647 if (caa_unlikely(!ht->split_count))
648 return;
649 index = ht_get_split_count_index(hash);
650 split_count = uatomic_add_return(&ht->split_count[index].add, 1);
651 if (caa_likely(split_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))
652 return;
653 /* Only if number of add multiple of 1UL << COUNT_COMMIT_ORDER */
654
655 dbg_printf("add split count %lu\n", split_count);
656 count = uatomic_add_return(&ht->count,
657 1UL << COUNT_COMMIT_ORDER);
658 if (caa_likely(count & (count - 1)))
659 return;
660 /* Only if global count is power of 2 */
661
662 if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) < size)
663 return;
664 dbg_printf("add set global %ld\n", count);
665 cds_lfht_resize_lazy_count(ht, size,
666 count >> (CHAIN_LEN_TARGET - 1));
667 }
668
669 static
670 void ht_count_del(struct cds_lfht *ht, unsigned long size, unsigned long hash)
671 {
672 unsigned long split_count;
673 int index;
674 long count;
675
676 if (caa_unlikely(!ht->split_count))
677 return;
678 index = ht_get_split_count_index(hash);
679 split_count = uatomic_add_return(&ht->split_count[index].del, 1);
680 if (caa_likely(split_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))
681 return;
682 /* Only if number of deletes multiple of 1UL << COUNT_COMMIT_ORDER */
683
684 dbg_printf("del split count %lu\n", split_count);
685 count = uatomic_add_return(&ht->count,
686 -(1UL << COUNT_COMMIT_ORDER));
687 if (caa_likely(count & (count - 1)))
688 return;
689 /* Only if global count is power of 2 */
690
691 if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) >= size)
692 return;
693 dbg_printf("del set global %ld\n", count);
694 /*
695 * Don't shrink table if the number of nodes is below a
696 * certain threshold.
697 */
698 if (count < (1UL << COUNT_COMMIT_ORDER) * (split_count_mask + 1))
699 return;
700 cds_lfht_resize_lazy_count(ht, size,
701 count >> (CHAIN_LEN_TARGET - 1));
702 }
703
704 static
705 void check_resize(struct cds_lfht *ht, unsigned long size, uint32_t chain_len)
706 {
707 unsigned long count;
708
709 if (!(ht->flags & CDS_LFHT_AUTO_RESIZE))
710 return;
711 count = uatomic_read(&ht->count);
712 /*
713 * Use bucket-local length for small table expand and for
714 * environments lacking per-cpu data support.
715 */
716 if (count >= (1UL << COUNT_COMMIT_ORDER))
717 return;
718 if (chain_len > 100)
719 dbg_printf("WARNING: large chain length: %u.\n",
720 chain_len);
721 if (chain_len >= CHAIN_LEN_RESIZE_THRESHOLD)
722 cds_lfht_resize_lazy_grow(ht, size,
723 cds_lfht_get_count_order_u32(chain_len - (CHAIN_LEN_TARGET - 1)));
724 }
725
726 static
727 struct cds_lfht_node *clear_flag(struct cds_lfht_node *node)
728 {
729 return (struct cds_lfht_node *) (((unsigned long) node) & ~FLAGS_MASK);
730 }
731
732 static
733 int is_removed(struct cds_lfht_node *node)
734 {
735 return ((unsigned long) node) & REMOVED_FLAG;
736 }
737
738 static
739 int is_bucket(struct cds_lfht_node *node)
740 {
741 return ((unsigned long) node) & BUCKET_FLAG;
742 }
743
744 static
745 struct cds_lfht_node *flag_bucket(struct cds_lfht_node *node)
746 {
747 return (struct cds_lfht_node *) (((unsigned long) node) | BUCKET_FLAG);
748 }
749
750 static
751 int is_removal_owner(struct cds_lfht_node *node)
752 {
753 return ((unsigned long) node) & REMOVAL_OWNER_FLAG;
754 }
755
756 static
757 struct cds_lfht_node *flag_removal_owner(struct cds_lfht_node *node)
758 {
759 return (struct cds_lfht_node *) (((unsigned long) node) | REMOVAL_OWNER_FLAG);
760 }
761
762 static
763 struct cds_lfht_node *flag_removed_or_removal_owner(struct cds_lfht_node *node)
764 {
765 return (struct cds_lfht_node *) (((unsigned long) node) | REMOVED_FLAG | REMOVAL_OWNER_FLAG);
766 }
767
768 static
769 struct cds_lfht_node *get_end(void)
770 {
771 return (struct cds_lfht_node *) END_VALUE;
772 }
773
774 static
775 int is_end(struct cds_lfht_node *node)
776 {
777 return clear_flag(node) == (struct cds_lfht_node *) END_VALUE;
778 }
779
780 static
781 unsigned long _uatomic_xchg_monotonic_increase(unsigned long *ptr,
782 unsigned long v)
783 {
784 unsigned long old1, old2;
785
786 old1 = uatomic_read(ptr);
787 do {
788 old2 = old1;
789 if (old2 >= v)
790 return old2;
791 } while ((old1 = uatomic_cmpxchg(ptr, old2, v)) != old2);
792 return old2;
793 }
794
795 static
796 void cds_lfht_alloc_bucket_table(struct cds_lfht *ht, unsigned long order)
797 {
798 return ht->mm->alloc_bucket_table(ht, order);
799 }
800
801 /*
802 * cds_lfht_free_bucket_table() should be called with decreasing order.
803 * When cds_lfht_free_bucket_table(0) is called, it means the whole
804 * lfht is destroyed.
805 */
806 static
807 void cds_lfht_free_bucket_table(struct cds_lfht *ht, unsigned long order)
808 {
809 return ht->mm->free_bucket_table(ht, order);
810 }
811
812 static inline
813 struct cds_lfht_node *bucket_at(struct cds_lfht *ht, unsigned long index)
814 {
815 return ht->bucket_at(ht, index);
816 }
817
818 static inline
819 struct cds_lfht_node *lookup_bucket(struct cds_lfht *ht, unsigned long size,
820 unsigned long hash)
821 {
822 assert(size > 0);
823 return bucket_at(ht, hash & (size - 1));
824 }
825
826 /*
827 * Remove all logically deleted nodes from a bucket up to a certain node key.
