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 | 73 | |
786ee85b MD |
74 | static void internal_urcu_lock(pthread_mutex_t *mutex) |
75 | { | |
76 | int ret; | |
77 | ||
78 | #ifndef DISTRUST_SIGNALS_EXTREME | |
79 | ret = pthread_mutex_lock(mutex); | |
80 | if (ret) { | |
81 | perror("Error in pthread mutex lock"); | |
82 | exit(-1); | |
83 | } | |
84 | #else /* #ifndef DISTRUST_SIGNALS_EXTREME */ | |
85 | while ((ret = pthread_mutex_trylock(mutex)) != 0) { | |
86 | if (ret != EBUSY && ret != EINTR) { | |
87 | printf("ret = %d, errno = %d\n", ret, errno); | |
88 | perror("Error in pthread mutex lock"); | |
89 | exit(-1); | |
90 | } | |
4ce9e4f2 | 91 | pthread_testcancel(); |
786ee85b MD |
92 | poll(NULL,0,10); |
93 | } | |
94 | #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */ | |
95 | } | |
96 | ||
97 | static void internal_urcu_unlock(pthread_mutex_t *mutex) | |
98 | { | |
99 | int ret; | |
100 | ||
101 | ret = pthread_mutex_unlock(mutex); | |
102 | if (ret) { | |
103 | perror("Error in pthread mutex unlock"); | |
104 | exit(-1); | |
105 | } | |
106 | } | |
107 | ||
04eb9c4f MD |
108 | /* |
109 | * Wake-up any waiting defer thread. Called from many concurrent threads. | |
110 | */ | |
111 | static void wake_up_defer(void) | |
112 | { | |
113 | if (unlikely(atomic_read(&defer_thread_futex) == -1)) { | |
114 | atomic_set(&defer_thread_futex, 0); | |
115 | futex(&defer_thread_futex, FUTEX_WAKE, 0, | |
116 | NULL, NULL, 0); | |
117 | } | |
118 | } | |
119 | ||
120 | static unsigned long rcu_defer_num_callbacks(void) | |
121 | { | |
122 | unsigned long num_items = 0, head; | |
123 | struct deferer_registry *index; | |
124 | ||
125 | internal_urcu_lock(&urcu_defer_mutex); | |
126 | for (index = registry; index < registry + num_deferers; index++) { | |
127 | head = LOAD_SHARED(index->defer_queue->head); | |
128 | num_items += head - index->defer_queue->tail; | |
129 | } | |
130 | internal_urcu_unlock(&urcu_defer_mutex); | |
131 | return num_items; | |
132 | } | |
133 | ||
134 | /* | |
135 | * Defer thread waiting. Single thread. | |
136 | */ | |
137 | static void wait_defer(void) | |
138 | { | |
139 | atomic_dec(&defer_thread_futex); | |
140 | smp_mb(); /* Write futex before read queue */ | |
141 | if (rcu_defer_num_callbacks()) { | |
142 | smp_mb(); /* Read queue before write futex */ | |
143 | /* Callbacks are queued, don't wait. */ | |
144 | atomic_set(&defer_thread_futex, 0); | |
145 | } else { | |
146 | smp_rmb(); /* Read queue before read futex */ | |
147 | if (atomic_read(&defer_thread_futex) == -1) | |
148 | futex(&defer_thread_futex, FUTEX_WAIT, -1, | |
149 | NULL, NULL, 0); | |
150 | } | |
151 | } | |
152 | ||
786ee85b MD |
153 | /* |
154 | * Must be called after Q.S. is reached. | |
155 | */ | |
156 | static void rcu_defer_barrier_queue(struct defer_queue *queue, | |
804b4375 | 157 | unsigned long head) |
786ee85b MD |
158 | { |
159 | unsigned long i; | |
804b4375 MD |
160 | void (*fct)(void *p); |
161 | void *p; | |
786ee85b MD |
162 | |
163 | /* | |
164 | * Tail is only modified when lock is held. | |
165 | * Head is only modified by owner thread. | |
166 | */ | |
167 | ||
804b4375 | 168 | for (i = queue->tail; i != head;) { |
786ee85b | 169 | smp_rmb(); /* read head before q[]. */ |
804b4375 MD |
