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. | |
56 | * - Per-CPU Split-counters are used to keep track of the number of | |
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 MD |
175 | /* |
176 | * Per-CPU split-counters lazily update the global counter each 1024 | |
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 | |
6ea6bc67 MD |
182 | #define CHAIN_LEN_TARGET 1 |
183 | #define CHAIN_LEN_RESIZE_THRESHOLD 3 | |
2ed95849 | 184 | |
cd95516d | 185 | /* |
76a73da8 | 186 | * Define the minimum table size. |
cd95516d | 187 | */ |
c9edd44a | 188 | #define MIN_TABLE_SIZE 1 |
cd95516d | 189 | |
4105056a MD |
190 | #if (CAA_BITS_PER_LONG == 32) |
191 | #define MAX_TABLE_ORDER 32 | |
192 | #else | |
193 | #define MAX_TABLE_ORDER 64 | |
194 | #endif | |
195 | ||
b7d619b0 MD |
196 | /* |
197 | * Minimum number of dummy nodes to touch per thread to parallelize grow/shrink. | |
198 | */ | |
6083a889 MD |
199 | #define MIN_PARTITION_PER_THREAD_ORDER 12 |
200 | #define MIN_PARTITION_PER_THREAD (1UL << MIN_PARTITION_PER_THREAD_ORDER) | |
b7d619b0 | 201 | |
4105056a MD |
202 | #ifndef min |
203 | #define min(a, b) ((a) < (b) ? (a) : (b)) | |
204 | #endif | |
205 | ||
abc490a1 MD |
206 | #ifndef max |
207 | #define max(a, b) ((a) > (b) ? (a) : (b)) | |
208 | #endif | |
2ed95849 | 209 | |
d95bd160 MD |
210 | /* |
211 | * The removed flag needs to be updated atomically with the pointer. | |
48ed1c18 | 212 | * It indicates that no node must attach to the node scheduled for |
b198f0fd | 213 | * removal, and that node garbage collection must be performed. |
d95bd160 MD |
214 | * The dummy flag does not require to be updated atomically with the |
215 | * pointer, but it is added as a pointer low bit flag to save space. | |
216 | */ | |
d37166c6 | 217 | #define REMOVED_FLAG (1UL << 0) |
b198f0fd MD |
218 | #define DUMMY_FLAG (1UL << 1) |
219 | #define FLAGS_MASK ((1UL << 2) - 1) | |
d37166c6 | 220 | |
bb7b2f26 | 221 | /* Value of the end pointer. Should not interact with flags. */ |
f9c80341 | 222 | #define END_VALUE NULL |
bb7b2f26 | 223 | |
df44348d | 224 | struct ht_items_count { |
860d07e8 | 225 | unsigned long add, del; |
df44348d MD |
226 | } __attribute__((aligned(CAA_CACHE_LINE_SIZE))); |
227 | ||
1475579c | 228 | struct rcu_level { |
0d14ceb2 | 229 | /* Note: manually update allocation length when adding a field */ |
1475579c MD |
230 | struct _cds_lfht_node nodes[0]; |
231 | }; | |
232 | ||
395270b6 | 233 | struct rcu_table { |
4105056a | 234 | unsigned long size; /* always a power of 2, shared (RCU) */ |
f9830efd | 235 | unsigned long resize_target; |
11519af6 | 236 | int resize_initiated; |
4105056a | 237 | struct rcu_level *tbl[MAX_TABLE_ORDER]; |
395270b6 MD |
238 | }; |
239 | ||
14044b37 | 240 | struct cds_lfht { |
4105056a | 241 | struct rcu_table t; |
14044b37 MD |
242 | cds_lfht_hash_fct hash_fct; |
243 | cds_lfht_compare_fct compare_fct; | |
732ad076 | 244 | unsigned long hash_seed; |
b8af5011 | 245 | int flags; |
5f511391 MD |
246 | /* |
247 | * We need to put the work threads offline (QSBR) when taking this | |
248 | * mutex, because we use synchronize_rcu within this mutex critical | |
249 | * section, which waits on read-side critical sections, and could | |
250 | * therefore cause grace-period deadlock if we hold off RCU G.P. | |
251 | * completion. | |
252 | */ | |
464a1ec9 | 253 | pthread_mutex_t resize_mutex; /* resize mutex: add/del mutex */ |
33c7c748 | 254 | unsigned int in_progress_resize, in_progress_destroy; |
14044b37 | 255 | void (*cds_lfht_call_rcu)(struct rcu_head *head, |
abc490a1 | 256 | void (*func)(struct rcu_head *head)); |
1475579c | 257 | void (*cds_lfht_synchronize_rcu)(void); |
01dbfa62 MD |
258 | void (*cds_lfht_rcu_read_lock)(void); |
259 | void (*cds_lfht_rcu_read_unlock)(void); | |
5f511391 MD |
260 | void (*cds_lfht_rcu_thread_offline)(void); |
261 | void (*cds_lfht_rcu_thread_online)(void); | |
b7d619b0 MD |
262 | void (*cds_lfht_rcu_register_thread)(void); |
263 | void (*cds_lfht_rcu_unregister_thread)(void); | |
264 | pthread_attr_t *resize_attr; /* Resize threads attributes */ | |
7de5ccfd | 265 | long count; /* global approximate item count */ |
df44348d | 266 | struct ht_items_count *percpu_count; /* per-cpu item count */ |
2ed95849 MD |
267 | }; |
268 | ||
abc490a1 MD |
269 | struct rcu_resize_work { |
270 | struct rcu_head head; | |
14044b37 | 271 | struct cds_lfht *ht; |
abc490a1 | 272 | }; |
2ed95849 | 273 | |
b7d619b0 | 274 | struct partition_resize_work { |
1af6e26e | 275 | pthread_t thread_id; |
b7d619b0 MD |
276 | struct cds_lfht *ht; |
277 | unsigned long i, start, len; | |
278 | void (*fct)(struct cds_lfht *ht, unsigned long i, | |
279 | unsigned long start, unsigned long len); | |
280 | }; | |
281 | ||
76a73da8 | 282 | static |
83beee94 MD |
283 | void _cds_lfht_add(struct cds_lfht *ht, |
284 | unsigned long size, | |
285 | struct cds_lfht_node *node, | |
286 | struct cds_lfht_iter *unique_ret, | |
287 | int dummy); | |
48ed1c18 | 288 | |
abc490a1 MD |
289 | /* |
290 | * Algorithm to reverse bits in a word by lookup table, extended to | |
291 | * 64-bit words. | |
f9830efd | 292 | * Source: |
abc490a1 | 293 | * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable |
f9830efd | 294 | * Originally from Public Domain. |
abc490a1 MD |
295 | */ |
296 | ||
297 | static const uint8_t BitReverseTable256[256] = | |
2ed95849 | 298 | { |
abc490a1 MD |
299 | #define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64 |
300 | #define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16) | |
301 | #define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 ) | |
302 | R6(0), R6(2), R6(1), R6(3) | |
303 | }; | |
304 | #undef R2 | |
305 | #undef R4 | |
306 | #undef R6 | |
2ed95849 | 307 | |
abc490a1 MD |
308 | static |
309 | uint8_t bit_reverse_u8(uint8_t v) | |
310 | { | |
311 | return BitReverseTable256[v]; | |
312 | } | |
ab7d5fc6 | 313 | |
abc490a1 MD |
314 | static __attribute__((unused)) |
315 | uint32_t bit_reverse_u32(uint32_t v) | |
316 | { | |
317 | return ((uint32_t) bit_reverse_u8(v) << 24) | | |
318 | ((uint32_t) bit_reverse_u8(v >> 8) << 16) | | |
319 | ((uint32_t) bit_reverse_u8(v >> 16) << 8) | | |
320 | ((uint32_t) bit_reverse_u8(v >> 24)); | |
2ed95849 MD |
321 | } |
322 | ||
abc490a1 MD |
323 | static __attribute__((unused)) |
324 | uint64_t bit_reverse_u64(uint64_t v) | |
2ed95849 | 325 | { |
abc490a1 MD |
326 | return ((uint64_t) bit_reverse_u8(v) << 56) | |
327 | ((uint64_t) bit_reverse_u8(v >> 8) << 48) | | |
328 | ((uint64_t) bit_reverse_u8(v >> 16) << 40) | | |
329 | ((uint64_t) bit_reverse_u8(v >> 24) << 32) | | |
330 | ((uint64_t) bit_reverse_u8(v >> 32) << 24) | | |
331 | ((uint64_t) bit_reverse_u8(v >> 40) << 16) | | |
332 | ((uint64_t) bit_reverse_u8(v >> 48) << 8) | | |
333 | ((uint64_t) bit_reverse_u8(v >> 56)); | |
334 | } | |
335 | ||
336 | static | |
337 | unsigned long bit_reverse_ulong(unsigned long v) | |
338 | { | |
339 | #if (CAA_BITS_PER_LONG == 32) | |
340 | return bit_reverse_u32(v); | |
341 | #else | |
342 | return bit_reverse_u64(v); | |
343 | #endif | |
344 | } | |
345 | ||
f9830efd | 346 | /* |
24365af7 MD |
347 | * fls: returns the position of the most significant bit. |
348 | * Returns 0 if no bit is set, else returns the position of the most | |
349 | * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit). | |
f9830efd | 350 | */ |
24365af7 MD |
351 | #if defined(__i386) || defined(__x86_64) |
352 | static inline | |
353 | unsigned int fls_u32(uint32_t x) | |
f9830efd | 354 | { |
24365af7 MD |
355 | int r; |
356 | ||
357 | asm("bsrl %1,%0\n\t" | |
358 | "jnz 1f\n\t" | |
359 | "movl $-1,%0\n\t" | |
360 | "1:\n\t" | |
361 | : "=r" (r) : "rm" (x)); | |
362 | return r + 1; | |
363 | } | |
364 | #define HAS_FLS_U32 | |
365 | #endif | |
366 | ||
367 | #if defined(__x86_64) | |
368 | static inline | |
369 | unsigned int fls_u64(uint64_t x) | |
370 | { | |
371 | long r; | |
