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 | |
94 | * this order. | |
95 | * - call_rcu is used to garbage-collect the old order table. | |
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. | |
1475579c MD |
99 | * |
100 | * A bit of ascii art explanation: | |
101 | * | |
102 | * Order index is the off-by-one compare to the actual power of 2 because | |
103 | * we use index 0 to deal with the 0 special-case. | |
104 | * | |
105 | * This shows the nodes for a small table ordered by reversed bits: | |
106 | * | |
107 | * bits reverse | |
108 | * 0 000 000 | |
109 | * 4 100 001 | |
110 | * 2 010 010 | |
111 | * 6 110 011 | |
112 | * 1 001 100 | |
113 | * 5 101 101 | |
114 | * 3 011 110 | |
115 | * 7 111 111 | |
116 | * | |
117 | * This shows the nodes in order of non-reversed bits, linked by | |
118 | * reversed-bit order. | |
119 | * | |
120 | * order bits reverse | |
121 | * 0 0 000 000 | |
122 | * | | |
f6fdd688 MD |
123 | * 1 | 1 001 100 <- <- |
124 | * | | | | | |
125 | * 2 | | 2 010 010 | | | |
126 | * | | | 3 011 110 | <- | | |
127 | * | | | | | | | | |
1475579c MD |
128 | * 3 -> | | | 4 100 001 | | |
129 | * -> | | 5 101 101 | | |
130 | * -> | 6 110 011 | |
131 | * -> 7 111 111 | |
e753ff5a MD |
132 | */ |
133 | ||
2ed95849 MD |
134 | #define _LGPL_SOURCE |
135 | #include <stdlib.h> | |
e0ba718a MD |
136 | #include <errno.h> |
137 | #include <assert.h> | |
138 | #include <stdio.h> | |
abc490a1 | 139 | #include <stdint.h> |
f000907d | 140 | #include <string.h> |
e0ba718a | 141 | |
15cfbec7 | 142 | #include "config.h" |
2ed95849 | 143 | #include <urcu.h> |
abc490a1 | 144 | #include <urcu-call-rcu.h> |
a42cc659 MD |
145 | #include <urcu/arch.h> |
146 | #include <urcu/uatomic.h> | |
a42cc659 | 147 | #include <urcu/compiler.h> |
abc490a1 | 148 | #include <urcu/rculfhash.h> |
5e28c532 | 149 | #include <stdio.h> |
464a1ec9 | 150 | #include <pthread.h> |
44395fb7 | 151 | |
f9830efd | 152 | #ifdef DEBUG |
f0c29ed7 | 153 | #define dbg_printf(fmt, args...) printf("[debug rculfhash] " fmt, ## args) |
f9830efd | 154 | #else |
e753ff5a | 155 | #define dbg_printf(fmt, args...) |
f9830efd MD |
156 | #endif |
157 | ||
f8994aee MD |
158 | /* |
159 | * Per-CPU split-counters lazily update the global counter each 1024 | |
160 | * addition/removal. It automatically keeps track of resize required. | |
161 | * We use the bucket length as indicator for need to expand for small | |
162 | * tables and machines lacking per-cpu data suppport. | |
163 | */ | |
164 | #define COUNT_COMMIT_ORDER 10 | |
6ea6bc67 MD |
165 | #define CHAIN_LEN_TARGET 1 |
166 | #define CHAIN_LEN_RESIZE_THRESHOLD 3 | |
2ed95849 | 167 | |
cd95516d | 168 | /* |
76a73da8 | 169 | * Define the minimum table size. |
cd95516d | 170 | */ |
c9edd44a | 171 | #define MIN_TABLE_SIZE 1 |
cd95516d | 172 | |
4105056a MD |
173 | #if (CAA_BITS_PER_LONG == 32) |
174 | #define MAX_TABLE_ORDER 32 | |
175 | #else | |
176 | #define MAX_TABLE_ORDER 64 | |
177 | #endif | |
178 | ||
b7d619b0 MD |
179 | /* |
180 | * Minimum number of dummy nodes to touch per thread to parallelize grow/shrink. | |
181 | */ | |
6083a889 MD |
182 | #define MIN_PARTITION_PER_THREAD_ORDER 12 |
183 | #define MIN_PARTITION_PER_THREAD (1UL << MIN_PARTITION_PER_THREAD_ORDER) | |
b7d619b0 | 184 | |
4105056a MD |
185 | #ifndef min |
186 | #define min(a, b) ((a) < (b) ? (a) : (b)) | |
187 | #endif | |
188 | ||
abc490a1 MD |
189 | #ifndef max |
190 | #define max(a, b) ((a) > (b) ? (a) : (b)) | |
191 | #endif | |
2ed95849 | 192 | |
d95bd160 MD |
193 | /* |
194 | * The removed flag needs to be updated atomically with the pointer. | |
48ed1c18 | 195 | * It indicates that no node must attach to the node scheduled for |
b198f0fd | 196 | * removal, and that node garbage collection must be performed. |
d95bd160 MD |
197 | * The dummy flag does not require to be updated atomically with the |
198 | * pointer, but it is added as a pointer low bit flag to save space. | |
199 | */ | |
d37166c6 | 200 | #define REMOVED_FLAG (1UL << 0) |
b198f0fd MD |
201 | #define DUMMY_FLAG (1UL << 1) |
202 | #define FLAGS_MASK ((1UL << 2) - 1) | |
d37166c6 | 203 | |
bb7b2f26 | 204 | /* Value of the end pointer. Should not interact with flags. */ |
f9c80341 | 205 | #define END_VALUE NULL |
bb7b2f26 | 206 | |
df44348d | 207 | struct ht_items_count { |
860d07e8 | 208 | unsigned long add, del; |
df44348d MD |
209 | } __attribute__((aligned(CAA_CACHE_LINE_SIZE))); |
210 | ||
1475579c | 211 | struct rcu_level { |
0d14ceb2 | 212 | /* Note: manually update allocation length when adding a field */ |
1475579c MD |
213 | struct _cds_lfht_node nodes[0]; |
214 | }; | |
215 | ||
395270b6 | 216 | struct rcu_table { |
4105056a | 217 | unsigned long size; /* always a power of 2, shared (RCU) */ |
f9830efd | 218 | unsigned long resize_target; |
11519af6 | 219 | int resize_initiated; |
4105056a | 220 | struct rcu_level *tbl[MAX_TABLE_ORDER]; |
395270b6 MD |
221 | }; |
222 | ||
14044b37 | 223 | struct cds_lfht { |
4105056a | 224 | struct rcu_table t; |
14044b37 MD |
225 | cds_lfht_hash_fct hash_fct; |
226 | cds_lfht_compare_fct compare_fct; | |
732ad076 | 227 | unsigned long hash_seed; |
b8af5011 | 228 | int flags; |
5f511391 MD |
229 | /* |
230 | * We need to put the work threads offline (QSBR) when taking this | |
231 | * mutex, because we use synchronize_rcu within this mutex critical | |
232 | * section, which waits on read-side critical sections, and could | |
233 | * therefore cause grace-period deadlock if we hold off RCU G.P. | |
234 | * completion. | |
235 | */ | |
464a1ec9 | 236 | pthread_mutex_t resize_mutex; /* resize mutex: add/del mutex */ |
33c7c748 | 237 | unsigned int in_progress_resize, in_progress_destroy; |
14044b37 | 238 | void (*cds_lfht_call_rcu)(struct rcu_head *head, |
abc490a1 | 239 | void (*func)(struct rcu_head *head)); |
1475579c | 240 | void (*cds_lfht_synchronize_rcu)(void); |
01dbfa62 MD |
241 | void (*cds_lfht_rcu_read_lock)(void); |
242 | void (*cds_lfht_rcu_read_unlock)(void); | |
5f511391 MD |
243 | void (*cds_lfht_rcu_thread_offline)(void); |
244 | void (*cds_lfht_rcu_thread_online)(void); | |
b7d619b0 MD |
245 | void (*cds_lfht_rcu_register_thread)(void); |
246 | void (*cds_lfht_rcu_unregister_thread)(void); | |
247 | pthread_attr_t *resize_attr; /* Resize threads attributes */ | |
7de5ccfd | 248 | long count; /* global approximate item count */ |
df44348d | 249 | struct ht_items_count *percpu_count; /* per-cpu item count */ |
2ed95849 MD |
250 | }; |
251 | ||
abc490a1 MD |
252 | struct rcu_resize_work { |
253 | struct rcu_head head; | |
14044b37 | 254 | struct cds_lfht *ht; |
abc490a1 | 255 | }; |
2ed95849 | 256 | |
b7d619b0 | 257 | struct partition_resize_work { |
1af6e26e | 258 | pthread_t thread_id; |
b7d619b0 MD |
259 | struct cds_lfht *ht; |
260 | unsigned long i, start, len; | |
261 | void (*fct)(struct cds_lfht *ht, unsigned long i, | |
262 | unsigned long start, unsigned long len); | |
263 | }; | |
264 | ||
48ed1c18 MD |
265 | enum add_mode { |
266 | ADD_DEFAULT = 0, | |
267 | ADD_UNIQUE = 1, | |
268 | ADD_REPLACE = 2, | |
269 | }; | |
270 | ||
76a73da8 MD |
271 | static |
272 | struct cds_lfht_node *_cds_lfht_add(struct cds_lfht *ht, | |
273 | unsigned long size, | |
274 | struct cds_lfht_node *node, | |
48ed1c18 MD |
275 | enum add_mode mode, int dummy); |
276 | ||
abc490a1 MD |
277 | /* |
278 | * Algorithm to reverse bits in a word by lookup table, extended to | |
279 | * 64-bit words. | |
f9830efd | 280 | * Source: |
abc490a1 | 281 | * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable |
f9830efd | 282 | * Originally from Public Domain. |
abc490a1 MD |
283 | */ |
284 | ||
285 | static const uint8_t BitReverseTable256[256] = | |
2ed95849 | 286 | { |
abc490a1 MD |
287 | #define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64 |
288 | #define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16) | |
289 | #define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 ) | |
290 | R6(0), R6(2), R6(1), R6(3) | |
291 | }; | |
292 | #undef R2 | |
293 | #undef R4 | |
294 | #undef R6 | |
2ed95849 | 295 | |
abc490a1 MD |
296 | static |
297 | uint8_t bit_reverse_u8(uint8_t v) | |
298 | { | |
