Commit | Line | Data |
---|---|---|
fa68aa62 MD |
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> | |
bec39940 | 7 | * Copyright 2011 - Lai Jiangshan <laijs@cn.fujitsu.com> |
fa68aa62 MD |
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 | |
cd5376da DG |
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. | |
fa68aa62 MD |
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 | |
cd5376da DG |
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. | |
fa68aa62 MD |
50 | * - Hash table nodes are contained within a split-ordered list. This |
51 | * list is ordered by incrementing reversed-bits-hash value. | |
bec39940 | 52 | * - An index of bucket nodes is kept. These bucket nodes are the hash |
cd5376da DG |
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. | |
fa68aa62 MD |
62 | * - The resize operation for larger tables (and available through an |
63 | * API) allows both expanding and shrinking the hash table. | |
d6b18934 | 64 | * - Split-counters are used to keep track of the number of |
fa68aa62 MD |
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 | |
cd5376da DG |
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 | |
fa68aa62 MD |
95 | * helping the remover unlink nodes from the list rather than to wait |
96 | * for it do to so. | |
cd5376da DG |
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, | |
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. | |
d6b18934 | 206 | * |
bec39940 | 207 | * Bucket node tables: |
d6b18934 | 208 | * |
bec39940 DG |
209 | * hash table hash table the last all bucket node tables |
210 | * order size bucket node 0 1 2 3 4 5 6(index) | |
d6b18934 DG |
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 | * | |
bec39940 | 220 | * When growing/shrinking, we only focus on the last bucket node table |
d6b18934 DG |
221 | * which size is (!order ? 1 : (1 << (order -1))). |
222 | * | |
223 | * Example for growing/shrinking: | |
bec39940 DG |
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 | |
d6b18934 | 226 | * |
fa68aa62 | 227 | * A bit of ascii art explanation: |
bec39940 | 228 | * |
cd5376da DG |
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. | |
bec39940 | 231 | * |
fa68aa62 | 232 | * This shows the nodes for a small table ordered by reversed bits: |
bec39940 | 233 | * |
fa68aa62 MD |
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 | |
bec39940 DG |
243 | * |
244 | * This shows the nodes in order of non-reversed bits, linked by | |
fa68aa62 | 245 | * reversed-bit order. |
bec39940 | 246 | * |
fa68aa62 MD |
247 | * order bits reverse |
248 | * 0 0 000 000 | |
d6b18934 DG |
249 | * 1 | 1 001 100 <- |
250 | * 2 | | 2 010 010 <- | | |
fa68aa62 | 251 | * | | | 3 011 110 | <- | |
fa68aa62 MD |
252 | * 3 -> | | | 4 100 001 | | |
253 | * -> | | 5 101 101 | | |
254 | * -> | 6 110 011 | |
255 | * -> 7 111 111 | |
256 | */ | |
257 | ||
258 | #define _LGPL_SOURCE | |
cd5376da | 259 | #define _GNU_SOURCE |
fa68aa62 MD |
260 | #include <stdlib.h> |
261 | #include <errno.h> | |
262 | #include <assert.h> | |
263 | #include <stdio.h> | |
264 | #include <stdint.h> | |
265 | #include <string.h> | |
cd5376da | 266 | #include <sched.h> |
e4e50330 | 267 | #include <unistd.h> |
fa68aa62 | 268 | |
cd5376da | 269 | #include "config.h" |
e4e50330 | 270 | #include <urcu-pointer.h> |
fa68aa62 MD |
271 | #include <urcu-call-rcu.h> |
272 | #include <urcu/arch.h> | |
273 | #include <urcu/uatomic.h> | |
274 | #include <urcu/compiler.h> | |
fa68aa62 MD |
275 | #include <stdio.h> |
276 | #include <pthread.h> | |
277 | ||
f6a9efaa | 278 | #include "rculfhash.h" |
bec39940 | 279 | #include "rculfhash-internal.h" |
afccc240 | 280 | #include <urcu-flavor.h> |
fa68aa62 | 281 | |
a0daceaa MD |
282 | #include <common/common.h> |
283 | ||
9ef70f87 | 284 | /* |
cd5376da DG |
285 | * We need to lock pthread exit, which deadlocks __nptl_setxid in the runas |
286 | * clone. This work-around will be allowed to be removed when runas.c gets | |
287 | * changed to do an exec() before issuing seteuid/setegid. See | |
288 | * http://sourceware.org/bugzilla/show_bug.cgi?id=10184 for details. | |
9ef70f87 MD |
289 | */ |
290 | pthread_mutex_t lttng_libc_state_lock = PTHREAD_MUTEX_INITIALIZER; | |
291 | ||
fa68aa62 | 292 | /* |
d6b18934 | 293 | * Split-counters lazily update the global counter each 1024 |
fa68aa62 MD |
294 | * addition/removal. It automatically keeps track of resize required. |
295 | * We use the bucket length as indicator for need to expand for small | |
296 | * tables and machines lacking per-cpu data suppport. | |
297 | */ | |
298 | #define COUNT_COMMIT_ORDER 10 | |
d6b18934 | 299 | #define DEFAULT_SPLIT_COUNT_MASK 0xFUL |
fa68aa62 MD |
300 | #define CHAIN_LEN_TARGET 1 |
301 | #define CHAIN_LEN_RESIZE_THRESHOLD 3 | |
302 | ||
303 | /* | |
304 | * Define the minimum table size. | |
305 | */ | |
bec39940 DG |
306 | #define MIN_TABLE_ORDER 0 |
307 | #define MIN_TABLE_SIZE (1UL << MIN_TABLE_ORDER) | |
fa68aa62 MD |
308 | |
309 | /* | |
bec39940 | 310 | * Minimum number of bucket nodes to touch per thread to parallelize grow/shrink. |
fa68aa62 MD |
311 | */ |
312 | #define MIN_PARTITION_PER_THREAD_ORDER 12 | |
313 | #define MIN_PARTITION_PER_THREAD (1UL << MIN_PARTITION_PER_THREAD_ORDER) | |
314 | ||
fa68aa62 MD |
315 | /* |
316 | * The removed flag needs to be updated atomically with the pointer. | |
317 | * It indicates that no node must attach to the node scheduled for | |
318 | * removal, and that node garbage collection must be performed. | |
bec39940 | 319 | * The bucket flag does not require to be updated atomically with the |
fa68aa62 | 320 | * pointer, but it is added as a pointer low bit flag to save space. |
cd5376da DG |
321 | * The "removal owner" flag is used to detect which of the "del" |
322 | * operation that has set the "removed flag" gets to return the removed | |
323 | * node to its caller. Note that the replace operation does not need to | |
324 | * iteract with the "removal owner" flag, because it validates that | |
325 | * the "removed" flag is not set before performing its cmpxchg. | |
fa68aa62 MD |
326 | */ |
327 | #define REMOVED_FLAG (1UL << 0) | |
bec39940 DG |
328 | #define BUCKET_FLAG (1UL << 1) |
329 | #define REMOVAL_OWNER_FLAG (1UL << 2) | |
330 | #define FLAGS_MASK ((1UL << 3) - 1) | |
fa68aa62 MD |
331 | |
332 | /* Value of the end pointer. Should not interact with flags. */ | |
333 | #define END_VALUE NULL | |
334 | ||
d6b18934 DG |
335 | /* |
336 | * ht_items_count: Split-counters counting the number of node addition | |
337 | * and removal in the table. Only used if the CDS_LFHT_ACCOUNTING flag | |
338 | * is set at hash table creation. | |
339 | * | |
340 | * These are free-running counters, never reset to zero. They count the | |
341 | * number of add/remove, and trigger every (1 << COUNT_COMMIT_ORDER) | |
342 | * operations to update the global counter. We choose a power-of-2 value | |
343 | * for the trigger to deal with 32 or 64-bit overflow of the counter. | |
344 | */ | |
fa68aa62 MD |
345 | struct ht_items_count { |
346 | unsigned long add, del; | |
347 | } __attribute__((aligned(CAA_CACHE_LINE_SIZE))); | |
348 | ||
d6b18934 DG |
349 | /* |
350 | * rcu_resize_work: Contains arguments passed to RCU worker thread | |
351 | * responsible for performing lazy resize. | |
352 | */ | |
fa68aa62 MD |
353 | struct rcu_resize_work { |
354 | struct rcu_head head; | |
355 | struct cds_lfht *ht; | |
356 | }; | |
357 | ||
d6b18934 DG |
358 | /* |
359 | * partition_resize_work: Contains arguments passed to worker threads | |
360 | * executing the hash table resize on partitions of the hash table | |
361 | * assigned to each processor's worker thread. | |
362 | */ | |
fa68aa62 | 363 | struct partition_resize_work { |
d6b18934 | 364 | pthread_t thread_id; |
fa68aa62 MD |
365 | struct cds_lfht *ht; |
366 | unsigned long i, start, len; | |
367 | void (*fct)(struct cds_lfht *ht, unsigned long i, | |
368 | unsigned long start, unsigned long len); | |
369 | }; | |
370 | ||
fa68aa62 MD |
371 | /* |
372 | * Algorithm to reverse bits in a word by lookup table, extended to | |
373 | * 64-bit words. | |
374 | * Source: | |
375 | * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable | |
376 | * Originally from Public Domain. | |
377 | */ | |
378 | ||
bec39940 | 379 | static const uint8_t BitReverseTable256[256] = |
fa68aa62 MD |
380 | { |
381 | #define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64 | |
382 | #define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16) | |
383 | #define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 ) | |
384 | R6(0), R6(2), R6(1), R6(3) | |
385 | }; | |
386 | #undef R2 | |
387 | #undef R4 | |
388 | #undef R6 | |
389 | ||
390 | static | |
391 | uint8_t bit_reverse_u8(uint8_t v) | |
392 | { | |
393 | return BitReverseTable256[v]; | |
394 | } | |
395 | ||
396 | static __attribute__((unused)) | |
397 | uint32_t bit_reverse_u32(uint32_t v) | |
398 | { | |
bec39940 DG |
399 | return ((uint32_t) bit_reverse_u8(v) << 24) | |
400 | ((uint32_t) bit_reverse_u8(v >> 8) << 16) | | |
401 | ((uint32_t) bit_reverse_u8(v >> 16) << 8) | | |
fa68aa62 MD |
402 | ((uint32_t) bit_reverse_u8(v >> 24)); |