372 | ||
373 | asm("bsrq %1,%0\n\t" | |
374 | "jnz 1f\n\t" | |
375 | "movq $-1,%0\n\t" | |
376 | "1:\n\t" | |
377 | : "=r" (r) : "rm" (x)); | |
378 | return r + 1; | |
379 | } | |
380 | #define HAS_FLS_U64 | |
381 | #endif | |
382 | ||
383 | #ifndef HAS_FLS_U64 | |
384 | static __attribute__((unused)) | |
385 | unsigned int fls_u64(uint64_t x) | |
386 | { | |
387 | unsigned int r = 64; | |
388 | ||
389 | if (!x) | |
390 | return 0; | |
391 | ||
392 | if (!(x & 0xFFFFFFFF00000000ULL)) { | |
393 | x <<= 32; | |
394 | r -= 32; | |
395 | } | |
396 | if (!(x & 0xFFFF000000000000ULL)) { | |
397 | x <<= 16; | |
398 | r -= 16; | |
399 | } | |
400 | if (!(x & 0xFF00000000000000ULL)) { | |
401 | x <<= 8; | |
402 | r -= 8; | |
403 | } | |
404 | if (!(x & 0xF000000000000000ULL)) { | |
405 | x <<= 4; | |
406 | r -= 4; | |
407 | } | |
408 | if (!(x & 0xC000000000000000ULL)) { | |
409 | x <<= 2; | |
410 | r -= 2; | |
411 | } | |
412 | if (!(x & 0x8000000000000000ULL)) { | |
413 | x <<= 1; | |
414 | r -= 1; | |
415 | } | |
416 | return r; | |
417 | } | |
418 | #endif | |
419 | ||
420 | #ifndef HAS_FLS_U32 | |
421 | static __attribute__((unused)) | |
422 | unsigned int fls_u32(uint32_t x) | |
423 | { | |
424 | unsigned int r = 32; | |
f9830efd | 425 | |
24365af7 MD |
426 | if (!x) |
427 | return 0; | |
428 | if (!(x & 0xFFFF0000U)) { | |
429 | x <<= 16; | |
430 | r -= 16; | |
431 | } | |
432 | if (!(x & 0xFF000000U)) { | |
433 | x <<= 8; | |
434 | r -= 8; | |
435 | } | |
436 | if (!(x & 0xF0000000U)) { | |
437 | x <<= 4; | |
438 | r -= 4; | |
439 | } | |
440 | if (!(x & 0xC0000000U)) { | |
441 | x <<= 2; | |
442 | r -= 2; | |
443 | } | |
444 | if (!(x & 0x80000000U)) { | |
445 | x <<= 1; | |
446 | r -= 1; | |
447 | } | |
448 | return r; | |
449 | } | |
450 | #endif | |
451 | ||
452 | unsigned int fls_ulong(unsigned long x) | |
f9830efd | 453 | { |
24365af7 MD |
454 | #if (CAA_BITS_PER_lONG == 32) |
455 | return fls_u32(x); | |
456 | #else | |
457 | return fls_u64(x); | |
458 | #endif | |
459 | } | |
f9830efd | 460 | |
920f8ef6 LJ |
461 | /* |
462 | * Return the minimum order for which x <= (1UL << order). | |
463 | * Return -1 if x is 0. | |
464 | */ | |
24365af7 MD |
465 | int get_count_order_u32(uint32_t x) |
466 | { | |
920f8ef6 LJ |
467 | if (!x) |
468 | return -1; | |
24365af7 | 469 | |
920f8ef6 | 470 | return fls_u32(x - 1); |
24365af7 MD |
471 | } |
472 | ||
920f8ef6 LJ |
473 | /* |
474 | * Return the minimum order for which x <= (1UL << order). | |
475 | * Return -1 if x is 0. | |
476 | */ | |
24365af7 MD |
477 | int get_count_order_ulong(unsigned long x) |
478 | { | |
920f8ef6 LJ |
479 | if (!x) |
480 | return -1; | |
24365af7 | 481 | |
920f8ef6 | 482 | return fls_ulong(x - 1); |
f9830efd MD |
483 | } |
484 | ||
98808fb1 MD |
485 | #ifdef POISON_FREE |
486 | #define poison_free(ptr) \ | |
487 | do { \ | |
488 | memset(ptr, 0x42, sizeof(*(ptr))); \ | |
489 | free(ptr); \ | |
490 | } while (0) | |
491 | #else | |
492 | #define poison_free(ptr) free(ptr) | |
493 | #endif | |
494 | ||
f9830efd | 495 | static |
4105056a | 496 | void cds_lfht_resize_lazy(struct cds_lfht *ht, unsigned long size, int growth); |
f9830efd | 497 | |
df44348d MD |
498 | /* |
499 | * If the sched_getcpu() and sysconf(_SC_NPROCESSORS_CONF) calls are | |
500 | * available, then we support hash table item accounting. | |
501 | * In the unfortunate event the number of CPUs reported would be | |
502 | * inaccurate, we use modulo arithmetic on the number of CPUs we got. | |
503 | */ | |
df44348d MD |
504 | #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) |
505 | ||
f8994aee | 506 | static |
4105056a | 507 | void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size, |
f8994aee MD |
508 | unsigned long count); |
509 | ||
df44348d MD |
510 | static long nr_cpus_mask = -1; |
511 | ||
512 | static | |
513 | struct ht_items_count *alloc_per_cpu_items_count(void) | |
514 | { | |
515 | struct ht_items_count *count; | |
516 | ||
517 | switch (nr_cpus_mask) { | |
518 | case -2: | |
519 | return NULL; | |
520 | case -1: | |
521 | { | |
522 | long maxcpus; | |
523 | ||
524 | maxcpus = sysconf(_SC_NPROCESSORS_CONF); | |
525 | if (maxcpus <= 0) { | |
526 | nr_cpus_mask = -2; | |
527 | return NULL; | |
528 | } | |
529 | /* | |
530 | * round up number of CPUs to next power of two, so we | |
531 | * can use & for modulo. | |
532 | */ | |
533 | maxcpus = 1UL << get_count_order_ulong(maxcpus); | |
534 | nr_cpus_mask = maxcpus - 1; | |
535 | } | |
536 | /* Fall-through */ | |
537 | default: | |
538 | return calloc(nr_cpus_mask + 1, sizeof(*count)); | |
539 | } | |
540 | } | |
541 | ||
542 | static | |
543 | void free_per_cpu_items_count(struct ht_items_count *count) | |
544 | { | |
98808fb1 | 545 | poison_free(count); |
df44348d MD |
546 | } |
547 | ||
548 | static | |
549 | int ht_get_cpu(void) | |
550 | { | |
551 | int cpu; | |
552 | ||
553 | assert(nr_cpus_mask >= 0); | |
554 | cpu = sched_getcpu(); | |
555 | if (unlikely(cpu < 0)) | |
556 | return cpu; | |
557 | else | |
558 | return cpu & nr_cpus_mask; | |
559 | } | |
560 | ||
561 | static | |
4105056a | 562 | void ht_count_add(struct cds_lfht *ht, unsigned long size) |
df44348d | 563 | { |
3171717f | 564 | unsigned long percpu_count; |
df44348d MD |
565 | int cpu; |
566 | ||
567 | if (unlikely(!ht->percpu_count)) | |
3171717f | 568 | return; |
df44348d MD |
569 | cpu = ht_get_cpu(); |
570 | if (unlikely(cpu < 0)) | |
3171717f MD |
571 | return; |
572 | percpu_count = uatomic_add_return(&ht->percpu_count[cpu].add, 1); | |
df44348d | 573 | if (unlikely(!(percpu_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) { |
e3ecefd6 | 574 | long count; |
df44348d MD |
575 | |
576 | dbg_printf("add percpu %lu\n", percpu_count); | |
577 | count = uatomic_add_return(&ht->count, | |
578 | 1UL << COUNT_COMMIT_ORDER); | |
579 | /* If power of 2 */ | |
580 | if (!(count & (count - 1))) { | |
4105056a | 581 | if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) < size) |
f8994aee | 582 | return; |
e3ecefd6 | 583 | dbg_printf("add set global %ld\n", count); |
4105056a | 584 | cds_lfht_resize_lazy_count(ht, size, |
6ea6bc67 | 585 | count >> (CHAIN_LEN_TARGET - 1)); |
df44348d MD |
586 | } |
587 | } | |
588 | } | |
589 | ||
590 | static | |
860d07e8 | 591 | void ht_count_del(struct cds_lfht *ht, unsigned long size) |
df44348d MD |
592 | { |
593 | unsigned long percpu_count; | |
3171717f | 594 | int cpu; |
df44348d | 595 | |
3171717f MD |
596 | if (unlikely(!ht->percpu_count)) |
597 | return; | |
598 | cpu = ht_get_cpu(); | |
599 | if (unlikely(cpu < 0)) | |
600 | return; | |
80d90c06 | 601 | percpu_count = uatomic_add_return(&ht->percpu_count[cpu].del, 1); |
df44348d | 602 | if (unlikely(!(percpu_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) { |
e3ecefd6 | 603 | long count; |
df44348d | 604 | |
860d07e8 | 605 | dbg_printf("del percpu %lu\n", percpu_count); |
df44348d | 606 | count = uatomic_add_return(&ht->count, |
3171717f | 607 | -(1UL << COUNT_COMMIT_ORDER)); |
df44348d MD |
608 | /* If power of 2 */ |
609 | if (!(count & (count - 1))) { | |
4105056a | 610 | if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) >= size) |
f8994aee | 611 | return; |
e3ecefd6 MD |
612 | dbg_printf("del set global %ld\n", count); |
613 | /* | |
c941bb9e | 614 | * Don't shrink table if the number of nodes is below a |
e3ecefd6 MD |
615 | * certain threshold. |
616 | */ | |
c941bb9e | 617 | if (count < (1UL << COUNT_COMMIT_ORDER) * (nr_cpus_mask + 1)) |
e3ecefd6 | 618 | return; |
4105056a | 619 | cds_lfht_resize_lazy_count(ht, size, |
6ea6bc67 | 620 | count >> (CHAIN_LEN_TARGET - 1)); |
df44348d MD |
621 | } |
622 | } | |
623 | } | |
624 | ||
625 | #else /* #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */ | |
626 | ||
91452a6a | 627 | static const long nr_cpus_mask = -2; |
df44348d MD |
628 | |
629 | static | |
630 | struct ht_items_count *alloc_per_cpu_items_count(void) | |
