Commit | Line | Data |
---|---|---|
5e28c532 | 1 | /* |
abc490a1 MD |
2 | * rculfhash.c |
3 | * | |
1475579c | 4 | * Userspace RCU library - Lock-Free Resizable RCU Hash Table |
abc490a1 MD |
5 | * |
6 | * Copyright 2010-2011 - Mathieu Desnoyers <mathieu.desnoyers@efficios.com> | |
7 | * | |
8 | * This library is free software; you can redistribute it and/or | |
9 | * modify it under the terms of the GNU Lesser General Public | |
10 | * License as published by the Free Software Foundation; either | |
11 | * version 2.1 of the License, or (at your option) any later version. | |
12 | * | |
13 | * This library is distributed in the hope that it will be useful, | |
14 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
15 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | |
16 | * Lesser General Public License for more details. | |
17 | * | |
18 | * You should have received a copy of the GNU Lesser General Public | |
19 | * License along with this library; if not, write to the Free Software | |
20 | * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA | |
5e28c532 MD |
21 | */ |
22 | ||
e753ff5a MD |
23 | /* |
24 | * Based on the following articles: | |
25 | * - Ori Shalev and Nir Shavit. Split-ordered lists: Lock-free | |
26 | * extensible hash tables. J. ACM 53, 3 (May 2006), 379-405. | |
27 | * - Michael, M. M. High performance dynamic lock-free hash tables | |
28 | * and list-based sets. In Proceedings of the fourteenth annual ACM | |
29 | * symposium on Parallel algorithms and architectures, ACM Press, | |
30 | * (2002), 73-82. | |
31 | * | |
1475579c | 32 | * Some specificities of this Lock-Free Resizable RCU Hash Table |
e753ff5a MD |
33 | * implementation: |
34 | * | |
35 | * - RCU read-side critical section allows readers to perform hash | |
36 | * table lookups and use the returned objects safely by delaying | |
37 | * memory reclaim of a grace period. | |
38 | * - Add and remove operations are lock-free, and do not need to | |
39 | * allocate memory. They need to be executed within RCU read-side | |
40 | * critical section to ensure the objects they read are valid and to | |
41 | * deal with the cmpxchg ABA problem. | |
42 | * - add and add_unique operations are supported. add_unique checks if | |
43 | * the node key already exists in the hash table. It ensures no key | |
44 | * duplicata exists. | |
45 | * - The resize operation executes concurrently with add/remove/lookup. | |
46 | * - Hash table nodes are contained within a split-ordered list. This | |
47 | * list is ordered by incrementing reversed-bits-hash value. | |
48 | * - An index of dummy nodes is kept. These dummy nodes are the hash | |
49 | * table "buckets", and they are also chained together in the | |
50 | * split-ordered list, which allows recursive expansion. | |
1475579c MD |
51 | * - The resize operation for small tables only allows expanding the hash table. |
52 | * It is triggered automatically by detecting long chains in the add | |
53 | * operation. | |
54 | * - The resize operation for larger tables (and available through an | |
55 | * API) allows both expanding and shrinking the hash table. | |
4c42f1b8 | 56 | * - Split-counters are used to keep track of the number of |
1475579c | 57 | * nodes within the hash table for automatic resize triggering. |
e753ff5a MD |
58 | * - Resize operation initiated by long chain detection is executed by a |
59 | * call_rcu thread, which keeps lock-freedom of add and remove. | |
60 | * - Resize operations are protected by a mutex. | |
61 | * - The removal operation is split in two parts: first, a "removed" | |
62 | * flag is set in the next pointer within the node to remove. Then, | |
63 | * a "garbage collection" is performed in the bucket containing the | |
64 | * removed node (from the start of the bucket up to the removed node). | |
65 | * All encountered nodes with "removed" flag set in their next | |
66 | * pointers are removed from the linked-list. If the cmpxchg used for | |
67 | * removal fails (due to concurrent garbage-collection or concurrent | |
68 | * add), we retry from the beginning of the bucket. This ensures that | |
69 | * the node with "removed" flag set is removed from the hash table | |
70 | * (not visible to lookups anymore) before the RCU read-side critical | |
71 | * section held across removal ends. Furthermore, this ensures that | |
72 | * the node with "removed" flag set is removed from the linked-list | |
73 | * before its memory is reclaimed. Only the thread which removal | |
74 | * successfully set the "removed" flag (with a cmpxchg) into a node's | |
75 | * next pointer is considered to have succeeded its removal (and thus | |
76 | * owns the node to reclaim). Because we garbage-collect starting from | |
77 | * an invariant node (the start-of-bucket dummy node) up to the | |
78 | * "removed" node (or find a reverse-hash that is higher), we are sure | |
79 | * that a successful traversal of the chain leads to a chain that is | |
80 | * present in the linked-list (the start node is never removed) and | |
81 | * that is does not contain the "removed" node anymore, even if | |
82 | * concurrent delete/add operations are changing the structure of the | |
83 | * list concurrently. | |
29e669f6 MD |
84 | * - The add operation performs gargage collection of buckets if it |
85 | * encounters nodes with removed flag set in the bucket where it wants | |
86 | * to add its new node. This ensures lock-freedom of add operation by | |
87 | * helping the remover unlink nodes from the list rather than to wait | |
88 | * for it do to so. | |
e753ff5a MD |
89 | * - A RCU "order table" indexed by log2(hash index) is copied and |
90 | * expanded by the resize operation. This order table allows finding | |
91 | * the "dummy node" tables. | |
92 | * - There is one dummy node table per hash index order. The size of | |
93 | * each dummy node table is half the number of hashes contained in | |
93d46c39 LJ |
94 | * this order (except for order 0). |
95 | * - synchronzie_rcu is used to garbage-collect the old dummy node table. | |
e753ff5a MD |
96 | * - The per-order dummy node tables contain a compact version of the |
97 | * hash table nodes. These tables are invariant after they are | |
98 | * populated into the hash table. | |
93d46c39 LJ |
99 | * |
100 | * Dummy node tables: | |
101 | * | |
102 | * hash table hash table the last all dummy node tables | |
103 | * order size dummy node 0 1 2 3 4 5 6(index) | |
104 | * table size | |
105 | * 0 1 1 1 | |
106 | * 1 2 1 1 1 | |
107 | * 2 4 2 1 1 2 | |
108 | * 3 8 4 1 1 2 4 | |
109 | * 4 16 8 1 1 2 4 8 | |
110 | * 5 32 16 1 1 2 4 8 16 | |
111 | * 6 64 32 1 1 2 4 8 16 32 | |
112 | * | |
113 | * When growing/shrinking, we only focus on the last dummy node table | |
114 | * which size is (!order ? 1 : (1 << (order -1))). | |
115 | * | |
116 | * Example for growing/shrinking: | |
117 | * grow hash table from order 5 to 6: init the index=6 dummy node table | |
118 | * shrink hash table from order 6 to 5: fini the index=6 dummy node table | |
119 | * | |
1475579c MD |
120 | * A bit of ascii art explanation: |
121 | * | |
122 | * Order index is the off-by-one compare to the actual power of 2 because | |
123 | * we use index 0 to deal with the 0 special-case. | |
124 | * | |
125 | * This shows the nodes for a small table ordered by reversed bits: | |
126 | * | |
127 | * bits reverse | |
128 | * 0 000 000 | |
129 | * 4 100 001 | |
130 | * 2 010 010 | |
131 | * 6 110 011 | |
132 | * 1 001 100 | |
133 | * 5 101 101 | |
134 | * 3 011 110 | |
135 | * 7 111 111 | |
136 | * | |
137 | * This shows the nodes in order of non-reversed bits, linked by | |
138 | * reversed-bit order. | |
139 | * | |
140 | * order bits reverse | |
141 | * 0 0 000 000 | |
0adc36a8 LJ |
142 | * 1 | 1 001 100 <- |
143 | * 2 | | 2 010 010 <- | | |
f6fdd688 | 144 | * | | | 3 011 110 | <- | |
1475579c MD |
145 | * 3 -> | | | 4 100 001 | | |
146 | * -> | | 5 101 101 | | |
147 | * -> | 6 110 011 | |
148 | * -> 7 111 111 | |
e753ff5a MD |
149 | */ |
150 | ||
2ed95849 MD |
151 | #define _LGPL_SOURCE |
152 | #include <stdlib.h> | |
e0ba718a MD |
153 | #include <errno.h> |
154 | #include <assert.h> | |
155 | #include <stdio.h> | |
abc490a1 | 156 | #include <stdint.h> |
f000907d | 157 | #include <string.h> |
e0ba718a | 158 | |
15cfbec7 | 159 | #include "config.h" |
2ed95849 | 160 | #include <urcu.h> |
abc490a1 | 161 | #include <urcu-call-rcu.h> |
a42cc659 MD |
162 | #include <urcu/arch.h> |
163 | #include <urcu/uatomic.h> | |
a42cc659 | 164 | #include <urcu/compiler.h> |
abc490a1 | 165 | #include <urcu/rculfhash.h> |
5e28c532 | 166 | #include <stdio.h> |
464a1ec9 | 167 | #include <pthread.h> |
44395fb7 | 168 | |
f9830efd | 169 | #ifdef DEBUG |
f0c29ed7 | 170 | #define dbg_printf(fmt, args...) printf("[debug rculfhash] " fmt, ## args) |
f9830efd | 171 | #else |
e753ff5a | 172 | #define dbg_printf(fmt, args...) |
f9830efd MD |
173 | #endif |
174 | ||
f8994aee | 175 | /* |
4c42f1b8 | 176 | * Split-counters lazily update the global counter each 1024 |
f8994aee MD |
177 | * addition/removal. It automatically keeps track of resize required. |
178 | * We use the bucket length as indicator for need to expand for small | |
179 | * tables and machines lacking per-cpu data suppport. | |
180 | */ | |
181 | #define COUNT_COMMIT_ORDER 10 | |
4ddbb355 | 182 | #define DEFAULT_SPLIT_COUNT_MASK 0xFUL |
6ea6bc67 MD |
183 | #define CHAIN_LEN_TARGET 1 |
184 | #define CHAIN_LEN_RESIZE_THRESHOLD 3 | |
2ed95849 | 185 | |
cd95516d | 186 | /* |
76a73da8 | 187 | * Define the minimum table size. |
cd95516d | 188 | */ |
c9edd44a | 189 | #define MIN_TABLE_SIZE 1 |
cd95516d | 190 | |
4105056a MD |