299 | return BitReverseTable256[v]; | |
300 | } | |
ab7d5fc6 | 301 | |
abc490a1 MD |
302 | static __attribute__((unused)) |
303 | uint32_t bit_reverse_u32(uint32_t v) | |
304 | { | |
305 | return ((uint32_t) bit_reverse_u8(v) << 24) | | |
306 | ((uint32_t) bit_reverse_u8(v >> 8) << 16) | | |
307 | ((uint32_t) bit_reverse_u8(v >> 16) << 8) | | |
308 | ((uint32_t) bit_reverse_u8(v >> 24)); | |
2ed95849 MD |
309 | } |
310 | ||
abc490a1 MD |
311 | static __attribute__((unused)) |
312 | uint64_t bit_reverse_u64(uint64_t v) | |
2ed95849 | 313 | { |
abc490a1 MD |
314 | return ((uint64_t) bit_reverse_u8(v) << 56) | |
315 | ((uint64_t) bit_reverse_u8(v >> 8) << 48) | | |
316 | ((uint64_t) bit_reverse_u8(v >> 16) << 40) | | |
317 | ((uint64_t) bit_reverse_u8(v >> 24) << 32) | | |
318 | ((uint64_t) bit_reverse_u8(v >> 32) << 24) | | |
319 | ((uint64_t) bit_reverse_u8(v >> 40) << 16) | | |
320 | ((uint64_t) bit_reverse_u8(v >> 48) << 8) | | |
321 | ((uint64_t) bit_reverse_u8(v >> 56)); | |
322 | } | |
323 | ||
324 | static | |
325 | unsigned long bit_reverse_ulong(unsigned long v) | |
326 | { | |
327 | #if (CAA_BITS_PER_LONG == 32) | |
328 | return bit_reverse_u32(v); | |
329 | #else | |
330 | return bit_reverse_u64(v); | |
331 | #endif | |
332 | } | |
333 | ||
f9830efd | 334 | /* |
24365af7 MD |
335 | * fls: returns the position of the most significant bit. |
336 | * Returns 0 if no bit is set, else returns the position of the most | |
337 | * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit). | |
f9830efd | 338 | */ |
24365af7 MD |
339 | #if defined(__i386) || defined(__x86_64) |
340 | static inline | |
341 | unsigned int fls_u32(uint32_t x) | |
f9830efd | 342 | { |
24365af7 MD |
343 | int r; |
344 | ||
345 | asm("bsrl %1,%0\n\t" | |
346 | "jnz 1f\n\t" | |
347 | "movl $-1,%0\n\t" | |
348 | "1:\n\t" | |
349 | : "=r" (r) : "rm" (x)); | |
350 | return r + 1; | |
351 | } | |
352 | #define HAS_FLS_U32 | |
353 | #endif | |
354 | ||
355 | #if defined(__x86_64) | |
356 | static inline | |
357 | unsigned int fls_u64(uint64_t x) | |
358 | { | |
359 | long r; | |
360 | ||
361 | asm("bsrq %1,%0\n\t" | |
362 | "jnz 1f\n\t" | |
363 | "movq $-1,%0\n\t" | |
364 | "1:\n\t" | |
365 | : "=r" (r) : "rm" (x)); | |
366 | return r + 1; | |
367 | } | |
368 | #define HAS_FLS_U64 | |
369 | #endif | |
370 | ||
371 | #ifndef HAS_FLS_U64 | |
372 | static __attribute__((unused)) | |
373 | unsigned int fls_u64(uint64_t x) | |
374 | { | |
375 | unsigned int r = 64; | |
376 | ||
377 | if (!x) | |
378 | return 0; | |
379 | ||
380 | if (!(x & 0xFFFFFFFF00000000ULL)) { | |
381 | x <<= 32; | |
382 | r -= 32; | |
383 | } | |
384 | if (!(x & 0xFFFF000000000000ULL)) { | |
385 | x <<= 16; | |
386 | r -= 16; | |
387 | } | |
388 | if (!(x & 0xFF00000000000000ULL)) { | |
389 | x <<= 8; | |
390 | r -= 8; | |
391 | } | |
392 | if (!(x & 0xF000000000000000ULL)) { | |
393 | x <<= 4; | |
394 | r -= 4; | |
395 | } | |
396 | if (!(x & 0xC000000000000000ULL)) { | |
397 | x <<= 2; | |
398 | r -= 2; | |
399 | } | |
400 | if (!(x & 0x8000000000000000ULL)) { | |
401 | x <<= 1; | |
402 | r -= 1; | |
403 | } | |
404 | return r; | |
405 | } | |
406 | #endif | |
407 | ||
408 | #ifndef HAS_FLS_U32 | |
409 | static __attribute__((unused)) | |
410 | unsigned int fls_u32(uint32_t x) | |
411 | { | |
412 | unsigned int r = 32; | |
f9830efd | 413 | |
24365af7 MD |
414 | if (!x) |
415 | return 0; | |
416 | if (!(x & 0xFFFF0000U)) { | |
417 | x <<= 16; | |
418 | r -= 16; | |
419 | } | |
420 | if (!(x & 0xFF000000U)) { | |
421 | x <<= 8; | |
422 | r -= 8; | |
423 | } | |
424 | if (!(x & 0xF0000000U)) { | |
425 | x <<= 4; | |
426 | r -= 4; | |
427 | } | |
428 | if (!(x & 0xC0000000U)) { | |
429 | x <<= 2; | |
430 | r -= 2; | |
431 | } | |
432 | if (!(x & 0x80000000U)) { | |
433 | x <<= 1; | |
434 | r -= 1; | |
435 | } | |
436 | return r; | |
437 | } | |
438 | #endif | |
439 | ||
440 | unsigned int fls_ulong(unsigned long x) | |
f9830efd | 441 | { |
24365af7 MD |
442 | #if (CAA_BITS_PER_lONG == 32) |
443 | return fls_u32(x); | |
444 | #else | |
445 | return fls_u64(x); | |
446 | #endif | |
447 | } | |
f9830efd | 448 | |
920f8ef6 LJ |
449 | /* |
450 | * Return the minimum order for which x <= (1UL << order). | |
451 | * Return -1 if x is 0. | |
452 | */ | |
24365af7 MD |
453 | int get_count_order_u32(uint32_t x) |
454 | { | |
920f8ef6 LJ |
455 | if (!x) |
456 | return -1; | |
24365af7 | 457 | |
920f8ef6 | 458 | return fls_u32(x - 1); |
24365af7 MD |
459 | } |
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_ulong(unsigned long x) |
466 | { | |
920f8ef6 LJ |
467 | if (!x) |
468 | return -1; | |
24365af7 | 469 | |
920f8ef6 | 470 | return fls_ulong(x - 1); |
f9830efd MD |
471 | } |
472 | ||
98808fb1 MD |
473 | #ifdef POISON_FREE |
474 | #define poison_free(ptr) \ | |
475 | do { \ | |
476 | memset(ptr, 0x42, sizeof(*(ptr))); \ | |
477 | free(ptr); \ | |
478 | } while (0) | |
479 | #else | |
480 | #define poison_free(ptr) free(ptr) | |
481 | #endif | |
482 | ||
f9830efd | 483 | static |
4105056a | 484 | void cds_lfht_resize_lazy(struct cds_lfht *ht, unsigned long size, int growth); |
f9830efd | 485 | |
df44348d MD |
486 | /* |
487 | * If the sched_getcpu() and sysconf(_SC_NPROCESSORS_CONF) calls are | |
488 | * available, then we support hash table item accounting. | |
489 | * In the unfortunate event the number of CPUs reported would be | |
490 | * inaccurate, we use modulo arithmetic on the number of CPUs we got. | |
491 | */ | |
df44348d MD |
492 | #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) |
493 | ||
f8994aee | 494 | static |
4105056a | 495 | void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size, |
f8994aee MD |
496 | unsigned long count); |
497 | ||
df44348d MD |
498 | static long nr_cpus_mask = -1; |
499 | ||
500 | static | |
501 | struct ht_items_count *alloc_per_cpu_items_count(void) | |
502 | { | |
503 | struct ht_items_count *count; | |
504 | ||
505 | switch (nr_cpus_mask) { | |
506 | case -2: | |
507 | return NULL; | |
508 | case -1: | |
509 | { | |
510 | long maxcpus; | |
511 | ||
512 | maxcpus = sysconf(_SC_NPROCESSORS_CONF); | |
513 | if (maxcpus <= 0) { | |
514 | nr_cpus_mask = -2; | |
515 | return NULL; | |
516 | } | |
517 | /* | |
518 | * round up number of CPUs to next power of two, so we | |
519 | * can use & for modulo. | |
520 | */ | |
521 | maxcpus = 1UL << get_count_order_ulong(maxcpus); | |
522 | nr_cpus_mask = maxcpus - 1; | |
523 | } | |
524 | /* Fall-through */ | |
525 | default: | |
526 | return calloc(nr_cpus_mask + 1, sizeof(*count)); | |
527 | } | |
528 | } | |
529 | ||
530 | static | |
531 | void free_per_cpu_items_count(struct ht_items_count *count) | |
532 | { | |
98808fb1 | 533 | poison_free(count); |
df44348d MD |
534 | } |
535 | ||
536 | static | |
537 | int ht_get_cpu(void) | |
538 | { | |
539 | int cpu; | |
540 | ||
541 | assert(nr_cpus_mask >= 0); | |
542 | cpu = sched_getcpu(); | |
543 | if (unlikely(cpu < 0)) | |
544 | return cpu; | |
545 | else | |
546 | return cpu & nr_cpus_mask; | |
547 | } | |
548 | ||
549 | static | |
4105056a | 550 | void ht_count_add(struct cds_lfht *ht, unsigned long size) |
df44348d | 551 | { |
3171717f | 552 | unsigned long percpu_count; |
df44348d MD |
553 | int cpu; |
554 | ||
555 | if (unlikely(!ht->percpu_count)) | |
3171717f | 556 | return; |
df44348d MD |
557 | cpu = ht_get_cpu(); |
558 | if (unlikely(cpu < 0)) | |
3171717f MD |
559 | return; |
560 | percpu_count = uatomic_add_return(&ht->percpu_count[cpu].add, 1); | |
df44348d | 561 | if (unlikely(!(percpu_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) { |
e3ecefd6 | 562 | long count; |
df44348d MD |
563 | |
564 | dbg_printf("add percpu %lu\n", percpu_count); | |
565 | count = uatomic_add_return(&ht->count, | |
566 | 1UL << COUNT_COMMIT_ORDER); | |
567 | /* If power of 2 */ | |
568 | if (!(count & (count - 1))) { | |
4105056a | 569 | if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) < size) |
f8994aee | 570 | return; |
e3ecefd6 | 571 | dbg_printf("add set global %ld\n", count); |
4105056a | 572 | cds_lfht_resize_lazy_count(ht, size, |
6ea6bc67 | 573 | count >> (CHAIN_LEN_TARGET - 1)); |
df44348d MD |
574 | } |
575 | } | |
576 | } | |
577 | ||
578 | static | |