631 | { | |
632 | return NULL; | |
633 | } | |
634 | ||
635 | static | |
636 | void free_per_cpu_items_count(struct ht_items_count *count) | |
637 | { | |
638 | } | |
639 | ||
640 | static | |
4105056a | 641 | void ht_count_add(struct cds_lfht *ht, unsigned long size) |
df44348d MD |
642 | { |
643 | } | |
644 | ||
645 | static | |
860d07e8 | 646 | void ht_count_del(struct cds_lfht *ht, unsigned long size) |
df44348d MD |
647 | { |
648 | } | |
649 | ||
650 | #endif /* #else #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */ | |
651 | ||
652 | ||
f9830efd | 653 | static |
4105056a | 654 | void check_resize(struct cds_lfht *ht, unsigned long size, uint32_t chain_len) |
f9830efd | 655 | { |
f8994aee MD |
656 | unsigned long count; |
657 | ||
b8af5011 MD |
658 | if (!(ht->flags & CDS_LFHT_AUTO_RESIZE)) |
659 | return; | |
f8994aee MD |
660 | count = uatomic_read(&ht->count); |
661 | /* | |
662 | * Use bucket-local length for small table expand and for | |
663 | * environments lacking per-cpu data support. | |
664 | */ | |
665 | if (count >= (1UL << COUNT_COMMIT_ORDER)) | |
666 | return; | |
24365af7 | 667 | if (chain_len > 100) |
f0c29ed7 | 668 | dbg_printf("WARNING: large chain length: %u.\n", |
24365af7 | 669 | chain_len); |
3390d470 | 670 | if (chain_len >= CHAIN_LEN_RESIZE_THRESHOLD) |
4105056a | 671 | cds_lfht_resize_lazy(ht, size, |
01370f0b | 672 | get_count_order_u32(chain_len - (CHAIN_LEN_TARGET - 1))); |
f9830efd MD |
673 | } |
674 | ||
abc490a1 | 675 | static |
14044b37 | 676 | struct cds_lfht_node *clear_flag(struct cds_lfht_node *node) |
abc490a1 | 677 | { |
14044b37 | 678 | return (struct cds_lfht_node *) (((unsigned long) node) & ~FLAGS_MASK); |
abc490a1 MD |
679 | } |
680 | ||
681 | static | |
14044b37 | 682 | int is_removed(struct cds_lfht_node *node) |
abc490a1 | 683 | { |
d37166c6 | 684 | return ((unsigned long) node) & REMOVED_FLAG; |
abc490a1 MD |
685 | } |
686 | ||
687 | static | |
14044b37 | 688 | struct cds_lfht_node *flag_removed(struct cds_lfht_node *node) |
abc490a1 | 689 | { |
14044b37 | 690 | return (struct cds_lfht_node *) (((unsigned long) node) | REMOVED_FLAG); |
abc490a1 MD |
691 | } |
692 | ||
f5596c94 | 693 | static |
14044b37 | 694 | int is_dummy(struct cds_lfht_node *node) |
f5596c94 MD |
695 | { |
696 | return ((unsigned long) node) & DUMMY_FLAG; | |
697 | } | |
698 | ||
699 | static | |
14044b37 | 700 | struct cds_lfht_node *flag_dummy(struct cds_lfht_node *node) |
f5596c94 | 701 | { |
14044b37 | 702 | return (struct cds_lfht_node *) (((unsigned long) node) | DUMMY_FLAG); |
f5596c94 | 703 | } |
bb7b2f26 MD |
704 | |
705 | static | |
706 | struct cds_lfht_node *get_end(void) | |
707 | { | |
708 | return (struct cds_lfht_node *) END_VALUE; | |
709 | } | |
710 | ||
711 | static | |
712 | int is_end(struct cds_lfht_node *node) | |
713 | { | |
714 | return clear_flag(node) == (struct cds_lfht_node *) END_VALUE; | |
715 | } | |
716 | ||
abc490a1 | 717 | static |
f9830efd | 718 | unsigned long _uatomic_max(unsigned long *ptr, unsigned long v) |
abc490a1 MD |
719 | { |
720 | unsigned long old1, old2; | |
721 | ||
722 | old1 = uatomic_read(ptr); | |
723 | do { | |
724 | old2 = old1; | |
725 | if (old2 >= v) | |
f9830efd | 726 | return old2; |
abc490a1 | 727 | } while ((old1 = uatomic_cmpxchg(ptr, old2, v)) != old2); |
f9830efd | 728 | return v; |
abc490a1 MD |
729 | } |
730 | ||
f4a9cc0b LJ |
731 | static |
732 | struct _cds_lfht_node *lookup_bucket(struct cds_lfht *ht, unsigned long size, | |
733 | unsigned long hash) | |
734 | { | |
735 | unsigned long index, order; | |
736 | ||
737 | assert(size > 0); | |
738 | index = hash & (size - 1); | |
a4ea2223 LJ |
739 | /* |
740 | * equivalent to get_count_order_ulong(index + 1), but optimizes | |
741 | * away the non-existing 0 special-case for | |
742 | * get_count_order_ulong. | |
743 | */ | |
744 | order = fls_ulong(index); | |
f4a9cc0b LJ |
745 | |
746 | dbg_printf("lookup hash %lu index %lu order %lu aridx %lu\n", | |
747 | hash, index, order, index & (!order ? 0 : ((1UL << (order - 1)) - 1))); | |
748 | ||
749 | return &ht->t.tbl[order]->nodes[index & (!order ? 0 : ((1UL << (order - 1)) - 1))]; | |
750 | } | |
751 | ||
273399de MD |
752 | /* |
753 | * Remove all logically deleted nodes from a bucket up to a certain node key. | |
754 | */ | |
755 | static | |
f9c80341 | 756 | void _cds_lfht_gc_bucket(struct cds_lfht_node *dummy, struct cds_lfht_node *node) |
273399de | 757 | { |
14044b37 | 758 | struct cds_lfht_node *iter_prev, *iter, *next, *new_next; |
273399de | 759 | |
c90201ac MD |
760 | assert(!is_dummy(dummy)); |
761 | assert(!is_removed(dummy)); | |
762 | assert(!is_dummy(node)); | |
763 | assert(!is_removed(node)); | |
273399de MD |
764 | for (;;) { |
765 | iter_prev = dummy; | |
766 | /* We can always skip the dummy node initially */ | |
cc4fcb10 | 767 | iter = rcu_dereference(iter_prev->p.next); |
b4cb483f | 768 | assert(!is_removed(iter)); |
cc4fcb10 | 769 | assert(iter_prev->p.reverse_hash <= node->p.reverse_hash); |
bd4db153 MD |
770 | /* |
771 | * We should never be called with dummy (start of chain) | |
772 | * and logically removed node (end of path compression | |
773 | * marker) being the actual same node. This would be a | |
774 | * bug in the algorithm implementation. | |
775 | */ | |
776 | assert(dummy != node); | |
273399de | 777 | for (;;) { |
bb7b2f26 | 778 | if (unlikely(is_end(iter))) |
f9c80341 | 779 | return; |
76412f24 | 780 | if (likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash)) |
f9c80341 | 781 | return; |
cc4fcb10 | 782 | next = rcu_dereference(clear_flag(iter)->p.next); |
b198f0fd | 783 | if (likely(is_removed(next))) |
273399de | 784 | break; |
b453eae1 | 785 | iter_prev = clear_flag(iter); |
273399de MD |
786 | iter = next; |
787 | } | |
b198f0fd | 788 | assert(!is_removed(iter)); |
f5596c94 MD |
789 | if (is_dummy(iter)) |
790 | new_next = flag_dummy(clear_flag(next)); | |
791 | else | |
792 | new_next = clear_flag(next); | |
793 | (void) uatomic_cmpxchg(&iter_prev->p.next, iter, new_next); | |
273399de | 794 | } |
f9c80341 | 795 | return; |
273399de MD |
796 | } |
797 | ||
9357c415 MD |
798 | static |
799 | int _cds_lfht_replace(struct cds_lfht *ht, unsigned long size, | |
800 | struct cds_lfht_node *old_node, | |
3fb86f26 | 801 | struct cds_lfht_node *old_next, |
9357c415 MD |
802 | struct cds_lfht_node *new_node) |
803 | { | |
3fb86f26 | 804 | struct cds_lfht_node *dummy, *ret_next; |
9357c415 | 805 | struct _cds_lfht_node *lookup; |
9357c415 MD |
806 | |
807 | if (!old_node) /* Return -ENOENT if asked to replace NULL node */ | |
7801dadd | 808 | return -ENOENT; |
9357c415 MD |
809 | |
810 | assert(!is_removed(old_node)); | |
811 | assert(!is_dummy(old_node)); | |
812 | assert(!is_removed(new_node)); | |
813 | assert(!is_dummy(new_node)); | |
814 | assert(new_node != old_node); | |
3fb86f26 | 815 | for (;;) { |
9357c415 | 816 | /* Insert after node to be replaced */ |
9357c415 MD |
817 | if (is_removed(old_next)) { |
818 | /* | |
819 | * Too late, the old node has been removed under us | |
820 | * between lookup and replace. Fail. | |
821 | */ | |
7801dadd | 822 | return -ENOENT; |
9357c415 MD |
823 | } |
824 | assert(!is_dummy(old_next)); | |
825 | assert(new_node != clear_flag(old_next)); | |
826 | new_node->p.next = clear_flag(old_next); | |
827 | /* | |
828 | * Here is the whole trick for lock-free replace: we add | |
829 | * the replacement node _after_ the node we want to | |
830 | * replace by atomically setting its next pointer at the | |
831 | * same time we set its removal flag. Given that | |
832 | * the lookups/get next use an iterator aware of the | |
833 | * next pointer, they will either skip the old node due | |
834 | * to the removal flag and see the new node, or use | |
835 | * the old node, but will not see the new one. | |
836 | */ | |
837 | ret_next = uatomic_cmpxchg(&old_node->p.next, | |
838 | old_next, flag_removed(new_node)); | |