403 | } | |
404 | ||
405 | static __attribute__((unused)) | |
406 | uint64_t bit_reverse_u64(uint64_t v) | |
407 | { | |
bec39940 DG |
408 | return ((uint64_t) bit_reverse_u8(v) << 56) | |
409 | ((uint64_t) bit_reverse_u8(v >> 8) << 48) | | |
fa68aa62 MD |
410 | ((uint64_t) bit_reverse_u8(v >> 16) << 40) | |
411 | ((uint64_t) bit_reverse_u8(v >> 24) << 32) | | |
bec39940 DG |
412 | ((uint64_t) bit_reverse_u8(v >> 32) << 24) | |
413 | ((uint64_t) bit_reverse_u8(v >> 40) << 16) | | |
fa68aa62 MD |
414 | ((uint64_t) bit_reverse_u8(v >> 48) << 8) | |
415 | ((uint64_t) bit_reverse_u8(v >> 56)); | |
416 | } | |
417 | ||
418 | static | |
419 | unsigned long bit_reverse_ulong(unsigned long v) | |
420 | { | |
421 | #if (CAA_BITS_PER_LONG == 32) | |
422 | return bit_reverse_u32(v); | |
423 | #else | |
424 | return bit_reverse_u64(v); | |
425 | #endif | |
426 | } | |
427 | ||
428 | /* | |
429 | * fls: returns the position of the most significant bit. | |
430 | * Returns 0 if no bit is set, else returns the position of the most | |
431 | * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit). | |
432 | */ | |
433 | #if defined(__i386) || defined(__x86_64) | |
434 | static inline | |
435 | unsigned int fls_u32(uint32_t x) | |
436 | { | |
437 | int r; | |
438 | ||
439 | asm("bsrl %1,%0\n\t" | |
440 | "jnz 1f\n\t" | |
441 | "movl $-1,%0\n\t" | |
442 | "1:\n\t" | |
443 | : "=r" (r) : "rm" (x)); | |
444 | return r + 1; | |
445 | } | |
446 | #define HAS_FLS_U32 | |
447 | #endif | |
448 | ||
449 | #if defined(__x86_64) | |
450 | static inline | |
451 | unsigned int fls_u64(uint64_t x) | |
452 | { | |
453 | long r; | |
454 | ||
455 | asm("bsrq %1,%0\n\t" | |
456 | "jnz 1f\n\t" | |
457 | "movq $-1,%0\n\t" | |
458 | "1:\n\t" | |
459 | : "=r" (r) : "rm" (x)); | |
460 | return r + 1; | |
461 | } | |
462 | #define HAS_FLS_U64 | |
463 | #endif | |
464 | ||
465 | #ifndef HAS_FLS_U64 | |
466 | static __attribute__((unused)) | |
467 | unsigned int fls_u64(uint64_t x) | |
468 | { | |
469 | unsigned int r = 64; | |
470 | ||
471 | if (!x) | |
472 | return 0; | |
473 | ||
474 | if (!(x & 0xFFFFFFFF00000000ULL)) { | |
475 | x <<= 32; | |
476 | r -= 32; | |
477 | } | |
478 | if (!(x & 0xFFFF000000000000ULL)) { | |
479 | x <<= 16; | |
480 | r -= 16; | |
481 | } | |
482 | if (!(x & 0xFF00000000000000ULL)) { | |
483 | x <<= 8; | |
484 | r -= 8; | |
485 | } | |
486 | if (!(x & 0xF000000000000000ULL)) { | |
487 | x <<= 4; | |
488 | r -= 4; | |
489 | } | |
490 | if (!(x & 0xC000000000000000ULL)) { | |
491 | x <<= 2; | |
492 | r -= 2; | |
493 | } | |
494 | if (!(x & 0x8000000000000000ULL)) { | |
495 | x <<= 1; | |
496 | r -= 1; | |
497 | } | |
498 | return r; | |
499 | } | |
500 | #endif | |
501 | ||
502 | #ifndef HAS_FLS_U32 | |
503 | static __attribute__((unused)) | |
504 | unsigned int fls_u32(uint32_t x) | |
505 | { | |
506 | unsigned int r = 32; | |
507 | ||
508 | if (!x) | |
509 | return 0; | |
510 | if (!(x & 0xFFFF0000U)) { | |
511 | x <<= 16; | |
512 | r -= 16; | |
513 | } | |
514 | if (!(x & 0xFF000000U)) { | |
515 | x <<= 8; | |
516 | r -= 8; | |
517 | } | |
518 | if (!(x & 0xF0000000U)) { | |
519 | x <<= 4; | |
520 | r -= 4; | |
521 | } | |
522 | if (!(x & 0xC0000000U)) { | |
523 | x <<= 2; | |
524 | r -= 2; | |
525 | } | |
526 | if (!(x & 0x80000000U)) { | |
527 | x <<= 1; | |
528 | r -= 1; | |
529 | } | |
530 | return r; | |
531 | } | |
532 | #endif | |
533 | ||
bec39940 | 534 | unsigned int cds_lfht_fls_ulong(unsigned long x) |
fa68aa62 | 535 | { |
d6b18934 | 536 | #if (CAA_BITS_PER_LONG == 32) |
fa68aa62 MD |
537 | return fls_u32(x); |
538 | #else | |
539 | return fls_u64(x); | |
540 | #endif | |
541 | } | |
542 | ||
d6b18934 DG |
543 | /* |
544 | * Return the minimum order for which x <= (1UL << order). | |
545 | * Return -1 if x is 0. | |
546 | */ | |
bec39940 | 547 | int cds_lfht_get_count_order_u32(uint32_t x) |
fa68aa62 | 548 | { |
d6b18934 DG |
549 | if (!x) |
550 | return -1; | |
fa68aa62 | 551 | |
d6b18934 | 552 | return fls_u32(x - 1); |
fa68aa62 MD |
553 | } |
554 | ||
d6b18934 DG |
555 | /* |
556 | * Return the minimum order for which x <= (1UL << order). | |
557 | * Return -1 if x is 0. | |
558 | */ | |
bec39940 | 559 | int cds_lfht_get_count_order_ulong(unsigned long x) |
fa68aa62 | 560 | { |
d6b18934 DG |
561 | if (!x) |
562 | return -1; | |
fa68aa62 | 563 | |
bec39940 | 564 | return cds_lfht_fls_ulong(x - 1); |
fa68aa62 MD |
565 | } |
566 | ||
fa68aa62 | 567 | static |
bec39940 | 568 | void cds_lfht_resize_lazy_grow(struct cds_lfht *ht, unsigned long size, int growth); |
fa68aa62 | 569 | |
fa68aa62 MD |
570 | static |
571 | void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size, | |
572 | unsigned long count); | |
573 | ||
574 | static long nr_cpus_mask = -1; | |
d6b18934 | 575 | static long split_count_mask = -1; |
ec18533b | 576 | static int split_count_order = -1; |
d6b18934 DG |
577 | |
578 | #if defined(HAVE_SYSCONF) | |
579 | static void ht_init_nr_cpus_mask(void) | |
580 | { | |
581 | long maxcpus; | |
582 | ||
583 | maxcpus = sysconf(_SC_NPROCESSORS_CONF); | |
584 | if (maxcpus <= 0) { | |
585 | nr_cpus_mask = -2; | |
586 | return; | |
587 | } | |
588 | /* | |
589 | * round up number of CPUs to next power of two, so we | |
590 | * can use & for modulo. | |
591 | */ | |
bec39940 | 592 | maxcpus = 1UL << cds_lfht_get_count_order_ulong(maxcpus); |
d6b18934 DG |
593 | nr_cpus_mask = maxcpus - 1; |
594 | } | |
595 | #else /* #if defined(HAVE_SYSCONF) */ | |
596 | static void ht_init_nr_cpus_mask(void) | |
597 | { | |
598 | nr_cpus_mask = -2; | |
599 | } | |
600 | #endif /* #else #if defined(HAVE_SYSCONF) */ | |
fa68aa62 MD |
601 | |
602 | static | |
d6b18934 | 603 | void alloc_split_items_count(struct cds_lfht *ht) |
fa68aa62 MD |
604 | { |
605 | struct ht_items_count *count; | |
606 | ||
d6b18934 DG |
607 | if (nr_cpus_mask == -1) { |
608 | ht_init_nr_cpus_mask(); | |
609 | if (nr_cpus_mask < 0) | |
610 | split_count_mask = DEFAULT_SPLIT_COUNT_MASK; | |
611 | else | |
612 | split_count_mask = nr_cpus_mask; | |
ec18533b MD |
613 | split_count_order = |
614 | cds_lfht_get_count_order_ulong(split_count_mask + 1); | |
fa68aa62 | 615 | } |
d6b18934 DG |
616 | |
617 | assert(split_count_mask >= 0); | |
618 | ||
619 | if (ht->flags & CDS_LFHT_ACCOUNTING) { | |
620 | ht->split_count = calloc(split_count_mask + 1, sizeof(*count)); | |
621 | assert(ht->split_count); | |
622 | } else { | |
623 | ht->split_count = NULL; | |
fa68aa62 MD |
624 | } |
625 | } | |
626 | ||
627 | static | |
d6b18934 | 628 | void free_split_items_count(struct cds_lfht *ht) |
fa68aa62 | 629 | { |
d6b18934 | 630 | poison_free(ht->split_count); |
fa68aa62 MD |
631 | } |
632 | ||
8f0044bf | 633 | #if defined(HAVE_SCHED_GETCPU) && !defined(VALGRIND) |
fa68aa62 | 634 | static |
d6b18934 | 635 | int ht_get_split_count_index(unsigned long hash) |
fa68aa62 MD |
636 | { |
637 | int cpu; | |
638 | ||
d6b18934 | 639 | assert(split_count_mask >= 0); |
fa68aa62 | 640 | cpu = sched_getcpu(); |
6e59ae26 | 641 | if (caa_unlikely(cpu < 0)) |
d6b18934 | 642 | return hash & split_count_mask; |
fa68aa62 | 643 | else |
d6b18934 | 644 | return cpu & split_count_mask; |
fa68aa62 | 645 | } |
d6b18934 DG |
646 | #else /* #if defined(HAVE_SCHED_GETCPU) */ |
647 | static | |
648 | int ht_get_split_count_index(unsigned long hash) | |
649 | { | |
650 | return hash & split_count_mask; | |
651 | } | |
652 | #endif /* #else #if defined(HAVE_SCHED_GETCPU) */ | |
fa68aa62 MD |
653 | |
654 | static | |
d6b18934 | 655 | void ht_count_add(struct cds_lfht *ht, unsigned long size, unsigned long hash) |
fa68aa62 | 656 | { |
d6b18934 DG |
657 | unsigned long split_count; |
658 | int index; | |
bec39940 | 659 | long count; |
fa68aa62 | 660 | |
6e59ae26 | 661 | if (caa_unlikely(!ht->split_count)) |
fa68aa62 | 662 | return; |
d6b18934 DG |
663 | index = ht_get_split_count_index(hash); |
664 | split_count = uatomic_add_return(&ht->split_count[index].add, 1); | |
bec39940 DG |
665 | if (caa_likely(split_count & ((1UL << COUNT_COMMIT_ORDER) - 1))) |
666 | return; | |
667 | /* Only if number of add multiple of 1UL << COUNT_COMMIT_ORDER */ | |
668 | ||
669 | dbg_printf("add split count %lu\n", split_count); | |
670 | count = uatomic_add_return(&ht->count, | |
671 | 1UL << COUNT_COMMIT_ORDER); | |
672 | if (caa_likely(count & (count - 1))) | |
673 | return; | |
674 | /* Only if global count is power of 2 */ | |
675 | ||
676 | if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) < size) | |
677 | return; | |
678 | dbg_printf("add set global %ld\n", count); | |
679 | cds_lfht_resize_lazy_count(ht, size, | |
680 | count >> (CHAIN_LEN_TARGET - 1)); | |
fa68aa62 MD |
681 | } |
682 | ||
683 | static | |
d6b18934 | 684 | void ht_count_del(struct cds_lfht *ht, unsigned long size, unsigned long hash) |
fa68aa62 | 685 | { |
d6b18934 DG |
686 | unsigned long split_count; |
687 | int index; | |
bec39940 | 688 | long count; |
fa68aa62 | 689 | |
6e59ae26 | 690 | if (caa_unlikely(!ht->split_count)) |
fa68aa62 | 691 | return; |
d6b18934 DG |
692 | index = ht_get_split_count_index(hash); |
693 | split_count = uatomic_add_return(&ht->split_count[index].del, 1); | |
bec39940 DG |
