Commit | Line | Data |
---|---|---|
786ee85b MD |
1 | /* |
2 | * urcu-defer.c | |
3 | * | |
4 | * Userspace RCU library - batch memory reclamation | |
5 | * | |
6 | * Copyright (c) 2009 Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca> | |
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 | |
21 | */ | |
22 | ||
23 | #include <stdio.h> | |
24 | #include <pthread.h> | |
25 | #include <signal.h> | |
26 | #include <assert.h> | |
27 | #include <stdlib.h> | |
28 | #include <string.h> | |
29 | #include <errno.h> | |
30 | #include <poll.h> | |
4ce9e4f2 MD |
31 | #include <linux/futex.h> |
32 | #include <sys/time.h> | |
33 | #include <syscall.h> | |
34 | #include <unistd.h> | |
786ee85b MD |
35 | |
36 | #include "urcu-defer-static.h" | |
37 | /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */ | |
38 | #include "urcu-defer.h" | |
39 | ||
4ce9e4f2 MD |
40 | #define futex(...) syscall(__NR_futex, __VA_ARGS__) |
41 | ||
786ee85b MD |
42 | void __attribute__((destructor)) urcu_defer_exit(void); |
43 | ||
44 | extern void synchronize_rcu(void); | |
45 | ||
46 | /* | |
47 | * urcu_defer_mutex nests inside defer_thread_mutex. | |
48 | */ | |
49 | static pthread_mutex_t urcu_defer_mutex = PTHREAD_MUTEX_INITIALIZER; | |
50 | static pthread_mutex_t defer_thread_mutex = PTHREAD_MUTEX_INITIALIZER; | |
51 | ||
3dce4bfa | 52 | static int defer_thread_futex; |
4ce9e4f2 | 53 | |
786ee85b MD |
54 | /* |
55 | * Written to only by each individual deferer. Read by both the deferer and | |
56 | * the reclamation tread. | |
57 | */ | |
3dce4bfa | 58 | static struct defer_queue __thread defer_queue; |
786ee85b MD |
59 | |
60 | /* Thread IDs of registered deferers */ | |
61 | #define INIT_NUM_THREADS 4 | |
62 | ||
63 | struct deferer_registry { | |
64 | pthread_t tid; | |
65 | struct defer_queue *defer_queue; | |
66 | unsigned long last_head; | |
67 | }; | |
68 | ||
69 | static struct deferer_registry *registry; | |
70 | static int num_deferers, alloc_deferers; | |
71 | ||
72 | static pthread_t tid_defer; | |
4ce9e4f2 MD |
73 | |
74 | /* | |
75 | * Wake-up any waiting defer thread. Called from many concurrent threads. | |
76 | */ | |
2c22932b | 77 | static void wake_up_defer(void) |
4ce9e4f2 | 78 | { |
3dce4bfa | 79 | if (unlikely(atomic_read(&defer_thread_futex) == -1)) { |
4ce9e4f2 | 80 | atomic_set(&defer_thread_futex, 0); |
3dce4bfa MD |
81 | futex(&defer_thread_futex, FUTEX_WAKE, 0, |
82 | NULL, NULL, 0); | |
83 | } | |
4ce9e4f2 MD |
84 | } |
85 | ||
86 | /* | |
87 | * Defer thread waiting. Single thread. | |
88 | */ | |
89 | static void wait_defer(void) | |
90 | { | |
91 | atomic_dec(&defer_thread_futex); | |
92 | if (atomic_read(&defer_thread_futex) == -1) | |
93 | futex(&defer_thread_futex, FUTEX_WAIT, -1, | |
94 | NULL, NULL, 0); | |
95 | } | |
786ee85b MD |
96 | |
97 | static void internal_urcu_lock(pthread_mutex_t *mutex) | |
98 | { | |
99 | int ret; | |
100 | ||
101 | #ifndef DISTRUST_SIGNALS_EXTREME | |
102 | ret = pthread_mutex_lock(mutex); | |
103 | if (ret) { | |
104 | perror("Error in pthread mutex lock"); | |
105 | exit(-1); | |
106 | } | |
107 | #else /* #ifndef DISTRUST_SIGNALS_EXTREME */ | |
108 | while ((ret = pthread_mutex_trylock(mutex)) != 0) { | |
109 | if (ret != EBUSY && ret != EINTR) { | |
110 | printf("ret = %d, errno = %d\n", ret, errno); | |
111 | perror("Error in pthread mutex lock"); | |
112 | exit(-1); | |
113 | } | |
4ce9e4f2 | 114 | pthread_testcancel(); |
786ee85b MD |
115 | poll(NULL,0,10); |
116 | } | |
117 | #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */ | |
118 | } | |
119 | ||
120 | static void internal_urcu_unlock(pthread_mutex_t *mutex) | |
121 | { | |
122 | int ret; | |
123 | ||
124 | ret = pthread_mutex_unlock(mutex); | |
125 | if (ret) { | |
126 | perror("Error in pthread mutex unlock"); | |
127 | exit(-1); | |
128 | } | |
129 | } | |
130 | ||
131 | /* | |
132 | * Must be called after Q.S. is reached. | |
133 | */ | |
134 | static void rcu_defer_barrier_queue(struct defer_queue *queue, | |
804b4375 | 135 | unsigned long head) |
786ee85b MD |
136 | { |
137 | unsigned long i; | |
804b4375 MD |
138 | void (*fct)(void *p); |
139 | void *p; | |
786ee85b MD |
140 | |
141 | /* | |
142 | * Tail is only modified when lock is held. | |
143 | * Head is only modified by owner thread. | |
144 | */ | |
145 | ||
804b4375 | 146 | for (i = queue->tail; i != head;) { |
786ee85b | 147 | smp_rmb(); /* read head before q[]. */ |
804b4375 MD |
148 | p = LOAD_SHARED(queue->q[i++ & DEFER_QUEUE_MASK]); |
149 | if (unlikely(DQ_IS_FCT_BIT(p))) { | |
804b4375 MD |
150 | DQ_CLEAR_FCT_BIT(p); |
151 | queue->last_fct_out = p; | |
152 | p = LOAD_SHARED(queue->q[i++ & DEFER_QUEUE_MASK]); | |
153 | } else if (unlikely(p == DQ_FCT_MARK)) { | |
804b4375 MD |
154 | p = LOAD_SHARED(queue->q[i++ & DEFER_QUEUE_MASK]); |
155 | queue->last_fct_out = p; | |
156 | p = LOAD_SHARED(queue->q[i++ & DEFER_QUEUE_MASK]); | |
29cdb8d8 | 157 | } |
804b4375 | 158 | fct = queue->last_fct_out; |
804b4375 | 159 | fct(p); |
786ee85b MD |
160 | } |
161 | smp_mb(); /* push tail after having used q[] */ | |
162 | STORE_SHARED(queue->tail, i); | |
163 | } | |
164 | ||
165 | static void _rcu_defer_barrier_thread(void) | |
166 | { | |
0d0e6c21 | 167 | unsigned long head, num_items; |
786ee85b MD |
168 | |
169 | head = defer_queue.head; | |
0d0e6c21 MD |
170 | num_items = head - defer_queue.tail; |
171 | if (unlikely(!num_items)) | |
172 | return; | |
786ee85b MD |
173 | synchronize_rcu(); |
174 | rcu_defer_barrier_queue(&defer_queue, head); | |
175 | } | |
176 | ||
177 | ||
178 | void rcu_defer_barrier_thread(void) | |
179 | { | |
180 | internal_urcu_lock(&urcu_defer_mutex); | |
181 | _rcu_defer_barrier_thread(); | |
182 | internal_urcu_unlock(&urcu_defer_mutex); | |
183 | } | |
184 | ||
0d0e6c21 MD |
185 | /* |
186 | * rcu_defer_barrier - Execute all queued rcu callbacks. | |
187 | * | |
188 | * Execute all RCU callbacks queued before rcu_defer_barrier() execution. | |
189 | * All callbacks queued on the local thread prior to a rcu_defer_barrier() call | |
190 | * are guaranteed to be executed. | |
191 | * Callbacks queued by other threads concurrently with rcu_defer_barrier() | |
192 | * execution are not guaranteed to be executed in the current batch (could | |
193 | * be left for the next batch). These callbacks queued by other threads are only | |
194 | * guaranteed to be executed if there is explicit synchronization between | |
195 | * the thread adding to the queue and the thread issuing the defer_barrier call. | |
196 | */ | |
197 | ||
786ee85b MD |
198 | void rcu_defer_barrier(void) |
199 | { | |
200 | struct deferer_registry *index; | |
0d0e6c21 | 201 | unsigned long num_items = 0; |
786ee85b MD |
202 | |
203 | if (!registry) | |
204 | return; | |
205 | ||
206 | internal_urcu_lock(&urcu_defer_mutex); | |
0d0e6c21 | 207 | for (index = registry; index < registry + num_deferers; index++) { |
786ee85b | 208 | index->last_head = LOAD_SHARED(index->defer_queue->head); |
0d0e6c21 MD |
209 | num_items += index->last_head - index->defer_queue->tail; |
210 | } | |
211 | if (likely(!num_items)) { | |
212 | /* | |