3fb86f26 | 839 | if (ret_next == old_next) |
7801dadd | 840 | break; /* We performed the replacement. */ |
3fb86f26 LJ |
841 | old_next = ret_next; |
842 | } | |
9357c415 | 843 | |
9357c415 MD |
844 | /* |
845 | * Ensure that the old node is not visible to readers anymore: | |
846 | * lookup for the node, and remove it (along with any other | |
847 | * logically removed node) if found. | |
848 | */ | |
f4a9cc0b | 849 | lookup = lookup_bucket(ht, size, bit_reverse_ulong(old_node->p.reverse_hash)); |
9357c415 MD |
850 | dummy = (struct cds_lfht_node *) lookup; |
851 | _cds_lfht_gc_bucket(dummy, new_node); | |
7801dadd LJ |
852 | |
853 | assert(is_removed(rcu_dereference(old_node->p.next))); | |
854 | return 0; | |
9357c415 MD |
855 | } |
856 | ||
83beee94 MD |
857 | /* |
858 | * A non-NULL unique_ret pointer uses the "add unique" (or uniquify) add | |
859 | * mode. A NULL unique_ret allows creation of duplicate keys. | |
860 | */ | |
abc490a1 | 861 | static |
83beee94 MD |
862 | void _cds_lfht_add(struct cds_lfht *ht, |
863 | unsigned long size, | |
864 | struct cds_lfht_node *node, | |
865 | struct cds_lfht_iter *unique_ret, | |
866 | int dummy) | |
abc490a1 | 867 | { |
14044b37 | 868 | struct cds_lfht_node *iter_prev, *iter, *next, *new_node, *new_next, |
960c9e4f | 869 | *return_node; |
14044b37 | 870 | struct _cds_lfht_node *lookup; |
abc490a1 | 871 | |
c90201ac MD |
872 | assert(!is_dummy(node)); |
873 | assert(!is_removed(node)); | |
4105056a | 874 | if (!size) { |
f5596c94 | 875 | assert(dummy); |
83beee94 | 876 | assert(!unique_ret); |
bb7b2f26 | 877 | node->p.next = flag_dummy(get_end()); |
83beee94 | 878 | return; /* Initial first add (head) */ |
18117871 | 879 | } |
f4a9cc0b | 880 | lookup = lookup_bucket(ht, size, bit_reverse_ulong(node->p.reverse_hash)); |
abc490a1 | 881 | for (;;) { |
adc0de68 | 882 | uint32_t chain_len = 0; |
abc490a1 | 883 | |
11519af6 MD |
884 | /* |
885 | * iter_prev points to the non-removed node prior to the | |
886 | * insert location. | |
11519af6 | 887 | */ |
14044b37 | 888 | iter_prev = (struct cds_lfht_node *) lookup; |
11519af6 | 889 | /* We can always skip the dummy node initially */ |
cc4fcb10 MD |
890 | iter = rcu_dereference(iter_prev->p.next); |
891 | assert(iter_prev->p.reverse_hash <= node->p.reverse_hash); | |
abc490a1 | 892 | for (;;) { |
bb7b2f26 | 893 | if (unlikely(is_end(iter))) |
273399de | 894 | goto insert; |
76412f24 | 895 | if (likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash)) |
273399de | 896 | goto insert; |
238cc06e | 897 | |
194fdbd1 LJ |
898 | /* dummy node is the first node of the identical-hash-value chain */ |
899 | if (dummy && clear_flag(iter)->p.reverse_hash == node->p.reverse_hash) | |
900 | goto insert; | |
238cc06e | 901 | |
cc4fcb10 | 902 | next = rcu_dereference(clear_flag(iter)->p.next); |
b198f0fd | 903 | if (unlikely(is_removed(next))) |
9dba85be | 904 | goto gc_node; |
238cc06e LJ |
905 | |
906 | /* uniquely add */ | |
83beee94 | 907 | if (unique_ret |
1b81fe1a | 908 | && !is_dummy(next) |
238cc06e LJ |
909 | && clear_flag(iter)->p.reverse_hash == node->p.reverse_hash) { |
910 | struct cds_lfht_iter d_iter = { .node = node, .next = iter, }; | |
911 | ||
912 | /* | |
913 | * uniquely adding inserts the node as the first | |
914 | * node of the identical-hash-value node chain. | |
915 | * | |
916 | * This semantic ensures no duplicated keys | |
917 | * should ever be observable in the table | |
918 | * (including observe one node by one node | |
919 | * by forward iterations) | |
920 | */ | |
921 | cds_lfht_next_duplicate(ht, &d_iter); | |
922 | if (!d_iter.node) | |
923 | goto insert; | |
924 | ||
925 | *unique_ret = d_iter; | |
83beee94 | 926 | return; |
48ed1c18 | 927 | } |
238cc06e | 928 | |
11519af6 | 929 | /* Only account for identical reverse hash once */ |
24365af7 MD |
930 | if (iter_prev->p.reverse_hash != clear_flag(iter)->p.reverse_hash |
931 | && !is_dummy(next)) | |
4105056a | 932 | check_resize(ht, size, ++chain_len); |
11519af6 | 933 | iter_prev = clear_flag(iter); |
273399de | 934 | iter = next; |
abc490a1 | 935 | } |
48ed1c18 | 936 | |
273399de | 937 | insert: |
7ec59d3b | 938 | assert(node != clear_flag(iter)); |
11519af6 | 939 | assert(!is_removed(iter_prev)); |
c90201ac | 940 | assert(!is_removed(iter)); |
f000907d | 941 | assert(iter_prev != node); |
f9c80341 | 942 | if (!dummy) |
1b81fe1a | 943 | node->p.next = clear_flag(iter); |
f9c80341 MD |
944 | else |
945 | node->p.next = flag_dummy(clear_flag(iter)); | |
f5596c94 MD |
946 | if (is_dummy(iter)) |
947 | new_node = flag_dummy(node); | |
948 | else | |
949 | new_node = node; | |
cc4fcb10 | 950 | if (uatomic_cmpxchg(&iter_prev->p.next, iter, |
48ed1c18 | 951 | new_node) != iter) { |
273399de | 952 | continue; /* retry */ |
48ed1c18 | 953 | } else { |
83beee94 | 954 | return_node = node; |
960c9e4f | 955 | goto end; |
48ed1c18 MD |
956 | } |
957 | ||
9dba85be MD |
958 | gc_node: |
959 | assert(!is_removed(iter)); | |
f5596c94 MD |
960 | if (is_dummy(iter)) |
961 | new_next = flag_dummy(clear_flag(next)); | |
962 | else | |
963 | new_next = clear_flag(next); | |
964 | (void) uatomic_cmpxchg(&iter_prev->p.next, iter, new_next); | |
273399de | 965 | /* retry */ |
464a1ec9 | 966 | } |
9357c415 | 967 | end: |
83beee94 MD |
968 | if (unique_ret) { |
969 | unique_ret->node = return_node; | |
970 | /* unique_ret->next left unset, never used. */ | |
971 | } | |
abc490a1 | 972 | } |
464a1ec9 | 973 | |
abc490a1 | 974 | static |
860d07e8 | 975 | int _cds_lfht_del(struct cds_lfht *ht, unsigned long size, |
4105056a | 976 | struct cds_lfht_node *node, |
b198f0fd | 977 | int dummy_removal) |
abc490a1 | 978 | { |
14044b37 MD |
979 | struct cds_lfht_node *dummy, *next, *old; |
980 | struct _cds_lfht_node *lookup; | |
5e28c532 | 981 | |
9357c415 | 982 | if (!node) /* Return -ENOENT if asked to delete NULL node */ |
743f9143 | 983 | return -ENOENT; |
9357c415 | 984 | |
7ec59d3b | 985 | /* logically delete the node */ |
c90201ac MD |
986 | assert(!is_dummy(node)); |
987 | assert(!is_removed(node)); | |
cc4fcb10 | 988 | old = rcu_dereference(node->p.next); |
7ec59d3b | 989 | do { |
48ed1c18 MD |
990 | struct cds_lfht_node *new_next; |
991 | ||
7ec59d3b | 992 | next = old; |
76412f24 | 993 | if (unlikely(is_removed(next))) |
743f9143 | 994 | return -ENOENT; |
1475579c MD |
995 | if (dummy_removal) |
996 | assert(is_dummy(next)); | |
997 | else | |
998 | assert(!is_dummy(next)); | |
48ed1c18 | 999 | new_next = flag_removed(next); |
48ed1c18 | 1000 | old = uatomic_cmpxchg(&node->p.next, next, new_next); |
7ec59d3b | 1001 | } while (old != next); |
7ec59d3b | 1002 | /* We performed the (logical) deletion. */ |
7ec59d3b MD |
1003 | |
1004 | /* | |
1005 | * Ensure that the node is not visible to readers anymore: lookup for | |
273399de MD |
1006 | * the node, and remove it (along with any other logically removed node) |
1007 | * if found. | |
11519af6 | 1008 | */ |
f4a9cc0b | 1009 | lookup = lookup_bucket(ht, size, bit_reverse_ulong(node->p.reverse_hash)); |
14044b37 | 1010 | dummy = (struct cds_lfht_node *) lookup; |
f9c80341 | 1011 | _cds_lfht_gc_bucket(dummy, node); |
743f9143 LJ |
1012 | |
1013 | assert(is_removed(rcu_dereference(node->p.next))); | |
1014 | return 0; | |
abc490a1 | 1015 | } |
2ed95849 | 1016 | |
b7d619b0 MD |
1017 | static |
1018 | void *partition_resize_thread(void *arg) | |
1019 | { | |
1020 | struct partition_resize_work *work = arg; | |
1021 | ||
1022 | work->ht->cds_lfht_rcu_register_thread(); | |
1023 | work->fct(work->ht, work->i, work->start, work->len); | |
1024 | work->ht->cds_lfht_rcu_unregister_thread(); | |
1025 | return NULL; | |
1026 | } | |
1027 | ||
1028 | static | |
1029 | void partition_resize_helper(struct cds_lfht *ht, unsigned long i, | |
1030 | unsigned long len, | |
1031 | void (*fct)(struct cds_lfht *ht, unsigned long i, | |
1032 | unsigned long start, unsigned long len)) | |