694 | if (caa_likely(split_count & ((1UL << COUNT_COMMIT_ORDER) - 1))) |
695 | return; | |
696 | /* Only if number of deletes multiple of 1UL << COUNT_COMMIT_ORDER */ | |
697 | ||
698 | dbg_printf("del split count %lu\n", split_count); | |
699 | count = uatomic_add_return(&ht->count, | |
700 | -(1UL << COUNT_COMMIT_ORDER)); | |
701 | if (caa_likely(count & (count - 1))) | |
702 | return; | |
703 | /* Only if global count is power of 2 */ | |
704 | ||
705 | if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) >= size) | |
706 | return; | |
707 | dbg_printf("del set global %ld\n", count); | |
708 | /* | |
709 | * Don't shrink table if the number of nodes is below a | |
710 | * certain threshold. | |
711 | */ | |
712 | if (count < (1UL << COUNT_COMMIT_ORDER) * (split_count_mask + 1)) | |
713 | return; | |
714 | cds_lfht_resize_lazy_count(ht, size, | |
715 | count >> (CHAIN_LEN_TARGET - 1)); | |
fa68aa62 MD |
716 | } |
717 | ||
fa68aa62 MD |
718 | static |
719 | void check_resize(struct cds_lfht *ht, unsigned long size, uint32_t chain_len) | |
720 | { | |
721 | unsigned long count; | |
722 | ||
723 | if (!(ht->flags & CDS_LFHT_AUTO_RESIZE)) | |
724 | return; | |
725 | count = uatomic_read(&ht->count); | |
726 | /* | |
727 | * Use bucket-local length for small table expand and for | |
728 | * environments lacking per-cpu data support. | |
729 | */ | |
ec18533b | 730 | if (count >= (1UL << (COUNT_COMMIT_ORDER + split_count_order))) |
fa68aa62 MD |
731 | return; |
732 | if (chain_len > 100) | |
733 | dbg_printf("WARNING: large chain length: %u.\n", | |
734 | chain_len); | |
9d8ad8e2 MD |
735 | if (chain_len >= CHAIN_LEN_RESIZE_THRESHOLD) { |
736 | int growth; | |
737 | ||
738 | /* | |
739 | * Ideal growth calculated based on chain length. | |
740 | */ | |
741 | growth = cds_lfht_get_count_order_u32(chain_len | |
742 | - (CHAIN_LEN_TARGET - 1)); | |
743 | if ((ht->flags & CDS_LFHT_ACCOUNTING) | |
ec18533b MD |
744 | && (size << growth) |
745 | >= (1UL << (COUNT_COMMIT_ORDER | |
746 | + split_count_order))) { | |
9d8ad8e2 MD |
747 | /* |
748 | * If ideal growth expands the hash table size | |
749 | * beyond the "small hash table" sizes, use the | |
750 | * maximum small hash table size to attempt | |
751 | * expanding the hash table. This only applies | |
752 | * when node accounting is available, otherwise | |
753 | * the chain length is used to expand the hash | |
754 | * table in every case. | |
755 | */ | |
ec18533b MD |
756 | growth = COUNT_COMMIT_ORDER + split_count_order |
757 | - cds_lfht_get_count_order_ulong(size); | |
9d8ad8e2 MD |
758 | if (growth <= 0) |
759 | return; | |
760 | } | |
761 | cds_lfht_resize_lazy_grow(ht, size, growth); | |
762 | } | |
fa68aa62 MD |
763 | } |
764 | ||
765 | static | |
766 | struct cds_lfht_node *clear_flag(struct cds_lfht_node *node) | |
767 | { | |
768 | return (struct cds_lfht_node *) (((unsigned long) node) & ~FLAGS_MASK); | |
769 | } | |
770 | ||
771 | static | |
772 | int is_removed(struct cds_lfht_node *node) | |
773 | { | |
774 | return ((unsigned long) node) & REMOVED_FLAG; | |
775 | } | |
776 | ||
fa68aa62 | 777 | static |
bec39940 | 778 | int is_bucket(struct cds_lfht_node *node) |
fa68aa62 | 779 | { |
bec39940 | 780 | return ((unsigned long) node) & BUCKET_FLAG; |
fa68aa62 MD |
781 | } |
782 | ||
783 | static | |
bec39940 | 784 | struct cds_lfht_node *flag_bucket(struct cds_lfht_node *node) |
fa68aa62 | 785 | { |
bec39940 DG |
786 | return (struct cds_lfht_node *) (((unsigned long) node) | BUCKET_FLAG); |
787 | } | |
788 | ||
789 | static | |
790 | int is_removal_owner(struct cds_lfht_node *node) | |
791 | { | |
792 | return ((unsigned long) node) & REMOVAL_OWNER_FLAG; | |
793 | } | |
794 | ||
795 | static | |
796 | struct cds_lfht_node *flag_removal_owner(struct cds_lfht_node *node) | |
797 | { | |
798 | return (struct cds_lfht_node *) (((unsigned long) node) | REMOVAL_OWNER_FLAG); | |
fa68aa62 MD |
799 | } |
800 | ||
cd5376da DG |
801 | static |
802 | struct cds_lfht_node *flag_removed_or_removal_owner(struct cds_lfht_node *node) | |
803 | { | |
804 | return (struct cds_lfht_node *) (((unsigned long) node) | REMOVED_FLAG | REMOVAL_OWNER_FLAG); | |
805 | } | |
806 | ||
fa68aa62 MD |
807 | static |
808 | struct cds_lfht_node *get_end(void) | |
809 | { | |
810 | return (struct cds_lfht_node *) END_VALUE; | |
811 | } | |
812 | ||
813 | static | |
814 | int is_end(struct cds_lfht_node *node) | |
815 | { | |
816 | return clear_flag(node) == (struct cds_lfht_node *) END_VALUE; | |
817 | } | |
818 | ||
819 | static | |
bec39940 DG |
820 | unsigned long _uatomic_xchg_monotonic_increase(unsigned long *ptr, |
821 | unsigned long v) | |
fa68aa62 MD |
822 | { |
823 | unsigned long old1, old2; | |
824 | ||
825 | old1 = uatomic_read(ptr); | |
826 | do { | |
827 | old2 = old1; | |
828 | if (old2 >= v) | |
829 | return old2; | |
830 | } while ((old1 = uatomic_cmpxchg(ptr, old2, v)) != old2); | |
bec39940 | 831 | return old2; |
fa68aa62 MD |
832 | } |
833 | ||
834 | static | |
bec39940 | 835 | void cds_lfht_alloc_bucket_table(struct cds_lfht *ht, unsigned long order) |
fa68aa62 | 836 | { |
bec39940 DG |
837 | return ht->mm->alloc_bucket_table(ht, order); |
838 | } | |
d6b18934 | 839 | |
bec39940 DG |
840 | /* |
841 | * cds_lfht_free_bucket_table() should be called with decreasing order. | |
842 | * When cds_lfht_free_bucket_table(0) is called, it means the whole | |
843 | * lfht is destroyed. | |
844 | */ | |
845 | static | |
846 | void cds_lfht_free_bucket_table(struct cds_lfht *ht, unsigned long order) | |
847 | { | |
848 | return ht->mm->free_bucket_table(ht, order); | |
849 | } | |
d6b18934 | 850 | |
bec39940 DG |
851 | static inline |
852 | struct cds_lfht_node *bucket_at(struct cds_lfht *ht, unsigned long index) | |
853 | { | |
854 | return ht->bucket_at(ht, index); | |
855 | } | |
856 | ||
857 | static inline | |
858 | struct cds_lfht_node *lookup_bucket(struct cds_lfht *ht, unsigned long size, | |
859 | unsigned long hash) | |
860 | { | |
861 | assert(size > 0); | |
862 | return bucket_at(ht, hash & (size - 1)); | |
fa68aa62 MD |
863 | } |
864 | ||
865 | /* | |
866 | * Remove all logically deleted nodes from a bucket up to a certain node key. | |
867 | */ | |
868 | static | |
bec39940 | 869 | void _cds_lfht_gc_bucket(struct cds_lfht_node *bucket, struct cds_lfht_node *node) |
fa68aa62 MD |
870 | { |
871 | struct cds_lfht_node *iter_prev, *iter, *next, *new_next; | |
872 | ||
bec39940 DG |
873 | assert(!is_bucket(bucket)); |
874 | assert(!is_removed(bucket)); | |
875 | assert(!is_bucket(node)); | |
fa68aa62 MD |
876 | assert(!is_removed(node)); |
877 | for (;;) { | |
bec39940 DG |
878 | iter_prev = bucket; |
879 | /* We can always skip the bucket node initially */ | |
880 | iter = rcu_dereference(iter_prev->next); | |
d6b18934 | 881 | assert(!is_removed(iter)); |
bec39940 | 882 | assert(iter_prev->reverse_hash <= node->reverse_hash); |
fa68aa62 | 883 | /* |
bec39940 | 884 | * We should never be called with bucket (start of chain) |
fa68aa62 MD |
885 | * and logically removed node (end of path compression |
886 | * marker) being the actual same node. This would be a | |
887 | * bug in the algorithm implementation. | |
888 | */ | |
bec39940 | 889 | assert(bucket != node); |
fa68aa62 | 890 | for (;;) { |
6e59ae26 | 891 | if (caa_unlikely(is_end(iter))) |
fa68aa62 | 892 | return; |
bec39940 | 893 | if (caa_likely(clear_flag(iter)->reverse_hash > node->reverse_hash)) |
fa68aa62 | 894 | return; |
bec39940 | 895 | next = rcu_dereference(clear_flag(iter)->next); |
6e59ae26 | 896 | if (caa_likely(is_removed(next))) |
fa68aa62 MD |
897 | break; |
898 | iter_prev = clear_flag(iter); | |
899 | iter = next; | |
900 | } | |
901 | assert(!is_removed(iter)); | |
bec39940 DG |
902 | if (is_bucket(iter)) |
903 | new_next = flag_bucket(clear_flag(next)); | |
fa68aa62 MD |
904 | else |
905 | new_next = clear_flag(next); | |
bec39940 | 906 | (void) uatomic_cmpxchg(&iter_prev->next, iter, new_next); |
fa68aa62 | 907 | } |
fa68aa62 MD |
908 | } |
909 | ||
910 | static | |
911 | int _cds_lfht_replace(struct cds_lfht *ht, unsigned long size, | |
912 | struct cds_lfht_node *old_node, | |
d6b18934 | 913 | struct cds_lfht_node *old_next, |
fa68aa62 MD |
914 | struct cds_lfht_node *new_node) |
915 | { | |
bec39940 | 916 | struct cds_lfht_node *bucket, *ret_next; |
fa68aa62 MD |
917 | |
918 | if (!old_node) /* Return -ENOENT if asked to replace NULL node */ | |
d6b18934 | 919 | return -ENOENT; |
fa68aa62 MD |
920 | |
921 | assert(!is_removed(old_node)); | |
bec39940 | 922 | assert(!is_bucket(old_node)); |
fa68aa62 | 923 | assert(!is_removed(new_node)); |
bec39940 | 924 | assert(!is_bucket(new_node)); |
fa68aa62 | 925 | assert(new_node != old_node); |
d6b18934 | 926 | for (;;) { |
fa68aa62 | 927 | /* Insert after node to be replaced */ |
fa68aa62 MD |
928 | if (is_removed(old_next)) { |
929 | /* | |
930 | * Too late, the old node has been removed under us | |
931 | * between lookup and replace. Fail. | |
932 | */ | |
d6b18934 | 933 | return -ENOENT; |
fa68aa62 | 934 | } |
bec39940 DG |
935 | assert(old_next == clear_flag(old_next)); |
936 | assert(new_node != old_next); | |
cd5376da DG |
937 | /* |