191 | #if (CAA_BITS_PER_LONG == 32) |
192 | #define MAX_TABLE_ORDER 32 | |
193 | #else | |
194 | #define MAX_TABLE_ORDER 64 | |
195 | #endif | |
196 | ||
b7d619b0 MD |
197 | /* |
198 | * Minimum number of dummy nodes to touch per thread to parallelize grow/shrink. | |
199 | */ | |
6083a889 MD |
200 | #define MIN_PARTITION_PER_THREAD_ORDER 12 |
201 | #define MIN_PARTITION_PER_THREAD (1UL << MIN_PARTITION_PER_THREAD_ORDER) | |
b7d619b0 | 202 | |
4105056a MD |
203 | #ifndef min |
204 | #define min(a, b) ((a) < (b) ? (a) : (b)) | |
205 | #endif | |
206 | ||
abc490a1 MD |
207 | #ifndef max |
208 | #define max(a, b) ((a) > (b) ? (a) : (b)) | |
209 | #endif | |
2ed95849 | 210 | |
d95bd160 MD |
211 | /* |
212 | * The removed flag needs to be updated atomically with the pointer. | |
48ed1c18 | 213 | * It indicates that no node must attach to the node scheduled for |
b198f0fd | 214 | * removal, and that node garbage collection must be performed. |
d95bd160 MD |
215 | * The dummy flag does not require to be updated atomically with the |
216 | * pointer, but it is added as a pointer low bit flag to save space. | |
217 | */ | |
d37166c6 | 218 | #define REMOVED_FLAG (1UL << 0) |
b198f0fd MD |
219 | #define DUMMY_FLAG (1UL << 1) |
220 | #define FLAGS_MASK ((1UL << 2) - 1) | |
d37166c6 | 221 | |
bb7b2f26 | 222 | /* Value of the end pointer. Should not interact with flags. */ |
f9c80341 | 223 | #define END_VALUE NULL |
bb7b2f26 | 224 | |
7f52427b MD |
225 | /* |
226 | * ht_items_count: Split-counters counting the number of node addition | |
227 | * and removal in the table. Only used if the CDS_LFHT_ACCOUNTING flag | |
228 | * is set at hash table creation. | |
229 | * | |
230 | * These are free-running counters, never reset to zero. They count the | |
231 | * number of add/remove, and trigger every (1 << COUNT_COMMIT_ORDER) | |
232 | * operations to update the global counter. We choose a power-of-2 value | |
233 | * for the trigger to deal with 32 or 64-bit overflow of the counter. | |
234 | */ | |
df44348d | 235 | struct ht_items_count { |
860d07e8 | 236 | unsigned long add, del; |
df44348d MD |
237 | } __attribute__((aligned(CAA_CACHE_LINE_SIZE))); |
238 | ||
7f52427b MD |
239 | /* |
240 | * rcu_level: Contains the per order-index-level dummy node table. The | |
241 | * size of each dummy node table is half the number of hashes contained | |
242 | * in this order (except for order 0). The minimum allocation size | |
243 | * parameter allows combining the dummy node arrays of the lowermost | |
244 | * levels to improve cache locality for small index orders. | |
245 | */ | |
1475579c | 246 | struct rcu_level { |
0d14ceb2 | 247 | /* Note: manually update allocation length when adding a field */ |
1475579c MD |
248 | struct _cds_lfht_node nodes[0]; |
249 | }; | |
250 | ||
7f52427b MD |
251 | /* |
252 | * rcu_table: Contains the size and desired new size if a resize | |
253 | * operation is in progress, as well as the statically-sized array of | |
254 | * rcu_level pointers. | |
255 | */ | |
395270b6 | 256 | struct rcu_table { |
4105056a | 257 | unsigned long size; /* always a power of 2, shared (RCU) */ |
f9830efd | 258 | unsigned long resize_target; |
11519af6 | 259 | int resize_initiated; |
4105056a | 260 | struct rcu_level *tbl[MAX_TABLE_ORDER]; |
395270b6 MD |
261 | }; |
262 | ||
7f52427b MD |
263 | /* |
264 | * cds_lfht: Top-level data structure representing a lock-free hash | |
265 | * table. Defined in the implementation file to make it be an opaque | |
266 | * cookie to users. | |
267 | */ | |
14044b37 | 268 | struct cds_lfht { |
4105056a | 269 | struct rcu_table t; |
14044b37 MD |
270 | cds_lfht_hash_fct hash_fct; |
271 | cds_lfht_compare_fct compare_fct; | |
5488222b LJ |
272 | unsigned long min_alloc_order; |
273 | unsigned long min_alloc_size; | |
732ad076 | 274 | unsigned long hash_seed; |
b8af5011 | 275 | int flags; |
5f511391 MD |
276 | /* |
277 | * We need to put the work threads offline (QSBR) when taking this | |
278 | * mutex, because we use synchronize_rcu within this mutex critical | |
279 | * section, which waits on read-side critical sections, and could | |
280 | * therefore cause grace-period deadlock if we hold off RCU G.P. | |
281 | * completion. | |
282 | */ | |
464a1ec9 | 283 | pthread_mutex_t resize_mutex; /* resize mutex: add/del mutex */ |
33c7c748 | 284 | unsigned int in_progress_resize, in_progress_destroy; |
14044b37 | 285 | void (*cds_lfht_call_rcu)(struct rcu_head *head, |
abc490a1 | 286 | void (*func)(struct rcu_head *head)); |
1475579c | 287 | void (*cds_lfht_synchronize_rcu)(void); |
01dbfa62 MD |
288 | void (*cds_lfht_rcu_read_lock)(void); |
289 | void (*cds_lfht_rcu_read_unlock)(void); | |
5f511391 MD |
290 | void (*cds_lfht_rcu_thread_offline)(void); |
291 | void (*cds_lfht_rcu_thread_online)(void); | |
b7d619b0 MD |
292 | void (*cds_lfht_rcu_register_thread)(void); |
293 | void (*cds_lfht_rcu_unregister_thread)(void); | |
294 | pthread_attr_t *resize_attr; /* Resize threads attributes */ | |
7de5ccfd | 295 | long count; /* global approximate item count */ |
4c42f1b8 | 296 | struct ht_items_count *split_count; /* split item count */ |
2ed95849 MD |
297 | }; |
298 | ||
7f52427b MD |
299 | /* |
300 | * rcu_resize_work: Contains arguments passed to RCU worker thread | |
301 | * responsible for performing lazy resize. | |
302 | */ | |
abc490a1 MD |
303 | struct rcu_resize_work { |
304 | struct rcu_head head; | |
14044b37 | 305 | struct cds_lfht *ht; |
abc490a1 | 306 | }; |
2ed95849 | 307 | |
7f52427b MD |
308 | /* |
309 | * partition_resize_work: Contains arguments passed to worker threads | |
310 | * executing the hash table resize on partitions of the hash table | |
311 | * assigned to each processor's worker thread. | |
312 | */ | |
b7d619b0 | 313 | struct partition_resize_work { |
1af6e26e | 314 | pthread_t thread_id; |
b7d619b0 MD |
315 | struct cds_lfht *ht; |
316 | unsigned long i, start, len; | |
317 | void (*fct)(struct cds_lfht *ht, unsigned long i, | |
318 | unsigned long start, unsigned long len); | |
319 | }; | |
320 | ||
76a73da8 | 321 | static |
83beee94 MD |
322 | void _cds_lfht_add(struct cds_lfht *ht, |
323 | unsigned long size, | |
324 | struct cds_lfht_node *node, | |
325 | struct cds_lfht_iter *unique_ret, | |
326 | int dummy); | |
48ed1c18 | 327 | |
abc490a1 MD |
328 | /* |
329 | * Algorithm to reverse bits in a word by lookup table, extended to | |
330 | * 64-bit words. | |
f9830efd | 331 | * Source: |
abc490a1 | 332 | * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable |
f9830efd | 333 | * Originally from Public Domain. |
abc490a1 MD |
334 | */ |
335 | ||
336 | static const uint8_t BitReverseTable256[256] = | |
2ed95849 | 337 | { |
abc490a1 MD |
338 | #define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64 |
339 | #define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16) | |
340 | #define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 ) | |
341 | R6(0), R6(2), R6(1), R6(3) | |
342 | }; | |
343 | #undef R2 | |
344 | #undef R4 | |
345 | #undef R6 | |
2ed95849 | 346 | |
abc490a1 MD |
347 | static |
348 | uint8_t bit_reverse_u8(uint8_t v) | |
349 | { | |
350 | return BitReverseTable256[v]; | |
351 | } | |
ab7d5fc6 | 352 | |
abc490a1 MD |
353 | static __attribute__((unused)) |
354 | uint32_t bit_reverse_u32(uint32_t v) | |
355 | { | |
356 | return ((uint32_t) bit_reverse_u8(v) << 24) | | |
357 | ((uint32_t) bit_reverse_u8(v >> 8) << 16) | | |
358 | ((uint32_t) bit_reverse_u8(v >> 16) << 8) | | |
359 | ((uint32_t) bit_reverse_u8(v >> 24)); | |
2ed95849 MD |
360 | } |
361 | ||
abc490a1 MD |
362 | static __attribute__((unused)) |
363 | uint64_t bit_reverse_u64(uint64_t v) | |
2ed95849 | 364 | { |
abc490a1 MD |
365 | return ((uint64_t) bit_reverse_u8(v) << 56) | |
366 | ((uint64_t) bit_reverse_u8(v >> 8) << 48) | | |
367 | ((uint64_t) bit_reverse_u8(v >> 16) << 40) | | |
368 | ((uint64_t) bit_reverse_u8(v >> 24) << 32) | | |
369 | ((uint64_t) bit_reverse_u8(v >> 32) << 24) | | |
370 | ((uint64_t) bit_reverse_u8(v >> 40) << 16) | | |
371 | ((uint64_t) bit_reverse_u8(v >> 48) << 8) | | |
372 | ((uint64_t) bit_reverse_u8(v >> 56)); | |
373 | } | |
374 | ||
375 | static | |
376 | unsigned long bit_reverse_ulong(unsigned long v) | |
377 | { | |
378 | #if (CAA_BITS_PER_LONG == 32) | |
379 | return bit_reverse_u32(v); | |
380 | #else | |
381 | return bit_reverse_u64(v); | |
382 | #endif | |
383 | } | |
384 | ||
f9830efd | 385 | /* |
24365af7 MD |
386 | * fls: returns the position of the most significant bit. |
387 | * Returns 0 if no bit is set, else returns the position of the most | |
388 | * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit). | |
f9830efd | 389 | */ |
24365af7 MD |
390 | #if defined(__i386) || defined(__x86_64) |
391 | static inline | |
392 | unsigned int fls_u32(uint32_t x) | |
f9830efd | 393 | { |
24365af7 MD |
394 | int r; |
395 | ||
396 | asm("bsrl %1,%0\n\t" | |
397 | "jnz 1f\n\t" | |
398 | "movl $-1,%0\n\t" | |
399 | "1:\n\t" | |
400 | : "=r" (r) : "rm" (x)); | |
401 | return r + 1; | |
402 | } | |
403 | #define HAS_FLS_U32 | |
404 | #endif | |
405 | ||
406 | #if defined(__x86_64) | |
407 | static inline | |
408 | unsigned int fls_u64(uint64_t x) | |
409 | { | |
410 | long r; | |
411 | ||
412 | asm("bsrq %1,%0\n\t" | |
413 | "jnz 1f\n\t" | |
414 | "movq $-1,%0\n\t" | |
415 | "1:\n\t" | |
416 | : "=r" (r) : "rm" (x)); | |
417 | return r + 1; | |
418 | } | |
419 | #define HAS_FLS_U64 | |
420 | #endif | |
421 | ||
422 | #ifndef HAS_FLS_U64 | |
423 | static __attribute__((unused)) | |
424 | unsigned int fls_u64(uint64_t x) | |
425 | { | |
426 | unsigned int r = 64; | |
427 | ||
428 | if (!x) | |
429 | return 0; | |
430 | ||