1033 | { | |
1034 | unsigned long partition_len; | |
1035 | struct partition_resize_work *work; | |
6083a889 MD |
1036 | int thread, ret; |
1037 | unsigned long nr_threads; | |
b7d619b0 | 1038 | |
6083a889 MD |
1039 | /* |
1040 | * Note: nr_cpus_mask + 1 is always power of 2. | |
1041 | * We spawn just the number of threads we need to satisfy the minimum | |
1042 | * partition size, up to the number of CPUs in the system. | |
1043 | */ | |
91452a6a MD |
1044 | if (nr_cpus_mask > 0) { |
1045 | nr_threads = min(nr_cpus_mask + 1, | |
1046 | len >> MIN_PARTITION_PER_THREAD_ORDER); | |
1047 | } else { | |
1048 | nr_threads = 1; | |
1049 | } | |
6083a889 MD |
1050 | partition_len = len >> get_count_order_ulong(nr_threads); |
1051 | work = calloc(nr_threads, sizeof(*work)); | |
b7d619b0 | 1052 | assert(work); |
6083a889 MD |
1053 | for (thread = 0; thread < nr_threads; thread++) { |
1054 | work[thread].ht = ht; | |
1055 | work[thread].i = i; | |
1056 | work[thread].len = partition_len; | |
1057 | work[thread].start = thread * partition_len; | |
1058 | work[thread].fct = fct; | |
1af6e26e | 1059 | ret = pthread_create(&(work[thread].thread_id), ht->resize_attr, |
6083a889 | 1060 | partition_resize_thread, &work[thread]); |
b7d619b0 MD |
1061 | assert(!ret); |
1062 | } | |
6083a889 | 1063 | for (thread = 0; thread < nr_threads; thread++) { |
1af6e26e | 1064 | ret = pthread_join(work[thread].thread_id, NULL); |
b7d619b0 MD |
1065 | assert(!ret); |
1066 | } | |
1067 | free(work); | |
b7d619b0 MD |
1068 | } |
1069 | ||
e8de508e MD |
1070 | /* |
1071 | * Holding RCU read lock to protect _cds_lfht_add against memory | |
1072 | * reclaim that could be performed by other call_rcu worker threads (ABA | |
1073 | * problem). | |
9ee0fc9a | 1074 | * |
b7d619b0 | 1075 | * When we reach a certain length, we can split this population phase over |
9ee0fc9a MD |
1076 | * many worker threads, based on the number of CPUs available in the system. |
1077 | * This should therefore take care of not having the expand lagging behind too | |
1078 | * many concurrent insertion threads by using the scheduler's ability to | |
1079 | * schedule dummy node population fairly with insertions. | |
e8de508e | 1080 | */ |
4105056a | 1081 | static |
b7d619b0 MD |
1082 | void init_table_populate_partition(struct cds_lfht *ht, unsigned long i, |
1083 | unsigned long start, unsigned long len) | |
4105056a MD |
1084 | { |
1085 | unsigned long j; | |
1086 | ||
1087 | ht->cds_lfht_rcu_read_lock(); | |
b7d619b0 | 1088 | for (j = start; j < start + len; j++) { |
4105056a MD |
1089 | struct cds_lfht_node *new_node = |
1090 | (struct cds_lfht_node *) &ht->t.tbl[i]->nodes[j]; | |
1091 | ||
dc1da8f6 | 1092 | dbg_printf("init populate: i %lu j %lu hash %lu\n", |
4105056a | 1093 | i, j, !i ? 0 : (1UL << (i - 1)) + j); |
dc1da8f6 MD |
1094 | new_node->p.reverse_hash = |
1095 | bit_reverse_ulong(!i ? 0 : (1UL << (i - 1)) + j); | |
83beee94 MD |
1096 | _cds_lfht_add(ht, !i ? 0 : (1UL << (i - 1)), |
1097 | new_node, NULL, 1); | |
4105056a MD |
1098 | } |
1099 | ht->cds_lfht_rcu_read_unlock(); | |
b7d619b0 MD |
1100 | } |
1101 | ||
1102 | static | |
1103 | void init_table_populate(struct cds_lfht *ht, unsigned long i, | |
1104 | unsigned long len) | |
1105 | { | |
1106 | assert(nr_cpus_mask != -1); | |
6083a889 | 1107 | if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD) { |
b7d619b0 MD |
1108 | ht->cds_lfht_rcu_thread_online(); |
1109 | init_table_populate_partition(ht, i, 0, len); | |
1110 | ht->cds_lfht_rcu_thread_offline(); | |
1111 | return; | |
1112 | } | |
1113 | partition_resize_helper(ht, i, len, init_table_populate_partition); | |
4105056a MD |
1114 | } |
1115 | ||
abc490a1 | 1116 | static |
4105056a | 1117 | void init_table(struct cds_lfht *ht, |
93d46c39 | 1118 | unsigned long first_order, unsigned long last_order) |
24365af7 | 1119 | { |
93d46c39 | 1120 | unsigned long i; |
24365af7 | 1121 | |
93d46c39 LJ |
1122 | dbg_printf("init table: first_order %lu last_order %lu\n", |
1123 | first_order, last_order); | |
1124 | for (i = first_order; i <= last_order; i++) { | |
4105056a | 1125 | unsigned long len; |
24365af7 MD |
1126 | |
1127 | len = !i ? 1 : 1UL << (i - 1); | |
f0c29ed7 | 1128 | dbg_printf("init order %lu len: %lu\n", i, len); |
4d676753 MD |
1129 | |
1130 | /* Stop expand if the resize target changes under us */ | |
1131 | if (CMM_LOAD_SHARED(ht->t.resize_target) < (!i ? 1 : (1UL << i))) | |
1132 | break; | |
1133 | ||
0d14ceb2 | 1134 | ht->t.tbl[i] = calloc(1, len * sizeof(struct _cds_lfht_node)); |
b7d619b0 | 1135 | assert(ht->t.tbl[i]); |
4105056a | 1136 | |
4105056a | 1137 | /* |
dc1da8f6 MD |
1138 | * Set all dummy nodes reverse hash values for a level and |
1139 | * link all dummy nodes into the table. | |
4105056a | 1140 | */ |
dc1da8f6 | 1141 | init_table_populate(ht, i, len); |
4105056a | 1142 | |
f9c80341 MD |
1143 | /* |
1144 | * Update table size. | |
1145 | */ | |
1146 | cmm_smp_wmb(); /* populate data before RCU size */ | |
1147 | CMM_STORE_SHARED(ht->t.size, !i ? 1 : (1UL << i)); | |
1148 | ||
4105056a MD |
1149 | dbg_printf("init new size: %lu\n", !i ? 1 : (1UL << i)); |
1150 | if (CMM_LOAD_SHARED(ht->in_progress_destroy)) | |
1151 | break; | |
1152 | } | |
1153 | } | |
1154 | ||
e8de508e MD |
1155 | /* |
1156 | * Holding RCU read lock to protect _cds_lfht_remove against memory | |
1157 | * reclaim that could be performed by other call_rcu worker threads (ABA | |
1158 | * problem). | |
1159 | * For a single level, we logically remove and garbage collect each node. | |
1160 | * | |
1161 | * As a design choice, we perform logical removal and garbage collection on a | |
1162 | * node-per-node basis to simplify this algorithm. We also assume keeping good | |
1163 | * cache locality of the operation would overweight possible performance gain | |
1164 | * that could be achieved by batching garbage collection for multiple levels. | |
1165 | * However, this would have to be justified by benchmarks. | |
1166 | * | |
1167 | * Concurrent removal and add operations are helping us perform garbage | |
1168 | * collection of logically removed nodes. We guarantee that all logically | |
1169 | * removed nodes have been garbage-collected (unlinked) before call_rcu is | |
1170 | * invoked to free a hole level of dummy nodes (after a grace period). | |
1171 | * | |
1172 | * Logical removal and garbage collection can therefore be done in batch or on a | |
1173 | * node-per-node basis, as long as the guarantee above holds. | |
9ee0fc9a | 1174 | * |
b7d619b0 MD |
1175 | * When we reach a certain length, we can split this removal over many worker |
1176 | * threads, based on the number of CPUs available in the system. This should | |
1177 | * take care of not letting resize process lag behind too many concurrent | |
9ee0fc9a | 1178 | * updater threads actively inserting into the hash table. |
e8de508e | 1179 | */ |
4105056a | 1180 | static |
b7d619b0 MD |
1181 | void remove_table_partition(struct cds_lfht *ht, unsigned long i, |
1182 | unsigned long start, unsigned long len) | |
4105056a MD |
1183 | { |
1184 | unsigned long j; | |
1185 | ||
1186 | ht->cds_lfht_rcu_read_lock(); | |
b7d619b0 | 1187 | for (j = start; j < start + len; j++) { |
4105056a MD |
1188 | struct cds_lfht_node *fini_node = |
1189 | (struct cds_lfht_node *) &ht->t.tbl[i]->nodes[j]; | |
1190 | ||
1191 | dbg_printf("remove entry: i %lu j %lu hash %lu\n", | |
1192 | i, j, !i ? 0 : (1UL << (i - 1)) + j); | |
1193 | fini_node->p.reverse_hash = | |
1194 | bit_reverse_ulong(!i ? 0 : (1UL << (i - 1)) + j); | |
860d07e8 | 1195 | (void) _cds_lfht_del(ht, !i ? 0 : (1UL << (i - 1)), |
b198f0fd | 1196 | fini_node, 1); |
abc490a1 | 1197 | } |
4105056a | 1198 | ht->cds_lfht_rcu_read_unlock(); |
b7d619b0 MD |
1199 | } |
1200 | ||
1201 | static | |
1202 | void remove_table(struct cds_lfht *ht, unsigned long i, unsigned long len) | |
1203 | { | |
1204 | ||