170 | p = LOAD_SHARED(queue->q[i++ & DEFER_QUEUE_MASK]); |
171 | if (unlikely(DQ_IS_FCT_BIT(p))) { | |
804b4375 MD |
172 | DQ_CLEAR_FCT_BIT(p); |
173 | queue->last_fct_out = p; | |
174 | p = LOAD_SHARED(queue->q[i++ & DEFER_QUEUE_MASK]); | |
175 | } else if (unlikely(p == DQ_FCT_MARK)) { | |
804b4375 MD |
176 | p = LOAD_SHARED(queue->q[i++ & DEFER_QUEUE_MASK]); |
177 | queue->last_fct_out = p; | |
178 | p = LOAD_SHARED(queue->q[i++ & DEFER_QUEUE_MASK]); | |
29cdb8d8 | 179 | } |
804b4375 | 180 | fct = queue->last_fct_out; |
804b4375 | 181 | fct(p); |
786ee85b MD |
182 | } |
183 | smp_mb(); /* push tail after having used q[] */ | |
184 | STORE_SHARED(queue->tail, i); | |
185 | } | |
186 | ||
187 | static void _rcu_defer_barrier_thread(void) | |
188 | { | |
0d0e6c21 | 189 | unsigned long head, num_items; |
786ee85b MD |
190 | |
191 | head = defer_queue.head; | |
0d0e6c21 MD |
192 | num_items = head - defer_queue.tail; |
193 | if (unlikely(!num_items)) | |
194 | return; | |
786ee85b MD |
195 | synchronize_rcu(); |
196 | rcu_defer_barrier_queue(&defer_queue, head); | |
197 | } | |
198 | ||
199 | ||
200 | void rcu_defer_barrier_thread(void) | |
201 | { | |
202 | internal_urcu_lock(&urcu_defer_mutex); | |
203 | _rcu_defer_barrier_thread(); | |
204 | internal_urcu_unlock(&urcu_defer_mutex); | |
205 | } | |
206 | ||
0d0e6c21 MD |
207 | /* |
208 | * rcu_defer_barrier - Execute all queued rcu callbacks. | |
209 | * | |
210 | * Execute all RCU callbacks queued before rcu_defer_barrier() execution. | |
211 | * All callbacks queued on the local thread prior to a rcu_defer_barrier() call | |
212 | * are guaranteed to be executed. | |
213 | * Callbacks queued by other threads concurrently with rcu_defer_barrier() | |
214 | * execution are not guaranteed to be executed in the current batch (could | |
215 | * be left for the next batch). These callbacks queued by other threads are only | |
216 | * guaranteed to be executed if there is explicit synchronization between | |
217 | * the thread adding to the queue and the thread issuing the defer_barrier call. | |
218 | */ | |
219 | ||
786ee85b MD |
220 | void rcu_defer_barrier(void) |
221 | { | |
222 | struct deferer_registry *index; | |
0d0e6c21 | 223 | unsigned long num_items = 0; |
786ee85b MD |
224 | |
225 | if (!registry) | |
226 | return; | |
227 | ||
228 | internal_urcu_lock(&urcu_defer_mutex); | |
0d0e6c21 | 229 | for (index = registry; index < registry + num_deferers; index++) { |
786ee85b | 230 | index->last_head = LOAD_SHARED(index->defer_queue->head); |
0d0e6c21 MD |
231 | num_items += index->last_head - index->defer_queue->tail; |
232 | } | |
233 | if (likely(!num_items)) { | |
234 | /* | |
235 | * We skip the grace period because there are no queued | |
236 | * callbacks to execute. | |
237 | */ | |
238 | goto end; | |
239 | } | |
786ee85b MD |
240 | synchronize_rcu(); |
241 | for (index = registry; index < registry + num_deferers; index++) | |
242 | rcu_defer_barrier_queue(index->defer_queue, | |
243 | index->last_head); | |
0d0e6c21 | 244 | end: |
786ee85b MD |
245 | internal_urcu_unlock(&urcu_defer_mutex); |
246 | } | |
247 | ||
2c22932b MD |
248 | /* |
249 | * _rcu_defer_queue - Queue a RCU callback. | |
250 | */ | |
251 | void _rcu_defer_queue(void (*fct)(void *p), void *p) | |
252 | { | |