860d07e8 | 579 | void ht_count_del(struct cds_lfht *ht, unsigned long size) |
df44348d MD |
580 | { |
581 | unsigned long percpu_count; | |
3171717f | 582 | int cpu; |
df44348d | 583 | |
3171717f MD |
584 | if (unlikely(!ht->percpu_count)) |
585 | return; | |
586 | cpu = ht_get_cpu(); | |
587 | if (unlikely(cpu < 0)) | |
588 | return; | |
80d90c06 | 589 | percpu_count = uatomic_add_return(&ht->percpu_count[cpu].del, 1); |
df44348d | 590 | if (unlikely(!(percpu_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) { |
e3ecefd6 | 591 | long count; |
df44348d | 592 | |
860d07e8 | 593 | dbg_printf("del percpu %lu\n", percpu_count); |
df44348d | 594 | count = uatomic_add_return(&ht->count, |
3171717f | 595 | -(1UL << COUNT_COMMIT_ORDER)); |
df44348d MD |
596 | /* If power of 2 */ |
597 | if (!(count & (count - 1))) { | |
4105056a | 598 | if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) >= size) |
f8994aee | 599 | return; |
e3ecefd6 MD |
600 | dbg_printf("del set global %ld\n", count); |
601 | /* | |
c941bb9e | 602 | * Don't shrink table if the number of nodes is below a |
e3ecefd6 MD |
603 | * certain threshold. |
604 | */ | |
c941bb9e | 605 | if (count < (1UL << COUNT_COMMIT_ORDER) * (nr_cpus_mask + 1)) |
e3ecefd6 | 606 | return; |
4105056a | 607 | cds_lfht_resize_lazy_count(ht, size, |
6ea6bc67 | 608 | count >> (CHAIN_LEN_TARGET - 1)); |
df44348d MD |
609 | } |
610 | } | |
611 | } | |
612 | ||
613 | #else /* #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */ | |
614 | ||
91452a6a | 615 | static const long nr_cpus_mask = -2; |
df44348d MD |
616 | |
617 | static | |
618 | struct ht_items_count *alloc_per_cpu_items_count(void) | |
619 | { | |
620 | return NULL; | |
621 | } | |
622 | ||
623 | static | |
624 | void free_per_cpu_items_count(struct ht_items_count *count) | |
625 | { | |
626 | } | |
627 | ||
628 | static | |
4105056a | 629 | void ht_count_add(struct cds_lfht *ht, unsigned long size) |
df44348d MD |
630 | { |
631 | } | |
632 | ||
633 | static | |
860d07e8 | 634 | void ht_count_del(struct cds_lfht *ht, unsigned long size) |
df44348d MD |
635 | { |
636 | } | |
637 | ||
638 | #endif /* #else #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */ | |
639 | ||
640 | ||
f9830efd | 641 | static |
4105056a | 642 | void check_resize(struct cds_lfht *ht, unsigned long size, uint32_t chain_len) |
f9830efd | 643 | { |
f8994aee MD |
644 | unsigned long count; |
645 | ||
b8af5011 MD |
646 | if (!(ht->flags & CDS_LFHT_AUTO_RESIZE)) |
647 | return; | |
f8994aee MD |
648 | count = uatomic_read(&ht->count); |
649 | /* | |
650 | * Use bucket-local length for small table expand and for | |
651 | * environments lacking per-cpu data support. | |
652 | */ | |
653 | if (count >= (1UL << COUNT_COMMIT_ORDER)) | |
654 | return; | |
24365af7 | 655 | if (chain_len > 100) |
f0c29ed7 | 656 | dbg_printf("WARNING: large chain length: %u.\n", |
24365af7 | 657 | chain_len); |
3390d470 | 658 | if (chain_len >= CHAIN_LEN_RESIZE_THRESHOLD) |
4105056a | 659 | cds_lfht_resize_lazy(ht, size, |
01370f0b | 660 | get_count_order_u32(chain_len - (CHAIN_LEN_TARGET - 1))); |
f9830efd MD |
661 | } |
662 | ||
abc490a1 | 663 | static |
14044b37 | 664 | struct cds_lfht_node *clear_flag(struct cds_lfht_node *node) |
abc490a1 | 665 | { |
14044b37 | 666 | return (struct cds_lfht_node *) (((unsigned long) node) & ~FLAGS_MASK); |
abc490a1 MD |
667 | } |
668 | ||
669 | static | |
14044b37 | 670 | int is_removed(struct cds_lfht_node *node) |
abc490a1 | 671 | { |
d37166c6 | 672 | return ((unsigned long) node) & REMOVED_FLAG; |
abc490a1 MD |
673 | } |
674 | ||
675 | static | |
14044b37 | 676 | struct cds_lfht_node *flag_removed(struct cds_lfht_node *node) |
abc490a1 | 677 | { |
14044b37 | 678 | return (struct cds_lfht_node *) (((unsigned long) node) | REMOVED_FLAG); |
abc490a1 MD |
679 | } |
680 | ||
f5596c94 | 681 | static |
14044b37 | 682 | int is_dummy(struct cds_lfht_node *node) |
f5596c94 MD |
683 | { |
684 | return ((unsigned long) node) & DUMMY_FLAG; | |
685 | } | |
686 | ||
687 | static | |
14044b37 | 688 | struct cds_lfht_node *flag_dummy(struct cds_lfht_node *node) |
f5596c94 | 689 | { |
14044b37 | 690 | return (struct cds_lfht_node *) (((unsigned long) node) | DUMMY_FLAG); |
f5596c94 | 691 | } |
bb7b2f26 MD |
692 | |
693 | static | |
694 | struct cds_lfht_node *get_end(void) | |
695 | { | |
696 | return (struct cds_lfht_node *) END_VALUE; | |
697 | } | |
698 | ||
699 | static | |
700 | int is_end(struct cds_lfht_node *node) | |
701 | { | |
702 | return clear_flag(node) == (struct cds_lfht_node *) END_VALUE; | |
703 | } | |
704 | ||
abc490a1 | 705 | static |
f9830efd | 706 | unsigned long _uatomic_max(unsigned long *ptr, unsigned long v) |
abc490a1 MD |
707 | { |
708 | unsigned long old1, old2; | |
709 | ||
710 | old1 = uatomic_read(ptr); | |
711 | do { | |
712 | old2 = old1; | |
713 | if (old2 >= v) | |
f9830efd | 714 | return old2; |
abc490a1 | 715 | } while ((old1 = uatomic_cmpxchg(ptr, old2, v)) != old2); |
f9830efd | 716 | return v; |
abc490a1 MD |
717 | } |
718 | ||
f4a9cc0b LJ |
719 | static |
720 | struct _cds_lfht_node *lookup_bucket(struct cds_lfht *ht, unsigned long size, | |
721 | unsigned long hash) | |
722 | { | |
723 | unsigned long index, order; | |
724 | ||
725 | assert(size > 0); | |
726 | index = hash & (size - 1); | |
a4ea2223 LJ |
727 | /* |
728 | * equivalent to get_count_order_ulong(index + 1), but optimizes | |
729 | * away the non-existing 0 special-case for | |
730 | * get_count_order_ulong. | |
731 | */ | |
732 | order = fls_ulong(index); | |
f4a9cc0b LJ |
733 | |
734 | dbg_printf("lookup hash %lu index %lu order %lu aridx %lu\n", | |
735 | hash, index, order, index & (!order ? 0 : ((1UL << (order - 1)) - 1))); | |
736 | ||
737 | return &ht->t.tbl[order]->nodes[index & (!order ? 0 : ((1UL << (order - 1)) - 1))]; | |
738 | } | |
739 | ||
273399de MD |
740 | /* |
741 | * Remove all logically deleted nodes from a bucket up to a certain node key. | |
742 | */ | |
743 | static | |
f9c80341 | 744 | void _cds_lfht_gc_bucket(struct cds_lfht_node *dummy, struct cds_lfht_node *node) |
273399de | 745 | { |
14044b37 | 746 | struct cds_lfht_node *iter_prev, *iter, *next, *new_next; |
273399de | 747 | |
c90201ac MD |
748 | assert(!is_dummy(dummy)); |
749 | assert(!is_removed(dummy)); | |
750 | assert(!is_dummy(node)); | |
751 | assert(!is_removed(node)); | |
273399de MD |
752 | for (;;) { |
753 | iter_prev = dummy; | |
754 | /* We can always skip the dummy node initially */ | |
cc4fcb10 MD |
755 | iter = rcu_dereference(iter_prev->p.next); |
756 | assert(iter_prev->p.reverse_hash <= node->p.reverse_hash); | |
bd4db153 MD |
757 | /* |
758 | * We should never be called with dummy (start of chain) | |
759 | * and logically removed node (end of path compression | |
760 | * marker) being the actual same node. This would be a | |
761 | * bug in the algorithm implementation. | |
762 | */ | |
763 | assert(dummy != node); | |
273399de | 764 | for (;;) { |
bb7b2f26 | 765 | if (unlikely(is_end(iter))) |
f9c80341 | 766 | return; |
76412f24 | 767 | if (likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash)) |
f9c80341 | 768 | return; |
cc4fcb10 | 769 | next = rcu_dereference(clear_flag(iter)->p.next); |
b198f0fd | 770 | if (likely(is_removed(next))) |
273399de | 771 | break; |
b453eae1 | 772 | iter_prev = clear_flag(iter); |
273399de MD |
773 | iter = next; |
774 | } | |
b198f0fd | 775 | assert(!is_removed(iter)); |
f5596c94 MD |
776 | if (is_dummy(iter)) |
777 | new_next = flag_dummy(clear_flag(next)); | |
778 | else | |
779 | new_next = clear_flag(next); | |
48ed1c18 MD |
780 | if (is_removed(iter)) |
781 | new_next = flag_removed(new_next); | |
f5596c94 | 782 | (void) uatomic_cmpxchg(&iter_prev->p.next, iter, new_next); |
273399de | 783 | } |
f9c80341 | 784 | return; |
273399de MD |
785 | } |
786 | ||
9357c415 MD |
787 | static |
788 | int _cds_lfht_replace(struct cds_lfht *ht, unsigned long size, | |
789 | struct cds_lfht_node *old_node, | |
790 | struct cds_lfht_node *ret_next, | |
791 | struct cds_lfht_node *new_node) | |
792 | { | |
793 | struct cds_lfht_node *dummy, *old_next; | |
794 | struct _cds_lfht_node *lookup; | |
795 | int flagged = 0; | |