431 | if (!(x & 0xFFFFFFFF00000000ULL)) { | |
432 | x <<= 32; | |
433 | r -= 32; | |
434 | } | |
435 | if (!(x & 0xFFFF000000000000ULL)) { | |
436 | x <<= 16; | |
437 | r -= 16; | |
438 | } | |
439 | if (!(x & 0xFF00000000000000ULL)) { | |
440 | x <<= 8; | |
441 | r -= 8; | |
442 | } | |
443 | if (!(x & 0xF000000000000000ULL)) { | |
444 | x <<= 4; | |
445 | r -= 4; | |
446 | } | |
447 | if (!(x & 0xC000000000000000ULL)) { | |
448 | x <<= 2; | |
449 | r -= 2; | |
450 | } | |
451 | if (!(x & 0x8000000000000000ULL)) { | |
452 | x <<= 1; | |
453 | r -= 1; | |
454 | } | |
455 | return r; | |
456 | } | |
457 | #endif | |
458 | ||
459 | #ifndef HAS_FLS_U32 | |
460 | static __attribute__((unused)) | |
461 | unsigned int fls_u32(uint32_t x) | |
462 | { | |
463 | unsigned int r = 32; | |
f9830efd | 464 | |
24365af7 MD |
465 | if (!x) |
466 | return 0; | |
467 | if (!(x & 0xFFFF0000U)) { | |
468 | x <<= 16; | |
469 | r -= 16; | |
470 | } | |
471 | if (!(x & 0xFF000000U)) { | |
472 | x <<= 8; | |
473 | r -= 8; | |
474 | } | |
475 | if (!(x & 0xF0000000U)) { | |
476 | x <<= 4; | |
477 | r -= 4; | |
478 | } | |
479 | if (!(x & 0xC0000000U)) { | |
480 | x <<= 2; | |
481 | r -= 2; | |
482 | } | |
483 | if (!(x & 0x80000000U)) { | |
484 | x <<= 1; | |
485 | r -= 1; | |
486 | } | |
487 | return r; | |
488 | } | |
489 | #endif | |
490 | ||
491 | unsigned int fls_ulong(unsigned long x) | |
f9830efd | 492 | { |
6887cc5e | 493 | #if (CAA_BITS_PER_LONG == 32) |
24365af7 MD |
494 | return fls_u32(x); |
495 | #else | |
496 | return fls_u64(x); | |
497 | #endif | |
498 | } | |
f9830efd | 499 | |
920f8ef6 LJ |
500 | /* |
501 | * Return the minimum order for which x <= (1UL << order). | |
502 | * Return -1 if x is 0. | |
503 | */ | |
24365af7 MD |
504 | int get_count_order_u32(uint32_t x) |
505 | { | |
920f8ef6 LJ |
506 | if (!x) |
507 | return -1; | |
24365af7 | 508 | |
920f8ef6 | 509 | return fls_u32(x - 1); |
24365af7 MD |
510 | } |
511 | ||
920f8ef6 LJ |
512 | /* |
513 | * Return the minimum order for which x <= (1UL << order). | |
514 | * Return -1 if x is 0. | |
515 | */ | |
24365af7 MD |
516 | int get_count_order_ulong(unsigned long x) |
517 | { | |
920f8ef6 LJ |
518 | if (!x) |
519 | return -1; | |
24365af7 | 520 | |
920f8ef6 | 521 | return fls_ulong(x - 1); |
f9830efd MD |
522 | } |
523 | ||
98808fb1 | 524 | #ifdef POISON_FREE |
5afadd12 LJ |
525 | #define poison_free(ptr) \ |
526 | do { \ | |
527 | if (ptr) { \ | |
528 | memset(ptr, 0x42, sizeof(*(ptr))); \ | |
529 | free(ptr); \ | |
530 | } \ | |
98808fb1 MD |
531 | } while (0) |
532 | #else | |
533 | #define poison_free(ptr) free(ptr) | |
534 | #endif | |
535 | ||
f9830efd | 536 | static |
ab65b890 | 537 | void cds_lfht_resize_lazy_grow(struct cds_lfht *ht, unsigned long size, int growth); |
f9830efd | 538 | |
f8994aee | 539 | static |
4105056a | 540 | void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size, |
f8994aee MD |
541 | unsigned long count); |
542 | ||
df44348d | 543 | static long nr_cpus_mask = -1; |
4c42f1b8 LJ |
544 | static long split_count_mask = -1; |
545 | ||
4ddbb355 | 546 | #if defined(HAVE_SYSCONF) |
4c42f1b8 LJ |
547 | static void ht_init_nr_cpus_mask(void) |
548 | { | |
549 | long maxcpus; | |
550 | ||
551 | maxcpus = sysconf(_SC_NPROCESSORS_CONF); | |
552 | if (maxcpus <= 0) { | |
553 | nr_cpus_mask = -2; | |
554 | return; | |
555 | } | |
556 | /* | |
557 | * round up number of CPUs to next power of two, so we | |
558 | * can use & for modulo. | |
559 | */ | |
560 | maxcpus = 1UL << get_count_order_ulong(maxcpus); | |
561 | nr_cpus_mask = maxcpus - 1; | |
562 | } | |
4ddbb355 LJ |
563 | #else /* #if defined(HAVE_SYSCONF) */ |
564 | static void ht_init_nr_cpus_mask(void) | |
565 | { | |
566 | nr_cpus_mask = -2; | |
567 | } | |
568 | #endif /* #else #if defined(HAVE_SYSCONF) */ | |
df44348d MD |
569 | |
570 | static | |
5afadd12 | 571 | void alloc_split_items_count(struct cds_lfht *ht) |
df44348d MD |
572 | { |
573 | struct ht_items_count *count; | |
574 | ||
4c42f1b8 LJ |
575 | if (nr_cpus_mask == -1) { |
576 | ht_init_nr_cpus_mask(); | |
4ddbb355 LJ |
577 | if (nr_cpus_mask < 0) |
578 | split_count_mask = DEFAULT_SPLIT_COUNT_MASK; | |
579 | else | |
580 | split_count_mask = nr_cpus_mask; | |
df44348d | 581 | } |
4c42f1b8 | 582 | |
4ddbb355 | 583 | assert(split_count_mask >= 0); |
5afadd12 LJ |
584 | |
585 | if (ht->flags & CDS_LFHT_ACCOUNTING) { | |
586 | ht->split_count = calloc(split_count_mask + 1, sizeof(*count)); | |
587 | assert(ht->split_count); | |
588 | } else { | |
589 | ht->split_count = NULL; | |
590 | } | |
df44348d MD |
591 | } |
592 | ||
593 | static | |
5afadd12 | 594 | void free_split_items_count(struct cds_lfht *ht) |
df44348d | 595 | { |
5afadd12 | 596 | poison_free(ht->split_count); |
df44348d MD |
597 | } |
598 | ||
14360f1c | 599 | #if defined(HAVE_SCHED_GETCPU) |
df44348d | 600 | static |
14360f1c | 601 | int ht_get_split_count_index(unsigned long hash) |
df44348d MD |
602 | { |
603 | int cpu; | |
604 | ||
4c42f1b8 | 605 | assert(split_count_mask >= 0); |
df44348d | 606 | cpu = sched_getcpu(); |
8ed51e04 | 607 | if (caa_unlikely(cpu < 0)) |
14360f1c | 608 | return hash & split_count_mask; |
df44348d | 609 | else |
4c42f1b8 | 610 | return cpu & split_count_mask; |
df44348d | 611 | } |
14360f1c LJ |
612 | #else /* #if defined(HAVE_SCHED_GETCPU) */ |
613 | static | |
614 | int ht_get_split_count_index(unsigned long hash) | |
615 | { | |
616 | return hash & split_count_mask; | |
617 | } | |
618 | #endif /* #else #if defined(HAVE_SCHED_GETCPU) */ | |
df44348d MD |
619 | |
620 | static | |
14360f1c | 621 | void ht_count_add(struct cds_lfht *ht, unsigned long size, unsigned long hash) |
df44348d | 622 | { |
4c42f1b8 LJ |
623 | unsigned long split_count; |
624 | int index; | |
df44348d | 625 | |
8ed51e04 | 626 | if (caa_unlikely(!ht->split_count)) |
3171717f | 627 | return; |
14360f1c | 628 | index = ht_get_split_count_index(hash); |
4c42f1b8 | 629 | split_count = uatomic_add_return(&ht->split_count[index].add, 1); |
8ed51e04 | 630 | if (caa_unlikely(!(split_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) { |
e3ecefd6 | 631 | long count; |
df44348d | 632 | |
4c42f1b8 | 633 | dbg_printf("add split count %lu\n", split_count); |
df44348d MD |
634 | count = uatomic_add_return(&ht->count, |
635 | 1UL << COUNT_COMMIT_ORDER); | |
636 | /* If power of 2 */ | |
637 | if (!(count & (count - 1))) { | |
4105056a | 638 | if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) < size) |
f8994aee | 639 | return; |
e3ecefd6 | 640 | dbg_printf("add set global %ld\n", count); |
4105056a | 641 | cds_lfht_resize_lazy_count(ht, size, |
6ea6bc67 | 642 | count >> (CHAIN_LEN_TARGET - 1)); |
df44348d MD |
643 | } |
644 | } | |
645 | } | |
646 | ||
647 | static | |
14360f1c | 648 | void ht_count_del(struct cds_lfht *ht, unsigned long size, unsigned long hash) |
df44348d | 649 | { |
4c42f1b8 LJ |
650 | unsigned long split_count; |
651 | int index; | |
df44348d | 652 | |
8ed51e04 | 653 | if (caa_unlikely(!ht->split_count)) |
3171717f | 654 | return; |
14360f1c | 655 | index = ht_get_split_count_index(hash); |
4c42f1b8 | 656 | split_count = uatomic_add_return(&ht->split_count[index].del, 1); |
8ed51e04 | 657 | if (caa_unlikely(!(split_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) { |
e3ecefd6 | 658 | long count; |
df44348d | 659 | |
4c42f1b8 | 660 | dbg_printf("del split count %lu\n", split_count); |
df44348d | 661 | count = uatomic_add_return(&ht->count, |
3171717f | 662 | -(1UL << COUNT_COMMIT_ORDER)); |
df44348d MD |
663 | /* If power of 2 */ |
664 | if (!(count & (count - 1))) { | |
4105056a | 665 | if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) >= size) |
f8994aee | 666 | return; |
e3ecefd6 MD |
667 | dbg_printf("del set global %ld\n", count); |
668 | /* | |
c941bb9e | 669 | * Don't shrink table if the number of nodes is below a |
e3ecefd6 MD |
670 | * certain threshold. |
671 | */ | |
4c42f1b8 | 672 | if (count < (1UL << COUNT_COMMIT_ORDER) * (split_count_mask + 1)) |
e3ecefd6 | 673 | return; |
4105056a | 674 | cds_lfht_resize_lazy_count(ht, size, |
6ea6bc67 | 675 | count >> (CHAIN_LEN_TARGET - 1)); |
df44348d MD |
676 | } |
677 | } | |
678 | } | |
679 | ||
f9830efd | 680 | static |
4105056a | 681 | void check_resize(struct cds_lfht *ht, unsigned long size, uint32_t chain_len) |
f9830efd | 682 | { |
f8994aee MD |
683 | unsigned long count; |
684 | ||
b8af5011 MD |
685 | if (!(ht->flags & CDS_LFHT_AUTO_RESIZE)) |
686 | return; | |
f8994aee MD |
687 | count = uatomic_read(&ht->count); |
688 | /* | |
689 | * Use bucket-local length for small table expand and for | |
690 | * environments lacking per-cpu data support. | |
691 | */ | |
692 | if (count >= (1UL << COUNT_COMMIT_ORDER)) | |
693 | return; | |
24365af7 | 694 | if (chain_len > 100) |
f0c29ed7 | 695 | dbg_printf("WARNING: large chain length: %u.\n", |
24365af7 | 696 | chain_len); |
3390d470 | 697 | if (chain_len >= CHAIN_LEN_RESIZE_THRESHOLD) |
ab65b890 | 698 | cds_lfht_resize_lazy_grow(ht, size, |
01370f0b | 699 | get_count_order_u32(chain_len - (CHAIN_LEN_TARGET - 1))); |
f9830efd MD |
700 | } |
701 | ||
abc490a1 | 702 | static |
14044b37 | 703 | struct cds_lfht_node *clear_flag(struct cds_lfht_node *node) |
abc490a1 | 704 | { |
14044b37 | 705 | return (struct cds_lfht_node *) (((unsigned long) node) & ~FLAGS_MASK); |
abc490a1 MD |
706 | } |
707 | ||
708 | static | |
14044b37 | 709 | int is_removed(struct cds_lfht_node *node) |
abc490a1 | 710 | { |
d37166c6 | 711 | return ((unsigned long) node) & REMOVED_FLAG; |
abc490a1 MD |
712 | } |
713 | ||
714 | static | |
14044b37 | 715 | struct cds_lfht_node *flag_removed(struct cds_lfht_node *node) |
abc490a1 | 716 | { |