938 | * REMOVAL_OWNER flag is _NEVER_ set before the REMOVED | |
939 | * flag. It is either set atomically at the same time | |
940 | * (replace) or after (del). | |
941 | */ | |
942 | assert(!is_removal_owner(old_next)); | |
bec39940 | 943 | new_node->next = old_next; |
fa68aa62 MD |
944 | /* |
945 | * Here is the whole trick for lock-free replace: we add | |
946 | * the replacement node _after_ the node we want to | |
947 | * replace by atomically setting its next pointer at the | |
948 | * same time we set its removal flag. Given that | |
949 | * the lookups/get next use an iterator aware of the | |
950 | * next pointer, they will either skip the old node due | |
951 | * to the removal flag and see the new node, or use | |
952 | * the old node, but will not see the new one. | |
bec39940 DG |
953 | * This is a replacement of a node with another node |
954 | * that has the same value: we are therefore not | |
cd5376da DG |
955 | * removing a value from the hash table. We set both the |
956 | * REMOVED and REMOVAL_OWNER flags atomically so we own | |
957 | * the node after successful cmpxchg. | |
fa68aa62 | 958 | */ |
bec39940 | 959 | ret_next = uatomic_cmpxchg(&old_node->next, |
cd5376da | 960 | old_next, flag_removed_or_removal_owner(new_node)); |
d6b18934 DG |
961 | if (ret_next == old_next) |
962 | break; /* We performed the replacement. */ | |
963 | old_next = ret_next; | |
964 | } | |
fa68aa62 MD |
965 | |
966 | /* | |
967 | * Ensure that the old node is not visible to readers anymore: | |
968 | * lookup for the node, and remove it (along with any other | |
969 | * logically removed node) if found. | |
970 | */ | |
bec39940 DG |
971 | bucket = lookup_bucket(ht, size, bit_reverse_ulong(old_node->reverse_hash)); |
972 | _cds_lfht_gc_bucket(bucket, new_node); | |
d6b18934 | 973 | |
cd5376da | 974 | assert(is_removed(CMM_LOAD_SHARED(old_node->next))); |
d6b18934 | 975 | return 0; |
fa68aa62 MD |
976 | } |
977 | ||
d6b18934 DG |
978 | /* |
979 | * A non-NULL unique_ret pointer uses the "add unique" (or uniquify) add | |
980 | * mode. A NULL unique_ret allows creation of duplicate keys. | |
981 | */ | |
fa68aa62 | 982 | static |
d6b18934 | 983 | void _cds_lfht_add(struct cds_lfht *ht, |
bec39940 DG |
984 | unsigned long hash, |
985 | cds_lfht_match_fct match, | |
986 | const void *key, | |
d6b18934 DG |
987 | unsigned long size, |
988 | struct cds_lfht_node *node, | |
989 | struct cds_lfht_iter *unique_ret, | |
bec39940 | 990 | int bucket_flag) |
fa68aa62 MD |
991 | { |
992 | struct cds_lfht_node *iter_prev, *iter, *next, *new_node, *new_next, | |
d6b18934 | 993 | *return_node; |
bec39940 | 994 | struct cds_lfht_node *bucket; |
fa68aa62 | 995 | |
bec39940 | 996 | assert(!is_bucket(node)); |
fa68aa62 | 997 | assert(!is_removed(node)); |
bec39940 | 998 | bucket = lookup_bucket(ht, size, hash); |
fa68aa62 MD |
999 | for (;;) { |
1000 | uint32_t chain_len = 0; | |
1001 | ||
1002 | /* | |
1003 | * iter_prev points to the non-removed node prior to the | |
1004 | * insert location. | |
1005 | */ | |
bec39940 DG |
1006 | iter_prev = bucket; |
1007 | /* We can always skip the bucket node initially */ | |
1008 | iter = rcu_dereference(iter_prev->next); | |
1009 | assert(iter_prev->reverse_hash <= node->reverse_hash); | |
fa68aa62 | 1010 | for (;;) { |
6e59ae26 | 1011 | if (caa_unlikely(is_end(iter))) |
fa68aa62 | 1012 | goto insert; |
bec39940 | 1013 | if (caa_likely(clear_flag(iter)->reverse_hash > node->reverse_hash)) |
fa68aa62 | 1014 | goto insert; |
d6b18934 | 1015 | |
bec39940 DG |
1016 | /* bucket node is the first node of the identical-hash-value chain */ |
1017 | if (bucket_flag && clear_flag(iter)->reverse_hash == node->reverse_hash) | |
d6b18934 DG |
1018 | goto insert; |
1019 | ||
bec39940 | 1020 | next = rcu_dereference(clear_flag(iter)->next); |
6e59ae26 | 1021 | if (caa_unlikely(is_removed(next))) |
fa68aa62 | 1022 | goto gc_node; |
d6b18934 DG |
1023 | |
1024 | /* uniquely add */ | |
1025 | if (unique_ret | |
bec39940 DG |
1026 | && !is_bucket(next) |
1027 | && clear_flag(iter)->reverse_hash == node->reverse_hash) { | |
d6b18934 DG |
1028 | struct cds_lfht_iter d_iter = { .node = node, .next = iter, }; |
1029 | ||
1030 | /* | |
1031 | * uniquely adding inserts the node as the first | |
1032 | * node of the identical-hash-value node chain. | |
1033 | * | |
1034 | * This semantic ensures no duplicated keys | |
1035 | * should ever be observable in the table | |
cd5376da DG |
1036 | * (including traversing the table node by |
1037 | * node by forward iterations) | |
d6b18934 | 1038 | */ |
bec39940 | 1039 | cds_lfht_next_duplicate(ht, match, key, &d_iter); |
d6b18934 DG |
1040 | if (!d_iter.node) |
1041 | goto insert; | |
1042 | ||
1043 | *unique_ret = d_iter; | |
1044 | return; | |
fa68aa62 | 1045 | } |
d6b18934 | 1046 | |
fa68aa62 | 1047 | /* Only account for identical reverse hash once */ |
bec39940 DG |
1048 | if (iter_prev->reverse_hash != clear_flag(iter)->reverse_hash |
1049 | && !is_bucket(next)) | |
fa68aa62 MD |
1050 | check_resize(ht, size, ++chain_len); |
1051 | iter_prev = clear_flag(iter); | |
1052 | iter = next; | |
1053 | } | |
1054 | ||
1055 | insert: | |
1056 | assert(node != clear_flag(iter)); | |
1057 | assert(!is_removed(iter_prev)); | |
1058 | assert(!is_removed(iter)); | |
1059 | assert(iter_prev != node); | |
bec39940 DG |
1060 | if (!bucket_flag) |
1061 | node->next = clear_flag(iter); | |
fa68aa62 | 1062 | else |
bec39940 DG |
1063 | node->next = flag_bucket(clear_flag(iter)); |
1064 | if (is_bucket(iter)) | |
1065 | new_node = flag_bucket(node); | |
fa68aa62 MD |
1066 | else |
1067 | new_node = node; | |
bec39940 | 1068 | if (uatomic_cmpxchg(&iter_prev->next, iter, |
fa68aa62 MD |
1069 | new_node) != iter) { |
1070 | continue; /* retry */ | |
1071 | } else { | |
d6b18934 DG |
1072 | return_node = node; |
1073 | goto end; | |
fa68aa62 MD |
1074 | } |
1075 | ||
1076 | gc_node: | |
1077 | assert(!is_removed(iter)); | |
bec39940 DG |
1078 | if (is_bucket(iter)) |
1079 | new_next = flag_bucket(clear_flag(next)); | |
fa68aa62 MD |
1080 | else |
1081 | new_next = clear_flag(next); | |
bec39940 | 1082 | (void) uatomic_cmpxchg(&iter_prev->next, iter, new_next); |
fa68aa62 MD |
1083 | /* retry */ |
1084 | } | |
fa68aa62 | 1085 | end: |
d6b18934 DG |
1086 | if (unique_ret) { |
1087 | unique_ret->node = return_node; | |
1088 | /* unique_ret->next left unset, never used. */ | |
1089 | } | |
fa68aa62 MD |
1090 | } |
1091 | ||
1092 | static | |
1093 | int _cds_lfht_del(struct cds_lfht *ht, unsigned long size, | |
bec39940 | 1094 | struct cds_lfht_node *node) |
fa68aa62 | 1095 | { |
bec39940 | 1096 | struct cds_lfht_node *bucket, *next; |
fa68aa62 MD |
1097 | |
1098 | if (!node) /* Return -ENOENT if asked to delete NULL node */ | |
d6b18934 | 1099 | return -ENOENT; |
fa68aa62 MD |
1100 | |
1101 | /* logically delete the node */ | |
bec39940 | 1102 | assert(!is_bucket(node)); |
fa68aa62 | 1103 | assert(!is_removed(node)); |
bec39940 | 1104 | assert(!is_removal_owner(node)); |
fa68aa62 | 1105 | |
bec39940 DG |
1106 | /* |
1107 | * We are first checking if the node had previously been | |
1108 | * logically removed (this check is not atomic with setting the | |
1109 | * logical removal flag). Return -ENOENT if the node had | |
1110 | * previously been removed. | |
1111 | */ | |
cd5376da | 1112 | next = CMM_LOAD_SHARED(node->next); /* next is not dereferenced */ |
bec39940 DG |
1113 | if (caa_unlikely(is_removed(next))) |
1114 | return -ENOENT; | |
1115 | assert(!is_bucket(next)); | |
cd5376da DG |
1116 | /* |
1117 | * The del operation semantic guarantees a full memory barrier | |
1118 | * before the uatomic_or atomic commit of the deletion flag. | |
1119 | */ | |
1120 | cmm_smp_mb__before_uatomic_or(); | |
bec39940 DG |
1121 | /* |
1122 | * We set the REMOVED_FLAG unconditionally. Note that there may | |
1123 | * be more than one concurrent thread setting this flag. | |
1124 | * Knowing which wins the race will be known after the garbage | |
1125 | * collection phase, stay tuned! | |
1126 | */ | |
1127 | uatomic_or(&node->next, REMOVED_FLAG); | |
fa68aa62 | 1128 | /* We performed the (logical) deletion. */ |
fa68aa62 MD |
1129 | |
1130 | /* | |
1131 | * Ensure that the node is not visible to readers anymore: lookup for | |
1132 | * the node, and remove it (along with any other logically removed node) | |
1133 | * if found. | |
1134 | */ | |
bec39940 DG |
1135 | bucket = lookup_bucket(ht, size, bit_reverse_ulong(node->reverse_hash)); |
1136 | _cds_lfht_gc_bucket(bucket, node); | |
d6b18934 | 1137 | |
cd5376da | 1138 | assert(is_removed(CMM_LOAD_SHARED(node->next))); |
bec39940 DG |
1139 | /* |
1140 | * Last phase: atomically exchange node->next with a version | |
1141 | * having "REMOVAL_OWNER_FLAG" set. If the returned node->next | |
1142 | * pointer did _not_ have "REMOVAL_OWNER_FLAG" set, we now own | |
1143 | * the node and win the removal race. | |
1144 | * It is interesting to note that all "add" paths are forbidden | |
1145 | * to change the next pointer starting from the point where the | |
1146 | * REMOVED_FLAG is set, so here using a read, followed by a | |
1147 | * xchg() suffice to guarantee that the xchg() will ever only | |