828 */
829 static
830 void _cds_lfht_gc_bucket(struct cds_lfht_node *bucket, struct cds_lfht_node *node)
831 {
832 struct cds_lfht_node *iter_prev, *iter, *next, *new_next;
833
834 assert(!is_bucket(bucket));
835 assert(!is_removed(bucket));
836 assert(!is_removal_owner(bucket));
837 assert(!is_bucket(node));
838 assert(!is_removed(node));
839 assert(!is_removal_owner(node));
840 for (;;) {
841 iter_prev = bucket;
842 /* We can always skip the bucket node initially */
843 iter = rcu_dereference(iter_prev->next);
844 assert(!is_removed(iter));
845 assert(!is_removal_owner(iter));
846 assert(iter_prev->reverse_hash <= node->reverse_hash);
847 /*
848 * We should never be called with bucket (start of chain)
849 * and logically removed node (end of path compression
850 * marker) being the actual same node. This would be a
851 * bug in the algorithm implementation.
852 */
853 assert(bucket != node);
854 for (;;) {
855 if (caa_unlikely(is_end(iter)))
856 return;
857 if (caa_likely(clear_flag(iter)->reverse_hash > node->reverse_hash))
858 return;
859 next = rcu_dereference(clear_flag(iter)->next);
860 if (caa_likely(is_removed(next)))
861 break;
862 iter_prev = clear_flag(iter);
863 iter = next;
864 }
865 assert(!is_removed(iter));
866 assert(!is_removal_owner(iter));
867 if (is_bucket(iter))
868 new_next = flag_bucket(clear_flag(next));
869 else
870 new_next = clear_flag(next);
871 (void) uatomic_cmpxchg(&iter_prev->next, iter, new_next);
872 }
873 }
874
875 static
876 int _cds_lfht_replace(struct cds_lfht *ht, unsigned long size,
877 struct cds_lfht_node *old_node,
878 struct cds_lfht_node *old_next,
879 struct cds_lfht_node *new_node)
880 {
881 struct cds_lfht_node *bucket, *ret_next;
882
883 if (!old_node) /* Return -ENOENT if asked to replace NULL node */
884 return -ENOENT;
885
886 assert(!is_removed(old_node));
887 assert(!is_removal_owner(old_node));
888 assert(!is_bucket(old_node));
889 assert(!is_removed(new_node));
890 assert(!is_removal_owner(new_node));
891 assert(!is_bucket(new_node));
892 assert(new_node != old_node);
893 for (;;) {
894 /* Insert after node to be replaced */
895 if (is_removed(old_next)) {
896 /*
897 * Too late, the old node has been removed under us
898 * between lookup and replace. Fail.
899 */
900 return -ENOENT;
901 }
902 assert(old_next == clear_flag(old_next));
903 assert(new_node != old_next);
904 /*
905 * REMOVAL_OWNER flag is _NEVER_ set before the REMOVED
906 * flag. It is either set atomically at the same time
907 * (replace) or after (del).
908 */
909 assert(!is_removal_owner(old_next));
910 new_node->next = old_next;
911 /*
912 * Here is the whole trick for lock-free replace: we add
913 * the replacement node _after_ the node we want to
914 * replace by atomically setting its next pointer at the
915 * same time we set its removal flag. Given that
916 * the lookups/get next use an iterator aware of the
917 * next pointer, they will either skip the old node due
918 * to the removal flag and see the new node, or use
919 * the old node, but will not see the new one.
920 * This is a replacement of a node with another node
921 * that has the same value: we are therefore not
922 * removing a value from the hash table. We set both the
923 * REMOVED and REMOVAL_OWNER flags atomically so we own
924 * the node after successful cmpxchg.
925 */
926 ret_next = uatomic_cmpxchg(&old_node->next,
927 old_next, flag_removed_or_removal_owner(new_node));
928 if (ret_next == old_next)
929 break; /* We performed the replacement. */
930 old_next = ret_next;
931 }
932
933 /*
934 * Ensure that the old node is not visible to readers anymore:
935 * lookup for the node, and remove it (along with any other
936 * logically removed node) if found.
937 */
938 bucket = lookup_bucket(ht, size, bit_reverse_ulong(old_node->reverse_hash));
939 _cds_lfht_gc_bucket(bucket, new_node);
940
941 assert(is_removed(CMM_LOAD_SHARED(old_node->next)));
942 return 0;
943 }
944
945 /*
946 * A non-NULL unique_ret pointer uses the "add unique" (or uniquify) add
947 * mode. A NULL unique_ret allows creation of duplicate keys.
948 */
949 static
950 void _cds_lfht_add(struct cds_lfht *ht,
951 unsigned long hash,
952 cds_lfht_match_fct match,
953 const void *key,
954 unsigned long size,
955 struct cds_lfht_node *node,
956 struct cds_lfht_iter *unique_ret,
957 int bucket_flag)
958 {
959 struct cds_lfht_node *iter_prev, *iter, *next, *new_node, *new_next,
960 *return_node;
961 struct cds_lfht_node *bucket;
962
963 assert(!is_bucket(node));
964 assert(!is_removed(node));
965 assert(!is_removal_owner(node));
966 bucket = lookup_bucket(ht, size, hash);
967 for (;;) {
968 uint32_t chain_len = 0;
969
970 /*
971 * iter_prev points to the non-removed node prior to the
972 * insert location.
973 */
974 iter_prev = bucket;
975 /* We can always skip the bucket node initially */
976 iter = rcu_dereference(iter_prev->next);
977 assert(iter_prev->reverse_hash <= node->reverse_hash);
978 for (;;) {
979 if (caa_unlikely(is_end(iter)))
980 goto insert;
981 if (caa_likely(clear_flag(iter)->reverse_hash > node->reverse_hash))
982 goto insert;
983
984 /* bucket node is the first node of the identical-hash-value chain */
985 if (bucket_flag && clear_flag(iter)->reverse_hash == node->reverse_hash)
986 goto insert;
987
988 next = rcu_dereference(clear_flag(iter)->next);
989 if (caa_unlikely(is_removed(next)))
990 goto gc_node;
991
992 /* uniquely add */
993 if (unique_ret
994 && !is_bucket(next)
995 && clear_flag(iter)->reverse_hash == node->reverse_hash) {
996 struct cds_lfht_iter d_iter = { .node = node, .next = iter, };
997
998 /*
999 * uniquely adding inserts the node as the first
1000 * node of the identical-hash-value node chain.