14044b37 | 717 | return (struct cds_lfht_node *) (((unsigned long) node) | REMOVED_FLAG); |
abc490a1 MD |
718 | } |
719 | ||
f5596c94 | 720 | static |
14044b37 | 721 | int is_dummy(struct cds_lfht_node *node) |
f5596c94 MD |
722 | { |
723 | return ((unsigned long) node) & DUMMY_FLAG; | |
724 | } | |
725 | ||
726 | static | |
14044b37 | 727 | struct cds_lfht_node *flag_dummy(struct cds_lfht_node *node) |
f5596c94 | 728 | { |
14044b37 | 729 | return (struct cds_lfht_node *) (((unsigned long) node) | DUMMY_FLAG); |
f5596c94 | 730 | } |
bb7b2f26 MD |
731 | |
732 | static | |
733 | struct cds_lfht_node *get_end(void) | |
734 | { | |
735 | return (struct cds_lfht_node *) END_VALUE; | |
736 | } | |
737 | ||
738 | static | |
739 | int is_end(struct cds_lfht_node *node) | |
740 | { | |
741 | return clear_flag(node) == (struct cds_lfht_node *) END_VALUE; | |
742 | } | |
743 | ||
abc490a1 | 744 | static |
ab65b890 LJ |
745 | unsigned long _uatomic_xchg_monotonic_increase(unsigned long *ptr, |
746 | unsigned long v) | |
abc490a1 MD |
747 | { |
748 | unsigned long old1, old2; | |
749 | ||
750 | old1 = uatomic_read(ptr); | |
751 | do { | |
752 | old2 = old1; | |
753 | if (old2 >= v) | |
f9830efd | 754 | return old2; |
abc490a1 | 755 | } while ((old1 = uatomic_cmpxchg(ptr, old2, v)) != old2); |
ab65b890 | 756 | return old2; |
abc490a1 MD |
757 | } |
758 | ||
f4a9cc0b LJ |
759 | static |
760 | struct _cds_lfht_node *lookup_bucket(struct cds_lfht *ht, unsigned long size, | |
761 | unsigned long hash) | |
762 | { | |
763 | unsigned long index, order; | |
764 | ||
765 | assert(size > 0); | |
766 | index = hash & (size - 1); | |
ef6e6171 LJ |
767 | |
768 | if (index < ht->min_alloc_size) { | |
769 | dbg_printf("lookup hash %lu index %lu order 0 aridx 0\n", | |
770 | hash, index); | |
771 | return &ht->t.tbl[0]->nodes[index]; | |
772 | } | |
a4ea2223 LJ |
773 | /* |
774 | * equivalent to get_count_order_ulong(index + 1), but optimizes | |
775 | * away the non-existing 0 special-case for | |
776 | * get_count_order_ulong. | |
777 | */ | |
778 | order = fls_ulong(index); | |
f4a9cc0b | 779 | dbg_printf("lookup hash %lu index %lu order %lu aridx %lu\n", |
ef6e6171 LJ |
780 | hash, index, order, index & ((1UL << (order - 1)) - 1)); |
781 | return &ht->t.tbl[order]->nodes[index & ((1UL << (order - 1)) - 1)]; | |
f4a9cc0b LJ |
782 | } |
783 | ||
273399de MD |
784 | /* |
785 | * Remove all logically deleted nodes from a bucket up to a certain node key. | |
786 | */ | |
787 | static | |
f9c80341 | 788 | void _cds_lfht_gc_bucket(struct cds_lfht_node *dummy, struct cds_lfht_node *node) |
273399de | 789 | { |
14044b37 | 790 | struct cds_lfht_node *iter_prev, *iter, *next, *new_next; |
273399de | 791 | |
c90201ac MD |
792 | assert(!is_dummy(dummy)); |
793 | assert(!is_removed(dummy)); | |
794 | assert(!is_dummy(node)); | |
795 | assert(!is_removed(node)); | |
273399de MD |
796 | for (;;) { |
797 | iter_prev = dummy; | |
798 | /* We can always skip the dummy node initially */ | |
cc4fcb10 | 799 | iter = rcu_dereference(iter_prev->p.next); |
b4cb483f | 800 | assert(!is_removed(iter)); |
cc4fcb10 | 801 | assert(iter_prev->p.reverse_hash <= node->p.reverse_hash); |
bd4db153 MD |
802 | /* |
803 | * We should never be called with dummy (start of chain) | |
804 | * and logically removed node (end of path compression | |
805 | * marker) being the actual same node. This would be a | |
806 | * bug in the algorithm implementation. | |
807 | */ | |
808 | assert(dummy != node); | |
273399de | 809 | for (;;) { |
8ed51e04 | 810 | if (caa_unlikely(is_end(iter))) |
f9c80341 | 811 | return; |
8ed51e04 | 812 | if (caa_likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash)) |
f9c80341 | 813 | return; |
cc4fcb10 | 814 | next = rcu_dereference(clear_flag(iter)->p.next); |
8ed51e04 | 815 | if (caa_likely(is_removed(next))) |
273399de | 816 | break; |
b453eae1 | 817 | iter_prev = clear_flag(iter); |
273399de MD |
818 | iter = next; |
819 | } | |
b198f0fd | 820 | assert(!is_removed(iter)); |
f5596c94 MD |
821 | if (is_dummy(iter)) |
822 | new_next = flag_dummy(clear_flag(next)); | |
823 | else | |
824 | new_next = clear_flag(next); | |
825 | (void) uatomic_cmpxchg(&iter_prev->p.next, iter, new_next); | |
273399de | 826 | } |
f9c80341 | 827 | return; |
273399de MD |
828 | } |
829 | ||
9357c415 MD |
830 | static |
831 | int _cds_lfht_replace(struct cds_lfht *ht, unsigned long size, | |
832 | struct cds_lfht_node *old_node, | |
3fb86f26 | 833 | struct cds_lfht_node *old_next, |
9357c415 MD |
834 | struct cds_lfht_node *new_node) |
835 | { | |
3fb86f26 | 836 | struct cds_lfht_node *dummy, *ret_next; |
9357c415 | 837 | struct _cds_lfht_node *lookup; |
9357c415 MD |
838 | |
839 | if (!old_node) /* Return -ENOENT if asked to replace NULL node */ | |
7801dadd | 840 | return -ENOENT; |
9357c415 MD |
841 | |
842 | assert(!is_removed(old_node)); | |
843 | assert(!is_dummy(old_node)); | |
844 | assert(!is_removed(new_node)); | |
845 | assert(!is_dummy(new_node)); | |
846 | assert(new_node != old_node); | |
3fb86f26 | 847 | for (;;) { |
9357c415 | 848 | /* Insert after node to be replaced */ |
9357c415 MD |
849 | if (is_removed(old_next)) { |
850 | /* | |
851 | * Too late, the old node has been removed under us | |
852 | * between lookup and replace. Fail. | |
853 | */ | |
7801dadd | 854 | return -ENOENT; |
9357c415 MD |
855 | } |
856 | assert(!is_dummy(old_next)); | |
857 | assert(new_node != clear_flag(old_next)); | |
858 | new_node->p.next = clear_flag(old_next); | |
859 | /* | |
860 | * Here is the whole trick for lock-free replace: we add | |
861 | * the replacement node _after_ the node we want to | |
862 | * replace by atomically setting its next pointer at the | |
863 | * same time we set its removal flag. Given that | |
864 | * the lookups/get next use an iterator aware of the | |
865 | * next pointer, they will either skip the old node due | |
866 | * to the removal flag and see the new node, or use | |
867 | * the old node, but will not see the new one. | |
868 | */ | |
869 | ret_next = uatomic_cmpxchg(&old_node->p.next, | |
870 | old_next, flag_removed(new_node)); | |
3fb86f26 | 871 | if (ret_next == old_next) |
7801dadd | 872 | break; /* We performed the replacement. */ |
3fb86f26 LJ |
873 | old_next = ret_next; |
874 | } | |
9357c415 | 875 | |
9357c415 MD |
876 | /* |
877 | * Ensure that the old node is not visible to readers anymore: | |
878 | * lookup for the node, and remove it (along with any other | |
879 | * logically removed node) if found. | |
880 | */ | |
f4a9cc0b | 881 | lookup = lookup_bucket(ht, size, bit_reverse_ulong(old_node->p.reverse_hash)); |
9357c415 MD |
882 | dummy = (struct cds_lfht_node *) lookup; |
883 | _cds_lfht_gc_bucket(dummy, new_node); | |
7801dadd LJ |
884 | |
885 | assert(is_removed(rcu_dereference(old_node->p.next))); | |
886 | return 0; | |
9357c415 MD |
887 | } |
888 | ||
83beee94 MD |
889 | /* |
890 | * A non-NULL unique_ret pointer uses the "add unique" (or uniquify) add | |
891 | * mode. A NULL unique_ret allows creation of duplicate keys. | |
892 | */ | |
abc490a1 | 893 | static |
83beee94 MD |
894 | void _cds_lfht_add(struct cds_lfht *ht, |
895 | unsigned long size, | |
896 | struct cds_lfht_node *node, | |
897 | struct cds_lfht_iter *unique_ret, | |
898 | int dummy) | |
abc490a1 | 899 | { |
14044b37 | 900 | struct cds_lfht_node *iter_prev, *iter, *next, *new_node, *new_next, |
960c9e4f | 901 | *return_node; |
14044b37 | 902 | struct _cds_lfht_node *lookup; |
abc490a1 | 903 | |
c90201ac MD |
904 | assert(!is_dummy(node)); |
905 | assert(!is_removed(node)); | |
f4a9cc0b | 906 | lookup = lookup_bucket(ht, size, bit_reverse_ulong(node->p.reverse_hash)); |
abc490a1 | 907 | for (;;) { |
adc0de68 | 908 | uint32_t chain_len = 0; |
abc490a1 | 909 | |
11519af6 MD |
910 | /* |
911 | * iter_prev points to the non-removed node prior to the | |
912 | * insert location. | |
11519af6 | 913 | */ |
14044b37 | 914 | iter_prev = (struct cds_lfht_node *) lookup; |
11519af6 | 915 | /* We can always skip the dummy node initially */ |
cc4fcb10 MD |
916 | iter = rcu_dereference(iter_prev->p.next); |
917 | assert(iter_prev->p.reverse_hash <= node->p.reverse_hash); | |
abc490a1 | 918 | for (;;) { |
8ed51e04 | 919 | if (caa_unlikely(is_end(iter))) |
273399de | 920 | goto insert; |
8ed51e04 | 921 | if (caa_likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash)) |
273399de | 922 | goto insert; |
238cc06e | 923 | |
194fdbd1 LJ |
924 | /* dummy node is the first node of the identical-hash-value chain */ |
925 | if (dummy && clear_flag(iter)->p.reverse_hash == node->p.reverse_hash) | |
926 | goto insert; | |
238cc06e | 927 | |
cc4fcb10 | 928 | next = rcu_dereference(clear_flag(iter)->p.next); |
8ed51e04 | 929 | if (caa_unlikely(is_removed(next))) |
9dba85be | 930 | goto gc_node; |
238cc06e LJ |
931 | |
932 | /* uniquely add */ | |
83beee94 | 933 | if (unique_ret |
1b81fe1a | 934 | && !is_dummy(next) |
238cc06e LJ |
935 | && clear_flag(iter)->p.reverse_hash == node->p.reverse_hash) { |
936 | struct cds_lfht_iter d_iter = { .node = node, .next = iter, }; | |
937 | ||
938 | /* | |
939 | * uniquely adding inserts the node as the first | |
940 | * node of the identical-hash-value node chain. | |
941 | * | |
942 | * This semantic ensures no duplicated keys | |
943 | * should ever be observable in the table | |
944 | * (including observe one node by one node | |
945 | * by forward iterations) | |
946 | */ | |
947 | cds_lfht_next_duplicate(ht, &d_iter); | |
948 | if (!d_iter.node) | |
949 | goto insert; | |
950 | ||
951 | *unique_ret = d_iter; | |
83beee94 | 952 | return; |
48ed1c18 | 953 | } |
238cc06e | 954 | |
11519af6 | 955 | /* Only account for identical reverse hash once */ |