1148 | * set the "REMOVAL_OWNER_FLAG" (or change nothing if the flag | |
1149 | * was already set). | |
1150 | */ | |
1151 | if (!is_removal_owner(uatomic_xchg(&node->next, | |
1152 | flag_removal_owner(node->next)))) | |
1153 | return 0; | |
1154 | else | |
1155 | return -ENOENT; | |
fa68aa62 MD |
1156 | } |
1157 | ||
1158 | static | |
1159 | void *partition_resize_thread(void *arg) | |
1160 | { | |
1161 | struct partition_resize_work *work = arg; | |
1162 | ||
bec39940 | 1163 | work->ht->flavor->register_thread(); |
fa68aa62 | 1164 | work->fct(work->ht, work->i, work->start, work->len); |
bec39940 | 1165 | work->ht->flavor->unregister_thread(); |
fa68aa62 MD |
1166 | return NULL; |
1167 | } | |
1168 | ||
1169 | static | |
1170 | void partition_resize_helper(struct cds_lfht *ht, unsigned long i, | |
1171 | unsigned long len, | |
1172 | void (*fct)(struct cds_lfht *ht, unsigned long i, | |
1173 | unsigned long start, unsigned long len)) | |
1174 | { | |
1175 | unsigned long partition_len; | |
1176 | struct partition_resize_work *work; | |
1177 | int thread, ret; | |
1178 | unsigned long nr_threads; | |
fa68aa62 MD |
1179 | |
1180 | /* | |
1181 | * Note: nr_cpus_mask + 1 is always power of 2. | |
1182 | * We spawn just the number of threads we need to satisfy the minimum | |
1183 | * partition size, up to the number of CPUs in the system. | |
1184 | */ | |
f6a9efaa DG |
1185 | if (nr_cpus_mask > 0) { |
1186 | nr_threads = min(nr_cpus_mask + 1, | |
1187 | len >> MIN_PARTITION_PER_THREAD_ORDER); | |
1188 | } else { | |
1189 | nr_threads = 1; | |
1190 | } | |
bec39940 | 1191 | partition_len = len >> cds_lfht_get_count_order_ulong(nr_threads); |
fa68aa62 | 1192 | work = calloc(nr_threads, sizeof(*work)); |
fa68aa62 MD |
1193 | assert(work); |
1194 | for (thread = 0; thread < nr_threads; thread++) { | |
1195 | work[thread].ht = ht; | |
1196 | work[thread].i = i; | |
1197 | work[thread].len = partition_len; | |
1198 | work[thread].start = thread * partition_len; | |
1199 | work[thread].fct = fct; | |
d6b18934 | 1200 | ret = pthread_create(&(work[thread].thread_id), ht->resize_attr, |
fa68aa62 MD |
1201 | partition_resize_thread, &work[thread]); |
1202 | assert(!ret); | |
1203 | } | |
1204 | for (thread = 0; thread < nr_threads; thread++) { | |
d6b18934 | 1205 | ret = pthread_join(work[thread].thread_id, NULL); |
fa68aa62 MD |
1206 | assert(!ret); |
1207 | } | |
1208 | free(work); | |
fa68aa62 MD |
1209 | } |
1210 | ||
1211 | /* | |
1212 | * Holding RCU read lock to protect _cds_lfht_add against memory | |
1213 | * reclaim that could be performed by other call_rcu worker threads (ABA | |
1214 | * problem). | |
1215 | * | |
1216 | * When we reach a certain length, we can split this population phase over | |
1217 | * many worker threads, based on the number of CPUs available in the system. | |
1218 | * This should therefore take care of not having the expand lagging behind too | |
1219 | * many concurrent insertion threads by using the scheduler's ability to | |
bec39940 | 1220 | * schedule bucket node population fairly with insertions. |
fa68aa62 MD |
1221 | */ |
1222 | static | |
1223 | void init_table_populate_partition(struct cds_lfht *ht, unsigned long i, | |
1224 | unsigned long start, unsigned long len) | |
1225 | { | |
bec39940 | 1226 | unsigned long j, size = 1UL << (i - 1); |
fa68aa62 | 1227 | |
bec39940 DG |
1228 | assert(i > MIN_TABLE_ORDER); |
1229 | ht->flavor->read_lock(); | |
1230 | for (j = size + start; j < size + start + len; j++) { | |
1231 | struct cds_lfht_node *new_node = bucket_at(ht, j); | |
fa68aa62 | 1232 | |
bec39940 DG |
1233 | assert(j >= size && j < (size << 1)); |
1234 | dbg_printf("init populate: order %lu index %lu hash %lu\n", | |
1235 | i, j, j); | |
1236 | new_node->reverse_hash = bit_reverse_ulong(j); | |
1237 | _cds_lfht_add(ht, j, NULL, NULL, size, new_node, NULL, 1); | |
fa68aa62 | 1238 | } |
bec39940 | 1239 | ht->flavor->read_unlock(); |
fa68aa62 MD |
1240 | } |
1241 | ||
1242 | static | |
1243 | void init_table_populate(struct cds_lfht *ht, unsigned long i, | |
1244 | unsigned long len) | |
1245 | { | |
1246 | assert(nr_cpus_mask != -1); | |
1247 | if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD) { | |
bec39940 | 1248 | ht->flavor->thread_online(); |
fa68aa62 | 1249 | init_table_populate_partition(ht, i, 0, len); |
bec39940 | 1250 | ht->flavor->thread_offline(); |
fa68aa62 MD |
1251 | return; |
1252 | } | |
1253 | partition_resize_helper(ht, i, len, init_table_populate_partition); | |
1254 | } | |
1255 | ||
1256 | static | |
1257 | void init_table(struct cds_lfht *ht, | |
d6b18934 | 1258 | unsigned long first_order, unsigned long last_order) |
fa68aa62 | 1259 | { |
d6b18934 | 1260 | unsigned long i; |
fa68aa62 | 1261 | |
d6b18934 DG |
1262 | dbg_printf("init table: first_order %lu last_order %lu\n", |
1263 | first_order, last_order); | |
bec39940 | 1264 | assert(first_order > MIN_TABLE_ORDER); |
d6b18934 | 1265 | for (i = first_order; i <= last_order; i++) { |
fa68aa62 MD |
1266 | unsigned long len; |
1267 | ||
d6b18934 | 1268 | len = 1UL << (i - 1); |
fa68aa62 MD |
1269 | dbg_printf("init order %lu len: %lu\n", i, len); |
1270 | ||
1271 | /* Stop expand if the resize target changes under us */ | |
bec39940 | 1272 | if (CMM_LOAD_SHARED(ht->resize_target) < (1UL << i)) |
fa68aa62 MD |
1273 | break; |
1274 | ||
bec39940 | 1275 | cds_lfht_alloc_bucket_table(ht, i); |
fa68aa62 MD |
1276 | |
1277 | /* | |
bec39940 DG |
1278 | * Set all bucket nodes reverse hash values for a level and |
1279 | * link all bucket nodes into the table. | |
fa68aa62 MD |
1280 | */ |
1281 | init_table_populate(ht, i, len); | |
1282 | ||
1283 | /* | |
1284 | * Update table size. | |
1285 | */ | |
1286 | cmm_smp_wmb(); /* populate data before RCU size */ | |
bec39940 | 1287 | CMM_STORE_SHARED(ht->size, 1UL << i); |
fa68aa62 | 1288 | |
d6b18934 | 1289 | dbg_printf("init new size: %lu\n", 1UL << i); |
fa68aa62 MD |
1290 | if (CMM_LOAD_SHARED(ht->in_progress_destroy)) |
1291 | break; | |
1292 | } | |
1293 | } | |
1294 | ||
1295 | /* | |
1296 | * Holding RCU read lock to protect _cds_lfht_remove against memory | |
1297 | * reclaim that could be performed by other call_rcu worker threads (ABA | |
1298 | * problem). | |
1299 | * For a single level, we logically remove and garbage collect each node. | |
1300 | * | |
1301 | * As a design choice, we perform logical removal and garbage collection on a | |
1302 | * node-per-node basis to simplify this algorithm. We also assume keeping good | |
1303 | * cache locality of the operation would overweight possible performance gain | |
1304 | * that could be achieved by batching garbage collection for multiple levels. | |
1305 | * However, this would have to be justified by benchmarks. | |
1306 | * | |
1307 | * Concurrent removal and add operations are helping us perform garbage | |
1308 | * collection of logically removed nodes. We guarantee that all logically | |
1309 | * removed nodes have been garbage-collected (unlinked) before call_rcu is | |
bec39940 | 1310 | * invoked to free a hole level of bucket nodes (after a grace period). |
fa68aa62 | 1311 | * |
cd5376da DG |
1312 | * Logical removal and garbage collection can therefore be done in batch |
1313 | * or on a node-per-node basis, as long as the guarantee above holds. | |
fa68aa62 MD |
1314 | * |
1315 | * When we reach a certain length, we can split this removal over many worker | |
1316 | * threads, based on the number of CPUs available in the system. This should | |
1317 | * take care of not letting resize process lag behind too many concurrent | |
1318 | * updater threads actively inserting into the hash table. | |
1319 | */ | |
1320 | static | |
1321 | void remove_table_partition(struct cds_lfht *ht, unsigned long i, | |
1322 | unsigned long start, unsigned long len) | |
1323 | { | |
bec39940 | 1324 | unsigned long j, size = 1UL << (i - 1); |
fa68aa62 | 1325 | |
bec39940 DG |
1326 | assert(i > MIN_TABLE_ORDER); |
1327 | ht->flavor->read_lock(); | |
1328 | for (j = size + start; j < size + start + len; j++) { | |
1329 | struct cds_lfht_node *fini_bucket = bucket_at(ht, j); | |
1330 | struct cds_lfht_node *parent_bucket = bucket_at(ht, j - size); | |
fa68aa62 | 1331 | |
bec39940 DG |
1332 | assert(j >= size && j < (size << 1)); |
1333 | dbg_printf("remove entry: order %lu index %lu hash %lu\n", | |
1334 | i, j, j); | |
1335 | /* Set the REMOVED_FLAG to freeze the ->next for gc */ | |
1336 | uatomic_or(&fini_bucket->next, REMOVED_FLAG); | |
1337 | _cds_lfht_gc_bucket(parent_bucket, fini_bucket); | |
fa68aa62 | 1338 | } |
bec39940 | 1339 | ht->flavor->read_unlock(); |
fa68aa62 MD |
1340 | } |
1341 | ||
1342 | static | |
1343 | void remove_table(struct cds_lfht *ht, unsigned long i, unsigned long len) | |
1344 | { | |
1345 | ||
1346 | assert(nr_cpus_mask != -1); | |
1347 | if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD) { | |
bec39940 | 1348 | ht->flavor->thread_online(); |
fa68aa62 | 1349 | remove_table_partition(ht, i, 0, len); |
bec39940 | 1350 | ht->flavor->thread_offline(); |
fa68aa62 MD |
1351 | return; |
1352 | } | |
1353 | partition_resize_helper(ht, i, len, remove_table_partition); | |