9357c415 MD |
796 | |
797 | if (!old_node) /* Return -ENOENT if asked to replace NULL node */ | |
798 | goto end; | |
799 | ||
800 | assert(!is_removed(old_node)); | |
801 | assert(!is_dummy(old_node)); | |
802 | assert(!is_removed(new_node)); | |
803 | assert(!is_dummy(new_node)); | |
804 | assert(new_node != old_node); | |
805 | do { | |
806 | /* Insert after node to be replaced */ | |
807 | old_next = ret_next; | |
808 | if (is_removed(old_next)) { | |
809 | /* | |
810 | * Too late, the old node has been removed under us | |
811 | * between lookup and replace. Fail. | |
812 | */ | |
813 | goto end; | |
814 | } | |
815 | assert(!is_dummy(old_next)); | |
816 | assert(new_node != clear_flag(old_next)); | |
817 | new_node->p.next = clear_flag(old_next); | |
818 | /* | |
819 | * Here is the whole trick for lock-free replace: we add | |
820 | * the replacement node _after_ the node we want to | |
821 | * replace by atomically setting its next pointer at the | |
822 | * same time we set its removal flag. Given that | |
823 | * the lookups/get next use an iterator aware of the | |
824 | * next pointer, they will either skip the old node due | |
825 | * to the removal flag and see the new node, or use | |
826 | * the old node, but will not see the new one. | |
827 | */ | |
828 | ret_next = uatomic_cmpxchg(&old_node->p.next, | |
829 | old_next, flag_removed(new_node)); | |
830 | } while (ret_next != old_next); | |
831 | ||
832 | /* We performed the replacement. */ | |
833 | flagged = 1; | |
834 | ||
835 | /* | |
836 | * Ensure that the old node is not visible to readers anymore: | |
837 | * lookup for the node, and remove it (along with any other | |
838 | * logically removed node) if found. | |
839 | */ | |
f4a9cc0b | 840 | lookup = lookup_bucket(ht, size, bit_reverse_ulong(old_node->p.reverse_hash)); |
9357c415 MD |
841 | dummy = (struct cds_lfht_node *) lookup; |
842 | _cds_lfht_gc_bucket(dummy, new_node); | |
843 | end: | |
844 | /* | |
845 | * Only the flagging action indicated that we (and no other) | |
846 | * replaced the node from the hash table. | |
847 | */ | |
848 | if (flagged) { | |
849 | assert(is_removed(rcu_dereference(old_node->p.next))); | |
850 | return 0; | |
851 | } else { | |
852 | return -ENOENT; | |
853 | } | |
854 | } | |
855 | ||
abc490a1 | 856 | static |
4105056a MD |
857 | struct cds_lfht_node *_cds_lfht_add(struct cds_lfht *ht, |
858 | unsigned long size, | |
859 | struct cds_lfht_node *node, | |
48ed1c18 | 860 | enum add_mode mode, int dummy) |
abc490a1 | 861 | { |
14044b37 | 862 | struct cds_lfht_node *iter_prev, *iter, *next, *new_node, *new_next, |
960c9e4f | 863 | *return_node; |
14044b37 | 864 | struct _cds_lfht_node *lookup; |
abc490a1 | 865 | |
c90201ac MD |
866 | assert(!is_dummy(node)); |
867 | assert(!is_removed(node)); | |
4105056a | 868 | if (!size) { |
f5596c94 | 869 | assert(dummy); |
bb7b2f26 | 870 | node->p.next = flag_dummy(get_end()); |
18117871 MD |
871 | return node; /* Initial first add (head) */ |
872 | } | |
f4a9cc0b | 873 | lookup = lookup_bucket(ht, size, bit_reverse_ulong(node->p.reverse_hash)); |
abc490a1 | 874 | for (;;) { |
adc0de68 | 875 | uint32_t chain_len = 0; |
abc490a1 | 876 | |
11519af6 MD |
877 | /* |
878 | * iter_prev points to the non-removed node prior to the | |
879 | * insert location. | |
11519af6 | 880 | */ |
14044b37 | 881 | iter_prev = (struct cds_lfht_node *) lookup; |
11519af6 | 882 | /* We can always skip the dummy node initially */ |
cc4fcb10 MD |
883 | iter = rcu_dereference(iter_prev->p.next); |
884 | assert(iter_prev->p.reverse_hash <= node->p.reverse_hash); | |
abc490a1 | 885 | for (;;) { |
bb7b2f26 | 886 | if (unlikely(is_end(iter))) |
273399de | 887 | goto insert; |
76412f24 | 888 | if (likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash)) |
273399de | 889 | goto insert; |
cc4fcb10 | 890 | next = rcu_dereference(clear_flag(iter)->p.next); |
b198f0fd | 891 | if (unlikely(is_removed(next))) |
9dba85be | 892 | goto gc_node; |
48ed1c18 | 893 | if ((mode == ADD_UNIQUE || mode == ADD_REPLACE) |
1b81fe1a | 894 | && !is_dummy(next) |
ae450da3 | 895 | && clear_flag(iter)->p.reverse_hash == node->p.reverse_hash |
e43f23f8 MD |
896 | && !ht->compare_fct(node->key, node->key_len, |
897 | clear_flag(iter)->key, | |
48ed1c18 MD |
898 | clear_flag(iter)->key_len)) { |
899 | if (mode == ADD_UNIQUE) | |
900 | return clear_flag(iter); | |
901 | else /* mode == ADD_REPLACE */ | |
902 | goto replace; | |
903 | } | |
11519af6 | 904 | /* Only account for identical reverse hash once */ |
24365af7 MD |
905 | if (iter_prev->p.reverse_hash != clear_flag(iter)->p.reverse_hash |
906 | && !is_dummy(next)) | |
4105056a | 907 | check_resize(ht, size, ++chain_len); |
11519af6 | 908 | iter_prev = clear_flag(iter); |
273399de | 909 | iter = next; |
abc490a1 | 910 | } |
48ed1c18 | 911 | |
273399de | 912 | insert: |
7ec59d3b | 913 | assert(node != clear_flag(iter)); |
11519af6 | 914 | assert(!is_removed(iter_prev)); |
c90201ac | 915 | assert(!is_removed(iter)); |
f000907d | 916 | assert(iter_prev != node); |
f9c80341 | 917 | if (!dummy) |
1b81fe1a | 918 | node->p.next = clear_flag(iter); |
f9c80341 MD |
919 | else |
920 | node->p.next = flag_dummy(clear_flag(iter)); | |
f5596c94 MD |
921 | if (is_dummy(iter)) |
922 | new_node = flag_dummy(node); | |
923 | else | |
924 | new_node = node; | |
cc4fcb10 | 925 | if (uatomic_cmpxchg(&iter_prev->p.next, iter, |
48ed1c18 | 926 | new_node) != iter) { |
273399de | 927 | continue; /* retry */ |
48ed1c18 MD |
928 | } else { |
929 | if (mode == ADD_REPLACE) | |
930 | return_node = NULL; | |
931 | else /* ADD_DEFAULT and ADD_UNIQUE */ | |
932 | return_node = node; | |
960c9e4f | 933 | goto end; |
48ed1c18 MD |
934 | } |
935 | ||
936 | replace: | |
9357c415 MD |
937 | |
938 | if (!_cds_lfht_replace(ht, size, clear_flag(iter), next, | |
939 | node)) { | |
940 | return_node = clear_flag(iter); | |
941 | goto end; /* gc already done */ | |
48ed1c18 | 942 | } else { |
9357c415 | 943 | continue; /* retry */ |
48ed1c18 MD |
944 | } |
945 | ||
9dba85be MD |
946 | gc_node: |
947 | assert(!is_removed(iter)); | |
f5596c94 MD |
948 | if (is_dummy(iter)) |
949 | new_next = flag_dummy(clear_flag(next)); | |
950 | else | |
951 | new_next = clear_flag(next); | |
952 | (void) uatomic_cmpxchg(&iter_prev->p.next, iter, new_next); | |
273399de | 953 | /* retry */ |
464a1ec9 | 954 | } |
9357c415 | 955 | end: |
48ed1c18 | 956 | return return_node; |
abc490a1 | 957 | } |
464a1ec9 | 958 | |
abc490a1 | 959 | static |
860d07e8 | 960 | int _cds_lfht_del(struct cds_lfht *ht, unsigned long size, |
4105056a | 961 | struct cds_lfht_node *node, |
b198f0fd | 962 | int dummy_removal) |
abc490a1 | 963 | { |
14044b37 MD |
964 | struct cds_lfht_node *dummy, *next, *old; |
965 | struct _cds_lfht_node *lookup; | |
abc490a1 | 966 | int flagged = 0; |
5e28c532 | 967 | |
9357c415 MD |
968 | if (!node) /* Return -ENOENT if asked to delete NULL node */ |
969 | goto end; | |
970 | ||
7ec59d3b | 971 | /* logically delete the node */ |
c90201ac MD |
972 | assert(!is_dummy(node)); |
973 | assert(!is_removed(node)); | |
cc4fcb10 | 974 | old = rcu_dereference(node->p.next); |
7ec59d3b | 975 | do { |
48ed1c18 MD |
976 | struct cds_lfht_node *new_next; |
977 | ||
7ec59d3b | 978 | next = old; |
76412f24 | 979 | if (unlikely(is_removed(next))) |
7ec59d3b | 980 | goto end; |
1475579c MD |
981 | if (dummy_removal) |
982 | assert(is_dummy(next)); | |
983 | else | |
984 | assert(!is_dummy(next)); | |
48ed1c18 | 985 | new_next = flag_removed(next); |
48ed1c18 | 986 | old = uatomic_cmpxchg(&node->p.next, next, new_next); |
7ec59d3b MD |
987 | } while (old != next); |
988 | ||
989 | /* We performed the (logical) deletion. */ | |
990 | flagged = 1; | |
991 | ||
992 | /* | |
993 | * Ensure that the node is not visible to readers anymore: lookup for | |
273399de MD |
994 | * the node, and remove it (along with any other logically removed node) |
995 | * if found. | |
11519af6 | 996 | */ |
f4a9cc0b | 997 | lookup = lookup_bucket(ht, size, bit_reverse_ulong(node->p.reverse_hash)); |
14044b37 | 998 | dummy = (struct cds_lfht_node *) lookup; |
f9c80341 | 999 | _cds_lfht_gc_bucket(dummy, node); |