1205 | assert(nr_cpus_mask != -1); | |
6083a889 | 1206 | if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD) { |
b7d619b0 MD |
1207 | ht->cds_lfht_rcu_thread_online(); |
1208 | remove_table_partition(ht, i, 0, len); | |
1209 | ht->cds_lfht_rcu_thread_offline(); | |
1210 | return; | |
1211 | } | |
1212 | partition_resize_helper(ht, i, len, remove_table_partition); | |
2ed95849 MD |
1213 | } |
1214 | ||
1475579c | 1215 | static |
4105056a | 1216 | void fini_table(struct cds_lfht *ht, |
93d46c39 | 1217 | unsigned long first_order, unsigned long last_order) |
1475579c | 1218 | { |
93d46c39 | 1219 | long i; |
0d14ceb2 | 1220 | void *free_by_rcu = NULL; |
1475579c | 1221 | |
93d46c39 LJ |
1222 | dbg_printf("fini table: first_order %lu last_order %lu\n", |
1223 | first_order, last_order); | |
1475579c | 1224 | assert(first_order > 0); |
93d46c39 | 1225 | for (i = last_order; i >= first_order; i--) { |
4105056a | 1226 | unsigned long len; |
1475579c MD |
1227 | |
1228 | len = !i ? 1 : 1UL << (i - 1); | |
1229 | dbg_printf("fini order %lu len: %lu\n", i, len); | |
4105056a | 1230 | |
4d676753 MD |
1231 | /* Stop shrink if the resize target changes under us */ |
1232 | if (CMM_LOAD_SHARED(ht->t.resize_target) > (1UL << (i - 1))) | |
1233 | break; | |
1234 | ||
1235 | cmm_smp_wmb(); /* populate data before RCU size */ | |
1236 | CMM_STORE_SHARED(ht->t.size, 1UL << (i - 1)); | |
1237 | ||
1238 | /* | |
1239 | * We need to wait for all add operations to reach Q.S. (and | |
1240 | * thus use the new table for lookups) before we can start | |
1241 | * releasing the old dummy nodes. Otherwise their lookup will | |
1242 | * return a logically removed node as insert position. | |
1243 | */ | |
1244 | ht->cds_lfht_synchronize_rcu(); | |
0d14ceb2 LJ |
1245 | if (free_by_rcu) |
1246 | free(free_by_rcu); | |
4d676753 | 1247 | |
21263e21 | 1248 | /* |
4105056a MD |
1249 | * Set "removed" flag in dummy nodes about to be removed. |
1250 | * Unlink all now-logically-removed dummy node pointers. | |
1251 | * Concurrent add/remove operation are helping us doing | |
1252 | * the gc. | |
21263e21 | 1253 | */ |
4105056a MD |
1254 | remove_table(ht, i, len); |
1255 | ||
0d14ceb2 | 1256 | free_by_rcu = ht->t.tbl[i]; |
4105056a MD |
1257 | |
1258 | dbg_printf("fini new size: %lu\n", 1UL << i); | |
1475579c MD |
1259 | if (CMM_LOAD_SHARED(ht->in_progress_destroy)) |
1260 | break; | |
1261 | } | |
0d14ceb2 LJ |
1262 | |
1263 | if (free_by_rcu) { | |
1264 | ht->cds_lfht_synchronize_rcu(); | |
1265 | free(free_by_rcu); | |
1266 | } | |
1475579c MD |
1267 | } |
1268 | ||
7a9dcf9b | 1269 | struct cds_lfht *_cds_lfht_new(cds_lfht_hash_fct hash_fct, |
14044b37 MD |
1270 | cds_lfht_compare_fct compare_fct, |
1271 | unsigned long hash_seed, | |
1272 | unsigned long init_size, | |
b8af5011 | 1273 | int flags, |
14044b37 | 1274 | void (*cds_lfht_call_rcu)(struct rcu_head *head, |
1475579c | 1275 | void (*func)(struct rcu_head *head)), |
01dbfa62 MD |
1276 | void (*cds_lfht_synchronize_rcu)(void), |
1277 | void (*cds_lfht_rcu_read_lock)(void), | |
5f511391 MD |
1278 | void (*cds_lfht_rcu_read_unlock)(void), |
1279 | void (*cds_lfht_rcu_thread_offline)(void), | |
b7d619b0 MD |
1280 | void (*cds_lfht_rcu_thread_online)(void), |
1281 | void (*cds_lfht_rcu_register_thread)(void), | |
1282 | void (*cds_lfht_rcu_unregister_thread)(void), | |
1283 | pthread_attr_t *attr) | |
abc490a1 | 1284 | { |
14044b37 | 1285 | struct cds_lfht *ht; |
24365af7 | 1286 | unsigned long order; |
abc490a1 | 1287 | |
8129be4e | 1288 | /* init_size must be power of two */ |
49619ea0 | 1289 | if (init_size && (init_size & (init_size - 1))) |
8129be4e | 1290 | return NULL; |
14044b37 | 1291 | ht = calloc(1, sizeof(struct cds_lfht)); |
b7d619b0 | 1292 | assert(ht); |
abc490a1 | 1293 | ht->hash_fct = hash_fct; |
732ad076 MD |
1294 | ht->compare_fct = compare_fct; |
1295 | ht->hash_seed = hash_seed; | |
14044b37 | 1296 | ht->cds_lfht_call_rcu = cds_lfht_call_rcu; |
1475579c | 1297 | ht->cds_lfht_synchronize_rcu = cds_lfht_synchronize_rcu; |
01dbfa62 MD |
1298 | ht->cds_lfht_rcu_read_lock = cds_lfht_rcu_read_lock; |
1299 | ht->cds_lfht_rcu_read_unlock = cds_lfht_rcu_read_unlock; | |
5f511391 MD |
1300 | ht->cds_lfht_rcu_thread_offline = cds_lfht_rcu_thread_offline; |
1301 | ht->cds_lfht_rcu_thread_online = cds_lfht_rcu_thread_online; | |
b7d619b0 MD |
1302 | ht->cds_lfht_rcu_register_thread = cds_lfht_rcu_register_thread; |
1303 | ht->cds_lfht_rcu_unregister_thread = cds_lfht_rcu_unregister_thread; | |
1304 | ht->resize_attr = attr; | |
df44348d | 1305 | ht->percpu_count = alloc_per_cpu_items_count(); |
abc490a1 MD |
1306 | /* this mutex should not nest in read-side C.S. */ |
1307 | pthread_mutex_init(&ht->resize_mutex, NULL); | |
93d46c39 | 1308 | order = get_count_order_ulong(max(init_size, MIN_TABLE_SIZE)); |
b8af5011 | 1309 | ht->flags = flags; |
5f511391 | 1310 | ht->cds_lfht_rcu_thread_offline(); |
f000907d | 1311 | pthread_mutex_lock(&ht->resize_mutex); |
93d46c39 | 1312 | ht->t.resize_target = 1UL << order; |
4105056a | 1313 | init_table(ht, 0, order); |
f000907d | 1314 | pthread_mutex_unlock(&ht->resize_mutex); |
5f511391 | 1315 | ht->cds_lfht_rcu_thread_online(); |
abc490a1 MD |
1316 | return ht; |
1317 | } | |
1318 | ||
adc0de68 MD |
1319 | void cds_lfht_lookup(struct cds_lfht *ht, void *key, size_t key_len, |
1320 | struct cds_lfht_iter *iter) | |
2ed95849 | 1321 | { |
bb7b2f26 | 1322 | struct cds_lfht_node *node, *next, *dummy_node; |
14044b37 | 1323 | struct _cds_lfht_node *lookup; |
f4a9cc0b | 1324 | unsigned long hash, reverse_hash, size; |
2ed95849 | 1325 | |
732ad076 | 1326 | hash = ht->hash_fct(key, key_len, ht->hash_seed); |
abc490a1 | 1327 | reverse_hash = bit_reverse_ulong(hash); |
464a1ec9 | 1328 | |
4105056a | 1329 | size = rcu_dereference(ht->t.size); |
f4a9cc0b | 1330 | lookup = lookup_bucket(ht, size, hash); |
bb7b2f26 MD |
1331 | dummy_node = (struct cds_lfht_node *) lookup; |
1332 | /* We can always skip the dummy node initially */ | |
1333 | node = rcu_dereference(dummy_node->p.next); | |
bb7b2f26 | 1334 | node = clear_flag(node); |
2ed95849 | 1335 | for (;;) { |
bb7b2f26 | 1336 | if (unlikely(is_end(node))) { |
96ad1112 | 1337 | node = next = NULL; |
abc490a1 | 1338 | break; |
bb7b2f26 | 1339 | } |
cc4fcb10 | 1340 | if (unlikely(node->p.reverse_hash > reverse_hash)) { |
96ad1112 | 1341 | node = next = NULL; |
abc490a1 | 1342 | break; |
2ed95849 | 1343 | } |
1b81fe1a | 1344 | next = rcu_dereference(node->p.next); |
adc0de68 | 1345 | if (likely(!is_removed(next)) |
1b81fe1a | 1346 | && !is_dummy(next) |
ae450da3 | 1347 | && clear_flag(node)->p.reverse_hash == reverse_hash |
49c2e2d6 | 1348 | && likely(!ht->compare_fct(node->key, node->key_len, key, key_len))) { |
273399de | 1349 | break; |
2ed95849 | 1350 | } |
1b81fe1a | 1351 | node = clear_flag(next); |
2ed95849 | 1352 | } |
1b81fe1a | 1353 | assert(!node || !is_dummy(rcu_dereference(node->p.next))); |
adc0de68 MD |
1354 | iter->node = node; |
1355 | iter->next = next; | |
abc490a1 | 1356 | } |
e0ba718a | 1357 | |
3883c0e5 | 1358 | void cds_lfht_next_duplicate(struct cds_lfht *ht, struct cds_lfht_iter *iter) |
a481e5ff | 1359 | { |
adc0de68 | 1360 | struct cds_lfht_node *node, *next; |
a481e5ff MD |
1361 | unsigned long reverse_hash; |
1362 | void *key; | |
1363 | size_t key_len; | |
1364 | ||
adc0de68 | 1365 | node = iter->node; |
a481e5ff MD |
1366 | reverse_hash = node->p.reverse_hash; |
1367 | key = node->key; | |
1368 | key_len = node->key_len; | |
adc0de68 | 1369 | next = iter->next; |
a481e5ff MD |
1370 | node = clear_flag(next); |
1371 | ||
1372 | for (;;) { | |
bb7b2f26 | 1373 | if (unlikely(is_end(node))) { |
96ad1112 | 1374 | node = next = NULL; |
a481e5ff | 1375 | break; |
bb7b2f26 | 1376 | } |
a481e5ff | 1377 | if (unlikely(node->p.reverse_hash > reverse_hash)) { |
96ad1112 | 1378 | node = next = NULL; |
a481e5ff MD |
1379 | break; |
1380 | } | |