1001 *
1002 * This semantic ensures no duplicated keys
1003 * should ever be observable in the table
1004 * (including traversing the table node by
1005 * node by forward iterations)
1006 */
1007 cds_lfht_next_duplicate(ht, match, key, &d_iter);
1008 if (!d_iter.node)
1009 goto insert;
1010
1011 *unique_ret = d_iter;
1012 return;
1013 }
1014
1015 /* Only account for identical reverse hash once */
1016 if (iter_prev->reverse_hash != clear_flag(iter)->reverse_hash
1017 && !is_bucket(next))
1018 check_resize(ht, size, ++chain_len);
1019 iter_prev = clear_flag(iter);
1020 iter = next;
1021 }
1022
1023 insert:
1024 assert(node != clear_flag(iter));
1025 assert(!is_removed(iter_prev));
1026 assert(!is_removal_owner(iter_prev));
1027 assert(!is_removed(iter));
1028 assert(!is_removal_owner(iter));
1029 assert(iter_prev != node);
1030 if (!bucket_flag)
1031 node->next = clear_flag(iter);
1032 else
1033 node->next = flag_bucket(clear_flag(iter));
1034 if (is_bucket(iter))
1035 new_node = flag_bucket(node);
1036 else
1037 new_node = node;
1038 if (uatomic_cmpxchg(&iter_prev->next, iter,
1039 new_node) != iter) {
1040 continue; /* retry */
1041 } else {
1042 return_node = node;
1043 goto end;
1044 }
1045
1046 gc_node:
1047 assert(!is_removed(iter));
1048 assert(!is_removal_owner(iter));
1049 if (is_bucket(iter))
1050 new_next = flag_bucket(clear_flag(next));
1051 else
1052 new_next = clear_flag(next);
1053 (void) uatomic_cmpxchg(&iter_prev->next, iter, new_next);
1054 /* retry */
1055 }
1056 end:
1057 if (unique_ret) {
1058 unique_ret->node = return_node;
1059 /* unique_ret->next left unset, never used. */
1060 }
1061 }
1062
1063 static
1064 int _cds_lfht_del(struct cds_lfht *ht, unsigned long size,
1065 struct cds_lfht_node *node)
1066 {
1067 struct cds_lfht_node *bucket, *next;
1068
1069 if (!node) /* Return -ENOENT if asked to delete NULL node */
1070 return -ENOENT;
1071
1072 /* logically delete the node */
1073 assert(!is_bucket(node));
1074 assert(!is_removed(node));
1075 assert(!is_removal_owner(node));
1076
1077 /*
1078 * We are first checking if the node had previously been
1079 * logically removed (this check is not atomic with setting the
1080 * logical removal flag). Return -ENOENT if the node had
1081 * previously been removed.
1082 */
1083 next = CMM_LOAD_SHARED(node->next); /* next is not dereferenced */
1084 if (caa_unlikely(is_removed(next)))
1085 return -ENOENT;
1086 assert(!is_bucket(next));
1087 /*
1088 * The del operation semantic guarantees a full memory barrier
1089 * before the uatomic_or atomic commit of the deletion flag.
1090 */
1091 cmm_smp_mb__before_uatomic_or();
1092 /*
1093 * We set the REMOVED_FLAG unconditionally. Note that there may
1094 * be more than one concurrent thread setting this flag.
1095 * Knowing which wins the race will be known after the garbage
1096 * collection phase, stay tuned!
1097 */
1098 uatomic_or(&node->next, REMOVED_FLAG);
1099 /* We performed the (logical) deletion. */
1100
1101 /*
1102 * Ensure that the node is not visible to readers anymore: lookup for
1103 * the node, and remove it (along with any other logically removed node)
1104 * if found.
1105 */
1106 bucket = lookup_bucket(ht, size, bit_reverse_ulong(node->reverse_hash));
1107 _cds_lfht_gc_bucket(bucket, node);
1108
1109 assert(is_removed(CMM_LOAD_SHARED(node->next)));
1110 /*
1111 * Last phase: atomically exchange node->next with a version
1112 * having "REMOVAL_OWNER_FLAG" set. If the returned node->next
1113 * pointer did _not_ have "REMOVAL_OWNER_FLAG" set, we now own
1114 * the node and win the removal race.
1115 * It is interesting to note that all "add" paths are forbidden
1116 * to change the next pointer starting from the point where the
1117 * REMOVED_FLAG is set, so here using a read, followed by a
1118 * xchg() suffice to guarantee that the xchg() will ever only
1119 * set the "REMOVAL_OWNER_FLAG" (or change nothing if the flag
1120 * was already set).
1121 */
1122 if (!is_removal_owner(uatomic_xchg(&node->next,
1123 flag_removal_owner(node->next))))
1124 return 0;
1125 else
1126 return -ENOENT;
1127 }
1128
1129 static
1130 void *partition_resize_thread(void *arg)
1131 {
1132 struct partition_resize_work *work = arg;
1133
1134 work->ht->flavor->register_thread();
1135 work->fct(work->ht, work->i, work->start, work->len);
1136 work->ht->flavor->unregister_thread();
1137 return NULL;
1138 }
1139
1140 static
1141 void partition_resize_helper(struct cds_lfht *ht, unsigned long i,
1142 unsigned long len,
1143 void (*fct)(struct cds_lfht *ht, unsigned long i,
1144 unsigned long start, unsigned long len))
1145 {
1146 unsigned long partition_len;
1147 struct partition_resize_work *work;
1148 int thread, ret;
1149 unsigned long nr_threads;
1150
1151 /*
1152 * Note: nr_cpus_mask + 1 is always power of 2.
1153 * We spawn just the number of threads we need to satisfy the minimum
1154 * partition size, up to the number of CPUs in the system.