2ed95849 | 1000 | end: |
11519af6 MD |
1001 | /* |
1002 | * Only the flagging action indicated that we (and no other) | |
1003 | * removed the node from the hash. | |
1004 | */ | |
7ec59d3b | 1005 | if (flagged) { |
cc4fcb10 | 1006 | assert(is_removed(rcu_dereference(node->p.next))); |
11519af6 | 1007 | return 0; |
9357c415 | 1008 | } else { |
11519af6 | 1009 | return -ENOENT; |
9357c415 | 1010 | } |
abc490a1 | 1011 | } |
2ed95849 | 1012 | |
b7d619b0 MD |
1013 | static |
1014 | void *partition_resize_thread(void *arg) | |
1015 | { | |
1016 | struct partition_resize_work *work = arg; | |
1017 | ||
1018 | work->ht->cds_lfht_rcu_register_thread(); | |
1019 | work->fct(work->ht, work->i, work->start, work->len); | |
1020 | work->ht->cds_lfht_rcu_unregister_thread(); | |
1021 | return NULL; | |
1022 | } | |
1023 | ||
1024 | static | |
1025 | void partition_resize_helper(struct cds_lfht *ht, unsigned long i, | |
1026 | unsigned long len, | |
1027 | void (*fct)(struct cds_lfht *ht, unsigned long i, | |
1028 | unsigned long start, unsigned long len)) | |
1029 | { | |
1030 | unsigned long partition_len; | |
1031 | struct partition_resize_work *work; | |
6083a889 MD |
1032 | int thread, ret; |
1033 | unsigned long nr_threads; | |
b7d619b0 | 1034 | |
6083a889 MD |
1035 | /* |
1036 | * Note: nr_cpus_mask + 1 is always power of 2. | |
1037 | * We spawn just the number of threads we need to satisfy the minimum | |
1038 | * partition size, up to the number of CPUs in the system. | |
1039 | */ | |
91452a6a MD |
1040 | if (nr_cpus_mask > 0) { |
1041 | nr_threads = min(nr_cpus_mask + 1, | |
1042 | len >> MIN_PARTITION_PER_THREAD_ORDER); | |
1043 | } else { | |
1044 | nr_threads = 1; | |
1045 | } | |
6083a889 MD |
1046 | partition_len = len >> get_count_order_ulong(nr_threads); |
1047 | work = calloc(nr_threads, sizeof(*work)); | |
b7d619b0 | 1048 | assert(work); |
6083a889 MD |
1049 | for (thread = 0; thread < nr_threads; thread++) { |
1050 | work[thread].ht = ht; | |
1051 | work[thread].i = i; | |
1052 | work[thread].len = partition_len; | |
1053 | work[thread].start = thread * partition_len; | |
1054 | work[thread].fct = fct; | |
1af6e26e | 1055 | ret = pthread_create(&(work[thread].thread_id), ht->resize_attr, |
6083a889 | 1056 | partition_resize_thread, &work[thread]); |
b7d619b0 MD |
1057 | assert(!ret); |
1058 | } | |
6083a889 | 1059 | for (thread = 0; thread < nr_threads; thread++) { |
1af6e26e | 1060 | ret = pthread_join(work[thread].thread_id, NULL); |
b7d619b0 MD |
1061 | assert(!ret); |
1062 | } | |
1063 | free(work); | |
b7d619b0 MD |
1064 | } |
1065 | ||
e8de508e MD |
1066 | /* |
1067 | * Holding RCU read lock to protect _cds_lfht_add against memory | |
1068 | * reclaim that could be performed by other call_rcu worker threads (ABA | |
1069 | * problem). | |
9ee0fc9a | 1070 | * |
b7d619b0 | 1071 | * When we reach a certain length, we can split this population phase over |
9ee0fc9a MD |
1072 | * many worker threads, based on the number of CPUs available in the system. |
1073 | * This should therefore take care of not having the expand lagging behind too | |
1074 | * many concurrent insertion threads by using the scheduler's ability to | |
1075 | * schedule dummy node population fairly with insertions. | |
e8de508e | 1076 | */ |
4105056a | 1077 | static |
b7d619b0 MD |
1078 | void init_table_populate_partition(struct cds_lfht *ht, unsigned long i, |
1079 | unsigned long start, unsigned long len) | |
4105056a MD |
1080 | { |
1081 | unsigned long j; | |
1082 | ||
1083 | ht->cds_lfht_rcu_read_lock(); | |
b7d619b0 | 1084 | for (j = start; j < start + len; j++) { |
4105056a MD |
1085 | struct cds_lfht_node *new_node = |
1086 | (struct cds_lfht_node *) &ht->t.tbl[i]->nodes[j]; | |
1087 | ||
dc1da8f6 | 1088 | dbg_printf("init populate: i %lu j %lu hash %lu\n", |
4105056a | 1089 | i, j, !i ? 0 : (1UL << (i - 1)) + j); |
dc1da8f6 MD |
1090 | new_node->p.reverse_hash = |
1091 | bit_reverse_ulong(!i ? 0 : (1UL << (i - 1)) + j); | |
4105056a | 1092 | (void) _cds_lfht_add(ht, !i ? 0 : (1UL << (i - 1)), |
48ed1c18 | 1093 | new_node, ADD_DEFAULT, 1); |
4105056a MD |
1094 | } |
1095 | ht->cds_lfht_rcu_read_unlock(); | |
b7d619b0 MD |
1096 | } |
1097 | ||
1098 | static | |
1099 | void init_table_populate(struct cds_lfht *ht, unsigned long i, | |
1100 | unsigned long len) | |
1101 | { | |
1102 | assert(nr_cpus_mask != -1); | |
6083a889 | 1103 | if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD) { |
b7d619b0 MD |
1104 | ht->cds_lfht_rcu_thread_online(); |
1105 | init_table_populate_partition(ht, i, 0, len); | |
1106 | ht->cds_lfht_rcu_thread_offline(); | |
1107 | return; | |
1108 | } | |
1109 | partition_resize_helper(ht, i, len, init_table_populate_partition); | |
4105056a MD |
1110 | } |
1111 | ||
abc490a1 | 1112 | static |
4105056a | 1113 | void init_table(struct cds_lfht *ht, |
24365af7 MD |
1114 | unsigned long first_order, unsigned long len_order) |
1115 | { | |
1116 | unsigned long i, end_order; | |
1117 | ||
f0c29ed7 | 1118 | dbg_printf("init table: first_order %lu end_order %lu\n", |
24365af7 MD |
1119 | first_order, first_order + len_order); |
1120 | end_order = first_order + len_order; | |
24365af7 | 1121 | for (i = first_order; i < end_order; i++) { |
4105056a | 1122 | unsigned long len; |
24365af7 MD |
1123 | |
1124 | len = !i ? 1 : 1UL << (i - 1); | |
f0c29ed7 | 1125 | dbg_printf("init order %lu len: %lu\n", i, len); |
4d676753 MD |
1126 | |
1127 | /* Stop expand if the resize target changes under us */ | |
1128 | if (CMM_LOAD_SHARED(ht->t.resize_target) < (!i ? 1 : (1UL << i))) | |
1129 | break; | |
1130 | ||
0d14ceb2 | 1131 | ht->t.tbl[i] = calloc(1, len * sizeof(struct _cds_lfht_node)); |
b7d619b0 | 1132 | assert(ht->t.tbl[i]); |
4105056a | 1133 | |
4105056a | 1134 | /* |
dc1da8f6 MD |
1135 | * Set all dummy nodes reverse hash values for a level and |
1136 | * link all dummy nodes into the table. | |
4105056a | 1137 | */ |
dc1da8f6 | 1138 | init_table_populate(ht, i, len); |
4105056a | 1139 | |
f9c80341 MD |
1140 | /* |
1141 | * Update table size. | |
1142 | */ | |
1143 | cmm_smp_wmb(); /* populate data before RCU size */ | |
1144 | CMM_STORE_SHARED(ht->t.size, !i ? 1 : (1UL << i)); | |
1145 | ||
4105056a MD |
1146 | dbg_printf("init new size: %lu\n", !i ? 1 : (1UL << i)); |
1147 | if (CMM_LOAD_SHARED(ht->in_progress_destroy)) | |
1148 | break; | |
1149 | } | |
1150 | } | |
1151 | ||
e8de508e MD |
1152 | /* |
1153 | * Holding RCU read lock to protect _cds_lfht_remove against memory | |
1154 | * reclaim that could be performed by other call_rcu worker threads (ABA | |
1155 | * problem). | |
1156 | * For a single level, we logically remove and garbage collect each node. | |
1157 | * | |
1158 | * As a design choice, we perform logical removal and garbage collection on a | |
1159 | * node-per-node basis to simplify this algorithm. We also assume keeping good | |
1160 | * cache locality of the operation would overweight possible performance gain | |
1161 | * that could be achieved by batching garbage collection for multiple levels. | |
1162 | * However, this would have to be justified by benchmarks. | |
1163 | * | |
1164 | * Concurrent removal and add operations are helping us perform garbage | |
1165 | * collection of logically removed nodes. We guarantee that all logically | |
1166 | * removed nodes have been garbage-collected (unlinked) before call_rcu is | |
1167 | * invoked to free a hole level of dummy nodes (after a grace period). | |
1168 | * | |
1169 | * Logical removal and garbage collection can therefore be done in batch or on a | |
1170 | * node-per-node basis, as long as the guarantee above holds. | |
9ee0fc9a | 1171 | * |
b7d619b0 MD |
1172 | * When we reach a certain length, we can split this removal over many worker |
1173 | * threads, based on the number of CPUs available in the system. This should | |
1174 | * take care of not letting resize process lag behind too many concurrent | |
9ee0fc9a | 1175 | * updater threads actively inserting into the hash table. |
e8de508e | 1176 | */ |
4105056a | 1177 | static |
b7d619b0 MD |
1178 | void remove_table_partition(struct cds_lfht *ht, unsigned long i, |