1381 | next = rcu_dereference(node->p.next); | |
adc0de68 | 1382 | if (likely(!is_removed(next)) |
a481e5ff MD |
1383 | && !is_dummy(next) |
1384 | && likely(!ht->compare_fct(node->key, node->key_len, key, key_len))) { | |
1385 | break; | |
1386 | } | |
1387 | node = clear_flag(next); | |
1388 | } | |
1389 | assert(!node || !is_dummy(rcu_dereference(node->p.next))); | |
adc0de68 MD |
1390 | iter->node = node; |
1391 | iter->next = next; | |
a481e5ff MD |
1392 | } |
1393 | ||
4e9b9fbf MD |
1394 | void cds_lfht_next(struct cds_lfht *ht, struct cds_lfht_iter *iter) |
1395 | { | |
1396 | struct cds_lfht_node *node, *next; | |
1397 | ||
853395e1 | 1398 | node = clear_flag(iter->next); |
4e9b9fbf MD |
1399 | for (;;) { |
1400 | if (unlikely(is_end(node))) { | |
1401 | node = next = NULL; | |
1402 | break; | |
1403 | } | |
1404 | next = rcu_dereference(node->p.next); | |
1405 | if (likely(!is_removed(next)) | |
1406 | && !is_dummy(next)) { | |
1407 | break; | |
1408 | } | |
1409 | node = clear_flag(next); | |
1410 | } | |
1411 | assert(!node || !is_dummy(rcu_dereference(node->p.next))); | |
1412 | iter->node = node; | |
1413 | iter->next = next; | |
1414 | } | |
1415 | ||
1416 | void cds_lfht_first(struct cds_lfht *ht, struct cds_lfht_iter *iter) | |
1417 | { | |
1418 | struct _cds_lfht_node *lookup; | |
1419 | ||
1420 | /* | |
1421 | * Get next after first dummy node. The first dummy node is the | |
1422 | * first node of the linked list. | |
1423 | */ | |
1424 | lookup = &ht->t.tbl[0]->nodes[0]; | |
853395e1 | 1425 | iter->next = lookup->next; |
4e9b9fbf MD |
1426 | cds_lfht_next(ht, iter); |
1427 | } | |
1428 | ||
14044b37 | 1429 | void cds_lfht_add(struct cds_lfht *ht, struct cds_lfht_node *node) |
abc490a1 | 1430 | { |
4105056a | 1431 | unsigned long hash, size; |
ab7d5fc6 | 1432 | |
49c2e2d6 | 1433 | hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed); |
cc4fcb10 | 1434 | node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash); |
2ed95849 | 1435 | |
4105056a | 1436 | size = rcu_dereference(ht->t.size); |
83beee94 | 1437 | _cds_lfht_add(ht, size, node, NULL, 0); |
4105056a | 1438 | ht_count_add(ht, size); |
3eca1b8c MD |
1439 | } |
1440 | ||
14044b37 | 1441 | struct cds_lfht_node *cds_lfht_add_unique(struct cds_lfht *ht, |
48ed1c18 | 1442 | struct cds_lfht_node *node) |
3eca1b8c | 1443 | { |
4105056a | 1444 | unsigned long hash, size; |
83beee94 | 1445 | struct cds_lfht_iter iter; |
3eca1b8c | 1446 | |
49c2e2d6 | 1447 | hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed); |
cc4fcb10 | 1448 | node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash); |
3eca1b8c | 1449 | |
4105056a | 1450 | size = rcu_dereference(ht->t.size); |
83beee94 MD |
1451 | _cds_lfht_add(ht, size, node, &iter, 0); |
1452 | if (iter.node == node) | |
4105056a | 1453 | ht_count_add(ht, size); |
83beee94 | 1454 | return iter.node; |
2ed95849 MD |
1455 | } |
1456 | ||
9357c415 | 1457 | struct cds_lfht_node *cds_lfht_add_replace(struct cds_lfht *ht, |
48ed1c18 MD |
1458 | struct cds_lfht_node *node) |
1459 | { | |
1460 | unsigned long hash, size; | |
83beee94 | 1461 | struct cds_lfht_iter iter; |
48ed1c18 MD |
1462 | |
1463 | hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed); | |
1464 | node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash); | |
1465 | ||
1466 | size = rcu_dereference(ht->t.size); | |
83beee94 MD |
1467 | for (;;) { |
1468 | _cds_lfht_add(ht, size, node, &iter, 0); | |
1469 | if (iter.node == node) { | |
1470 | ht_count_add(ht, size); | |
1471 | return NULL; | |
1472 | } | |
1473 | ||
1474 | if (!_cds_lfht_replace(ht, size, iter.node, iter.next, node)) | |
1475 | return iter.node; | |
1476 | } | |
48ed1c18 MD |
1477 | } |
1478 | ||
9357c415 MD |
1479 | int cds_lfht_replace(struct cds_lfht *ht, struct cds_lfht_iter *old_iter, |
1480 | struct cds_lfht_node *new_node) | |
1481 | { | |
1482 | unsigned long size; | |
1483 | ||
1484 | size = rcu_dereference(ht->t.size); | |
1485 | return _cds_lfht_replace(ht, size, old_iter->node, old_iter->next, | |
1486 | new_node); | |
1487 | } | |
1488 | ||
1489 | int cds_lfht_del(struct cds_lfht *ht, struct cds_lfht_iter *iter) | |
2ed95849 | 1490 | { |
4105056a | 1491 | unsigned long size; |
df44348d | 1492 | int ret; |
abc490a1 | 1493 | |
4105056a | 1494 | size = rcu_dereference(ht->t.size); |
9357c415 | 1495 | ret = _cds_lfht_del(ht, size, iter->node, 0); |
df44348d | 1496 | if (!ret) |
860d07e8 | 1497 | ht_count_del(ht, size); |
df44348d | 1498 | return ret; |
2ed95849 | 1499 | } |
ab7d5fc6 | 1500 | |
abc490a1 | 1501 | static |
14044b37 | 1502 | int cds_lfht_delete_dummy(struct cds_lfht *ht) |
674f7a69 | 1503 | { |
14044b37 MD |
1504 | struct cds_lfht_node *node; |
1505 | struct _cds_lfht_node *lookup; | |
4105056a | 1506 | unsigned long order, i, size; |
674f7a69 | 1507 | |
abc490a1 | 1508 | /* Check that the table is empty */ |
4105056a | 1509 | lookup = &ht->t.tbl[0]->nodes[0]; |
14044b37 | 1510 | node = (struct cds_lfht_node *) lookup; |
abc490a1 | 1511 | do { |
1b81fe1a MD |
1512 | node = clear_flag(node)->p.next; |
1513 | if (!is_dummy(node)) | |
abc490a1 | 1514 | return -EPERM; |
273399de | 1515 | assert(!is_removed(node)); |
bb7b2f26 | 1516 | } while (!is_end(node)); |
4105056a MD |
1517 | /* |
1518 | * size accessed without rcu_dereference because hash table is | |
1519 | * being destroyed. | |
1520 | */ | |
1521 | size = ht->t.size; | |
abc490a1 | 1522 | /* Internal sanity check: all nodes left should be dummy */ |
4105056a | 1523 | for (order = 0; order < get_count_order_ulong(size) + 1; order++) { |
24365af7 MD |
1524 | unsigned long len; |
1525 | ||
1526 | len = !order ? 1 : 1UL << (order - 1); | |
1527 | for (i = 0; i < len; i++) { | |
f0c29ed7 | 1528 | dbg_printf("delete order %lu i %lu hash %lu\n", |
24365af7 | 1529 | order, i, |
4105056a MD |
1530 | bit_reverse_ulong(ht->t.tbl[order]->nodes[i].reverse_hash)); |
1531 | assert(is_dummy(ht->t.tbl[order]->nodes[i].next)); | |
24365af7 | 1532 | } |
4105056a | 1533 | poison_free(ht->t.tbl[order]); |
674f7a69 | 1534 | } |
abc490a1 | 1535 | return 0; |
674f7a69 MD |
1536 | } |
1537 | ||
1538 | /* | |
1539 | * Should only be called when no more concurrent readers nor writers can | |
1540 | * possibly access the table. | |
1541 | */ | |
b7d619b0 | 1542 | int cds_lfht_destroy(struct cds_lfht *ht, pthread_attr_t **attr) |
674f7a69 | 1543 | { |
5e28c532 MD |
1544 | int ret; |
1545 | ||
848d4088 | 1546 | /* Wait for in-flight resize operations to complete */ |
24953e08 MD |
1547 | _CMM_STORE_SHARED(ht->in_progress_destroy, 1); |
1548 | cmm_smp_mb(); /* Store destroy before load resize */ | |
848d4088 MD |
1549 | while (uatomic_read(&ht->in_progress_resize)) |
1550 | poll(NULL, 0, 100); /* wait for 100ms */ | |
14044b37 | 1551 | ret = cds_lfht_delete_dummy(ht); |
abc490a1 MD |
1552 | if (ret) |
1553 | return ret; | |
df44348d | 1554 | free_per_cpu_items_count(ht->percpu_count); |
b7d619b0 MD |
1555 | if (attr) |
1556 | *attr = ht->resize_attr; | |
98808fb1 | 1557 | poison_free(ht); |
5e28c532 | 1558 | return ret; |
674f7a69 MD |
1559 | } |
1560 | ||
14044b37 | 1561 | void cds_lfht_count_nodes(struct cds_lfht *ht, |
d933dd0e | 1562 | long *approx_before, |
273399de | 1563 | unsigned long *count, |
973e5e1b | 1564 | unsigned long *removed, |
d933dd0e | 1565 | long *approx_after) |
273399de | 1566 | { |
14044b37 MD |
1567 | struct cds_lfht_node *node, *next; |
1568 | struct _cds_lfht_node *lookup; | |
24365af7 | 1569 | unsigned long nr_dummy = 0; |
273399de | 1570 | |
7ed7682f | 1571 | *approx_before = 0; |
973e5e1b MD |
1572 | if (nr_cpus_mask >= 0) { |
1573 | int i; | |
1574 | ||
1575 | for (i = 0; i < nr_cpus_mask + 1; i++) { | |
1576 | *approx_before += uatomic_read(&ht->percpu_count[i].add); | |
1577 | *approx_before -= uatomic_read(&ht->percpu_count[i].del); | |
1578 | } | |
1579 | } | |
1580 | ||
273399de MD |
1581 | *count = 0; |
1582 | *removed = 0; | |
1583 | ||