253 | unsigned long head, tail; | |
254 | ||
255 | /* | |
256 | * Head is only modified by ourself. Tail can be modified by reclamation | |
257 | * thread. | |
258 | */ | |
259 | head = defer_queue.head; | |
260 | tail = LOAD_SHARED(defer_queue.tail); | |
261 | ||
262 | /* | |
263 | * If queue is full, empty it ourself. | |
264 | * Worse-case: must allow 2 supplementary entries for fct pointer. | |
265 | */ | |
266 | if (unlikely(head - tail >= DEFER_QUEUE_SIZE - 2)) { | |
267 | assert(head - tail <= DEFER_QUEUE_SIZE); | |
268 | rcu_defer_barrier_thread(); | |
269 | assert(head - LOAD_SHARED(defer_queue.tail) == 0); | |
270 | } | |
271 | ||
272 | if (unlikely(defer_queue.last_fct_in != fct)) { | |
273 | defer_queue.last_fct_in = fct; | |
274 | if (unlikely(DQ_IS_FCT_BIT(fct) || fct == DQ_FCT_MARK)) { | |
275 | /* | |
276 | * If the function to encode is not aligned or the | |
277 | * marker, write DQ_FCT_MARK followed by the function | |
278 | * pointer. | |
279 | */ | |
280 | _STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK], | |
281 | DQ_FCT_MARK); | |
282 | _STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK], | |
283 | fct); | |
284 | } else { | |
285 | DQ_SET_FCT_BIT(fct); | |
286 | _STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK], | |
287 | fct); | |
288 | } | |
289 | } else { | |
290 | if (unlikely(DQ_IS_FCT_BIT(p) || p == DQ_FCT_MARK)) { | |
291 | /* | |
292 | * If the data to encode is not aligned or the marker, | |
293 | * write DQ_FCT_MARK followed by the function pointer. | |
294 | */ | |
295 | _STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK], | |
296 | DQ_FCT_MARK); | |
297 | _STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK], | |
298 | fct); | |
299 | } | |
300 | } | |
301 | _STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK], p); | |
302 | smp_wmb(); /* Publish new pointer before head */ | |
303 | /* Write q[] before head. */ | |
304 | STORE_SHARED(defer_queue.head, head); | |
04eb9c4f | 305 | smp_mb(); /* Write queue head before read futex */ |
2c22932b MD |
306 | /* |
307 | * Wake-up any waiting defer thread. | |
308 | */ | |
309 | wake_up_defer(); | |
310 | } | |
311 | ||
786ee85b MD |
312 | void *thr_defer(void *args) |
313 | { | |
314 | for (;;) { | |
4ce9e4f2 | 315 | pthread_testcancel(); |
4ce9e4f2 MD |
316 | /* |
317 | * "Be green". Don't wake up the CPU if there is no RCU work | |
318 | * to perform whatsoever. Aims at saving laptop battery life by | |
319 | * leaving the processor in sleep state when idle. | |
320 | */ | |
4ce9e4f2 | 321 | wait_defer(); |
4ce9e4f2 | 322 | /* Sleeping after wait_defer to let many callbacks enqueue */ |
71df5ef4 | 323 | poll(NULL,0,100); /* wait for 100ms */ |
786ee85b MD |
324 | rcu_defer_barrier(); |
325 | } | |
326 | ||
327 | return NULL; | |
328 | } | |
329 | ||
330 | /* | |
331 | * library wrappers to be used by non-LGPL compatible source code. | |
332 | */ | |
333 | ||
804b4375 | 334 | void rcu_defer_queue(void (*fct)(void *p), void *p) |
786ee85b | 335 | { |
804b4375 | 336 | _rcu_defer_queue(fct, p); |
786ee85b MD |
337 | } |
338 | ||
339 | static void rcu_add_deferer(pthread_t id) | |
340 | { | |
341 | struct deferer_registry *oldarray; | |
342 | ||
343 | if (!registry) { | |
344 | alloc_deferers = INIT_NUM_THREADS; | |