1179 | unsigned long start, unsigned long len) | |
4105056a MD |
1180 | { |
1181 | unsigned long j; | |
1182 | ||
1183 | ht->cds_lfht_rcu_read_lock(); | |
b7d619b0 | 1184 | for (j = start; j < start + len; j++) { |
4105056a MD |
1185 | struct cds_lfht_node *fini_node = |
1186 | (struct cds_lfht_node *) &ht->t.tbl[i]->nodes[j]; | |
1187 | ||
1188 | dbg_printf("remove entry: i %lu j %lu hash %lu\n", | |
1189 | i, j, !i ? 0 : (1UL << (i - 1)) + j); | |
1190 | fini_node->p.reverse_hash = | |
1191 | bit_reverse_ulong(!i ? 0 : (1UL << (i - 1)) + j); | |
860d07e8 | 1192 | (void) _cds_lfht_del(ht, !i ? 0 : (1UL << (i - 1)), |
b198f0fd | 1193 | fini_node, 1); |
abc490a1 | 1194 | } |
4105056a | 1195 | ht->cds_lfht_rcu_read_unlock(); |
b7d619b0 MD |
1196 | } |
1197 | ||
1198 | static | |
1199 | void remove_table(struct cds_lfht *ht, unsigned long i, unsigned long len) | |
1200 | { | |
1201 | ||
1202 | assert(nr_cpus_mask != -1); | |
6083a889 | 1203 | if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD) { |
b7d619b0 MD |
1204 | ht->cds_lfht_rcu_thread_online(); |
1205 | remove_table_partition(ht, i, 0, len); | |
1206 | ht->cds_lfht_rcu_thread_offline(); | |
1207 | return; | |
1208 | } | |
1209 | partition_resize_helper(ht, i, len, remove_table_partition); | |
2ed95849 MD |
1210 | } |
1211 | ||
1475579c | 1212 | static |
4105056a | 1213 | void fini_table(struct cds_lfht *ht, |
1475579c MD |
1214 | unsigned long first_order, unsigned long len_order) |
1215 | { | |
1216 | long i, end_order; | |
0d14ceb2 | 1217 | void *free_by_rcu = NULL; |
1475579c MD |
1218 | |
1219 | dbg_printf("fini table: first_order %lu end_order %lu\n", | |
1220 | first_order, first_order + len_order); | |
1221 | end_order = first_order + len_order; | |
1222 | assert(first_order > 0); | |
1475579c | 1223 | for (i = end_order - 1; i >= first_order; i--) { |
4105056a | 1224 | unsigned long len; |
1475579c MD |
1225 | |
1226 | len = !i ? 1 : 1UL << (i - 1); | |
1227 | dbg_printf("fini order %lu len: %lu\n", i, len); | |
4105056a | 1228 | |
4d676753 MD |
1229 | /* Stop shrink if the resize target changes under us */ |
1230 | if (CMM_LOAD_SHARED(ht->t.resize_target) > (1UL << (i - 1))) | |
1231 | break; | |
1232 | ||
1233 | cmm_smp_wmb(); /* populate data before RCU size */ | |
1234 | CMM_STORE_SHARED(ht->t.size, 1UL << (i - 1)); | |
1235 | ||
1236 | /* | |
1237 | * We need to wait for all add operations to reach Q.S. (and | |
1238 | * thus use the new table for lookups) before we can start | |
1239 | * releasing the old dummy nodes. Otherwise their lookup will | |
1240 | * return a logically removed node as insert position. | |
1241 | */ | |
1242 | ht->cds_lfht_synchronize_rcu(); | |
0d14ceb2 LJ |
1243 | if (free_by_rcu) |
1244 | free(free_by_rcu); | |
4d676753 | 1245 | |
21263e21 | 1246 | /* |
4105056a MD |
1247 | * Set "removed" flag in dummy nodes about to be removed. |
1248 | * Unlink all now-logically-removed dummy node pointers. | |
1249 | * Concurrent add/remove operation are helping us doing | |
1250 | * the gc. | |
21263e21 | 1251 | */ |
4105056a MD |
1252 | remove_table(ht, i, len); |
1253 | ||
0d14ceb2 | 1254 | free_by_rcu = ht->t.tbl[i]; |
4105056a MD |
1255 | |
1256 | dbg_printf("fini new size: %lu\n", 1UL << i); | |
1475579c MD |
1257 | if (CMM_LOAD_SHARED(ht->in_progress_destroy)) |
1258 | break; | |
1259 | } | |
0d14ceb2 LJ |
1260 | |
1261 | if (free_by_rcu) { | |
1262 | ht->cds_lfht_synchronize_rcu(); | |
1263 | free(free_by_rcu); | |
1264 | } | |
1475579c MD |
1265 | } |
1266 | ||
7a9dcf9b | 1267 | struct cds_lfht *_cds_lfht_new(cds_lfht_hash_fct hash_fct, |
14044b37 MD |
1268 | cds_lfht_compare_fct compare_fct, |
1269 | unsigned long hash_seed, | |
1270 | unsigned long init_size, | |
b8af5011 | 1271 | int flags, |
14044b37 | 1272 | void (*cds_lfht_call_rcu)(struct rcu_head *head, |
1475579c | 1273 | void (*func)(struct rcu_head *head)), |
01dbfa62 MD |
1274 | void (*cds_lfht_synchronize_rcu)(void), |
1275 | void (*cds_lfht_rcu_read_lock)(void), | |
5f511391 MD |
1276 | void (*cds_lfht_rcu_read_unlock)(void), |
1277 | void (*cds_lfht_rcu_thread_offline)(void), | |
b7d619b0 MD |
1278 | void (*cds_lfht_rcu_thread_online)(void), |
1279 | void (*cds_lfht_rcu_register_thread)(void), | |
1280 | void (*cds_lfht_rcu_unregister_thread)(void), | |
1281 | pthread_attr_t *attr) | |
abc490a1 | 1282 | { |
14044b37 | 1283 | struct cds_lfht *ht; |
24365af7 | 1284 | unsigned long order; |
abc490a1 | 1285 | |
8129be4e | 1286 | /* init_size must be power of two */ |
49619ea0 | 1287 | if (init_size && (init_size & (init_size - 1))) |
8129be4e | 1288 | return NULL; |
14044b37 | 1289 | ht = calloc(1, sizeof(struct cds_lfht)); |
b7d619b0 | 1290 | assert(ht); |
abc490a1 | 1291 | ht->hash_fct = hash_fct; |
732ad076 MD |
1292 | ht->compare_fct = compare_fct; |
1293 | ht->hash_seed = hash_seed; | |
14044b37 | 1294 | ht->cds_lfht_call_rcu = cds_lfht_call_rcu; |
1475579c | 1295 | ht->cds_lfht_synchronize_rcu = cds_lfht_synchronize_rcu; |
01dbfa62 MD |
1296 | ht->cds_lfht_rcu_read_lock = cds_lfht_rcu_read_lock; |
1297 | ht->cds_lfht_rcu_read_unlock = cds_lfht_rcu_read_unlock; | |
5f511391 MD |
1298 | ht->cds_lfht_rcu_thread_offline = cds_lfht_rcu_thread_offline; |
1299 | ht->cds_lfht_rcu_thread_online = cds_lfht_rcu_thread_online; | |
b7d619b0 MD |
1300 | ht->cds_lfht_rcu_register_thread = cds_lfht_rcu_register_thread; |
1301 | ht->cds_lfht_rcu_unregister_thread = cds_lfht_rcu_unregister_thread; | |
1302 | ht->resize_attr = attr; | |
df44348d | 1303 | ht->percpu_count = alloc_per_cpu_items_count(); |
abc490a1 MD |
1304 | /* this mutex should not nest in read-side C.S. */ |
1305 | pthread_mutex_init(&ht->resize_mutex, NULL); | |
cd95516d | 1306 | order = get_count_order_ulong(max(init_size, MIN_TABLE_SIZE)) + 1; |
b8af5011 | 1307 | ht->flags = flags; |
5f511391 | 1308 | ht->cds_lfht_rcu_thread_offline(); |
f000907d | 1309 | pthread_mutex_lock(&ht->resize_mutex); |
4d676753 | 1310 | ht->t.resize_target = 1UL << (order - 1); |
4105056a | 1311 | init_table(ht, 0, order); |
f000907d | 1312 | pthread_mutex_unlock(&ht->resize_mutex); |
5f511391 | 1313 | ht->cds_lfht_rcu_thread_online(); |
abc490a1 MD |
1314 | return ht; |
1315 | } | |
1316 | ||
adc0de68 MD |
1317 | void cds_lfht_lookup(struct cds_lfht *ht, void *key, size_t key_len, |
1318 | struct cds_lfht_iter *iter) | |
2ed95849 | 1319 | { |
bb7b2f26 | 1320 | struct cds_lfht_node *node, *next, *dummy_node; |
14044b37 | 1321 | struct _cds_lfht_node *lookup; |
f4a9cc0b | 1322 | unsigned long hash, reverse_hash, size; |
2ed95849 | 1323 | |
732ad076 | 1324 | hash = ht->hash_fct(key, key_len, ht->hash_seed); |
abc490a1 | 1325 | reverse_hash = bit_reverse_ulong(hash); |
464a1ec9 | 1326 | |
4105056a | 1327 | size = rcu_dereference(ht->t.size); |
f4a9cc0b | 1328 | lookup = lookup_bucket(ht, size, hash); |
bb7b2f26 MD |
1329 | dummy_node = (struct cds_lfht_node *) lookup; |
1330 | /* We can always skip the dummy node initially */ | |
1331 | node = rcu_dereference(dummy_node->p.next); | |
bb7b2f26 | 1332 | node = clear_flag(node); |
2ed95849 | 1333 | for (;;) { |
bb7b2f26 | 1334 | if (unlikely(is_end(node))) { |
96ad1112 | 1335 | node = next = NULL; |
abc490a1 | 1336 | break; |
bb7b2f26 | 1337 | } |
cc4fcb10 | 1338 | if (unlikely(node->p.reverse_hash > reverse_hash)) { |
96ad1112 | 1339 | node = next = NULL; |
abc490a1 | 1340 | break; |
2ed95849 | 1341 | } |
1b81fe1a | 1342 | next = rcu_dereference(node->p.next); |
adc0de68 | 1343 | if (likely(!is_removed(next)) |
1b81fe1a | 1344 | && !is_dummy(next) |
ae450da3 | 1345 | && clear_flag(node)->p.reverse_hash == reverse_hash |
49c2e2d6 | 1346 | && likely(!ht->compare_fct(node->key, node->key_len, key, key_len))) { |
273399de | 1347 | break; |
2ed95849 | 1348 | } |
1b81fe1a | 1349 | node = clear_flag(next); |
2ed95849 | 1350 | } |
1b81fe1a | 1351 | assert(!node || !is_dummy(rcu_dereference(node->p.next))); |
adc0de68 MD |
1352 | iter->node = node; |
1353 | iter->next = next; | |
abc490a1 | 1354 | } |
e0ba718a | 1355 | |
3883c0e5 | 1356 | void cds_lfht_next_duplicate(struct cds_lfht *ht, struct cds_lfht_iter *iter) |
a481e5ff | 1357 | { |
adc0de68 | 1358 | struct cds_lfht_node *node, *next; |
a481e5ff MD |