24365af7 | 1584 | /* Count non-dummy nodes in the table */ |
4105056a | 1585 | lookup = &ht->t.tbl[0]->nodes[0]; |
14044b37 | 1586 | node = (struct cds_lfht_node *) lookup; |
273399de | 1587 | do { |
cc4fcb10 | 1588 | next = rcu_dereference(node->p.next); |
b198f0fd | 1589 | if (is_removed(next)) { |
973e5e1b MD |
1590 | if (!is_dummy(next)) |
1591 | (*removed)++; | |
1592 | else | |
1593 | (nr_dummy)++; | |
1b81fe1a | 1594 | } else if (!is_dummy(next)) |
273399de | 1595 | (*count)++; |
24365af7 MD |
1596 | else |
1597 | (nr_dummy)++; | |
273399de | 1598 | node = clear_flag(next); |
bb7b2f26 | 1599 | } while (!is_end(node)); |
f0c29ed7 | 1600 | dbg_printf("number of dummy nodes: %lu\n", nr_dummy); |
7ed7682f | 1601 | *approx_after = 0; |
973e5e1b MD |
1602 | if (nr_cpus_mask >= 0) { |
1603 | int i; | |
1604 | ||
1605 | for (i = 0; i < nr_cpus_mask + 1; i++) { | |
1606 | *approx_after += uatomic_read(&ht->percpu_count[i].add); | |
1607 | *approx_after -= uatomic_read(&ht->percpu_count[i].del); | |
1608 | } | |
1609 | } | |
273399de MD |
1610 | } |
1611 | ||
1475579c | 1612 | /* called with resize mutex held */ |
abc490a1 | 1613 | static |
4105056a | 1614 | void _do_cds_lfht_grow(struct cds_lfht *ht, |
1475579c | 1615 | unsigned long old_size, unsigned long new_size) |
abc490a1 | 1616 | { |
1475579c | 1617 | unsigned long old_order, new_order; |
1475579c | 1618 | |
93d46c39 LJ |
1619 | old_order = get_count_order_ulong(old_size); |
1620 | new_order = get_count_order_ulong(new_size); | |
1a401918 LJ |
1621 | dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n", |
1622 | old_size, old_order, new_size, new_order); | |
1475579c | 1623 | assert(new_size > old_size); |
93d46c39 | 1624 | init_table(ht, old_order + 1, new_order); |
abc490a1 MD |
1625 | } |
1626 | ||
1627 | /* called with resize mutex held */ | |
1628 | static | |
4105056a | 1629 | void _do_cds_lfht_shrink(struct cds_lfht *ht, |
1475579c | 1630 | unsigned long old_size, unsigned long new_size) |
464a1ec9 | 1631 | { |
1475579c | 1632 | unsigned long old_order, new_order; |
464a1ec9 | 1633 | |
cd95516d | 1634 | new_size = max(new_size, MIN_TABLE_SIZE); |
93d46c39 LJ |
1635 | old_order = get_count_order_ulong(old_size); |
1636 | new_order = get_count_order_ulong(new_size); | |
1a401918 LJ |
1637 | dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n", |
1638 | old_size, old_order, new_size, new_order); | |
1475579c | 1639 | assert(new_size < old_size); |
1475579c | 1640 | |
4105056a | 1641 | /* Remove and unlink all dummy nodes to remove. */ |
93d46c39 | 1642 | fini_table(ht, new_order + 1, old_order); |
464a1ec9 MD |
1643 | } |
1644 | ||
1475579c MD |
1645 | |
1646 | /* called with resize mutex held */ | |
1647 | static | |
1648 | void _do_cds_lfht_resize(struct cds_lfht *ht) | |
1649 | { | |
1650 | unsigned long new_size, old_size; | |
4105056a MD |
1651 | |
1652 | /* | |
1653 | * Resize table, re-do if the target size has changed under us. | |
1654 | */ | |
1655 | do { | |
d2be3620 MD |
1656 | assert(uatomic_read(&ht->in_progress_resize)); |
1657 | if (CMM_LOAD_SHARED(ht->in_progress_destroy)) | |
1658 | break; | |
4105056a MD |
1659 | ht->t.resize_initiated = 1; |
1660 | old_size = ht->t.size; | |
1661 | new_size = CMM_LOAD_SHARED(ht->t.resize_target); | |
1662 | if (old_size < new_size) | |
1663 | _do_cds_lfht_grow(ht, old_size, new_size); | |
1664 | else if (old_size > new_size) | |
1665 | _do_cds_lfht_shrink(ht, old_size, new_size); | |
1666 | ht->t.resize_initiated = 0; | |
1667 | /* write resize_initiated before read resize_target */ | |
1668 | cmm_smp_mb(); | |
4d676753 | 1669 | } while (ht->t.size != CMM_LOAD_SHARED(ht->t.resize_target)); |
1475579c MD |
1670 | } |
1671 | ||
abc490a1 | 1672 | static |
4105056a | 1673 | unsigned long resize_target_update(struct cds_lfht *ht, unsigned long size, |
f9830efd | 1674 | int growth_order) |
464a1ec9 | 1675 | { |
4105056a MD |
1676 | return _uatomic_max(&ht->t.resize_target, |
1677 | size << growth_order); | |
464a1ec9 MD |
1678 | } |
1679 | ||
1475579c | 1680 | static |
4105056a | 1681 | void resize_target_update_count(struct cds_lfht *ht, |
b8af5011 | 1682 | unsigned long count) |
1475579c | 1683 | { |
cd95516d | 1684 | count = max(count, MIN_TABLE_SIZE); |
4105056a | 1685 | uatomic_set(&ht->t.resize_target, count); |
1475579c MD |
1686 | } |
1687 | ||
1688 | void cds_lfht_resize(struct cds_lfht *ht, unsigned long new_size) | |
464a1ec9 | 1689 | { |
4105056a MD |
1690 | resize_target_update_count(ht, new_size); |
1691 | CMM_STORE_SHARED(ht->t.resize_initiated, 1); | |
5f511391 | 1692 | ht->cds_lfht_rcu_thread_offline(); |
1475579c MD |
1693 | pthread_mutex_lock(&ht->resize_mutex); |
1694 | _do_cds_lfht_resize(ht); | |
1695 | pthread_mutex_unlock(&ht->resize_mutex); | |
5f511391 | 1696 | ht->cds_lfht_rcu_thread_online(); |
abc490a1 | 1697 | } |
464a1ec9 | 1698 | |
abc490a1 MD |
1699 | static |
1700 | void do_resize_cb(struct rcu_head *head) | |
1701 | { | |
1702 | struct rcu_resize_work *work = | |
1703 | caa_container_of(head, struct rcu_resize_work, head); | |
14044b37 | 1704 | struct cds_lfht *ht = work->ht; |
abc490a1 | 1705 | |
5f511391 | 1706 | ht->cds_lfht_rcu_thread_offline(); |
abc490a1 | 1707 | pthread_mutex_lock(&ht->resize_mutex); |
14044b37 | 1708 | _do_cds_lfht_resize(ht); |
abc490a1 | 1709 | pthread_mutex_unlock(&ht->resize_mutex); |
5f511391 | 1710 | ht->cds_lfht_rcu_thread_online(); |
98808fb1 | 1711 | poison_free(work); |
848d4088 MD |
1712 | cmm_smp_mb(); /* finish resize before decrement */ |
1713 | uatomic_dec(&ht->in_progress_resize); | |
464a1ec9 MD |
1714 | } |
1715 | ||
abc490a1 | 1716 | static |
4105056a | 1717 | void cds_lfht_resize_lazy(struct cds_lfht *ht, unsigned long size, int growth) |
ab7d5fc6 | 1718 | { |
abc490a1 | 1719 | struct rcu_resize_work *work; |
f9830efd | 1720 | unsigned long target_size; |
abc490a1 | 1721 | |
4105056a MD |
1722 | target_size = resize_target_update(ht, size, growth); |
1723 | /* Store resize_target before read resize_initiated */ | |
1724 | cmm_smp_mb(); | |
1725 | if (!CMM_LOAD_SHARED(ht->t.resize_initiated) && size < target_size) { | |
848d4088 | 1726 | uatomic_inc(&ht->in_progress_resize); |
59290e9d | 1727 | cmm_smp_mb(); /* increment resize count before load destroy */ |
ed35e6d8 MD |
1728 | if (CMM_LOAD_SHARED(ht->in_progress_destroy)) { |
1729 | uatomic_dec(&ht->in_progress_resize); | |
59290e9d | 1730 | return; |
ed35e6d8 | 1731 | } |
f9830efd MD |
1732 | work = malloc(sizeof(*work)); |
1733 | work->ht = ht; | |
14044b37 | 1734 | ht->cds_lfht_call_rcu(&work->head, do_resize_cb); |
4105056a | 1735 | CMM_STORE_SHARED(ht->t.resize_initiated, 1); |
f9830efd | 1736 | } |
ab7d5fc6 | 1737 | } |
3171717f | 1738 | |
f8994aee MD |
1739 | #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) |
1740 | ||
3171717f | 1741 | static |
4105056a | 1742 | void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size, |
3171717f MD |
1743 | unsigned long count) |
1744 | { | |
1745 | struct rcu_resize_work *work; | |
3171717f | 1746 | |
b8af5011 MD |
1747 | if (!(ht->flags & CDS_LFHT_AUTO_RESIZE)) |
1748 | return; | |
4105056a MD |
1749 | resize_target_update_count(ht, count); |
1750 | /* Store resize_target before read resize_initiated */ | |
1751 | cmm_smp_mb(); | |
1752 | if (!CMM_LOAD_SHARED(ht->t.resize_initiated)) { | |
3171717f | 1753 | uatomic_inc(&ht->in_progress_resize); |
59290e9d | 1754 | cmm_smp_mb(); /* increment resize count before load destroy */ |
ed35e6d8 MD |
1755 | if (CMM_LOAD_SHARED(ht->in_progress_destroy)) { |
1756 | uatomic_dec(&ht->in_progress_resize); | |
59290e9d | 1757 | return; |
ed35e6d8 | 1758 | } |
3171717f MD |
1759 | work = malloc(sizeof(*work)); |
1760 | work->ht = ht; | |
1761 | ht->cds_lfht_call_rcu(&work->head, do_resize_cb); | |
4105056a | 1762 | CMM_STORE_SHARED(ht->t.resize_initiated, 1); |
3171717f MD |
1763 | } |
1764 | } | |
f8994aee MD |
1765 | |
1766 | #endif |