345 | num_deferers = 0; | |
346 | registry = | |
347 | malloc(sizeof(struct deferer_registry) * alloc_deferers); | |
348 | } | |
349 | if (alloc_deferers < num_deferers + 1) { | |
350 | oldarray = registry; | |
351 | registry = malloc(sizeof(struct deferer_registry) | |
352 | * (alloc_deferers << 1)); | |
353 | memcpy(registry, oldarray, | |
354 | sizeof(struct deferer_registry) * alloc_deferers); | |
355 | alloc_deferers <<= 1; | |
356 | free(oldarray); | |
357 | } | |
358 | registry[num_deferers].tid = id; | |
359 | /* reference to the TLS of _this_ deferer thread. */ | |
360 | registry[num_deferers].defer_queue = &defer_queue; | |
804b4375 | 361 | registry[num_deferers].last_head = 0; |
786ee85b MD |
362 | num_deferers++; |
363 | } | |
364 | ||
365 | /* | |
366 | * Never shrink (implementation limitation). | |
367 | * This is O(nb threads). Eventually use a hash table. | |
368 | */ | |
369 | static void rcu_remove_deferer(pthread_t id) | |
370 | { | |
371 | struct deferer_registry *index; | |
372 | ||
373 | assert(registry != NULL); | |
374 | for (index = registry; index < registry + num_deferers; index++) { | |
375 | if (pthread_equal(index->tid, id)) { | |
376 | memcpy(index, ®istry[num_deferers - 1], | |
377 | sizeof(struct deferer_registry)); | |
378 | registry[num_deferers - 1].tid = 0; | |
379 | registry[num_deferers - 1].defer_queue = NULL; | |
804b4375 | 380 | registry[num_deferers - 1].last_head = 0; |
786ee85b MD |
381 | num_deferers--; |
382 | return; | |
383 | } | |
384 | } | |
385 | /* Hrm not found, forgot to register ? */ | |
386 | assert(0); | |
387 | } | |
388 | ||
389 | static void start_defer_thread(void) | |
390 | { | |
391 | int ret; | |
392 | ||
393 | ret = pthread_create(&tid_defer, NULL, thr_defer, | |
394 | NULL); | |
395 | assert(!ret); | |
396 | } | |
397 | ||
398 | static void stop_defer_thread(void) | |
399 | { | |
400 | int ret; | |
401 | void *tret; | |
402 | ||
4ce9e4f2 MD |
403 | pthread_cancel(tid_defer); |
404 | wake_up_defer(); | |
786ee85b MD |
405 | ret = pthread_join(tid_defer, &tret); |
406 | assert(!ret); | |
407 | } | |
408 | ||
409 | void rcu_defer_register_thread(void) | |
410 | { | |
411 | int deferers; | |
412 | ||
413 | internal_urcu_lock(&defer_thread_mutex); | |
414 | internal_urcu_lock(&urcu_defer_mutex); | |
415 | defer_queue.q = malloc(sizeof(void *) * DEFER_QUEUE_SIZE); | |
416 | rcu_add_deferer(pthread_self()); | |
417 | deferers = num_deferers; | |
418 | internal_urcu_unlock(&urcu_defer_mutex); | |
419 | ||
420 | if (deferers == 1) | |
421 | start_defer_thread(); | |
422 | internal_urcu_unlock(&defer_thread_mutex); | |
423 | } | |
424 | ||
425 | void rcu_defer_unregister_thread(void) | |
426 | { | |
427 | int deferers; | |
428 | ||
429 | internal_urcu_lock(&defer_thread_mutex); | |
430 | internal_urcu_lock(&urcu_defer_mutex); | |
431 | rcu_remove_deferer(pthread_self()); | |
432 | _rcu_defer_barrier_thread(); | |
433 | free(defer_queue.q); | |
434 | defer_queue.q = NULL; | |
435 | deferers = num_deferers; | |
436 | internal_urcu_unlock(&urcu_defer_mutex); | |
437 | ||
438 | if (deferers == 0) | |
439 | stop_defer_thread(); | |
440 | internal_urcu_unlock(&defer_thread_mutex); | |
441 | } | |
442 | ||
443 | void urcu_defer_exit(void) | |
444 | { | |
445 | free(registry); | |
446 | } |