Fix: call_rcu: teardown default call_rcu worker on application exit
[urcu.git] / src / urcu-defer-impl.h
1 #ifndef _URCU_DEFER_IMPL_H
2 #define _URCU_DEFER_IMPL_H
3
4 /*
5 * urcu-defer-impl.h
6 *
7 * Userspace RCU header - memory reclamation.
8 *
9 * TO BE INCLUDED ONLY FROM URCU LIBRARY CODE. See urcu-defer.h for linking
10 * dynamically with the userspace rcu reclamation library.
11 *
12 * Copyright (c) 2009 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
13 * Copyright (c) 2009 Paul E. McKenney, IBM Corporation.
14 *
15 * This library is free software; you can redistribute it and/or
16 * modify it under the terms of the GNU Lesser General Public
17 * License as published by the Free Software Foundation; either
18 * version 2.1 of the License, or (at your option) any later version.
19 *
20 * This library is distributed in the hope that it will be useful,
21 * but WITHOUT ANY WARRANTY; without even the implied warranty of
22 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
23 * Lesser General Public License for more details.
24 *
25 * You should have received a copy of the GNU Lesser General Public
26 * License along with this library; if not, write to the Free Software
27 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
28 *
29 * IBM's contributions to this file may be relicensed under LGPLv2 or later.
30 */
31
32 #include <stdlib.h>
33 #include <pthread.h>
34 #include <stdio.h>
35 #include <signal.h>
36 #include <string.h>
37 #include <errno.h>
38 #include <poll.h>
39 #include <sys/time.h>
40 #include <unistd.h>
41 #include <stdint.h>
42
43 #include "urcu/futex.h"
44
45 #include <urcu/assert.h>
46 #include <urcu/compiler.h>
47 #include <urcu/arch.h>
48 #include <urcu/uatomic.h>
49 #include <urcu/list.h>
50 #include <urcu/system.h>
51 #include <urcu/tls-compat.h>
52 #include "urcu-die.h"
53 #include "urcu-utils.h"
54
55 /*
56 * Number of entries in the per-thread defer queue. Must be power of 2.
57 */
58 #define DEFER_QUEUE_SIZE (1 << 12)
59 #define DEFER_QUEUE_MASK (DEFER_QUEUE_SIZE - 1)
60
61 /*
62 * Typically, data is aligned at least on the architecture size.
63 * Use lowest bit to indicate that the current callback is changing.
64 * Assumes that (void *)-2L is not used often. Used to encode non-aligned
65 * functions and non-aligned data using extra space.
66 * We encode the (void *)-2L fct as: -2L, fct, data.
67 * We encode the (void *)-2L data as either:
68 * fct | DQ_FCT_BIT, data (if fct is aligned), or
69 * -2L, fct, data (if fct is not aligned).
70 * Here, DQ_FCT_MARK == ~DQ_FCT_BIT. Required for the test order.
71 */
72 #define DQ_FCT_BIT (1 << 0)
73 #define DQ_IS_FCT_BIT(x) ((unsigned long)(x) & DQ_FCT_BIT)
74 #define DQ_SET_FCT_BIT(x) \
75 (x = (void *)((unsigned long)(x) | DQ_FCT_BIT))
76 #define DQ_CLEAR_FCT_BIT(x) \
77 (x = (void *)((unsigned long)(x) & ~DQ_FCT_BIT))
78 #define DQ_FCT_MARK ((void *)(~DQ_FCT_BIT))
79
80 /*
81 * This code section can only be included in LGPL 2.1 compatible source code.
82 * See below for the function call wrappers which can be used in code meant to
83 * be only linked with the Userspace RCU library. This comes with a small
84 * performance degradation on the read-side due to the added function calls.
85 * This is required to permit relinking with newer versions of the library.
86 */
87
88 /*
89 * defer queue.
90 * Contains pointers. Encoded to save space when same callback is often used.
91 * When looking up the next item:
92 * - if DQ_FCT_BIT is set, set the current callback to DQ_CLEAR_FCT_BIT(ptr)
93 * - next element contains pointer to data.
94 * - else if item == DQ_FCT_MARK
95 * - set the current callback to next element ptr
96 * - following next element contains pointer to data.