1359 | unsigned long reverse_hash; |
1360 | void *key; | |
1361 | size_t key_len; | |
1362 | ||
adc0de68 | 1363 | node = iter->node; |
a481e5ff MD |
1364 | reverse_hash = node->p.reverse_hash; |
1365 | key = node->key; | |
1366 | key_len = node->key_len; | |
adc0de68 | 1367 | next = iter->next; |
a481e5ff MD |
1368 | node = clear_flag(next); |
1369 | ||
1370 | for (;;) { | |
bb7b2f26 | 1371 | if (unlikely(is_end(node))) { |
96ad1112 | 1372 | node = next = NULL; |
a481e5ff | 1373 | break; |
bb7b2f26 | 1374 | } |
a481e5ff | 1375 | if (unlikely(node->p.reverse_hash > reverse_hash)) { |
96ad1112 | 1376 | node = next = NULL; |
a481e5ff MD |
1377 | break; |
1378 | } | |
1379 | next = rcu_dereference(node->p.next); | |
adc0de68 | 1380 | if (likely(!is_removed(next)) |
a481e5ff MD |
1381 | && !is_dummy(next) |
1382 | && likely(!ht->compare_fct(node->key, node->key_len, key, key_len))) { | |
1383 | break; | |
1384 | } | |
1385 | node = clear_flag(next); | |
1386 | } | |
1387 | assert(!node || !is_dummy(rcu_dereference(node->p.next))); | |
adc0de68 MD |
1388 | iter->node = node; |
1389 | iter->next = next; | |
a481e5ff MD |
1390 | } |
1391 | ||
4e9b9fbf MD |
1392 | void cds_lfht_next(struct cds_lfht *ht, struct cds_lfht_iter *iter) |
1393 | { | |
1394 | struct cds_lfht_node *node, *next; | |
1395 | ||
853395e1 | 1396 | node = clear_flag(iter->next); |
4e9b9fbf MD |
1397 | for (;;) { |
1398 | if (unlikely(is_end(node))) { | |
1399 | node = next = NULL; | |
1400 | break; | |
1401 | } | |
1402 | next = rcu_dereference(node->p.next); | |
1403 | if (likely(!is_removed(next)) | |
1404 | && !is_dummy(next)) { | |
1405 | break; | |
1406 | } | |
1407 | node = clear_flag(next); | |
1408 | } | |
1409 | assert(!node || !is_dummy(rcu_dereference(node->p.next))); | |
1410 | iter->node = node; | |
1411 | iter->next = next; | |
1412 | } | |
1413 | ||
1414 | void cds_lfht_first(struct cds_lfht *ht, struct cds_lfht_iter *iter) | |
1415 | { | |
1416 | struct _cds_lfht_node *lookup; | |
1417 | ||
1418 | /* | |
1419 | * Get next after first dummy node. The first dummy node is the | |
1420 | * first node of the linked list. | |
1421 | */ | |
1422 | lookup = &ht->t.tbl[0]->nodes[0]; | |
853395e1 | 1423 | iter->next = lookup->next; |
4e9b9fbf MD |
1424 | cds_lfht_next(ht, iter); |
1425 | } | |
1426 | ||
14044b37 | 1427 | void cds_lfht_add(struct cds_lfht *ht, struct cds_lfht_node *node) |
abc490a1 | 1428 | { |
4105056a | 1429 | unsigned long hash, size; |
ab7d5fc6 | 1430 | |
49c2e2d6 | 1431 | hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed); |
cc4fcb10 | 1432 | node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash); |
2ed95849 | 1433 | |
4105056a | 1434 | size = rcu_dereference(ht->t.size); |
48ed1c18 | 1435 | (void) _cds_lfht_add(ht, size, node, ADD_DEFAULT, 0); |
4105056a | 1436 | ht_count_add(ht, size); |
3eca1b8c MD |
1437 | } |
1438 | ||
14044b37 | 1439 | struct cds_lfht_node *cds_lfht_add_unique(struct cds_lfht *ht, |
48ed1c18 | 1440 | struct cds_lfht_node *node) |
3eca1b8c | 1441 | { |
4105056a | 1442 | unsigned long hash, size; |
df44348d | 1443 | struct cds_lfht_node *ret; |
3eca1b8c | 1444 | |
49c2e2d6 | 1445 | hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed); |
cc4fcb10 | 1446 | node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash); |
3eca1b8c | 1447 | |
4105056a | 1448 | size = rcu_dereference(ht->t.size); |
48ed1c18 | 1449 | ret = _cds_lfht_add(ht, size, node, ADD_UNIQUE, 0); |
17f31d1b | 1450 | if (ret == node) |
4105056a | 1451 | ht_count_add(ht, size); |
df44348d | 1452 | return ret; |
2ed95849 MD |
1453 | } |
1454 | ||
9357c415 | 1455 | struct cds_lfht_node *cds_lfht_add_replace(struct cds_lfht *ht, |
48ed1c18 MD |
1456 | struct cds_lfht_node *node) |
1457 | { | |
1458 | unsigned long hash, size; | |
1459 | struct cds_lfht_node *ret; | |
1460 | ||
1461 | hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed); | |
1462 | node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash); | |
1463 | ||
1464 | size = rcu_dereference(ht->t.size); | |
1465 | ret = _cds_lfht_add(ht, size, node, ADD_REPLACE, 0); | |
1466 | if (ret == NULL) | |
1467 | ht_count_add(ht, size); | |
1468 | return ret; | |
1469 | } | |
1470 | ||
9357c415 MD |
1471 | int cds_lfht_replace(struct cds_lfht *ht, struct cds_lfht_iter *old_iter, |
1472 | struct cds_lfht_node *new_node) | |
1473 | { | |
1474 | unsigned long size; | |
1475 | ||
1476 | size = rcu_dereference(ht->t.size); | |
1477 | return _cds_lfht_replace(ht, size, old_iter->node, old_iter->next, | |
1478 | new_node); | |
1479 | } | |
1480 | ||
1481 | int cds_lfht_del(struct cds_lfht *ht, struct cds_lfht_iter *iter) | |
2ed95849 | 1482 | { |
4105056a | 1483 | unsigned long size; |
df44348d | 1484 | int ret; |
abc490a1 | 1485 | |
4105056a | 1486 | size = rcu_dereference(ht->t.size); |
9357c415 | 1487 | ret = _cds_lfht_del(ht, size, iter->node, 0); |
df44348d | 1488 | if (!ret) |
860d07e8 | 1489 | ht_count_del(ht, size); |
df44348d | 1490 | return ret; |
2ed95849 | 1491 | } |
ab7d5fc6 | 1492 | |
abc490a1 | 1493 | static |
14044b37 | 1494 | int cds_lfht_delete_dummy(struct cds_lfht *ht) |
674f7a69 | 1495 | { |
14044b37 MD |
1496 | struct cds_lfht_node *node; |
1497 | struct _cds_lfht_node *lookup; | |
4105056a | 1498 | unsigned long order, i, size; |
674f7a69 | 1499 | |
abc490a1 | 1500 | /* Check that the table is empty */ |
4105056a | 1501 | lookup = &ht->t.tbl[0]->nodes[0]; |
14044b37 | 1502 | node = (struct cds_lfht_node *) lookup; |
abc490a1 | 1503 | do { |
1b81fe1a MD |
1504 | node = clear_flag(node)->p.next; |
1505 | if (!is_dummy(node)) | |
abc490a1 | 1506 | return -EPERM; |
273399de | 1507 | assert(!is_removed(node)); |
bb7b2f26 | 1508 | } while (!is_end(node)); |
4105056a MD |
1509 | /* |
1510 | * size accessed without rcu_dereference because hash table is | |
1511 | * being destroyed. | |
1512 | */ | |
1513 | size = ht->t.size; | |
abc490a1 | 1514 | /* Internal sanity check: all nodes left should be dummy */ |
4105056a | 1515 | for (order = 0; order < get_count_order_ulong(size) + 1; order++) { |
24365af7 MD |
1516 | unsigned long len; |
1517 | ||
1518 | len = !order ? 1 : 1UL << (order - 1); | |
1519 | for (i = 0; i < len; i++) { | |
f0c29ed7 | 1520 | dbg_printf("delete order %lu i %lu hash %lu\n", |
24365af7 | 1521 | order, i, |
4105056a MD |
1522 | bit_reverse_ulong(ht->t.tbl[order]->nodes[i].reverse_hash)); |
1523 | assert(is_dummy(ht->t.tbl[order]->nodes[i].next)); | |
24365af7 | 1524 | } |
4105056a | 1525 | poison_free(ht->t.tbl[order]); |
674f7a69 | 1526 | } |
abc490a1 | 1527 | return 0; |
674f7a69 MD |
1528 | } |
1529 | ||
1530 | /* | |
1531 | * Should only be called when no more concurrent readers nor writers can | |
1532 | * possibly access the table. | |
1533 | */ | |
b7d619b0 | 1534 | int cds_lfht_destroy(struct cds_lfht *ht, pthread_attr_t **attr) |
674f7a69 | 1535 | { |
5e28c532 MD |
1536 | int ret; |
1537 | ||
848d4088 | 1538 | /* Wait for in-flight resize operations to complete */ |
24953e08 MD |
1539 | _CMM_STORE_SHARED(ht->in_progress_destroy, 1); |
1540 | cmm_smp_mb(); /* Store destroy before load resize */ | |
848d4088 MD |
1541 | while (uatomic_read(&ht->in_progress_resize)) |
1542 | poll(NULL, 0, 100); /* wait for 100ms */ | |
14044b37 | 1543 | ret = cds_lfht_delete_dummy(ht); |
abc490a1 MD |
1544 | if (ret) |
1545 | return ret; | |
df44348d | 1546 | free_per_cpu_items_count(ht->percpu_count); |
b7d619b0 MD |
1547 | if (attr) |
1548 | *attr = ht->resize_attr; | |
98808fb1 | 1549 | poison_free(ht); |
5e28c532 | 1550 | return ret; |
674f7a69 MD |
1551 | } |
1552 | ||
14044b37 | 1553 | void cds_lfht_count_nodes(struct cds_lfht *ht, |
d933dd0e | 1554 | long *approx_before, |
273399de | 1555 | unsigned long *count, |
973e5e1b | 1556 | unsigned long *removed, |
d933dd0e | 1557 | long *approx_after) |
273399de | 1558 | { |
14044b37 MD |
1559 | struct cds_lfht_node *node, *next; |
1560 | struct _cds_lfht_node *lookup; | |
24365af7 | 1561 | unsigned long nr_dummy = 0; |
273399de | 1562 | |
7ed7682f | 1563 | *approx_before = 0; |
973e5e1b MD |
1564 | if (nr_cpus_mask >= 0) { |
1565 | int i; | |
1566 | ||
1567 | for (i = 0; i < nr_cpus_mask + 1; i++) { | |