1354 | } | |
1355 | ||
bec39940 DG |
1356 | /* |
1357 | * fini_table() is never called for first_order == 0, which is why | |
1358 | * free_by_rcu_order == 0 can be used as criterion to know if free must | |
1359 | * be called. | |
1360 | */ | |
fa68aa62 MD |
1361 | static |
1362 | void fini_table(struct cds_lfht *ht, | |
d6b18934 | 1363 | unsigned long first_order, unsigned long last_order) |
fa68aa62 | 1364 | { |
d6b18934 | 1365 | long i; |
bec39940 | 1366 | unsigned long free_by_rcu_order = 0; |
fa68aa62 | 1367 | |
d6b18934 DG |
1368 | dbg_printf("fini table: first_order %lu last_order %lu\n", |
1369 | first_order, last_order); | |
bec39940 | 1370 | assert(first_order > MIN_TABLE_ORDER); |
d6b18934 | 1371 | for (i = last_order; i >= first_order; i--) { |
fa68aa62 MD |
1372 | unsigned long len; |
1373 | ||
d6b18934 | 1374 | len = 1UL << (i - 1); |
fa68aa62 MD |
1375 | dbg_printf("fini order %lu len: %lu\n", i, len); |
1376 | ||
1377 | /* Stop shrink if the resize target changes under us */ | |
bec39940 | 1378 | if (CMM_LOAD_SHARED(ht->resize_target) > (1UL << (i - 1))) |
fa68aa62 MD |
1379 | break; |
1380 | ||
1381 | cmm_smp_wmb(); /* populate data before RCU size */ | |
bec39940 | 1382 | CMM_STORE_SHARED(ht->size, 1UL << (i - 1)); |
fa68aa62 MD |
1383 | |
1384 | /* | |
1385 | * We need to wait for all add operations to reach Q.S. (and | |
1386 | * thus use the new table for lookups) before we can start | |
bec39940 | 1387 | * releasing the old bucket nodes. Otherwise their lookup will |
fa68aa62 MD |
1388 | * return a logically removed node as insert position. |
1389 | */ | |
bec39940 DG |
1390 | ht->flavor->update_synchronize_rcu(); |
1391 | if (free_by_rcu_order) | |
1392 | cds_lfht_free_bucket_table(ht, free_by_rcu_order); | |
fa68aa62 MD |
1393 | |
1394 | /* | |
bec39940 DG |
1395 | * Set "removed" flag in bucket nodes about to be removed. |
1396 | * Unlink all now-logically-removed bucket node pointers. | |
fa68aa62 MD |
1397 | * Concurrent add/remove operation are helping us doing |
1398 | * the gc. | |
1399 | */ | |
1400 | remove_table(ht, i, len); | |
1401 | ||
bec39940 | 1402 | free_by_rcu_order = i; |
fa68aa62 MD |
1403 | |
1404 | dbg_printf("fini new size: %lu\n", 1UL << i); | |
1405 | if (CMM_LOAD_SHARED(ht->in_progress_destroy)) | |
1406 | break; | |
1407 | } | |
d6b18934 | 1408 | |
bec39940 DG |
1409 | if (free_by_rcu_order) { |
1410 | ht->flavor->update_synchronize_rcu(); | |
1411 | cds_lfht_free_bucket_table(ht, free_by_rcu_order); | |
d6b18934 DG |
1412 | } |
1413 | } | |
1414 | ||
1415 | static | |
bec39940 | 1416 | void cds_lfht_create_bucket(struct cds_lfht *ht, unsigned long size) |
d6b18934 | 1417 | { |
bec39940 DG |
1418 | struct cds_lfht_node *prev, *node; |
1419 | unsigned long order, len, i; | |
d6b18934 | 1420 | |
bec39940 | 1421 | cds_lfht_alloc_bucket_table(ht, 0); |
d6b18934 | 1422 | |
bec39940 DG |
1423 | dbg_printf("create bucket: order 0 index 0 hash 0\n"); |
1424 | node = bucket_at(ht, 0); | |
1425 | node->next = flag_bucket(get_end()); | |
1426 | node->reverse_hash = 0; | |
d6b18934 | 1427 | |
bec39940 | 1428 | for (order = 1; order < cds_lfht_get_count_order_ulong(size) + 1; order++) { |
d6b18934 | 1429 | len = 1UL << (order - 1); |
bec39940 | 1430 | cds_lfht_alloc_bucket_table(ht, order); |
d6b18934 | 1431 | |
bec39940 DG |
1432 | for (i = 0; i < len; i++) { |
1433 | /* | |
1434 | * Now, we are trying to init the node with the | |
1435 | * hash=(len+i) (which is also a bucket with the | |
1436 | * index=(len+i)) and insert it into the hash table, | |
1437 | * so this node has to be inserted after the bucket | |
1438 | * with the index=(len+i)&(len-1)=i. And because there | |
1439 | * is no other non-bucket node nor bucket node with | |
1440 | * larger index/hash inserted, so the bucket node | |
1441 | * being inserted should be inserted directly linked | |
1442 | * after the bucket node with index=i. | |
1443 | */ | |
1444 | prev = bucket_at(ht, i); | |
1445 | node = bucket_at(ht, len + i); | |
1446 | ||
1447 | dbg_printf("create bucket: order %lu index %lu hash %lu\n", | |
1448 | order, len + i, len + i); | |
1449 | node->reverse_hash = bit_reverse_ulong(len + i); | |
d6b18934 | 1450 | |
bec39940 DG |
1451 | /* insert after prev */ |
1452 | assert(is_bucket(prev->next)); | |
d6b18934 | 1453 | node->next = prev->next; |
bec39940 | 1454 | prev->next = flag_bucket(node); |
d6b18934 DG |
1455 | } |
1456 | } | |
fa68aa62 MD |
1457 | } |
1458 | ||
bec39940 DG |
1459 | struct cds_lfht *_cds_lfht_new(unsigned long init_size, |
1460 | unsigned long min_nr_alloc_buckets, | |
1461 | unsigned long max_nr_buckets, | |
fa68aa62 | 1462 | int flags, |
bec39940 DG |
1463 | const struct cds_lfht_mm_type *mm, |
1464 | const struct rcu_flavor_struct *flavor, | |
fa68aa62 MD |
1465 | pthread_attr_t *attr) |
1466 | { | |
1467 | struct cds_lfht *ht; | |
1468 | unsigned long order; | |
1469 | ||
bec39940 DG |
1470 | /* min_nr_alloc_buckets must be power of two */ |
1471 | if (!min_nr_alloc_buckets || (min_nr_alloc_buckets & (min_nr_alloc_buckets - 1))) | |
d6b18934 | 1472 | return NULL; |
bec39940 | 1473 | |
fa68aa62 | 1474 | /* init_size must be power of two */ |
d6b18934 | 1475 | if (!init_size || (init_size & (init_size - 1))) |
fa68aa62 | 1476 | return NULL; |
bec39940 DG |
1477 | |
1478 | /* | |
1479 | * Memory management plugin default. | |
1480 | */ | |
1481 | if (!mm) { | |
1482 | if (CAA_BITS_PER_LONG > 32 | |
1483 | && max_nr_buckets | |
1484 | && max_nr_buckets <= (1ULL << 32)) { | |
1485 | /* | |
1486 | * For 64-bit architectures, with max number of | |
1487 | * buckets small enough not to use the entire | |
1488 | * 64-bit memory mapping space (and allowing a | |
1489 | * fair number of hash table instances), use the | |
1490 | * mmap allocator, which is faster than the | |
1491 | * order allocator. | |
1492 | */ | |
1493 | mm = &cds_lfht_mm_mmap; | |
1494 | } else { | |
1495 | /* | |
1496 | * The fallback is to use the order allocator. | |
1497 | */ | |
1498 | mm = &cds_lfht_mm_order; | |
1499 | } | |
1500 | } | |
1501 | ||
1502 | /* max_nr_buckets == 0 for order based mm means infinite */ | |
1503 | if (mm == &cds_lfht_mm_order && !max_nr_buckets) | |
1504 | max_nr_buckets = 1UL << (MAX_TABLE_ORDER - 1); | |
1505 | ||
1506 | /* max_nr_buckets must be power of two */ | |
1507 | if (!max_nr_buckets || (max_nr_buckets & (max_nr_buckets - 1))) | |
1508 | return NULL; | |
1509 | ||
1510 | min_nr_alloc_buckets = max(min_nr_alloc_buckets, MIN_TABLE_SIZE); | |
1511 | init_size = max(init_size, MIN_TABLE_SIZE); | |
1512 | max_nr_buckets = max(max_nr_buckets, min_nr_alloc_buckets); | |
1513 | init_size = min(init_size, max_nr_buckets); | |
1514 | ||
1515 | ht = mm->alloc_cds_lfht(min_nr_alloc_buckets, max_nr_buckets); | |
fa68aa62 | 1516 | assert(ht); |
bec39940 DG |
1517 | assert(ht->mm == mm); |
1518 | assert(ht->bucket_at == mm->bucket_at); | |
1519 | ||
d6b18934 | 1520 | ht->flags = flags; |
bec39940 | 1521 | ht->flavor = flavor; |
fa68aa62 | 1522 | ht->resize_attr = attr; |
d6b18934 | 1523 | alloc_split_items_count(ht); |
fa68aa62 MD |
1524 | /* this mutex should not nest in read-side C.S. */ |
1525 | pthread_mutex_init(&ht->resize_mutex, NULL); | |
bec39940 DG |
1526 | order = cds_lfht_get_count_order_ulong(init_size); |
1527 | ht->resize_target = 1UL << order; | |
1528 | cds_lfht_create_bucket(ht, 1UL << order); | |
1529 | ht->size = 1UL << order; | |
fa68aa62 MD |
1530 | return ht; |
1531 | } | |
1532 | ||
bec39940 DG |
1533 | void cds_lfht_lookup(struct cds_lfht *ht, unsigned long hash, |
1534 | cds_lfht_match_fct match, const void *key, | |
fa68aa62 MD |
1535 | struct cds_lfht_iter *iter) |
1536 | { | |
bec39940 DG |
1537 | struct cds_lfht_node *node, *next, *bucket; |
1538 | unsigned long reverse_hash, size; | |
fa68aa62 | 1539 | |
fa68aa62 MD |
1540 | reverse_hash = bit_reverse_ulong(hash); |
1541 | ||
bec39940 DG |
1542 | size = rcu_dereference(ht->size); |
1543 | bucket = lookup_bucket(ht, size, hash); | |
1544 | /* We can always skip the bucket node initially */ | |
1545 | node = rcu_dereference(bucket->next); | |
fa68aa62 MD |
1546 | node = clear_flag(node); |
1547 | for (;;) { | |
6e59ae26 | 1548 | if (caa_unlikely(is_end(node))) { |
fa68aa62 MD |
1549 | node = next = NULL; |
1550 | break; | |
1551 | } | |
bec39940 | 1552 | if (caa_unlikely(node->reverse_hash > reverse_hash)) { |
fa68aa62 MD |
1553 | node = next = NULL; |
1554 | break; | |
1555 | } | |
bec39940 | 1556 | next = rcu_dereference(node->next); |
d6b18934 | 1557 | assert(node == clear_flag(node)); |
6e59ae26 | 1558 | if (caa_likely(!is_removed(next)) |
bec39940 DG |
1559 | && !is_bucket(next) |
1560 | && node->reverse_hash == reverse_hash | |
1561 | && caa_likely(match(node, key))) { | |
fa68aa62 MD |
1562 | break; |
1563 | } | |
1564 | node = clear_flag(next); | |
1565 | } | |
cd5376da | 1566 | assert(!node || !is_bucket(CMM_LOAD_SHARED(node->next))); |
fa68aa62 MD |
1567 | iter->node = node; |
1568 | iter->next = next; | |
1569 | } | |
1570 | ||
bec39940 DG |
1571 | void cds_lfht_next_duplicate(struct cds_lfht *ht, cds_lfht_match_fct match, |