24365af7 MD |
956 | if (iter_prev->p.reverse_hash != clear_flag(iter)->p.reverse_hash |
957 | && !is_dummy(next)) | |
4105056a | 958 | check_resize(ht, size, ++chain_len); |
11519af6 | 959 | iter_prev = clear_flag(iter); |
273399de | 960 | iter = next; |
abc490a1 | 961 | } |
48ed1c18 | 962 | |
273399de | 963 | insert: |
7ec59d3b | 964 | assert(node != clear_flag(iter)); |
11519af6 | 965 | assert(!is_removed(iter_prev)); |
c90201ac | 966 | assert(!is_removed(iter)); |
f000907d | 967 | assert(iter_prev != node); |
f9c80341 | 968 | if (!dummy) |
1b81fe1a | 969 | node->p.next = clear_flag(iter); |
f9c80341 MD |
970 | else |
971 | node->p.next = flag_dummy(clear_flag(iter)); | |
f5596c94 MD |
972 | if (is_dummy(iter)) |
973 | new_node = flag_dummy(node); | |
974 | else | |
975 | new_node = node; | |
cc4fcb10 | 976 | if (uatomic_cmpxchg(&iter_prev->p.next, iter, |
48ed1c18 | 977 | new_node) != iter) { |
273399de | 978 | continue; /* retry */ |
48ed1c18 | 979 | } else { |
83beee94 | 980 | return_node = node; |
960c9e4f | 981 | goto end; |
48ed1c18 MD |
982 | } |
983 | ||
9dba85be MD |
984 | gc_node: |
985 | assert(!is_removed(iter)); | |
f5596c94 MD |
986 | if (is_dummy(iter)) |
987 | new_next = flag_dummy(clear_flag(next)); | |
988 | else | |
989 | new_next = clear_flag(next); | |
990 | (void) uatomic_cmpxchg(&iter_prev->p.next, iter, new_next); | |
273399de | 991 | /* retry */ |
464a1ec9 | 992 | } |
9357c415 | 993 | end: |
83beee94 MD |
994 | if (unique_ret) { |
995 | unique_ret->node = return_node; | |
996 | /* unique_ret->next left unset, never used. */ | |
997 | } | |
abc490a1 | 998 | } |
464a1ec9 | 999 | |
abc490a1 | 1000 | static |
860d07e8 | 1001 | int _cds_lfht_del(struct cds_lfht *ht, unsigned long size, |
4105056a | 1002 | struct cds_lfht_node *node, |
b198f0fd | 1003 | int dummy_removal) |
abc490a1 | 1004 | { |
14044b37 MD |
1005 | struct cds_lfht_node *dummy, *next, *old; |
1006 | struct _cds_lfht_node *lookup; | |
5e28c532 | 1007 | |
9357c415 | 1008 | if (!node) /* Return -ENOENT if asked to delete NULL node */ |
743f9143 | 1009 | return -ENOENT; |
9357c415 | 1010 | |
7ec59d3b | 1011 | /* logically delete the node */ |
c90201ac MD |
1012 | assert(!is_dummy(node)); |
1013 | assert(!is_removed(node)); | |
cc4fcb10 | 1014 | old = rcu_dereference(node->p.next); |
7ec59d3b | 1015 | do { |
48ed1c18 MD |
1016 | struct cds_lfht_node *new_next; |
1017 | ||
7ec59d3b | 1018 | next = old; |
8ed51e04 | 1019 | if (caa_unlikely(is_removed(next))) |
743f9143 | 1020 | return -ENOENT; |
1475579c MD |
1021 | if (dummy_removal) |
1022 | assert(is_dummy(next)); | |
1023 | else | |
1024 | assert(!is_dummy(next)); | |
48ed1c18 | 1025 | new_next = flag_removed(next); |
48ed1c18 | 1026 | old = uatomic_cmpxchg(&node->p.next, next, new_next); |
7ec59d3b | 1027 | } while (old != next); |
7ec59d3b | 1028 | /* We performed the (logical) deletion. */ |
7ec59d3b MD |
1029 | |
1030 | /* | |
1031 | * Ensure that the node is not visible to readers anymore: lookup for | |
273399de MD |
1032 | * the node, and remove it (along with any other logically removed node) |
1033 | * if found. | |
11519af6 | 1034 | */ |
f4a9cc0b | 1035 | lookup = lookup_bucket(ht, size, bit_reverse_ulong(node->p.reverse_hash)); |
14044b37 | 1036 | dummy = (struct cds_lfht_node *) lookup; |
f9c80341 | 1037 | _cds_lfht_gc_bucket(dummy, node); |
743f9143 LJ |
1038 | |
1039 | assert(is_removed(rcu_dereference(node->p.next))); | |
1040 | return 0; | |
abc490a1 | 1041 | } |
2ed95849 | 1042 | |
b7d619b0 MD |
1043 | static |
1044 | void *partition_resize_thread(void *arg) | |
1045 | { | |
1046 | struct partition_resize_work *work = arg; | |
1047 | ||
1048 | work->ht->cds_lfht_rcu_register_thread(); | |
1049 | work->fct(work->ht, work->i, work->start, work->len); | |
1050 | work->ht->cds_lfht_rcu_unregister_thread(); | |
1051 | return NULL; | |
1052 | } | |
1053 | ||
1054 | static | |
1055 | void partition_resize_helper(struct cds_lfht *ht, unsigned long i, | |
1056 | unsigned long len, | |
1057 | void (*fct)(struct cds_lfht *ht, unsigned long i, | |
1058 | unsigned long start, unsigned long len)) | |
1059 | { | |
1060 | unsigned long partition_len; | |
1061 | struct partition_resize_work *work; | |
6083a889 MD |
1062 | int thread, ret; |
1063 | unsigned long nr_threads; | |
b7d619b0 | 1064 | |
6083a889 MD |
1065 | /* |
1066 | * Note: nr_cpus_mask + 1 is always power of 2. | |
1067 | * We spawn just the number of threads we need to satisfy the minimum | |
1068 | * partition size, up to the number of CPUs in the system. | |
1069 | */ | |
91452a6a MD |
1070 | if (nr_cpus_mask > 0) { |
1071 | nr_threads = min(nr_cpus_mask + 1, | |
1072 | len >> MIN_PARTITION_PER_THREAD_ORDER); | |
1073 | } else { | |
1074 | nr_threads = 1; | |
1075 | } | |
6083a889 MD |
1076 | partition_len = len >> get_count_order_ulong(nr_threads); |
1077 | work = calloc(nr_threads, sizeof(*work)); | |
b7d619b0 | 1078 | assert(work); |
6083a889 MD |
1079 | for (thread = 0; thread < nr_threads; thread++) { |
1080 | work[thread].ht = ht; | |
1081 | work[thread].i = i; | |
1082 | work[thread].len = partition_len; | |
1083 | work[thread].start = thread * partition_len; | |
1084 | work[thread].fct = fct; | |
1af6e26e | 1085 | ret = pthread_create(&(work[thread].thread_id), ht->resize_attr, |
6083a889 | 1086 | partition_resize_thread, &work[thread]); |
b7d619b0 MD |
1087 | assert(!ret); |
1088 | } | |
6083a889 | 1089 | for (thread = 0; thread < nr_threads; thread++) { |
1af6e26e | 1090 | ret = pthread_join(work[thread].thread_id, NULL); |
b7d619b0 MD |
1091 | assert(!ret); |
1092 | } | |
1093 | free(work); | |
b7d619b0 MD |
1094 | } |
1095 | ||
e8de508e MD |
1096 | /* |
1097 | * Holding RCU read lock to protect _cds_lfht_add against memory | |
1098 | * reclaim that could be performed by other call_rcu worker threads (ABA | |
1099 | * problem). | |
9ee0fc9a | 1100 | * |
b7d619b0 | 1101 | * When we reach a certain length, we can split this population phase over |
9ee0fc9a MD |
1102 | * many worker threads, based on the number of CPUs available in the system. |
1103 | * This should therefore take care of not having the expand lagging behind too | |
1104 | * many concurrent insertion threads by using the scheduler's ability to | |
1105 | * schedule dummy node population fairly with insertions. | |
e8de508e | 1106 | */ |
4105056a | 1107 | static |
b7d619b0 MD |
1108 | void init_table_populate_partition(struct cds_lfht *ht, unsigned long i, |
1109 | unsigned long start, unsigned long len) | |
4105056a MD |
1110 | { |
1111 | unsigned long j; | |
1112 | ||
5488222b | 1113 | assert(i > ht->min_alloc_order); |
4105056a | 1114 | ht->cds_lfht_rcu_read_lock(); |
b7d619b0 | 1115 | for (j = start; j < start + len; j++) { |
4105056a MD |
1116 | struct cds_lfht_node *new_node = |
1117 | (struct cds_lfht_node *) &ht->t.tbl[i]->nodes[j]; | |
1118 | ||
dc1da8f6 | 1119 | dbg_printf("init populate: i %lu j %lu hash %lu\n", |
4f6e90b7 | 1120 | i, j, (1UL << (i - 1)) + j); |
dc1da8f6 | 1121 | new_node->p.reverse_hash = |
4f6e90b7 LJ |
1122 | bit_reverse_ulong((1UL << (i - 1)) + j); |
1123 | _cds_lfht_add(ht, 1UL << (i - 1), | |
83beee94 | 1124 | new_node, NULL, 1); |
4105056a MD |
1125 | } |
1126 | ht->cds_lfht_rcu_read_unlock(); | |
b7d619b0 MD |
1127 | } |
1128 | ||
1129 | static | |
1130 | void init_table_populate(struct cds_lfht *ht, unsigned long i, | |
1131 | unsigned long len) | |
1132 | { | |
1133 | assert(nr_cpus_mask != -1); | |
6083a889 | 1134 | if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD) { |
b7d619b0 MD |
1135 | ht->cds_lfht_rcu_thread_online(); |
1136 | init_table_populate_partition(ht, i, 0, len); | |
1137 | ht->cds_lfht_rcu_thread_offline(); | |
1138 | return; | |
1139 | } | |
1140 | partition_resize_helper(ht, i, len, init_table_populate_partition); | |
4105056a MD |
1141 | } |
1142 | ||
abc490a1 | 1143 | static |
4105056a | 1144 | void init_table(struct cds_lfht *ht, |
93d46c39 | 1145 | unsigned long first_order, unsigned long last_order) |
24365af7 | 1146 | { |
93d46c39 | 1147 | unsigned long i; |
24365af7 | 1148 | |
93d46c39 LJ |
1149 | dbg_printf("init table: first_order %lu last_order %lu\n", |
1150 | first_order, last_order); | |
5488222b | 1151 | assert(first_order > ht->min_alloc_order); |
93d46c39 | 1152 | for (i = first_order; i <= last_order; i++) { |
4105056a | 1153 | unsigned long len; |
24365af7 | 1154 | |
4f6e90b7 | 1155 | len = 1UL << (i - 1); |
f0c29ed7 | 1156 | dbg_printf("init order %lu len: %lu\n", i, len); |
4d676753 MD |
1157 | |
1158 | /* Stop expand if the resize target changes under us */ | |
4f6e90b7 | 1159 | if (CMM_LOAD_SHARED(ht->t.resize_target) < (1UL << i)) |
4d676753 MD |
1160 | break; |
1161 | ||
0d14ceb2 | 1162 | ht->t.tbl[i] = calloc(1, len * sizeof(struct _cds_lfht_node)); |
b7d619b0 | 1163 | assert(ht->t.tbl[i]); |
4105056a | 1164 | |
4105056a | 1165 | /* |
dc1da8f6 MD |
1166 | * Set all dummy nodes reverse hash values for a level and |
1167 | * link all dummy nodes into the table. | |
4105056a | 1168 | */ |
dc1da8f6 | 1169 | init_table_populate(ht, i, len); |
4105056a | 1170 | |
f9c80341 MD |
1171 | /* |
1172 | * Update table size. | |
1173 | */ | |
1174 | cmm_smp_wmb(); /* populate data before RCU size */ | |
4f6e90b7 | 1175 | CMM_STORE_SHARED(ht->t.size, 1UL << i); |
f9c80341 | 1176 | |
4f6e90b7 | 1177 | dbg_printf("init new size: %lu\n", 1UL << i); |
4105056a MD |
1178 | if (CMM_LOAD_SHARED(ht->in_progress_destroy)) |
1179 | break; | |
1180 | } | |
1181 | } | |
1182 | ||
e8de508e MD |
1183 | /* |
1184 | * Holding RCU read lock to protect _cds_lfht_remove against memory | |
1185 | * reclaim that could be performed by other call_rcu worker threads (ABA | |