1155 */
1156 if (nr_cpus_mask > 0) {
1157 nr_threads = min(nr_cpus_mask + 1,
1158 len >> MIN_PARTITION_PER_THREAD_ORDER);
1159 } else {
1160 nr_threads = 1;
1161 }
1162 partition_len = len >> cds_lfht_get_count_order_ulong(nr_threads);
1163 work = calloc(nr_threads, sizeof(*work));
1164 assert(work);
1165 for (thread = 0; thread < nr_threads; thread++) {
1166 work[thread].ht = ht;
1167 work[thread].i = i;
1168 work[thread].len = partition_len;
1169 work[thread].start = thread * partition_len;
1170 work[thread].fct = fct;
1171 ret = pthread_create(&(work[thread].thread_id), ht->resize_attr,
1172 partition_resize_thread, &work[thread]);
1173 assert(!ret);
1174 }
1175 for (thread = 0; thread < nr_threads; thread++) {
1176 ret = pthread_join(work[thread].thread_id, NULL);
1177 assert(!ret);
1178 }
1179 free(work);
1180 }
1181
1182 /*
1183 * Holding RCU read lock to protect _cds_lfht_add against memory
1184 * reclaim that could be performed by other call_rcu worker threads (ABA
1185 * problem).
1186 *
1187 * When we reach a certain length, we can split this population phase over
1188 * many worker threads, based on the number of CPUs available in the system.
1189 * This should therefore take care of not having the expand lagging behind too
1190 * many concurrent insertion threads by using the scheduler's ability to
1191 * schedule bucket node population fairly with insertions.
1192 */
1193 static
1194 void init_table_populate_partition(struct cds_lfht *ht, unsigned long i,
1195 unsigned long start, unsigned long len)
1196 {
1197 unsigned long j, size = 1UL << (i - 1);
1198
1199 assert(i > MIN_TABLE_ORDER);
1200 ht->flavor->read_lock();
1201 for (j = size + start; j < size + start + len; j++) {
1202 struct cds_lfht_node *new_node = bucket_at(ht, j);
1203
1204 assert(j >= size && j < (size << 1));
1205 dbg_printf("init populate: order %lu index %lu hash %lu\n",
1206 i, j, j);
1207 new_node->reverse_hash = bit_reverse_ulong(j);
1208 _cds_lfht_add(ht, j, NULL, NULL, size, new_node, NULL, 1);
1209 }
1210 ht->flavor->read_unlock();
1211 }
1212
1213 static
1214 void init_table_populate(struct cds_lfht *ht, unsigned long i,
1215 unsigned long len)
1216 {
1217 assert(nr_cpus_mask != -1);
1218 if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD) {
1219 ht->flavor->thread_online();
1220 init_table_populate_partition(ht, i, 0, len);
1221 ht->flavor->thread_offline();
1222 return;
1223 }
1224 partition_resize_helper(ht, i, len, init_table_populate_partition);
1225 }
1226
1227 static
1228 void init_table(struct cds_lfht *ht,
1229 unsigned long first_order, unsigned long last_order)
1230 {
1231 unsigned long i;
1232
1233 dbg_printf("init table: first_order %lu last_order %lu\n",
1234 first_order, last_order);
1235 assert(first_order > MIN_TABLE_ORDER);
1236 for (i = first_order; i <= last_order; i++) {
1237 unsigned long len;
1238
1239 len = 1UL << (i - 1);
1240 dbg_printf("init order %lu len: %lu\n", i, len);
1241
1242 /* Stop expand if the resize target changes under us */
1243 if (CMM_LOAD_SHARED(ht->resize_target) < (1UL << i))
1244 break;
1245
1246 cds_lfht_alloc_bucket_table(ht, i);
1247
1248 /*
1249 * Set all bucket nodes reverse hash values for a level and
1250 * link all bucket nodes into the table.
1251 */
1252 init_table_populate(ht, i, len);
1253
1254 /*
1255 * Update table size.
1256 */
1257 cmm_smp_wmb(); /* populate data before RCU size */
1258 CMM_STORE_SHARED(ht->size, 1UL << i);
1259
1260 dbg_printf("init new size: %lu\n", 1UL << i);
1261 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1262 break;
1263 }
1264 }
1265
1266 /*
1267 * Holding RCU read lock to protect _cds_lfht_remove against memory
1268 * reclaim that could be performed by other call_rcu worker threads (ABA
1269 * problem).
1270 * For a single level, we logically remove and garbage collect each node.
1271 *
1272 * As a design choice, we perform logical removal and garbage collection on a
1273 * node-per-node basis to simplify this algorithm. We also assume keeping good
1274 * cache locality of the operation would overweight possible performance gain
1275 * that could be achieved by batching garbage collection for multiple levels.
1276 * However, this would have to be justified by benchmarks.
1277 *
1278 * Concurrent removal and add operations are helping us perform garbage
1279 * collection of logically removed nodes. We guarantee that all logically
1280 * removed nodes have been garbage-collected (unlinked) before call_rcu is
1281 * invoked to free a hole level of bucket nodes (after a grace period).
1282 *
1283 * Logical removal and garbage collection can therefore be done in batch
1284 * or on a node-per-node basis, as long as the guarantee above holds.
1285 *
1286 * When we reach a certain length, we can split this removal over many worker
1287 * threads, based on the number of CPUs available in the system. This should
1288 * take care of not letting resize process lag behind too many concurrent
1289 * updater threads actively inserting into the hash table.
1290 */
1291 static
1292 void remove_table_partition(struct cds_lfht *ht, unsigned long i,
1293 unsigned long start, unsigned long len)
1294 {
1295 unsigned long j, size = 1UL << (i - 1);
1296
1297 assert(i > MIN_TABLE_ORDER);
1298 ht->flavor->read_lock();
1299 for (j = size + start; j < size + start + len; j++) {
1300 struct cds_lfht_node *fini_bucket = bucket_at(ht, j);
1301 struct cds_lfht_node *parent_bucket = bucket_at(ht, j - size);
1302
1303 assert(j >= size && j < (size << 1));
1304 dbg_printf("remove entry: order %lu index %lu hash %lu\n",
1305 i, j, j);
1306 /* Set the REMOVED_FLAG to freeze the ->next for gc */
1307 uatomic_or(&fini_bucket->next, REMOVED_FLAG);
1308 _cds_lfht_gc_bucket(parent_bucket, fini_bucket);
1309 }
1310 ht->flavor->read_unlock();
1311 }
1312
1313 static
1314 void remove_table(struct cds_lfht *ht, unsigned long i, unsigned long len)
1315 {
1316
1317 assert(nr_cpus_mask != -1);
1318 if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD) {
1319 ht->flavor->thread_online();
1320 remove_table_partition(ht, i, 0, len);
1321 ht->flavor->thread_offline();
1322 return;
1323 }
1324 partition_resize_helper(ht, i, len, remove_table_partition);
1325 }
1326
1327 /*
1328 * fini_table() is never called for first_order == 0, which is why
1329 * free_by_rcu_order == 0 can be used as criterion to know if free must
1330 * be called.