213 | * We skip the grace period because there are no queued | |
214 | * callbacks to execute. | |
215 | */ | |
216 | goto end; | |
217 | } | |
786ee85b MD |
218 | synchronize_rcu(); |
219 | for (index = registry; index < registry + num_deferers; index++) | |
220 | rcu_defer_barrier_queue(index->defer_queue, | |
221 | index->last_head); | |
0d0e6c21 | 222 | end: |
786ee85b MD |
223 | internal_urcu_unlock(&urcu_defer_mutex); |
224 | } | |
225 | ||
2c22932b MD |
226 | /* |
227 | * _rcu_defer_queue - Queue a RCU callback. | |
228 | */ | |
229 | void _rcu_defer_queue(void (*fct)(void *p), void *p) | |
230 | { | |
231 | unsigned long head, tail; | |
232 | ||
233 | /* | |
234 | * Head is only modified by ourself. Tail can be modified by reclamation | |
235 | * thread. | |
236 | */ | |
237 | head = defer_queue.head; | |
238 | tail = LOAD_SHARED(defer_queue.tail); | |
239 | ||
240 | /* | |
241 | * If queue is full, empty it ourself. | |
242 | * Worse-case: must allow 2 supplementary entries for fct pointer. | |
243 | */ | |
244 | if (unlikely(head - tail >= DEFER_QUEUE_SIZE - 2)) { | |
245 | assert(head - tail <= DEFER_QUEUE_SIZE); | |
246 | rcu_defer_barrier_thread(); | |
247 | assert(head - LOAD_SHARED(defer_queue.tail) == 0); | |
248 | } | |
249 | ||
250 | if (unlikely(defer_queue.last_fct_in != fct)) { | |
251 | defer_queue.last_fct_in = fct; | |
252 | if (unlikely(DQ_IS_FCT_BIT(fct) || fct == DQ_FCT_MARK)) { | |
253 | /* | |
254 | * If the function to encode is not aligned or the | |
255 | * marker, write DQ_FCT_MARK followed by the function | |
256 | * pointer. | |
257 | */ | |
258 | _STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK], | |
259 | DQ_FCT_MARK); | |
260 | _STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK], | |
261 | fct); | |
262 | } else { | |
263 | DQ_SET_FCT_BIT(fct); | |
264 | _STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK], | |
265 | fct); | |
266 | } | |
267 | } else { | |
268 | if (unlikely(DQ_IS_FCT_BIT(p) || p == DQ_FCT_MARK)) { | |
269 | /* | |
270 | * If the data to encode is not aligned or the marker, | |
271 | * write DQ_FCT_MARK followed by the function pointer. | |
272 | */ | |
273 | _STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK], | |
274 | DQ_FCT_MARK); | |
275 | _STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK], | |
276 | fct); | |
277 | } | |
278 | } | |
279 | _STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK], p); | |
280 | smp_wmb(); /* Publish new pointer before head */ | |
281 | /* Write q[] before head. */ | |
282 | STORE_SHARED(defer_queue.head, head); | |
283 | /* | |
284 | * Wake-up any waiting defer thread. | |
285 | */ | |
286 | wake_up_defer(); | |
287 | } | |
288 | ||
786ee85b MD |
289 | void *thr_defer(void *args) |
290 | { | |
291 | for (;;) { | |
4ce9e4f2 | 292 | pthread_testcancel(); |
4ce9e4f2 MD |
293 | /* |
294 | * "Be green". Don't wake up the CPU if there is no RCU work | |
295 | * to perform whatsoever. Aims at saving laptop battery life by | |
296 | * leaving the processor in sleep state when idle. | |
297 | */ | |
4ce9e4f2 | 298 | wait_defer(); |
4ce9e4f2 | 299 | /* Sleeping after wait_defer to let many callbacks enqueue */ |
71df5ef4 | 300 | poll(NULL,0,100); /* wait for 100ms */ |
786ee85b MD |
301 | rcu_defer_barrier(); |
302 | } | |
303 | ||
304 | return NULL; | |
305 | } | |
306 | ||
307 | /* | |
308 | * library wrappers to be used by non-LGPL compatible source code. | |
309 | */ | |
310 | ||
804b4375 | 311 | void rcu_defer_queue(void (*fct)(void *p), void *p) |
786ee85b | 312 | { |