1186 | * problem). | |
1187 | * For a single level, we logically remove and garbage collect each node. | |
1188 | * | |
1189 | * As a design choice, we perform logical removal and garbage collection on a | |
1190 | * node-per-node basis to simplify this algorithm. We also assume keeping good | |
1191 | * cache locality of the operation would overweight possible performance gain | |
1192 | * that could be achieved by batching garbage collection for multiple levels. | |
1193 | * However, this would have to be justified by benchmarks. | |
1194 | * | |
1195 | * Concurrent removal and add operations are helping us perform garbage | |
1196 | * collection of logically removed nodes. We guarantee that all logically | |
1197 | * removed nodes have been garbage-collected (unlinked) before call_rcu is | |
1198 | * invoked to free a hole level of dummy nodes (after a grace period). | |
1199 | * | |
1200 | * Logical removal and garbage collection can therefore be done in batch or on a | |
1201 | * node-per-node basis, as long as the guarantee above holds. | |
9ee0fc9a | 1202 | * |
b7d619b0 MD |
1203 | * When we reach a certain length, we can split this removal over many worker |
1204 | * threads, based on the number of CPUs available in the system. This should | |
1205 | * take care of not letting resize process lag behind too many concurrent | |
9ee0fc9a | 1206 | * updater threads actively inserting into the hash table. |
e8de508e | 1207 | */ |
4105056a | 1208 | static |
b7d619b0 MD |
1209 | void remove_table_partition(struct cds_lfht *ht, unsigned long i, |
1210 | unsigned long start, unsigned long len) | |
4105056a MD |
1211 | { |
1212 | unsigned long j; | |
1213 | ||
5488222b | 1214 | assert(i > ht->min_alloc_order); |
4105056a | 1215 | ht->cds_lfht_rcu_read_lock(); |
b7d619b0 | 1216 | for (j = start; j < start + len; j++) { |
4105056a MD |
1217 | struct cds_lfht_node *fini_node = |
1218 | (struct cds_lfht_node *) &ht->t.tbl[i]->nodes[j]; | |
1219 | ||
1220 | dbg_printf("remove entry: i %lu j %lu hash %lu\n", | |
4f6e90b7 | 1221 | i, j, (1UL << (i - 1)) + j); |
4105056a | 1222 | fini_node->p.reverse_hash = |
4f6e90b7 LJ |
1223 | bit_reverse_ulong((1UL << (i - 1)) + j); |
1224 | (void) _cds_lfht_del(ht, 1UL << (i - 1), fini_node, 1); | |
abc490a1 | 1225 | } |
4105056a | 1226 | ht->cds_lfht_rcu_read_unlock(); |
b7d619b0 MD |
1227 | } |
1228 | ||
1229 | static | |
1230 | void remove_table(struct cds_lfht *ht, unsigned long i, unsigned long len) | |
1231 | { | |
1232 | ||
1233 | assert(nr_cpus_mask != -1); | |
6083a889 | 1234 | if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD) { |
b7d619b0 MD |
1235 | ht->cds_lfht_rcu_thread_online(); |
1236 | remove_table_partition(ht, i, 0, len); | |
1237 | ht->cds_lfht_rcu_thread_offline(); | |
1238 | return; | |
1239 | } | |
1240 | partition_resize_helper(ht, i, len, remove_table_partition); | |
2ed95849 MD |
1241 | } |
1242 | ||
1475579c | 1243 | static |
4105056a | 1244 | void fini_table(struct cds_lfht *ht, |
93d46c39 | 1245 | unsigned long first_order, unsigned long last_order) |
1475579c | 1246 | { |
93d46c39 | 1247 | long i; |
0d14ceb2 | 1248 | void *free_by_rcu = NULL; |
1475579c | 1249 | |
93d46c39 LJ |
1250 | dbg_printf("fini table: first_order %lu last_order %lu\n", |
1251 | first_order, last_order); | |
5488222b | 1252 | assert(first_order > ht->min_alloc_order); |
93d46c39 | 1253 | for (i = last_order; i >= first_order; i--) { |
4105056a | 1254 | unsigned long len; |
1475579c | 1255 | |
4f6e90b7 | 1256 | len = 1UL << (i - 1); |
1475579c | 1257 | dbg_printf("fini order %lu len: %lu\n", i, len); |
4105056a | 1258 | |
4d676753 MD |
1259 | /* Stop shrink if the resize target changes under us */ |
1260 | if (CMM_LOAD_SHARED(ht->t.resize_target) > (1UL << (i - 1))) | |
1261 | break; | |
1262 | ||
1263 | cmm_smp_wmb(); /* populate data before RCU size */ | |
1264 | CMM_STORE_SHARED(ht->t.size, 1UL << (i - 1)); | |
1265 | ||
1266 | /* | |
1267 | * We need to wait for all add operations to reach Q.S. (and | |
1268 | * thus use the new table for lookups) before we can start | |
1269 | * releasing the old dummy nodes. Otherwise their lookup will | |
1270 | * return a logically removed node as insert position. | |
1271 | */ | |
1272 | ht->cds_lfht_synchronize_rcu(); | |
0d14ceb2 LJ |
1273 | if (free_by_rcu) |
1274 | free(free_by_rcu); | |
4d676753 | 1275 | |
21263e21 | 1276 | /* |
4105056a MD |
1277 | * Set "removed" flag in dummy nodes about to be removed. |
1278 | * Unlink all now-logically-removed dummy node pointers. | |
1279 | * Concurrent add/remove operation are helping us doing | |
1280 | * the gc. | |
21263e21 | 1281 | */ |
4105056a MD |
1282 | remove_table(ht, i, len); |
1283 | ||
0d14ceb2 | 1284 | free_by_rcu = ht->t.tbl[i]; |
4105056a MD |
1285 | |
1286 | dbg_printf("fini new size: %lu\n", 1UL << i); | |
1475579c MD |
1287 | if (CMM_LOAD_SHARED(ht->in_progress_destroy)) |
1288 | break; | |
1289 | } | |
0d14ceb2 LJ |
1290 | |
1291 | if (free_by_rcu) { | |
1292 | ht->cds_lfht_synchronize_rcu(); | |
1293 | free(free_by_rcu); | |
1294 | } | |
1475579c MD |
1295 | } |
1296 | ||
ff0d69de LJ |
1297 | static |
1298 | void cds_lfht_create_dummy(struct cds_lfht *ht, unsigned long size) | |
1299 | { | |
1300 | struct _cds_lfht_node *prev, *node; | |
1301 | unsigned long order, len, i, j; | |
1302 | ||
5488222b | 1303 | ht->t.tbl[0] = calloc(1, ht->min_alloc_size * sizeof(struct _cds_lfht_node)); |
ff0d69de LJ |
1304 | assert(ht->t.tbl[0]); |
1305 | ||
1306 | dbg_printf("create dummy: order %lu index %lu hash %lu\n", 0, 0, 0); | |
1307 | ht->t.tbl[0]->nodes[0].next = flag_dummy(get_end()); | |
1308 | ht->t.tbl[0]->nodes[0].reverse_hash = 0; | |
1309 | ||
1310 | for (order = 1; order < get_count_order_ulong(size) + 1; order++) { | |
1311 | len = 1UL << (order - 1); | |
5488222b | 1312 | if (order <= ht->min_alloc_order) { |
eb631bf2 | 1313 | ht->t.tbl[order] = (struct rcu_level *) (ht->t.tbl[0]->nodes + len); |
5488222b LJ |
1314 | } else { |
1315 | ht->t.tbl[order] = calloc(1, len * sizeof(struct _cds_lfht_node)); | |
1316 | assert(ht->t.tbl[order]); | |
1317 | } | |
ff0d69de LJ |
1318 | |
1319 | i = 0; | |
1320 | prev = ht->t.tbl[i]->nodes; | |
1321 | for (j = 0; j < len; j++) { | |
1322 | if (j & (j - 1)) { /* Between power of 2 */ | |
1323 | prev++; | |
1324 | } else if (j) { /* At each power of 2 */ | |
1325 | i++; | |
1326 | prev = ht->t.tbl[i]->nodes; | |
1327 | } | |
1328 | ||
1329 | node = &ht->t.tbl[order]->nodes[j]; | |
1330 | dbg_printf("create dummy: order %lu index %lu hash %lu\n", | |
1331 | order, j, j + len); | |
1332 | node->next = prev->next; | |
1333 | assert(is_dummy(node->next)); | |
1334 | node->reverse_hash = bit_reverse_ulong(j + len); | |
1335 | prev->next = flag_dummy((struct cds_lfht_node *)node); | |
1336 | } | |
1337 | } | |
1338 | } | |
1339 | ||
7a9dcf9b | 1340 | struct cds_lfht *_cds_lfht_new(cds_lfht_hash_fct hash_fct, |
14044b37 MD |
1341 | cds_lfht_compare_fct compare_fct, |
1342 | unsigned long hash_seed, | |
1343 | unsigned long init_size, | |
5488222b | 1344 | unsigned long min_alloc_size, |
b8af5011 | 1345 | int flags, |
14044b37 | 1346 | void (*cds_lfht_call_rcu)(struct rcu_head *head, |
1475579c | 1347 | void (*func)(struct rcu_head *head)), |
01dbfa62 MD |
1348 | void (*cds_lfht_synchronize_rcu)(void), |
1349 | void (*cds_lfht_rcu_read_lock)(void), | |
5f511391 MD |
1350 | void (*cds_lfht_rcu_read_unlock)(void), |
1351 | void (*cds_lfht_rcu_thread_offline)(void), | |
b7d619b0 MD |
1352 | void (*cds_lfht_rcu_thread_online)(void), |
1353 | void (*cds_lfht_rcu_register_thread)(void), | |
1354 | void (*cds_lfht_rcu_unregister_thread)(void), | |
1355 | pthread_attr_t *attr) | |
abc490a1 | 1356 | { |
14044b37 | 1357 | struct cds_lfht *ht; |
24365af7 | 1358 | unsigned long order; |
abc490a1 | 1359 | |
5488222b LJ |
1360 | /* min_alloc_size must be power of two */ |
1361 | if (!min_alloc_size || (min_alloc_size & (min_alloc_size - 1))) | |
1362 | return NULL; | |
8129be4e | 1363 | /* init_size must be power of two */ |
5488222b | 1364 | if (!init_size || (init_size & (init_size - 1))) |
8129be4e | 1365 | return NULL; |
5488222b LJ |
1366 | min_alloc_size = max(min_alloc_size, MIN_TABLE_SIZE); |
1367 | init_size = max(init_size, min_alloc_size); | |
14044b37 | 1368 | ht = calloc(1, sizeof(struct cds_lfht)); |
b7d619b0 | 1369 | assert(ht); |
b5d6b20f | 1370 | ht->flags = flags; |
abc490a1 | 1371 | ht->hash_fct = hash_fct; |
732ad076 MD |
1372 | ht->compare_fct = compare_fct; |
1373 | ht->hash_seed = hash_seed; | |
14044b37 | 1374 | ht->cds_lfht_call_rcu = cds_lfht_call_rcu; |
1475579c | 1375 | ht->cds_lfht_synchronize_rcu = cds_lfht_synchronize_rcu; |
01dbfa62 MD |
1376 | ht->cds_lfht_rcu_read_lock = cds_lfht_rcu_read_lock; |
1377 | ht->cds_lfht_rcu_read_unlock = cds_lfht_rcu_read_unlock; | |
5f511391 MD |
1378 | ht->cds_lfht_rcu_thread_offline = cds_lfht_rcu_thread_offline; |
1379 | ht->cds_lfht_rcu_thread_online = cds_lfht_rcu_thread_online; | |
b7d619b0 MD |
1380 | ht->cds_lfht_rcu_register_thread = cds_lfht_rcu_register_thread; |
1381 | ht->cds_lfht_rcu_unregister_thread = cds_lfht_rcu_unregister_thread; | |
1382 | ht->resize_attr = attr; | |
5afadd12 | 1383 | alloc_split_items_count(ht); |
abc490a1 MD |
1384 | /* this mutex should not nest in read-side C.S. */ |
1385 | pthread_mutex_init(&ht->resize_mutex, NULL); | |
5488222b | 1386 | order = get_count_order_ulong(init_size); |
93d46c39 | 1387 | ht->t.resize_target = 1UL << order; |
5488222b LJ |
1388 | ht->min_alloc_size = min_alloc_size; |
1389 | ht->min_alloc_order = get_count_order_ulong(min_alloc_size); | |
bcbd36fc LJ |
1390 | cds_lfht_create_dummy(ht, 1UL << order); |
1391 | ht->t.size = 1UL << order; | |