1331 */
1332 static
1333 void fini_table(struct cds_lfht *ht,
1334 unsigned long first_order, unsigned long last_order)
1335 {
1336 long i;
1337 unsigned long free_by_rcu_order = 0;
1338
1339 dbg_printf("fini table: first_order %lu last_order %lu\n",
1340 first_order, last_order);
1341 assert(first_order > MIN_TABLE_ORDER);
1342 for (i = last_order; i >= first_order; i--) {
1343 unsigned long len;
1344
1345 len = 1UL << (i - 1);
1346 dbg_printf("fini order %lu len: %lu\n", i, len);
1347
1348 /* Stop shrink if the resize target changes under us */
1349 if (CMM_LOAD_SHARED(ht->resize_target) > (1UL << (i - 1)))
1350 break;
1351
1352 cmm_smp_wmb(); /* populate data before RCU size */
1353 CMM_STORE_SHARED(ht->size, 1UL << (i - 1));
1354
1355 /*
1356 * We need to wait for all add operations to reach Q.S. (and
1357 * thus use the new table for lookups) before we can start
1358 * releasing the old bucket nodes. Otherwise their lookup will
1359 * return a logically removed node as insert position.
1360 */
1361 ht->flavor->update_synchronize_rcu();
1362 if (free_by_rcu_order)
1363 cds_lfht_free_bucket_table(ht, free_by_rcu_order);
1364
1365 /*
1366 * Set "removed" flag in bucket nodes about to be removed.
1367 * Unlink all now-logically-removed bucket node pointers.
1368 * Concurrent add/remove operation are helping us doing
1369 * the gc.
1370 */
1371 remove_table(ht, i, len);
1372
1373 free_by_rcu_order = i;
1374
1375 dbg_printf("fini new size: %lu\n", 1UL << i);
1376 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1377 break;
1378 }
1379
1380 if (free_by_rcu_order) {
1381 ht->flavor->update_synchronize_rcu();
1382 cds_lfht_free_bucket_table(ht, free_by_rcu_order);
1383 }
1384 }
1385
1386 static
1387 void cds_lfht_create_bucket(struct cds_lfht *ht, unsigned long size)
1388 {
1389 struct cds_lfht_node *prev, *node;
1390 unsigned long order, len, i;
1391
1392 cds_lfht_alloc_bucket_table(ht, 0);
1393
1394 dbg_printf("create bucket: order 0 index 0 hash 0\n");
1395 node = bucket_at(ht, 0);
1396 node->next = flag_bucket(get_end());
1397 node->reverse_hash = 0;
1398
1399 for (order = 1; order < cds_lfht_get_count_order_ulong(size) + 1; order++) {
1400 len = 1UL << (order - 1);
1401 cds_lfht_alloc_bucket_table(ht, order);
1402
1403 for (i = 0; i < len; i++) {
1404 /*
1405 * Now, we are trying to init the node with the
1406 * hash=(len+i) (which is also a bucket with the
1407 * index=(len+i)) and insert it into the hash table,
1408 * so this node has to be inserted after the bucket
1409 * with the index=(len+i)&(len-1)=i. And because there
1410 * is no other non-bucket node nor bucket node with
1411 * larger index/hash inserted, so the bucket node
1412 * being inserted should be inserted directly linked
1413 * after the bucket node with index=i.
1414 */
1415 prev = bucket_at(ht, i);
1416 node = bucket_at(ht, len + i);
1417
1418 dbg_printf("create bucket: order %lu index %lu hash %lu\n",
1419 order, len + i, len + i);
1420 node->reverse_hash = bit_reverse_ulong(len + i);
1421
1422 /* insert after prev */
1423 assert(is_bucket(prev->next));
1424 node->next = prev->next;
1425 prev->next = flag_bucket(node);
1426 }
1427 }
1428 }
1429
1430 struct cds_lfht *_cds_lfht_new(unsigned long init_size,
1431 unsigned long min_nr_alloc_buckets,
1432 unsigned long max_nr_buckets,
1433 int flags,
1434 const struct cds_lfht_mm_type *mm,
1435 const struct rcu_flavor_struct *flavor,
1436 pthread_attr_t *attr)
1437 {
1438 struct cds_lfht *ht;
1439 unsigned long order;
1440
1441 /* min_nr_alloc_buckets must be power of two */
1442 if (!min_nr_alloc_buckets || (min_nr_alloc_buckets & (min_nr_alloc_buckets - 1)))
1443 return NULL;
1444
1445 /* init_size must be power of two */
1446 if (!init_size || (init_size & (init_size - 1)))
1447 return NULL;
1448
1449 /*
1450 * Memory management plugin default.
1451 */
1452 if (!mm) {
1453 if (CAA_BITS_PER_LONG > 32
1454 && max_nr_buckets
1455 && max_nr_buckets <= (1ULL << 32)) {
1456 /*
1457 * For 64-bit architectures, with max number of
1458 * buckets small enough not to use the entire
1459 * 64-bit memory mapping space (and allowing a
1460 * fair number of hash table instances), use the
1461 * mmap allocator, which is faster than the
1462 * order allocator.
1463 */
1464 mm = &cds_lfht_mm_mmap;
1465 } else {
1466 /*
1467 * The fallback is to use the order allocator.