1572 | const void *key, struct cds_lfht_iter *iter) | |
fa68aa62 MD |
1573 | { |
1574 | struct cds_lfht_node *node, *next; | |
1575 | unsigned long reverse_hash; | |
fa68aa62 MD |
1576 | |
1577 | node = iter->node; | |
bec39940 | 1578 | reverse_hash = node->reverse_hash; |
fa68aa62 MD |
1579 | next = iter->next; |
1580 | node = clear_flag(next); | |
1581 | ||
1582 | for (;;) { | |
6e59ae26 | 1583 | if (caa_unlikely(is_end(node))) { |
fa68aa62 MD |
1584 | node = next = NULL; |
1585 | break; | |
1586 | } | |
bec39940 | 1587 | if (caa_unlikely(node->reverse_hash > reverse_hash)) { |
fa68aa62 MD |
1588 | node = next = NULL; |
1589 | break; | |
1590 | } | |
bec39940 | 1591 | next = rcu_dereference(node->next); |
6e59ae26 | 1592 | if (caa_likely(!is_removed(next)) |
bec39940 DG |
1593 | && !is_bucket(next) |
1594 | && caa_likely(match(node, key))) { | |
fa68aa62 MD |
1595 | break; |
1596 | } | |
1597 | node = clear_flag(next); | |
1598 | } | |
cd5376da | 1599 | assert(!node || !is_bucket(CMM_LOAD_SHARED(node->next))); |
fa68aa62 MD |
1600 | iter->node = node; |
1601 | iter->next = next; | |
1602 | } | |
1603 | ||
f6a9efaa DG |
1604 | void cds_lfht_next(struct cds_lfht *ht, struct cds_lfht_iter *iter) |
1605 | { | |
1606 | struct cds_lfht_node *node, *next; | |
1607 | ||
1608 | node = clear_flag(iter->next); | |
1609 | for (;;) { | |
6e59ae26 | 1610 | if (caa_unlikely(is_end(node))) { |
f6a9efaa DG |
1611 | node = next = NULL; |
1612 | break; | |
1613 | } | |
bec39940 | 1614 | next = rcu_dereference(node->next); |
6e59ae26 | 1615 | if (caa_likely(!is_removed(next)) |
bec39940 | 1616 | && !is_bucket(next)) { |
f6a9efaa DG |
1617 | break; |
1618 | } | |
1619 | node = clear_flag(next); | |
1620 | } | |
cd5376da | 1621 | assert(!node || !is_bucket(CMM_LOAD_SHARED(node->next))); |
f6a9efaa DG |
1622 | iter->node = node; |
1623 | iter->next = next; | |
1624 | } | |
1625 | ||
1626 | void cds_lfht_first(struct cds_lfht *ht, struct cds_lfht_iter *iter) | |
1627 | { | |
f6a9efaa | 1628 | /* |
bec39940 | 1629 | * Get next after first bucket node. The first bucket node is the |
f6a9efaa DG |
1630 | * first node of the linked list. |
1631 | */ | |
bec39940 | 1632 | iter->next = bucket_at(ht, 0)->next; |
f6a9efaa DG |
1633 | cds_lfht_next(ht, iter); |
1634 | } | |
1635 | ||
bec39940 DG |
1636 | void cds_lfht_add(struct cds_lfht *ht, unsigned long hash, |
1637 | struct cds_lfht_node *node) | |
fa68aa62 | 1638 | { |
bec39940 | 1639 | unsigned long size; |
fa68aa62 | 1640 | |
bec39940 DG |
1641 | node->reverse_hash = bit_reverse_ulong(hash); |
1642 | size = rcu_dereference(ht->size); | |
1643 | _cds_lfht_add(ht, hash, NULL, NULL, size, node, NULL, 0); | |
d6b18934 | 1644 | ht_count_add(ht, size, hash); |
fa68aa62 MD |
1645 | } |
1646 | ||
1647 | struct cds_lfht_node *cds_lfht_add_unique(struct cds_lfht *ht, | |
bec39940 DG |
1648 | unsigned long hash, |
1649 | cds_lfht_match_fct match, | |
1650 | const void *key, | |
fa68aa62 MD |
1651 | struct cds_lfht_node *node) |
1652 | { | |
bec39940 | 1653 | unsigned long size; |
d6b18934 | 1654 | struct cds_lfht_iter iter; |
fa68aa62 | 1655 | |
bec39940 DG |
1656 | node->reverse_hash = bit_reverse_ulong(hash); |
1657 | size = rcu_dereference(ht->size); | |
1658 | _cds_lfht_add(ht, hash, match, key, size, node, &iter, 0); | |
d6b18934 DG |
1659 | if (iter.node == node) |
1660 | ht_count_add(ht, size, hash); | |
1661 | return iter.node; | |
fa68aa62 MD |
1662 | } |
1663 | ||
1664 | struct cds_lfht_node *cds_lfht_add_replace(struct cds_lfht *ht, | |
bec39940 DG |
1665 | unsigned long hash, |
1666 | cds_lfht_match_fct match, | |
1667 | const void *key, | |
fa68aa62 MD |
1668 | struct cds_lfht_node *node) |
1669 | { | |
bec39940 | 1670 | unsigned long size; |
d6b18934 | 1671 | struct cds_lfht_iter iter; |
fa68aa62 | 1672 | |
bec39940 DG |
1673 | node->reverse_hash = bit_reverse_ulong(hash); |
1674 | size = rcu_dereference(ht->size); | |
d6b18934 | 1675 | for (;;) { |
bec39940 | 1676 | _cds_lfht_add(ht, hash, match, key, size, node, &iter, 0); |
d6b18934 DG |
1677 | if (iter.node == node) { |
1678 | ht_count_add(ht, size, hash); | |
1679 | return NULL; | |
1680 | } | |
1681 | ||
1682 | if (!_cds_lfht_replace(ht, size, iter.node, iter.next, node)) | |
1683 | return iter.node; | |
1684 | } | |
fa68aa62 MD |
1685 | } |
1686 | ||
bec39940 DG |
1687 | int cds_lfht_replace(struct cds_lfht *ht, |
1688 | struct cds_lfht_iter *old_iter, | |
1689 | unsigned long hash, | |
1690 | cds_lfht_match_fct match, | |
1691 | const void *key, | |
fa68aa62 MD |
1692 | struct cds_lfht_node *new_node) |
1693 | { | |
1694 | unsigned long size; | |
1695 | ||
bec39940 DG |
1696 | new_node->reverse_hash = bit_reverse_ulong(hash); |
1697 | if (!old_iter->node) | |
1698 | return -ENOENT; | |
1699 | if (caa_unlikely(old_iter->node->reverse_hash != new_node->reverse_hash)) | |
1700 | return -EINVAL; | |
1701 | if (caa_unlikely(!match(old_iter->node, key))) | |
1702 | return -EINVAL; | |
1703 | size = rcu_dereference(ht->size); | |
fa68aa62 MD |
1704 | return _cds_lfht_replace(ht, size, old_iter->node, old_iter->next, |
1705 | new_node); | |
1706 | } | |
1707 | ||
bec39940 | 1708 | int cds_lfht_del(struct cds_lfht *ht, struct cds_lfht_node *node) |
fa68aa62 | 1709 | { |
d6b18934 | 1710 | unsigned long size, hash; |
fa68aa62 MD |
1711 | int ret; |
1712 | ||
bec39940 DG |
1713 | size = rcu_dereference(ht->size); |
1714 | ret = _cds_lfht_del(ht, size, node); | |
d6b18934 | 1715 | if (!ret) { |
bec39940 | 1716 | hash = bit_reverse_ulong(node->reverse_hash); |
d6b18934 DG |
1717 | ht_count_del(ht, size, hash); |
1718 | } | |
fa68aa62 MD |
1719 | return ret; |
1720 | } | |
1721 | ||
cd5376da DG |
1722 | int cds_lfht_is_node_deleted(struct cds_lfht_node *node) |
1723 | { | |
1724 | return is_removed(CMM_LOAD_SHARED(node->next)); | |
1725 | } | |
1726 | ||
fa68aa62 | 1727 | static |
bec39940 | 1728 | int cds_lfht_delete_bucket(struct cds_lfht *ht) |
fa68aa62 MD |
1729 | { |
1730 | struct cds_lfht_node *node; | |
fa68aa62 MD |
1731 | unsigned long order, i, size; |
1732 | ||
1733 | /* Check that the table is empty */ | |
bec39940 | 1734 | node = bucket_at(ht, 0); |
fa68aa62 | 1735 | do { |
bec39940 DG |
1736 | node = clear_flag(node)->next; |
1737 | if (!is_bucket(node)) | |
fa68aa62 MD |
1738 | return -EPERM; |
1739 | assert(!is_removed(node)); | |
1740 | } while (!is_end(node)); | |
1741 | /* | |
1742 | * size accessed without rcu_dereference because hash table is | |
1743 | * being destroyed. | |
1744 | */ | |
bec39940 | 1745 | size = ht->size; |
cd5376da | 1746 | /* Internal sanity check: all nodes left should be buckets */ |
bec39940 DG |
1747 | for (i = 0; i < size; i++) { |
1748 | node = bucket_at(ht, i); | |
1749 | dbg_printf("delete bucket: index %lu expected hash %lu hash %lu\n", | |
1750 | i, i, bit_reverse_ulong(node->reverse_hash)); | |
1751 | assert(is_bucket(node->next)); | |
1752 | } | |
fa68aa62 | 1753 | |
bec39940 DG |
1754 | for (order = cds_lfht_get_count_order_ulong(size); (long)order >= 0; order--) |
1755 | cds_lfht_free_bucket_table(ht, order); | |
d6b18934 | 1756 | |
fa68aa62 MD |
1757 | return 0; |
1758 | } | |
1759 | ||
1760 | /* | |
1761 | * Should only be called when no more concurrent readers nor writers can | |
1762 | * possibly access the table. | |
1763 | */ | |
1764 | int cds_lfht_destroy(struct cds_lfht *ht, pthread_attr_t **attr) | |
1765 | { | |
1766 | int ret; | |
3a1aff7a MD |
1767 | #ifdef rcu_read_ongoing_mb |
1768 | int was_online; | |
1769 | #endif | |
fa68aa62 MD |
1770 | |
1771 | /* Wait for in-flight resize operations to complete */ | |
d6b18934 DG |
1772 | _CMM_STORE_SHARED(ht->in_progress_destroy, 1); |
1773 | cmm_smp_mb(); /* Store destroy before load resize */ | |
3a1aff7a MD |
1774 | #ifdef rcu_read_ongoing_mb |
1775 | was_online = ht->flavor->read_ongoing(); | |
1776 | if (was_online) | |
1777 | ht->flavor->thread_offline(); | |
1778 | /* Calling with RCU read-side held is an error. */ | |
1779 | if (ht->flavor->read_ongoing()) { | |
1780 | ret = -EINVAL; | |
1781 | if (was_online) | |
1782 | ht->flavor->thread_online(); | |
1783 | goto end; | |
1784 | } | |
1785 | #endif | |
fa68aa62 MD |
1786 | while (uatomic_read(&ht->in_progress_resize)) |
1787 | poll(NULL, 0, 100); /* wait for 100ms */ | |
bec39940 | 1788 | ret = cds_lfht_delete_bucket(ht); |
fa68aa62 MD |
1789 | if (ret) |
1790 | return ret; | |
d6b18934 | 1791 | free_split_items_count(ht); |
fa68aa62 MD |
1792 | if (attr) |
1793 | *attr = ht->resize_attr; | |
1794 | poison_free(ht); | |
3a1aff7a MD |
1795 | #ifdef rcu_read_ongoing_mb |
1796 | end: | |
1797 | #endif | |
fa68aa62 MD |
1798 | return ret; |
1799 | } | |
1800 | ||
1801 | void cds_lfht_count_nodes(struct cds_lfht *ht, | |
1802 | long *approx_before, | |
1803 | unsigned long *count, | |
fa68aa62 MD |
1804 | long *approx_after) |
1805 | { | |
1806 | struct cds_lfht_node *node, *next; | |
bec39940 | 1807 | unsigned long nr_bucket = 0, nr_removed = 0; |
fa68aa62 MD |
1808 | |
1809 | *approx_before = 0; | |
d6b18934 | 1810 | if (ht->split_count) { |
fa68aa62 MD |
1811 | int i; |
1812 | ||
d6b18934 DG |
1813 | for (i = 0; i < split_count_mask + 1; i++) { |
1814 | *approx_before += uatomic_read(&ht->split_count[i].add); | |