abc490a1 MD |
1392 | return ht; |
1393 | } | |
1394 | ||
adc0de68 MD |
1395 | void cds_lfht_lookup(struct cds_lfht *ht, void *key, size_t key_len, |
1396 | struct cds_lfht_iter *iter) | |
2ed95849 | 1397 | { |
bb7b2f26 | 1398 | struct cds_lfht_node *node, *next, *dummy_node; |
14044b37 | 1399 | struct _cds_lfht_node *lookup; |
f4a9cc0b | 1400 | unsigned long hash, reverse_hash, size; |
2ed95849 | 1401 | |
732ad076 | 1402 | hash = ht->hash_fct(key, key_len, ht->hash_seed); |
abc490a1 | 1403 | reverse_hash = bit_reverse_ulong(hash); |
464a1ec9 | 1404 | |
4105056a | 1405 | size = rcu_dereference(ht->t.size); |
f4a9cc0b | 1406 | lookup = lookup_bucket(ht, size, hash); |
bb7b2f26 MD |
1407 | dummy_node = (struct cds_lfht_node *) lookup; |
1408 | /* We can always skip the dummy node initially */ | |
1409 | node = rcu_dereference(dummy_node->p.next); | |
bb7b2f26 | 1410 | node = clear_flag(node); |
2ed95849 | 1411 | for (;;) { |
8ed51e04 | 1412 | if (caa_unlikely(is_end(node))) { |
96ad1112 | 1413 | node = next = NULL; |
abc490a1 | 1414 | break; |
bb7b2f26 | 1415 | } |
8ed51e04 | 1416 | if (caa_unlikely(node->p.reverse_hash > reverse_hash)) { |
96ad1112 | 1417 | node = next = NULL; |
abc490a1 | 1418 | break; |
2ed95849 | 1419 | } |
1b81fe1a | 1420 | next = rcu_dereference(node->p.next); |
7f52427b | 1421 | assert(node == clear_flag(node)); |
8ed51e04 | 1422 | if (caa_likely(!is_removed(next)) |
1b81fe1a | 1423 | && !is_dummy(next) |
7f52427b | 1424 | && node->p.reverse_hash == reverse_hash |
8ed51e04 | 1425 | && caa_likely(!ht->compare_fct(node->key, node->key_len, key, key_len))) { |
273399de | 1426 | break; |
2ed95849 | 1427 | } |
1b81fe1a | 1428 | node = clear_flag(next); |
2ed95849 | 1429 | } |
1b81fe1a | 1430 | assert(!node || !is_dummy(rcu_dereference(node->p.next))); |
adc0de68 MD |
1431 | iter->node = node; |
1432 | iter->next = next; | |
abc490a1 | 1433 | } |
e0ba718a | 1434 | |
3883c0e5 | 1435 | void cds_lfht_next_duplicate(struct cds_lfht *ht, struct cds_lfht_iter *iter) |
a481e5ff | 1436 | { |
adc0de68 | 1437 | struct cds_lfht_node *node, *next; |
a481e5ff MD |
1438 | unsigned long reverse_hash; |
1439 | void *key; | |
1440 | size_t key_len; | |
1441 | ||
adc0de68 | 1442 | node = iter->node; |
a481e5ff MD |
1443 | reverse_hash = node->p.reverse_hash; |
1444 | key = node->key; | |
1445 | key_len = node->key_len; | |
adc0de68 | 1446 | next = iter->next; |
a481e5ff MD |
1447 | node = clear_flag(next); |
1448 | ||
1449 | for (;;) { | |
8ed51e04 | 1450 | if (caa_unlikely(is_end(node))) { |
96ad1112 | 1451 | node = next = NULL; |
a481e5ff | 1452 | break; |
bb7b2f26 | 1453 | } |
8ed51e04 | 1454 | if (caa_unlikely(node->p.reverse_hash > reverse_hash)) { |
96ad1112 | 1455 | node = next = NULL; |
a481e5ff MD |
1456 | break; |
1457 | } | |
1458 | next = rcu_dereference(node->p.next); | |
8ed51e04 | 1459 | if (caa_likely(!is_removed(next)) |
a481e5ff | 1460 | && !is_dummy(next) |
8ed51e04 | 1461 | && caa_likely(!ht->compare_fct(node->key, node->key_len, key, key_len))) { |
a481e5ff MD |
1462 | break; |
1463 | } | |
1464 | node = clear_flag(next); | |
1465 | } | |
1466 | assert(!node || !is_dummy(rcu_dereference(node->p.next))); | |
adc0de68 MD |
1467 | iter->node = node; |
1468 | iter->next = next; | |
a481e5ff MD |
1469 | } |
1470 | ||
4e9b9fbf MD |
1471 | void cds_lfht_next(struct cds_lfht *ht, struct cds_lfht_iter *iter) |
1472 | { | |
1473 | struct cds_lfht_node *node, *next; | |
1474 | ||
853395e1 | 1475 | node = clear_flag(iter->next); |
4e9b9fbf | 1476 | for (;;) { |
8ed51e04 | 1477 | if (caa_unlikely(is_end(node))) { |
4e9b9fbf MD |
1478 | node = next = NULL; |
1479 | break; | |
1480 | } | |
1481 | next = rcu_dereference(node->p.next); | |
8ed51e04 | 1482 | if (caa_likely(!is_removed(next)) |
4e9b9fbf MD |
1483 | && !is_dummy(next)) { |
1484 | break; | |
1485 | } | |
1486 | node = clear_flag(next); | |
1487 | } | |
1488 | assert(!node || !is_dummy(rcu_dereference(node->p.next))); | |
1489 | iter->node = node; | |
1490 | iter->next = next; | |
1491 | } | |
1492 | ||
1493 | void cds_lfht_first(struct cds_lfht *ht, struct cds_lfht_iter *iter) | |
1494 | { | |
1495 | struct _cds_lfht_node *lookup; | |
1496 | ||
1497 | /* | |
1498 | * Get next after first dummy node. The first dummy node is the | |
1499 | * first node of the linked list. | |
1500 | */ | |
1501 | lookup = &ht->t.tbl[0]->nodes[0]; | |
853395e1 | 1502 | iter->next = lookup->next; |
4e9b9fbf MD |
1503 | cds_lfht_next(ht, iter); |
1504 | } | |
1505 | ||
14044b37 | 1506 | void cds_lfht_add(struct cds_lfht *ht, struct cds_lfht_node *node) |
abc490a1 | 1507 | { |
4105056a | 1508 | unsigned long hash, size; |
ab7d5fc6 | 1509 | |
49c2e2d6 | 1510 | hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed); |
cc4fcb10 | 1511 | node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash); |
2ed95849 | 1512 | |
4105056a | 1513 | size = rcu_dereference(ht->t.size); |
83beee94 | 1514 | _cds_lfht_add(ht, size, node, NULL, 0); |
14360f1c | 1515 | ht_count_add(ht, size, hash); |
3eca1b8c MD |
1516 | } |
1517 | ||
14044b37 | 1518 | struct cds_lfht_node *cds_lfht_add_unique(struct cds_lfht *ht, |
48ed1c18 | 1519 | struct cds_lfht_node *node) |
3eca1b8c | 1520 | { |
4105056a | 1521 | unsigned long hash, size; |
83beee94 | 1522 | struct cds_lfht_iter iter; |
3eca1b8c | 1523 | |
49c2e2d6 | 1524 | hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed); |
cc4fcb10 | 1525 | node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash); |
3eca1b8c | 1526 | |
4105056a | 1527 | size = rcu_dereference(ht->t.size); |
83beee94 MD |
1528 | _cds_lfht_add(ht, size, node, &iter, 0); |
1529 | if (iter.node == node) | |
14360f1c | 1530 | ht_count_add(ht, size, hash); |
83beee94 | 1531 | return iter.node; |
2ed95849 MD |
1532 | } |
1533 | ||
9357c415 | 1534 | struct cds_lfht_node *cds_lfht_add_replace(struct cds_lfht *ht, |
48ed1c18 MD |
1535 | struct cds_lfht_node *node) |
1536 | { | |
1537 | unsigned long hash, size; | |
83beee94 | 1538 | struct cds_lfht_iter iter; |
48ed1c18 MD |
1539 | |
1540 | hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed); | |
1541 | node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash); | |
1542 | ||
1543 | size = rcu_dereference(ht->t.size); | |
83beee94 MD |
1544 | for (;;) { |
1545 | _cds_lfht_add(ht, size, node, &iter, 0); | |
1546 | if (iter.node == node) { | |
14360f1c | 1547 | ht_count_add(ht, size, hash); |
83beee94 MD |
1548 | return NULL; |
1549 | } | |
1550 | ||
1551 | if (!_cds_lfht_replace(ht, size, iter.node, iter.next, node)) | |
1552 | return iter.node; | |
1553 | } | |
48ed1c18 MD |
1554 | } |
1555 | ||
9357c415 MD |
1556 | int cds_lfht_replace(struct cds_lfht *ht, struct cds_lfht_iter *old_iter, |
1557 | struct cds_lfht_node *new_node) | |
1558 | { | |
1559 | unsigned long size; | |
1560 | ||
1561 | size = rcu_dereference(ht->t.size); | |
1562 | return _cds_lfht_replace(ht, size, old_iter->node, old_iter->next, | |
1563 | new_node); | |
1564 | } | |
1565 | ||
1566 | int cds_lfht_del(struct cds_lfht *ht, struct cds_lfht_iter *iter) | |
2ed95849 | 1567 | { |
14360f1c | 1568 | unsigned long size, hash; |
df44348d | 1569 | int ret; |
abc490a1 | 1570 | |
4105056a | 1571 | size = rcu_dereference(ht->t.size); |
9357c415 | 1572 | ret = _cds_lfht_del(ht, size, iter->node, 0); |
14360f1c LJ |
1573 | if (!ret) { |
1574 | hash = bit_reverse_ulong(iter->node->p.reverse_hash); | |
1575 | ht_count_del(ht, size, hash); | |
1576 | } | |
df44348d | 1577 | return ret; |
2ed95849 | 1578 | } |
ab7d5fc6 | 1579 | |
abc490a1 | 1580 | static |
14044b37 | 1581 | int cds_lfht_delete_dummy(struct cds_lfht *ht) |
674f7a69 | 1582 | { |
14044b37 MD |
1583 | struct cds_lfht_node *node; |
1584 | struct _cds_lfht_node *lookup; | |
4105056a | 1585 | unsigned long order, i, size; |
674f7a69 | 1586 | |
abc490a1 | 1587 | /* Check that the table is empty */ |
4105056a | 1588 | lookup = &ht->t.tbl[0]->nodes[0]; |
14044b37 | 1589 | node = (struct cds_lfht_node *) lookup; |
abc490a1 | 1590 | do { |
1b81fe1a MD |
1591 | node = clear_flag(node)->p.next; |
1592 | if (!is_dummy(node)) | |
abc490a1 | 1593 | return -EPERM; |
273399de | 1594 | assert(!is_removed(node)); |
bb7b2f26 | 1595 | } while (!is_end(node)); |
4105056a MD |
1596 | /* |
1597 | * size accessed without rcu_dereference because hash table is | |
1598 | * being destroyed. | |
1599 | */ | |
1600 | size = ht->t.size; | |
abc490a1 | 1601 | /* Internal sanity check: all nodes left should be dummy */ |
4105056a | 1602 | for (order = 0; order < get_count_order_ulong(size) + 1; order++) { |
24365af7 MD |
1603 | unsigned long len; |
1604 | ||
1605 | len = !order ? 1 : 1UL << (order - 1); | |
1606 | for (i = 0; i < len; i++) { | |
f0c29ed7 | 1607 | dbg_printf("delete order %lu i %lu hash %lu\n", |
24365af7 | 1608 | order, i, |
4105056a MD |
1609 | bit_reverse_ulong(ht->t.tbl[order]->nodes[i].reverse_hash)); |
1610 | assert(is_dummy(ht->t.tbl[order]->nodes[i].next)); | |
24365af7 | 1611 | } |
5488222b LJ |
1612 | |
1613 | if (order == ht->min_alloc_order) | |
1614 | poison_free(ht->t.tbl[0]); | |
1615 | else if (order > ht->min_alloc_order) | |
1616 | poison_free(ht->t.tbl[order]); | |
1617 | /* Nothing to delete for order < ht->min_alloc_order */ | |
674f7a69 | 1618 | } |
abc490a1 | 1619 | return 0; |
674f7a69 MD |
1620 | } |
1621 | ||
1622 | /* | |
1623 | * Should only be called when no more concurrent readers nor writers can | |
1624 | * possibly access the table. | |
1625 | */ | |
b7d619b0 | 1626 | int cds_lfht_destroy(struct cds_lfht *ht, pthread_attr_t **attr) |
674f7a69 | 1627 | { |
5e28c532 MD |
1628 | int ret; |
1629 | ||
848d4088 | 1630 | /* Wait for in-flight resize operations to complete */ |