1468 */
1469 mm = &cds_lfht_mm_order;
1470 }
1471 }
1472
1473 /* max_nr_buckets == 0 for order based mm means infinite */
1474 if (mm == &cds_lfht_mm_order && !max_nr_buckets)
1475 max_nr_buckets = 1UL << (MAX_TABLE_ORDER - 1);
1476
1477 /* max_nr_buckets must be power of two */
1478 if (!max_nr_buckets || (max_nr_buckets & (max_nr_buckets - 1)))
1479 return NULL;
1480
1481 min_nr_alloc_buckets = max(min_nr_alloc_buckets, MIN_TABLE_SIZE);
1482 init_size = max(init_size, MIN_TABLE_SIZE);
1483 max_nr_buckets = max(max_nr_buckets, min_nr_alloc_buckets);
1484 init_size = min(init_size, max_nr_buckets);
1485
1486 ht = mm->alloc_cds_lfht(min_nr_alloc_buckets, max_nr_buckets);
1487 assert(ht);
1488 assert(ht->mm == mm);
1489 assert(ht->bucket_at == mm->bucket_at);
1490
1491 ht->flags = flags;
1492 ht->flavor = flavor;
1493 ht->resize_attr = attr;
1494 alloc_split_items_count(ht);
1495 /* this mutex should not nest in read-side C.S. */
1496 pthread_mutex_init(&ht->resize_mutex, NULL);
1497 order = cds_lfht_get_count_order_ulong(init_size);
1498 ht->resize_target = 1UL << order;
1499 cds_lfht_create_bucket(ht, 1UL << order);
1500 ht->size = 1UL << order;
1501 return ht;
1502 }
1503
1504 void cds_lfht_lookup(struct cds_lfht *ht, unsigned long hash,
1505 cds_lfht_match_fct match, const void *key,
1506 struct cds_lfht_iter *iter)
1507 {
1508 struct cds_lfht_node *node, *next, *bucket;
1509 unsigned long reverse_hash, size;
1510
1511 reverse_hash = bit_reverse_ulong(hash);
1512
1513 size = rcu_dereference(ht->size);
1514 bucket = lookup_bucket(ht, size, hash);
1515 /* We can always skip the bucket node initially */
1516 node = rcu_dereference(bucket->next);
1517 node = clear_flag(node);
1518 for (;;) {
1519 if (caa_unlikely(is_end(node))) {
1520 node = next = NULL;
1521 break;
1522 }
1523 if (caa_unlikely(node->reverse_hash > reverse_hash)) {
1524 node = next = NULL;
1525 break;
1526 }
1527 next = rcu_dereference(node->next);
1528 assert(node == clear_flag(node));
1529 if (caa_likely(!is_removed(next))
1530 && !is_bucket(next)
1531 && node->reverse_hash == reverse_hash
1532 && caa_likely(match(node, key))) {
1533 break;
1534 }
1535 node = clear_flag(next);
1536 }
1537 assert(!node || !is_bucket(CMM_LOAD_SHARED(node->next)));
1538 iter->node = node;
1539 iter->next = next;
1540 }
1541
1542 void cds_lfht_next_duplicate(struct cds_lfht *ht, cds_lfht_match_fct match,
1543 const void *key, struct cds_lfht_iter *iter)
1544 {
1545 struct cds_lfht_node *node, *next;
1546 unsigned long reverse_hash;
1547
1548 node = iter->node;
1549 reverse_hash = node->reverse_hash;
1550 next = iter->next;
1551 node = clear_flag(next);
1552
1553 for (;;) {
1554 if (caa_unlikely(is_end(node))) {
1555 node = next = NULL;
1556 break;
1557 }
1558 if (caa_unlikely(node->reverse_hash > reverse_hash)) {
1559 node = next = NULL;
1560 break;
1561 }
1562 next = rcu_dereference(node->next);
1563 if (caa_likely(!is_removed(next))
1564 && !is_bucket(next)
1565 && caa_likely(match(node, key))) {
1566 break;
1567 }
1568 node = clear_flag(next);
1569 }
1570 assert(!node || !is_bucket(CMM_LOAD_SHARED(node->next)));
1571 iter->node = node;
1572 iter->next = next;
1573 }
1574
1575 void cds_lfht_next(struct cds_lfht *ht, struct cds_lfht_iter *iter)
1576 {
1577 struct cds_lfht_node *node, *next;
1578
1579 node = clear_flag(iter->next);
1580 for (;;) {
1581 if (caa_unlikely(is_end(node))) {
1582 node = next = NULL;
1583 break;
1584 }
1585 next = rcu_dereference(node->next);
1586 if (caa_likely(!is_removed(next))
1587 && !is_bucket(next)) {
1588 break;
1589 }
1590 node = clear_flag(next);
1591 }
1592 assert(!node || !is_bucket(CMM_LOAD_SHARED(node->next)));
1593 iter->node = node;
1594 iter->next = next;
1595 }
1596
1597 void cds_lfht_first(struct cds_lfht *ht, struct cds_lfht_iter *iter)
1598 {
1599 /*
1600 * Get next after first bucket node. The first bucket node is the
1601 * first node of the linked list.