804b4375 | 313 | _rcu_defer_queue(fct, p); |
786ee85b MD |
314 | } |
315 | ||
316 | static void rcu_add_deferer(pthread_t id) | |
317 | { | |
318 | struct deferer_registry *oldarray; | |
319 | ||
320 | if (!registry) { | |
321 | alloc_deferers = INIT_NUM_THREADS; | |
322 | num_deferers = 0; | |
323 | registry = | |
324 | malloc(sizeof(struct deferer_registry) * alloc_deferers); | |
325 | } | |
326 | if (alloc_deferers < num_deferers + 1) { | |
327 | oldarray = registry; | |
328 | registry = malloc(sizeof(struct deferer_registry) | |
329 | * (alloc_deferers << 1)); | |
330 | memcpy(registry, oldarray, | |
331 | sizeof(struct deferer_registry) * alloc_deferers); | |
332 | alloc_deferers <<= 1; | |
333 | free(oldarray); | |
334 | } | |
335 | registry[num_deferers].tid = id; | |
336 | /* reference to the TLS of _this_ deferer thread. */ | |
337 | registry[num_deferers].defer_queue = &defer_queue; | |
804b4375 | 338 | registry[num_deferers].last_head = 0; |
786ee85b MD |
339 | num_deferers++; |
340 | } | |
341 | ||
342 | /* | |
343 | * Never shrink (implementation limitation). | |
344 | * This is O(nb threads). Eventually use a hash table. | |
345 | */ | |
346 | static void rcu_remove_deferer(pthread_t id) | |
347 | { | |
348 | struct deferer_registry *index; | |
349 | ||
350 | assert(registry != NULL); | |
351 | for (index = registry; index < registry + num_deferers; index++) { | |
352 | if (pthread_equal(index->tid, id)) { | |
353 | memcpy(index, ®istry[num_deferers - 1], | |
354 | sizeof(struct deferer_registry)); | |
355 | registry[num_deferers - 1].tid = 0; | |
356 | registry[num_deferers - 1].defer_queue = NULL; | |
804b4375 | 357 | registry[num_deferers - 1].last_head = 0; |
786ee85b MD |
358 | num_deferers--; |
359 | return; | |
360 | } | |
361 | } | |
362 | /* Hrm not found, forgot to register ? */ | |
363 | assert(0); | |
364 | } | |
365 | ||
366 | static void start_defer_thread(void) | |
367 | { | |
368 | int ret; | |
369 | ||
370 | ret = pthread_create(&tid_defer, NULL, thr_defer, | |
371 | NULL); | |
372 | assert(!ret); | |
373 | } | |
374 | ||
375 | static void stop_defer_thread(void) | |
376 | { | |
377 | int ret; | |
378 | void *tret; | |
379 | ||
4ce9e4f2 MD |
380 | pthread_cancel(tid_defer); |
381 | wake_up_defer(); | |
786ee85b MD |
382 | ret = pthread_join(tid_defer, &tret); |
383 | assert(!ret); | |
384 | } | |
385 | ||
386 | void rcu_defer_register_thread(void) | |
387 | { | |
388 | int deferers; | |
389 | ||
390 | internal_urcu_lock(&defer_thread_mutex); | |
391 | internal_urcu_lock(&urcu_defer_mutex); | |
392 | defer_queue.q = malloc(sizeof(void *) * DEFER_QUEUE_SIZE); | |
393 | rcu_add_deferer(pthread_self()); | |
394 | deferers = num_deferers; | |
395 | internal_urcu_unlock(&urcu_defer_mutex); | |
396 | ||
397 | if (deferers == 1) | |
398 | start_defer_thread(); | |
399 | internal_urcu_unlock(&defer_thread_mutex); | |
400 | } | |
401 | ||
402 | void rcu_defer_unregister_thread(void) | |
403 | { | |
404 | int deferers; | |
405 | ||
406 | internal_urcu_lock(&defer_thread_mutex); | |
407 | internal_urcu_lock(&urcu_defer_mutex); | |
408 | rcu_remove_deferer(pthread_self()); | |
409 | _rcu_defer_barrier_thread(); | |
410 | free(defer_queue.q); | |
411 | defer_queue.q = NULL; | |
412 | deferers = num_deferers; | |
413 | internal_urcu_unlock(&urcu_defer_mutex); | |
414 | ||
415 | if (deferers == 0) | |
416 | stop_defer_thread(); | |
417 | internal_urcu_unlock(&defer_thread_mutex); | |
418 | } | |
419 | ||
420 | void urcu_defer_exit(void) | |
421 | { | |
422 | free(registry); | |
423 | } |