24953e08 MD |
1631 | _CMM_STORE_SHARED(ht->in_progress_destroy, 1); |
1632 | cmm_smp_mb(); /* Store destroy before load resize */ | |
848d4088 MD |
1633 | while (uatomic_read(&ht->in_progress_resize)) |
1634 | poll(NULL, 0, 100); /* wait for 100ms */ | |
14044b37 | 1635 | ret = cds_lfht_delete_dummy(ht); |
abc490a1 MD |
1636 | if (ret) |
1637 | return ret; | |
5afadd12 | 1638 | free_split_items_count(ht); |
b7d619b0 MD |
1639 | if (attr) |
1640 | *attr = ht->resize_attr; | |
98808fb1 | 1641 | poison_free(ht); |
5e28c532 | 1642 | return ret; |
674f7a69 MD |
1643 | } |
1644 | ||
14044b37 | 1645 | void cds_lfht_count_nodes(struct cds_lfht *ht, |
d933dd0e | 1646 | long *approx_before, |
273399de | 1647 | unsigned long *count, |
973e5e1b | 1648 | unsigned long *removed, |
d933dd0e | 1649 | long *approx_after) |
273399de | 1650 | { |
14044b37 MD |
1651 | struct cds_lfht_node *node, *next; |
1652 | struct _cds_lfht_node *lookup; | |
24365af7 | 1653 | unsigned long nr_dummy = 0; |
273399de | 1654 | |
7ed7682f | 1655 | *approx_before = 0; |
5afadd12 | 1656 | if (ht->split_count) { |
973e5e1b MD |
1657 | int i; |
1658 | ||
4c42f1b8 LJ |
1659 | for (i = 0; i < split_count_mask + 1; i++) { |
1660 | *approx_before += uatomic_read(&ht->split_count[i].add); | |
1661 | *approx_before -= uatomic_read(&ht->split_count[i].del); | |
973e5e1b MD |
1662 | } |
1663 | } | |
1664 | ||
273399de MD |
1665 | *count = 0; |
1666 | *removed = 0; | |
1667 | ||
24365af7 | 1668 | /* Count non-dummy nodes in the table */ |
4105056a | 1669 | lookup = &ht->t.tbl[0]->nodes[0]; |
14044b37 | 1670 | node = (struct cds_lfht_node *) lookup; |
273399de | 1671 | do { |
cc4fcb10 | 1672 | next = rcu_dereference(node->p.next); |
b198f0fd | 1673 | if (is_removed(next)) { |
973e5e1b MD |
1674 | if (!is_dummy(next)) |
1675 | (*removed)++; | |
1676 | else | |
1677 | (nr_dummy)++; | |
1b81fe1a | 1678 | } else if (!is_dummy(next)) |
273399de | 1679 | (*count)++; |
24365af7 MD |
1680 | else |
1681 | (nr_dummy)++; | |
273399de | 1682 | node = clear_flag(next); |
bb7b2f26 | 1683 | } while (!is_end(node)); |
f0c29ed7 | 1684 | dbg_printf("number of dummy nodes: %lu\n", nr_dummy); |
7ed7682f | 1685 | *approx_after = 0; |
5afadd12 | 1686 | if (ht->split_count) { |
973e5e1b MD |
1687 | int i; |
1688 | ||
4c42f1b8 LJ |
1689 | for (i = 0; i < split_count_mask + 1; i++) { |
1690 | *approx_after += uatomic_read(&ht->split_count[i].add); | |
1691 | *approx_after -= uatomic_read(&ht->split_count[i].del); | |
973e5e1b MD |
1692 | } |
1693 | } | |
273399de MD |
1694 | } |
1695 | ||
1475579c | 1696 | /* called with resize mutex held */ |
abc490a1 | 1697 | static |
4105056a | 1698 | void _do_cds_lfht_grow(struct cds_lfht *ht, |
1475579c | 1699 | unsigned long old_size, unsigned long new_size) |
abc490a1 | 1700 | { |
1475579c | 1701 | unsigned long old_order, new_order; |
1475579c | 1702 | |
93d46c39 LJ |
1703 | old_order = get_count_order_ulong(old_size); |
1704 | new_order = get_count_order_ulong(new_size); | |
1a401918 LJ |
1705 | dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n", |
1706 | old_size, old_order, new_size, new_order); | |
1475579c | 1707 | assert(new_size > old_size); |
93d46c39 | 1708 | init_table(ht, old_order + 1, new_order); |
abc490a1 MD |
1709 | } |
1710 | ||
1711 | /* called with resize mutex held */ | |
1712 | static | |
4105056a | 1713 | void _do_cds_lfht_shrink(struct cds_lfht *ht, |
1475579c | 1714 | unsigned long old_size, unsigned long new_size) |
464a1ec9 | 1715 | { |
1475579c | 1716 | unsigned long old_order, new_order; |
464a1ec9 | 1717 | |
5488222b | 1718 | new_size = max(new_size, ht->min_alloc_size); |
93d46c39 LJ |
1719 | old_order = get_count_order_ulong(old_size); |
1720 | new_order = get_count_order_ulong(new_size); | |
1a401918 LJ |
1721 | dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n", |
1722 | old_size, old_order, new_size, new_order); | |
1475579c | 1723 | assert(new_size < old_size); |
1475579c | 1724 | |
4105056a | 1725 | /* Remove and unlink all dummy nodes to remove. */ |
93d46c39 | 1726 | fini_table(ht, new_order + 1, old_order); |
464a1ec9 MD |
1727 | } |
1728 | ||
1475579c MD |
1729 | |
1730 | /* called with resize mutex held */ | |
1731 | static | |
1732 | void _do_cds_lfht_resize(struct cds_lfht *ht) | |
1733 | { | |
1734 | unsigned long new_size, old_size; | |
4105056a MD |
1735 | |
1736 | /* | |
1737 | * Resize table, re-do if the target size has changed under us. | |
1738 | */ | |
1739 | do { | |
d2be3620 MD |
1740 | assert(uatomic_read(&ht->in_progress_resize)); |
1741 | if (CMM_LOAD_SHARED(ht->in_progress_destroy)) | |
1742 | break; | |
4105056a MD |
1743 | ht->t.resize_initiated = 1; |
1744 | old_size = ht->t.size; | |
1745 | new_size = CMM_LOAD_SHARED(ht->t.resize_target); | |
1746 | if (old_size < new_size) | |
1747 | _do_cds_lfht_grow(ht, old_size, new_size); | |
1748 | else if (old_size > new_size) | |
1749 | _do_cds_lfht_shrink(ht, old_size, new_size); | |
1750 | ht->t.resize_initiated = 0; | |
1751 | /* write resize_initiated before read resize_target */ | |
1752 | cmm_smp_mb(); | |
4d676753 | 1753 | } while (ht->t.size != CMM_LOAD_SHARED(ht->t.resize_target)); |
1475579c MD |
1754 | } |
1755 | ||
abc490a1 | 1756 | static |
ab65b890 | 1757 | unsigned long resize_target_grow(struct cds_lfht *ht, unsigned long new_size) |
464a1ec9 | 1758 | { |
ab65b890 | 1759 | return _uatomic_xchg_monotonic_increase(&ht->t.resize_target, new_size); |
464a1ec9 MD |
1760 | } |
1761 | ||
1475579c | 1762 | static |
4105056a | 1763 | void resize_target_update_count(struct cds_lfht *ht, |
b8af5011 | 1764 | unsigned long count) |
1475579c | 1765 | { |
5488222b | 1766 | count = max(count, ht->min_alloc_size); |
4105056a | 1767 | uatomic_set(&ht->t.resize_target, count); |
1475579c MD |
1768 | } |
1769 | ||
1770 | void cds_lfht_resize(struct cds_lfht *ht, unsigned long new_size) | |
464a1ec9 | 1771 | { |
4105056a MD |
1772 | resize_target_update_count(ht, new_size); |
1773 | CMM_STORE_SHARED(ht->t.resize_initiated, 1); | |
5f511391 | 1774 | ht->cds_lfht_rcu_thread_offline(); |
1475579c MD |
1775 | pthread_mutex_lock(&ht->resize_mutex); |
1776 | _do_cds_lfht_resize(ht); | |
1777 | pthread_mutex_unlock(&ht->resize_mutex); | |
5f511391 | 1778 | ht->cds_lfht_rcu_thread_online(); |
abc490a1 | 1779 | } |
464a1ec9 | 1780 | |
abc490a1 MD |
1781 | static |
1782 | void do_resize_cb(struct rcu_head *head) | |
1783 | { | |
1784 | struct rcu_resize_work *work = | |
1785 | caa_container_of(head, struct rcu_resize_work, head); | |
14044b37 | 1786 | struct cds_lfht *ht = work->ht; |
abc490a1 | 1787 | |
5f511391 | 1788 | ht->cds_lfht_rcu_thread_offline(); |
abc490a1 | 1789 | pthread_mutex_lock(&ht->resize_mutex); |
14044b37 | 1790 | _do_cds_lfht_resize(ht); |
abc490a1 | 1791 | pthread_mutex_unlock(&ht->resize_mutex); |
5f511391 | 1792 | ht->cds_lfht_rcu_thread_online(); |
98808fb1 | 1793 | poison_free(work); |
848d4088 MD |
1794 | cmm_smp_mb(); /* finish resize before decrement */ |
1795 | uatomic_dec(&ht->in_progress_resize); | |
464a1ec9 MD |
1796 | } |
1797 | ||
abc490a1 | 1798 | static |
f1f119ee | 1799 | void __cds_lfht_resize_lazy_launch(struct cds_lfht *ht) |
ab7d5fc6 | 1800 | { |
abc490a1 MD |
1801 | struct rcu_resize_work *work; |
1802 | ||
4105056a MD |
1803 | /* Store resize_target before read resize_initiated */ |
1804 | cmm_smp_mb(); | |
ab65b890 | 1805 | if (!CMM_LOAD_SHARED(ht->t.resize_initiated)) { |
848d4088 | 1806 | uatomic_inc(&ht->in_progress_resize); |
59290e9d | 1807 | cmm_smp_mb(); /* increment resize count before load destroy */ |
ed35e6d8 MD |
1808 | if (CMM_LOAD_SHARED(ht->in_progress_destroy)) { |
1809 | uatomic_dec(&ht->in_progress_resize); | |
59290e9d | 1810 | return; |
ed35e6d8 | 1811 | } |
f9830efd MD |
1812 | work = malloc(sizeof(*work)); |
1813 | work->ht = ht; | |
14044b37 | 1814 | ht->cds_lfht_call_rcu(&work->head, do_resize_cb); |
4105056a | 1815 | CMM_STORE_SHARED(ht->t.resize_initiated, 1); |
f9830efd | 1816 | } |
ab7d5fc6 | 1817 | } |
3171717f | 1818 | |
f1f119ee LJ |
1819 | static |
1820 | void cds_lfht_resize_lazy_grow(struct cds_lfht *ht, unsigned long size, int growth) | |
1821 | { | |
1822 | unsigned long target_size = size << growth; | |
1823 | ||
1824 | if (resize_target_grow(ht, target_size) >= target_size) | |
1825 | return; | |
1826 | ||
1827 | __cds_lfht_resize_lazy_launch(ht); | |
1828 | } | |
1829 | ||
89bb121d LJ |
1830 | /* |
1831 | * We favor grow operations over shrink. A shrink operation never occurs | |
1832 | * if a grow operation is queued for lazy execution. A grow operation | |
1833 | * cancels any pending shrink lazy execution. | |
1834 | */ | |
3171717f | 1835 | static |
4105056a | 1836 | void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size, |
3171717f MD |
1837 | unsigned long count) |
1838 | { | |
b8af5011 MD |
1839 | if (!(ht->flags & CDS_LFHT_AUTO_RESIZE)) |
1840 | return; | |
89bb121d LJ |
1841 | count = max(count, ht->min_alloc_size); |
1842 | if (count == size) | |
1843 | return; /* Already the right size, no resize needed */ | |
1844 | if (count > size) { /* lazy grow */ | |
1845 | if (resize_target_grow(ht, count) >= count) | |
1846 | return; | |
1847 | } else { /* lazy shrink */ | |
1848 | for (;;) { | |
1849 | unsigned long s; | |
1850 | ||
1851 | s = uatomic_cmpxchg(&ht->t.resize_target, size, count); | |
1852 | if (s == size) | |
1853 | break; /* no resize needed */ | |
1854 | if (s > size) | |
1855 | return; /* growing is/(was just) in progress */ | |
1856 | if (s <= count) | |
1857 | return; /* some other thread do shrink */ | |
1858 | size = s; | |
1859 | } | |
1860 | } | |
f1f119ee | 1861 | __cds_lfht_resize_lazy_launch(ht); |
3171717f | 1862 | } |