1602 */
1603 iter->next = bucket_at(ht, 0)->next;
1604 cds_lfht_next(ht, iter);
1605 }
1606
1607 void cds_lfht_add(struct cds_lfht *ht, unsigned long hash,
1608 struct cds_lfht_node *node)
1609 {
1610 unsigned long size;
1611
1612 node->reverse_hash = bit_reverse_ulong(hash);
1613 size = rcu_dereference(ht->size);
1614 _cds_lfht_add(ht, hash, NULL, NULL, size, node, NULL, 0);
1615 ht_count_add(ht, size, hash);
1616 }
1617
1618 struct cds_lfht_node *cds_lfht_add_unique(struct cds_lfht *ht,
1619 unsigned long hash,
1620 cds_lfht_match_fct match,
1621 const void *key,
1622 struct cds_lfht_node *node)
1623 {
1624 unsigned long size;
1625 struct cds_lfht_iter iter;
1626
1627 node->reverse_hash = bit_reverse_ulong(hash);
1628 size = rcu_dereference(ht->size);
1629 _cds_lfht_add(ht, hash, match, key, size, node, &iter, 0);
1630 if (iter.node == node)
1631 ht_count_add(ht, size, hash);
1632 return iter.node;
1633 }
1634
1635 struct cds_lfht_node *cds_lfht_add_replace(struct cds_lfht *ht,
1636 unsigned long hash,
1637 cds_lfht_match_fct match,
1638 const void *key,
1639 struct cds_lfht_node *node)
1640 {
1641 unsigned long size;
1642 struct cds_lfht_iter iter;
1643
1644 node->reverse_hash = bit_reverse_ulong(hash);
1645 size = rcu_dereference(ht->size);
1646 for (;;) {
1647 _cds_lfht_add(ht, hash, match, key, size, node, &iter, 0);
1648 if (iter.node == node) {
1649 ht_count_add(ht, size, hash);
1650 return NULL;
1651 }
1652
1653 if (!_cds_lfht_replace(ht, size, iter.node, iter.next, node))
1654 return iter.node;
1655 }
1656 }
1657
1658 int cds_lfht_replace(struct cds_lfht *ht,
1659 struct cds_lfht_iter *old_iter,
1660 unsigned long hash,
1661 cds_lfht_match_fct match,
1662 const void *key,
1663 struct cds_lfht_node *new_node)
1664 {
1665 unsigned long size;
1666
1667 new_node->reverse_hash = bit_reverse_ulong(hash);
1668 if (!old_iter->node)
1669 return -ENOENT;
1670 if (caa_unlikely(old_iter->node->reverse_hash != new_node->reverse_hash))
1671 return -EINVAL;
1672 if (caa_unlikely(!match(old_iter->node, key)))
1673 return -EINVAL;
1674 size = rcu_dereference(ht->size);
1675 return _cds_lfht_replace(ht, size, old_iter->node, old_iter->next,
1676 new_node);
1677 }
1678
1679 int cds_lfht_del(struct cds_lfht *ht, struct cds_lfht_node *node)
1680 {
1681 unsigned long size;
1682 int ret;
1683
1684 size = rcu_dereference(ht->size);
1685 ret = _cds_lfht_del(ht, size, node);
1686 if (!ret) {
1687 unsigned long hash;
1688
1689 hash = bit_reverse_ulong(node->reverse_hash);
1690 ht_count_del(ht, size, hash);
1691 }
1692 return ret;
1693 }
1694
1695 int cds_lfht_is_node_deleted(struct cds_lfht_node *node)
1696 {
1697 return is_removed(CMM_LOAD_SHARED(node->next));
1698 }
1699
1700 static
1701 int cds_lfht_delete_bucket(struct cds_lfht *ht)
1702 {
1703 struct cds_lfht_node *node;
1704 unsigned long order, i, size;
1705
1706 /* Check that the table is empty */
1707 node = bucket_at(ht, 0);
1708 do {
1709 node = clear_flag(node)->next;
1710 if (!is_bucket(node))
1711 return -EPERM;
1712 assert(!is_removed(node));
1713 assert(!is_removal_owner(node));
1714 } while (!is_end(node));
1715 /*
1716 * size accessed without rcu_dereference because hash table is
1717 * being destroyed.
1718 */
1719 size = ht->size;
1720 /* Internal sanity check: all nodes left should be buckets */
1721 for (i = 0; i < size; i++) {
1722 node = bucket_at(ht, i);
1723 dbg_printf("delete bucket: index %lu expected hash %lu hash %lu\n",
1724 i, i, bit_reverse_ulong(node->reverse_hash));
1725 assert(is_bucket(node->next));
1726 }
1727
1728 for (order = cds_lfht_get_count_order_ulong(size); (long)order >= 0; order--)
1729 cds_lfht_free_bucket_table(ht, order);
1730
1731 return 0;
1732 }
1733
1734 /*
1735 * Should only be called when no more concurrent readers nor writers can
1736 * possibly access the table.
1737 */
1738 int cds_lfht_destroy(struct cds_lfht *ht, pthread_attr_t **attr)
1739 {
1740 int ret, was_online;
1741
1742 /* Wait for in-flight resize operations to complete */
1743 _CMM_STORE_SHARED(ht->in_progress_destroy, 1);
1744 cmm_smp_mb(); /* Store destroy before load resize */
1745 was_online = ht->flavor->read_ongoing();
1746 if (was_online)
1747 ht->flavor->thread_offline();
1748 while (uatomic_read(&ht->in_progress_resize))
1749 poll(NULL, 0, 100); /* wait for 100ms */
1750 if (was_online)
1751 ht->flavor->thread_online();
1752 ret = cds_lfht_delete_bucket(ht);
1753 if (ret)
1754 return ret;
1755 free_split_items_count(ht);
1756 if (attr)
1757 *attr = ht->resize_attr;
1758 poison_free(ht);
1759 return ret;
1760 }
1761
1762 void cds_lfht_count_nodes(struct cds_lfht *ht,
1763 long *approx_before,
1764 unsigned long *count,
1765 long *approx_after)
1766 {
1767 struct cds_lfht_node *node, *next;
1768 unsigned long nr_bucket = 0, nr_removed = 0;
1769
1770 *approx_before = 0;
1771 if (ht->split_count) {
1772 int i;
1773
1774 for (i = 0; i < split_count_mask + 1; i++) {
1775 *approx_before += uatomic_read(&ht->split_count[i].add);
1776 *approx_before -= uatomic_read(&ht->split_count[i].del);
1777 }
1778 }
1779
1780 *count = 0;
1781
1782 /* Count non-bucket nodes in the table */
1783 node = bucket_at(ht, 0);
1784 do {
1785 next = rcu_dereference(node->next);
1786 if (is_removed(next)) {
1787 if (!is_bucket(next))
1788 (nr_removed)++;
1789 else
1790 (nr_bucket)++;
1791 } else if (!is_bucket(next))
1792 (*count)++;
1793 else
1794 (nr_bucket)++;
1795 node = clear_flag(next);
1796 } while (!is_end(node));
1797 dbg_printf("number of logically removed nodes: %lu\n", nr_removed);
1798 dbg_printf("number of bucket nodes: %lu\n", nr_bucket);
1799 *approx_after = 0;
1800 if (ht->split_count) {
1801 int i;
1802
1803 for (i = 0; i < split_count_mask + 1; i++) {
1804 *approx_after += uatomic_read(&ht->split_count[i].add);
1805 *approx_after -= uatomic_read(&ht->split_count[i].del);
1806 }
1807 }
1808 }
1809
1810 /* called with resize mutex held */
1811 static
1812 void _do_cds_lfht_grow(struct cds_lfht *ht,
1813 unsigned long old_size, unsigned long new_size)
1814 {
1815 unsigned long old_order, new_order;
1816
1817 old_order = cds_lfht_get_count_order_ulong(old_size);
1818 new_order = cds_lfht_get_count_order_ulong(new_size);
1819 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1820 old_size, old_order, new_size, new_order);
1821 assert(new_size > old_size);
1822 init_table(ht, old_order + 1, new_order);
1823 }
1824
1825 /* called with resize mutex held */
1826 static
1827 void _do_cds_lfht_shrink(struct cds_lfht *ht,
1828 unsigned long old_size, unsigned long new_size)
1829 {
1830 unsigned long old_order, new_order;
1831
1832 new_size = max(new_size, MIN_TABLE_SIZE);
1833 old_order = cds_lfht_get_count_order_ulong(old_size);
1834 new_order = cds_lfht_get_count_order_ulong(new_size);
1835 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1836 old_size, old_order, new_size, new_order);
1837 assert(new_size < old_size);
1838
1839 /* Remove and unlink all bucket nodes to remove. */
1840 fini_table(ht, new_order + 1, old_order);
1841 }
1842
1843
1844 /* called with resize mutex held */
1845 static
1846 void _do_cds_lfht_resize(struct cds_lfht *ht)
1847 {
1848 unsigned long new_size, old_size;
1849
1850 /*
1851 * Resize table, re-do if the target size has changed under us.