1568 | *approx_before += uatomic_read(&ht->percpu_count[i].add); | |
1569 | *approx_before -= uatomic_read(&ht->percpu_count[i].del); | |
1570 | } | |
1571 | } | |
1572 | ||
273399de MD |
1573 | *count = 0; |
1574 | *removed = 0; | |
1575 | ||
24365af7 | 1576 | /* Count non-dummy nodes in the table */ |
4105056a | 1577 | lookup = &ht->t.tbl[0]->nodes[0]; |
14044b37 | 1578 | node = (struct cds_lfht_node *) lookup; |
273399de | 1579 | do { |
cc4fcb10 | 1580 | next = rcu_dereference(node->p.next); |
b198f0fd | 1581 | if (is_removed(next)) { |
973e5e1b MD |
1582 | if (!is_dummy(next)) |
1583 | (*removed)++; | |
1584 | else | |
1585 | (nr_dummy)++; | |
1b81fe1a | 1586 | } else if (!is_dummy(next)) |
273399de | 1587 | (*count)++; |
24365af7 MD |
1588 | else |
1589 | (nr_dummy)++; | |
273399de | 1590 | node = clear_flag(next); |
bb7b2f26 | 1591 | } while (!is_end(node)); |
f0c29ed7 | 1592 | dbg_printf("number of dummy nodes: %lu\n", nr_dummy); |
7ed7682f | 1593 | *approx_after = 0; |
973e5e1b MD |
1594 | if (nr_cpus_mask >= 0) { |
1595 | int i; | |
1596 | ||
1597 | for (i = 0; i < nr_cpus_mask + 1; i++) { | |
1598 | *approx_after += uatomic_read(&ht->percpu_count[i].add); | |
1599 | *approx_after -= uatomic_read(&ht->percpu_count[i].del); | |
1600 | } | |
1601 | } | |
273399de MD |
1602 | } |
1603 | ||
1475579c | 1604 | /* called with resize mutex held */ |
abc490a1 | 1605 | static |
4105056a | 1606 | void _do_cds_lfht_grow(struct cds_lfht *ht, |
1475579c | 1607 | unsigned long old_size, unsigned long new_size) |
abc490a1 | 1608 | { |
1475579c | 1609 | unsigned long old_order, new_order; |
1475579c MD |
1610 | |
1611 | old_order = get_count_order_ulong(old_size) + 1; | |
1612 | new_order = get_count_order_ulong(new_size) + 1; | |
1a401918 LJ |
1613 | dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n", |
1614 | old_size, old_order, new_size, new_order); | |
1475579c | 1615 | assert(new_size > old_size); |
4105056a | 1616 | init_table(ht, old_order, new_order - old_order); |
abc490a1 MD |
1617 | } |
1618 | ||
1619 | /* called with resize mutex held */ | |
1620 | static | |
4105056a | 1621 | void _do_cds_lfht_shrink(struct cds_lfht *ht, |
1475579c | 1622 | unsigned long old_size, unsigned long new_size) |
464a1ec9 | 1623 | { |
1475579c | 1624 | unsigned long old_order, new_order; |
464a1ec9 | 1625 | |
cd95516d | 1626 | new_size = max(new_size, MIN_TABLE_SIZE); |
24365af7 | 1627 | old_order = get_count_order_ulong(old_size) + 1; |
24365af7 | 1628 | new_order = get_count_order_ulong(new_size) + 1; |
1a401918 LJ |
1629 | dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n", |
1630 | old_size, old_order, new_size, new_order); | |
1475579c | 1631 | assert(new_size < old_size); |
1475579c | 1632 | |
4105056a MD |
1633 | /* Remove and unlink all dummy nodes to remove. */ |
1634 | fini_table(ht, new_order, old_order - new_order); | |
464a1ec9 MD |
1635 | } |
1636 | ||
1475579c MD |
1637 | |
1638 | /* called with resize mutex held */ | |
1639 | static | |
1640 | void _do_cds_lfht_resize(struct cds_lfht *ht) | |
1641 | { | |
1642 | unsigned long new_size, old_size; | |
4105056a MD |
1643 | |
1644 | /* | |
1645 | * Resize table, re-do if the target size has changed under us. | |
1646 | */ | |
1647 | do { | |
d2be3620 MD |
1648 | assert(uatomic_read(&ht->in_progress_resize)); |
1649 | if (CMM_LOAD_SHARED(ht->in_progress_destroy)) | |
1650 | break; | |
4105056a MD |
1651 | ht->t.resize_initiated = 1; |
1652 | old_size = ht->t.size; | |
1653 | new_size = CMM_LOAD_SHARED(ht->t.resize_target); | |
1654 | if (old_size < new_size) | |
1655 | _do_cds_lfht_grow(ht, old_size, new_size); | |
1656 | else if (old_size > new_size) | |
1657 | _do_cds_lfht_shrink(ht, old_size, new_size); | |
1658 | ht->t.resize_initiated = 0; | |
1659 | /* write resize_initiated before read resize_target */ | |
1660 | cmm_smp_mb(); | |
4d676753 | 1661 | } while (ht->t.size != CMM_LOAD_SHARED(ht->t.resize_target)); |
1475579c MD |
1662 | } |
1663 | ||
abc490a1 | 1664 | static |
4105056a | 1665 | unsigned long resize_target_update(struct cds_lfht *ht, unsigned long size, |
f9830efd | 1666 | int growth_order) |
464a1ec9 | 1667 | { |
4105056a MD |
1668 | return _uatomic_max(&ht->t.resize_target, |
1669 | size << growth_order); | |
464a1ec9 MD |
1670 | } |
1671 | ||
1475579c | 1672 | static |
4105056a | 1673 | void resize_target_update_count(struct cds_lfht *ht, |
b8af5011 | 1674 | unsigned long count) |
1475579c | 1675 | { |
cd95516d | 1676 | count = max(count, MIN_TABLE_SIZE); |
4105056a | 1677 | uatomic_set(&ht->t.resize_target, count); |
1475579c MD |
1678 | } |
1679 | ||
1680 | void cds_lfht_resize(struct cds_lfht *ht, unsigned long new_size) | |
464a1ec9 | 1681 | { |
4105056a MD |
1682 | resize_target_update_count(ht, new_size); |
1683 | CMM_STORE_SHARED(ht->t.resize_initiated, 1); | |
5f511391 | 1684 | ht->cds_lfht_rcu_thread_offline(); |
1475579c MD |
1685 | pthread_mutex_lock(&ht->resize_mutex); |
1686 | _do_cds_lfht_resize(ht); | |
1687 | pthread_mutex_unlock(&ht->resize_mutex); | |
5f511391 | 1688 | ht->cds_lfht_rcu_thread_online(); |
abc490a1 | 1689 | } |
464a1ec9 | 1690 | |
abc490a1 MD |
1691 | static |
1692 | void do_resize_cb(struct rcu_head *head) | |
1693 | { | |
1694 | struct rcu_resize_work *work = | |
1695 | caa_container_of(head, struct rcu_resize_work, head); | |
14044b37 | 1696 | struct cds_lfht *ht = work->ht; |
abc490a1 | 1697 | |
5f511391 | 1698 | ht->cds_lfht_rcu_thread_offline(); |
abc490a1 | 1699 | pthread_mutex_lock(&ht->resize_mutex); |
14044b37 | 1700 | _do_cds_lfht_resize(ht); |
abc490a1 | 1701 | pthread_mutex_unlock(&ht->resize_mutex); |
5f511391 | 1702 | ht->cds_lfht_rcu_thread_online(); |
98808fb1 | 1703 | poison_free(work); |
848d4088 MD |
1704 | cmm_smp_mb(); /* finish resize before decrement */ |
1705 | uatomic_dec(&ht->in_progress_resize); | |
464a1ec9 MD |
1706 | } |
1707 | ||
abc490a1 | 1708 | static |
4105056a | 1709 | void cds_lfht_resize_lazy(struct cds_lfht *ht, unsigned long size, int growth) |
ab7d5fc6 | 1710 | { |
abc490a1 | 1711 | struct rcu_resize_work *work; |
f9830efd | 1712 | unsigned long target_size; |
abc490a1 | 1713 | |
4105056a MD |
1714 | target_size = resize_target_update(ht, size, growth); |
1715 | /* Store resize_target before read resize_initiated */ | |
1716 | cmm_smp_mb(); | |
1717 | if (!CMM_LOAD_SHARED(ht->t.resize_initiated) && size < target_size) { | |
848d4088 | 1718 | uatomic_inc(&ht->in_progress_resize); |
59290e9d | 1719 | cmm_smp_mb(); /* increment resize count before load destroy */ |
ed35e6d8 MD |
1720 | if (CMM_LOAD_SHARED(ht->in_progress_destroy)) { |
1721 | uatomic_dec(&ht->in_progress_resize); | |
59290e9d | 1722 | return; |
ed35e6d8 | 1723 | } |
f9830efd MD |
1724 | work = malloc(sizeof(*work)); |
1725 | work->ht = ht; | |
14044b37 | 1726 | ht->cds_lfht_call_rcu(&work->head, do_resize_cb); |
4105056a | 1727 | CMM_STORE_SHARED(ht->t.resize_initiated, 1); |
f9830efd | 1728 | } |
ab7d5fc6 | 1729 | } |
3171717f | 1730 | |
f8994aee MD |
1731 | #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) |
1732 | ||
3171717f | 1733 | static |
4105056a | 1734 | void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size, |
3171717f MD |
1735 | unsigned long count) |
1736 | { | |
1737 | struct rcu_resize_work *work; | |
3171717f | 1738 | |
b8af5011 MD |
1739 | if (!(ht->flags & CDS_LFHT_AUTO_RESIZE)) |
1740 | return; | |
4105056a MD |
1741 | resize_target_update_count(ht, count); |
1742 | /* Store resize_target before read resize_initiated */ | |
1743 | cmm_smp_mb(); | |
1744 | if (!CMM_LOAD_SHARED(ht->t.resize_initiated)) { | |
3171717f | 1745 | uatomic_inc(&ht->in_progress_resize); |
59290e9d | 1746 | cmm_smp_mb(); /* increment resize count before load destroy */ |
ed35e6d8 MD |
1747 | if (CMM_LOAD_SHARED(ht->in_progress_destroy)) { |
1748 | uatomic_dec(&ht->in_progress_resize); | |
59290e9d | 1749 | return; |
ed35e6d8 | 1750 | } |
3171717f MD |
1751 | work = malloc(sizeof(*work)); |
1752 | work->ht = ht; | |
1753 | ht->cds_lfht_call_rcu(&work->head, do_resize_cb); | |
4105056a | 1754 | CMM_STORE_SHARED(ht->t.resize_initiated, 1); |
3171717f MD |
1755 | } |
1756 | } | |
f8994aee MD |
1757 | |
1758 | #endif |