1815 | *approx_before -= uatomic_read(&ht->split_count[i].del); | |
fa68aa62 MD |
1816 | } |
1817 | } | |
1818 | ||
1819 | *count = 0; | |
fa68aa62 | 1820 | |
bec39940 DG |
1821 | /* Count non-bucket nodes in the table */ |
1822 | node = bucket_at(ht, 0); | |
fa68aa62 | 1823 | do { |
bec39940 | 1824 | next = rcu_dereference(node->next); |
fa68aa62 | 1825 | if (is_removed(next)) { |
bec39940 DG |
1826 | if (!is_bucket(next)) |
1827 | (nr_removed)++; | |
fa68aa62 | 1828 | else |
bec39940 DG |
1829 | (nr_bucket)++; |
1830 | } else if (!is_bucket(next)) | |
fa68aa62 MD |
1831 | (*count)++; |
1832 | else | |
bec39940 | 1833 | (nr_bucket)++; |
fa68aa62 MD |
1834 | node = clear_flag(next); |
1835 | } while (!is_end(node)); | |
bec39940 DG |
1836 | dbg_printf("number of logically removed nodes: %lu\n", nr_removed); |
1837 | dbg_printf("number of bucket nodes: %lu\n", nr_bucket); | |
fa68aa62 | 1838 | *approx_after = 0; |
d6b18934 | 1839 | if (ht->split_count) { |
fa68aa62 MD |
1840 | int i; |
1841 | ||
d6b18934 DG |
1842 | for (i = 0; i < split_count_mask + 1; i++) { |
1843 | *approx_after += uatomic_read(&ht->split_count[i].add); | |
1844 | *approx_after -= uatomic_read(&ht->split_count[i].del); | |
fa68aa62 MD |
1845 | } |
1846 | } | |
1847 | } | |
1848 | ||
1849 | /* called with resize mutex held */ | |
1850 | static | |
1851 | void _do_cds_lfht_grow(struct cds_lfht *ht, | |
1852 | unsigned long old_size, unsigned long new_size) | |
1853 | { | |
1854 | unsigned long old_order, new_order; | |
1855 | ||
bec39940 DG |
1856 | old_order = cds_lfht_get_count_order_ulong(old_size); |
1857 | new_order = cds_lfht_get_count_order_ulong(new_size); | |
d6b18934 DG |
1858 | dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n", |
1859 | old_size, old_order, new_size, new_order); | |
fa68aa62 | 1860 | assert(new_size > old_size); |
d6b18934 | 1861 | init_table(ht, old_order + 1, new_order); |
fa68aa62 MD |
1862 | } |
1863 | ||
1864 | /* called with resize mutex held */ | |
1865 | static | |
1866 | void _do_cds_lfht_shrink(struct cds_lfht *ht, | |
1867 | unsigned long old_size, unsigned long new_size) | |
1868 | { | |
1869 | unsigned long old_order, new_order; | |
1870 | ||
bec39940 DG |
1871 | new_size = max(new_size, MIN_TABLE_SIZE); |
1872 | old_order = cds_lfht_get_count_order_ulong(old_size); | |
1873 | new_order = cds_lfht_get_count_order_ulong(new_size); | |
d6b18934 DG |
1874 | dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n", |
1875 | old_size, old_order, new_size, new_order); | |
fa68aa62 MD |
1876 | assert(new_size < old_size); |
1877 | ||
bec39940 | 1878 | /* Remove and unlink all bucket nodes to remove. */ |
d6b18934 | 1879 | fini_table(ht, new_order + 1, old_order); |
fa68aa62 MD |
1880 | } |
1881 | ||
1882 | ||
1883 | /* called with resize mutex held */ | |
1884 | static | |
1885 | void _do_cds_lfht_resize(struct cds_lfht *ht) | |
1886 | { | |
1887 | unsigned long new_size, old_size; | |
1888 | ||
1889 | /* | |
1890 | * Resize table, re-do if the target size has changed under us. | |
1891 | */ | |
1892 | do { | |
d6b18934 DG |
1893 | assert(uatomic_read(&ht->in_progress_resize)); |
1894 | if (CMM_LOAD_SHARED(ht->in_progress_destroy)) | |
1895 | break; | |
bec39940 DG |
1896 | ht->resize_initiated = 1; |
1897 | old_size = ht->size; | |
1898 | new_size = CMM_LOAD_SHARED(ht->resize_target); | |
fa68aa62 MD |
1899 | if (old_size < new_size) |
1900 | _do_cds_lfht_grow(ht, old_size, new_size); | |
1901 | else if (old_size > new_size) | |
1902 | _do_cds_lfht_shrink(ht, old_size, new_size); | |
bec39940 | 1903 | ht->resize_initiated = 0; |
fa68aa62 MD |
1904 | /* write resize_initiated before read resize_target */ |
1905 | cmm_smp_mb(); | |
bec39940 | 1906 | } while (ht->size != CMM_LOAD_SHARED(ht->resize_target)); |
fa68aa62 MD |
1907 | } |
1908 | ||
1909 | static | |
bec39940 | 1910 | unsigned long resize_target_grow(struct cds_lfht *ht, unsigned long new_size) |
fa68aa62 | 1911 | { |
bec39940 | 1912 | return _uatomic_xchg_monotonic_increase(&ht->resize_target, new_size); |
fa68aa62 MD |
1913 | } |
1914 | ||
1915 | static | |
1916 | void resize_target_update_count(struct cds_lfht *ht, | |
1917 | unsigned long count) | |
1918 | { | |
bec39940 DG |
1919 | count = max(count, MIN_TABLE_SIZE); |
1920 | count = min(count, ht->max_nr_buckets); | |
1921 | uatomic_set(&ht->resize_target, count); | |
fa68aa62 MD |
1922 | } |
1923 | ||
1924 | void cds_lfht_resize(struct cds_lfht *ht, unsigned long new_size) | |
1925 | { | |
3a1aff7a MD |
1926 | #ifdef rcu_read_ongoing_mb |
1927 | int was_online; | |
1928 | ||
1929 | was_online = ht->flavor->read_ongoing(); | |
1930 | if (was_online) | |
1931 | ht->flavor->thread_offline(); | |
1932 | /* Calling with RCU read-side held is an error. */ | |
1933 | if (ht->flavor->read_ongoing()) { | |
1934 | static int print_once; | |
1935 | ||
1936 | if (!CMM_LOAD_SHARED(print_once)) | |
1937 | fprintf(stderr, "[error] rculfhash: cds_lfht_resize " | |
1938 | "called with RCU read-side lock held.\n"); | |
1939 | CMM_STORE_SHARED(print_once, 1); | |
1940 | assert(0); | |
1941 | goto end; | |
1942 | } | |
1943 | #endif | |
fa68aa62 | 1944 | resize_target_update_count(ht, new_size); |
bec39940 | 1945 | CMM_STORE_SHARED(ht->resize_initiated, 1); |
fa68aa62 MD |
1946 | pthread_mutex_lock(&ht->resize_mutex); |
1947 | _do_cds_lfht_resize(ht); | |
1948 | pthread_mutex_unlock(&ht->resize_mutex); | |
3a1aff7a MD |
1949 | #ifdef rcu_read_ongoing_mb |
1950 | end: | |
1951 | if (was_online) | |
1952 | ht->flavor->thread_online(); | |
1953 | #endif | |
fa68aa62 MD |
1954 | } |
1955 | ||
1956 | static | |
1957 | void do_resize_cb(struct rcu_head *head) | |
1958 | { | |
1959 | struct rcu_resize_work *work = | |
1960 | caa_container_of(head, struct rcu_resize_work, head); | |
1961 | struct cds_lfht *ht = work->ht; | |
1962 | ||
bec39940 | 1963 | ht->flavor->thread_offline(); |
fa68aa62 MD |
1964 | pthread_mutex_lock(&ht->resize_mutex); |
1965 | _do_cds_lfht_resize(ht); | |
1966 | pthread_mutex_unlock(&ht->resize_mutex); | |
bec39940 | 1967 | ht->flavor->thread_online(); |
fa68aa62 MD |
1968 | poison_free(work); |
1969 | cmm_smp_mb(); /* finish resize before decrement */ | |
1970 | uatomic_dec(&ht->in_progress_resize); | |
1971 | } | |
1972 | ||
1973 | static | |
bec39940 | 1974 | void __cds_lfht_resize_lazy_launch(struct cds_lfht *ht) |
fa68aa62 MD |
1975 | { |
1976 | struct rcu_resize_work *work; | |
fa68aa62 | 1977 | |
fa68aa62 MD |
1978 | /* Store resize_target before read resize_initiated */ |
1979 | cmm_smp_mb(); | |
bec39940 | 1980 | if (!CMM_LOAD_SHARED(ht->resize_initiated)) { |
fa68aa62 | 1981 | uatomic_inc(&ht->in_progress_resize); |
d6b18934 DG |
1982 | cmm_smp_mb(); /* increment resize count before load destroy */ |
1983 | if (CMM_LOAD_SHARED(ht->in_progress_destroy)) { | |
1984 | uatomic_dec(&ht->in_progress_resize); | |
1985 | return; | |
1986 | } | |
a0daceaa | 1987 | work = zmalloc(sizeof(*work)); |
cd5376da DG |
1988 | if (work == NULL) { |
1989 | dbg_printf("error allocating resize work, bailing out\n"); | |
1990 | uatomic_dec(&ht->in_progress_resize); | |
1991 | return; | |
1992 | } | |
fa68aa62 | 1993 | work->ht = ht; |
bec39940 DG |
1994 | ht->flavor->update_call_rcu(&work->head, do_resize_cb); |
1995 | CMM_STORE_SHARED(ht->resize_initiated, 1); | |
fa68aa62 MD |
1996 | } |
1997 | } | |
1998 | ||
bec39940 DG |
1999 | static |
2000 | void cds_lfht_resize_lazy_grow(struct cds_lfht *ht, unsigned long size, int growth) | |
2001 | { | |
2002 | unsigned long target_size = size << growth; | |
2003 | ||
2004 | target_size = min(target_size, ht->max_nr_buckets); | |
2005 | if (resize_target_grow(ht, target_size) >= target_size) | |
2006 | return; | |
2007 | ||
2008 | __cds_lfht_resize_lazy_launch(ht); | |
2009 | } | |
2010 | ||
2011 | /* | |
2012 | * We favor grow operations over shrink. A shrink operation never occurs | |
2013 | * if a grow operation is queued for lazy execution. A grow operation | |
2014 | * cancels any pending shrink lazy execution. | |
2015 | */ | |
fa68aa62 MD |
2016 | static |
2017 | void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size, | |
2018 | unsigned long count) | |
2019 | { | |
fa68aa62 MD |
2020 | if (!(ht->flags & CDS_LFHT_AUTO_RESIZE)) |
2021 | return; | |
bec39940 DG |
2022 | count = max(count, MIN_TABLE_SIZE); |
2023 | count = min(count, ht->max_nr_buckets); | |
2024 | if (count == size) | |
2025 | return; /* Already the right size, no resize needed */ | |
2026 | if (count > size) { /* lazy grow */ | |
2027 | if (resize_target_grow(ht, count) >= count) | |
d6b18934 | 2028 | return; |
bec39940 DG |
2029 | } else { /* lazy shrink */ |
2030 | for (;;) { | |
2031 | unsigned long s; | |
2032 | ||
2033 | s = uatomic_cmpxchg(&ht->resize_target, size, count); | |
2034 | if (s == size) | |
2035 | break; /* no resize needed */ | |
2036 | if (s > size) | |
2037 | return; /* growing is/(was just) in progress */ | |
2038 | if (s <= count) | |
2039 | return; /* some other thread do shrink */ | |
2040 | size = s; | |
d6b18934 | 2041 | } |
fa68aa62 | 2042 | } |
bec39940 | 2043 | __cds_lfht_resize_lazy_launch(ht); |
fa68aa62 | 2044 | } |