1852 */
1853 do {
1854 assert(uatomic_read(&ht->in_progress_resize));
1855 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1856 break;
1857 ht->resize_initiated = 1;
1858 old_size = ht->size;
1859 new_size = CMM_LOAD_SHARED(ht->resize_target);
1860 if (old_size < new_size)
1861 _do_cds_lfht_grow(ht, old_size, new_size);
1862 else if (old_size > new_size)
1863 _do_cds_lfht_shrink(ht, old_size, new_size);
1864 ht->resize_initiated = 0;
1865 /* write resize_initiated before read resize_target */
1866 cmm_smp_mb();
1867 } while (ht->size != CMM_LOAD_SHARED(ht->resize_target));
1868 }
1869
1870 static
1871 unsigned long resize_target_grow(struct cds_lfht *ht, unsigned long new_size)
1872 {
1873 return _uatomic_xchg_monotonic_increase(&ht->resize_target, new_size);
1874 }
1875
1876 static
1877 void resize_target_update_count(struct cds_lfht *ht,
1878 unsigned long count)
1879 {
1880 count = max(count, MIN_TABLE_SIZE);
1881 count = min(count, ht->max_nr_buckets);
1882 uatomic_set(&ht->resize_target, count);
1883 }
1884
1885 void cds_lfht_resize(struct cds_lfht *ht, unsigned long new_size)
1886 {
1887 int was_online;
1888
1889 resize_target_update_count(ht, new_size);
1890 CMM_STORE_SHARED(ht->resize_initiated, 1);
1891 was_online = ht->flavor->read_ongoing();
1892 if (was_online)
1893 ht->flavor->thread_offline();
1894 pthread_mutex_lock(&ht->resize_mutex);
1895 _do_cds_lfht_resize(ht);
1896 pthread_mutex_unlock(&ht->resize_mutex);
1897 if (was_online)
1898 ht->flavor->thread_online();
1899 }
1900
1901 static
1902 void do_resize_cb(struct rcu_head *head)
1903 {
1904 struct rcu_resize_work *work =
1905 caa_container_of(head, struct rcu_resize_work, head);
1906 struct cds_lfht *ht = work->ht;
1907
1908 ht->flavor->thread_offline();
1909 pthread_mutex_lock(&ht->resize_mutex);
1910 _do_cds_lfht_resize(ht);
1911 pthread_mutex_unlock(&ht->resize_mutex);
1912 ht->flavor->thread_online();
1913 poison_free(work);
1914 cmm_smp_mb(); /* finish resize before decrement */
1915 uatomic_dec(&ht->in_progress_resize);
1916 }
1917
1918 static
1919 void __cds_lfht_resize_lazy_launch(struct cds_lfht *ht)
1920 {
1921 struct rcu_resize_work *work;
1922
1923 /* Store resize_target before read resize_initiated */
1924 cmm_smp_mb();
1925 if (!CMM_LOAD_SHARED(ht->resize_initiated)) {
1926 uatomic_inc(&ht->in_progress_resize);
1927 cmm_smp_mb(); /* increment resize count before load destroy */
1928 if (CMM_LOAD_SHARED(ht->in_progress_destroy)) {
1929 uatomic_dec(&ht->in_progress_resize);
1930 return;
1931 }
1932 work = malloc(sizeof(*work));
1933 if (work == NULL) {
1934 dbg_printf("error allocating resize work, bailing out\n");
1935 uatomic_dec(&ht->in_progress_resize);
1936 return;
1937 }
1938 work->ht = ht;
1939 ht->flavor->update_call_rcu(&work->head, do_resize_cb);
1940 CMM_STORE_SHARED(ht->resize_initiated, 1);
1941 }
1942 }
1943
1944 static
1945 void cds_lfht_resize_lazy_grow(struct cds_lfht *ht, unsigned long size, int growth)
1946 {
1947 unsigned long target_size = size << growth;
1948
1949 target_size = min(target_size, ht->max_nr_buckets);
1950 if (resize_target_grow(ht, target_size) >= target_size)
1951 return;
1952
1953 __cds_lfht_resize_lazy_launch(ht);
1954 }
1955
1956 /*
1957 * We favor grow operations over shrink. A shrink operation never occurs
1958 * if a grow operation is queued for lazy execution. A grow operation
1959 * cancels any pending shrink lazy execution.
1960 */
1961 static
1962 void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
1963 unsigned long count)
1964 {
1965 if (!(ht->flags & CDS_LFHT_AUTO_RESIZE))
1966 return;
1967 count = max(count, MIN_TABLE_SIZE);
1968 count = min(count, ht->max_nr_buckets);
1969 if (count == size)
1970 return; /* Already the right size, no resize needed */
1971 if (count > size) { /* lazy grow */
1972 if (resize_target_grow(ht, count) >= count)
1973 return;
1974 } else { /* lazy shrink */
1975 for (;;) {
1976 unsigned long s;
1977
1978 s = uatomic_cmpxchg(&ht->resize_target, size, count);
1979 if (s == size)
1980 break; /* no resize needed */
1981 if (s > size)
1982 return; /* growing is/(was just) in progress */
1983 if (s <= count)
1984 return; /* some other thread do shrink */
1985 size = s;
1986 }
1987 }
1988 __cds_lfht_resize_lazy_launch(ht);
1989 }
This page took 0.070448 seconds and 4 git commands to generate.