97 * - else current element contains data
98 */
99 struct defer_queue {
100 unsigned long head; /* add element at head */
101 void *last_fct_in; /* last fct pointer encoded */
102 unsigned long tail; /* next element to remove at tail */
103 void *last_fct_out; /* last fct pointer encoded */
104 void **q;
105 /* registry information */
106 unsigned long last_head;
107 struct cds_list_head list; /* list of thread queues */
108 };
109
110 /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */
111 #include <urcu/defer.h>
112
113 void __attribute__((destructor)) rcu_defer_exit(void);
114
115 extern void synchronize_rcu(void);
116
117 /*
118 * rcu_defer_mutex nests inside defer_thread_mutex.
119 */
120 static pthread_mutex_t rcu_defer_mutex = PTHREAD_MUTEX_INITIALIZER;
121 static pthread_mutex_t defer_thread_mutex = PTHREAD_MUTEX_INITIALIZER;
122
123 static int32_t defer_thread_futex;
124 static int32_t defer_thread_stop;
125
126 /*
127 * Written to only by each individual deferer. Read by both the deferer and
128 * the reclamation tread.
129 */
130 static DEFINE_URCU_TLS(struct defer_queue, defer_queue);
131 static CDS_LIST_HEAD(registry_defer);
132 static pthread_t tid_defer;
133
134 static void mutex_lock_defer(pthread_mutex_t *mutex)
135 {
136 int ret;
137
138 #ifndef DISTRUST_SIGNALS_EXTREME
139 ret = pthread_mutex_lock(mutex);
140 if (ret)
141 urcu_die(ret);
142 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
143 while ((ret = pthread_mutex_trylock(mutex)) != 0) {
144 if (ret != EBUSY && ret != EINTR)
145 urcu_die(ret);
146 (void) poll(NULL,0,10);
147 }
148 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
149 }
150
151 /*
152 * Wake-up any waiting defer thread. Called from many concurrent threads.
153 */
154 static void wake_up_defer(void)
155 {
156 if (caa_unlikely(uatomic_read(&defer_thread_futex) == -1)) {
157 uatomic_set(&defer_thread_futex, 0);
158 if (futex_noasync(&defer_thread_futex, FUTEX_WAKE, 1,
159 NULL, NULL, 0) < 0)
160 urcu_die(errno);
161 }
162 }
163
164 static unsigned long rcu_defer_num_callbacks(void)
165 {
166 unsigned long num_items = 0, head;
167 struct defer_queue *index;
168
169 mutex_lock_defer(&rcu_defer_mutex);
170 cds_list_for_each_entry(index, &registry_defer, list) {
171 head = CMM_LOAD_SHARED(index->head);
172 num_items += head - index->tail;
173 }
174 mutex_unlock(&rcu_defer_mutex);
175 return num_items;
176 }
177
178 /*
179 * Defer thread waiting. Single thread.
180 */
181 static void wait_defer(void)
182 {
183 uatomic_dec(&defer_thread_futex);
184 /* Write futex before read queue */
185 /* Write futex before read defer_thread_stop */
186 cmm_smp_mb();
187 if (_CMM_LOAD_SHARED(defer_thread_stop)) {
188 uatomic_set(&defer_thread_futex, 0);
189 pthread_exit(0);
190 }
191 if (rcu_defer_num_callbacks()) {
192 cmm_smp_mb(); /* Read queue before write futex */
193 /* Callbacks are queued, don't wait. */
194 uatomic_set(&defer_thread_futex, 0);
195 } else {
196 cmm_smp_rmb(); /* Read queue before read futex */
197 while (uatomic_read(&defer_thread_futex) == -1) {
198 if (!futex_noasync(&defer_thread_futex, FUTEX_WAIT, -1, NULL, NULL, 0)) {
199 /*
200 * Prior queued wakeups queued by unrelated code
201 * using the same address can cause futex wait to
202 * return 0 even through the futex value is still
203 * -1 (spurious wakeups). Check the value again
204 * in user-space to validate whether it really
205 * differs from -1.
206 */
207 continue;
208 }
209 switch (errno) {
210 case EAGAIN:
211 /* Value already changed. */
212 return;
213 case EINTR:
214 /* Retry if interrupted by signal. */
215 break; /* Get out of switch. Check again. */
216 default:
217 /* Unexpected error. */
218 urcu_die(errno);
219 }
220 }
221 }
222 }
223
224 /*
225 * Must be called after Q.S. is reached.
226 */
227 static void rcu_defer_barrier_queue(struct defer_queue *queue,
228 unsigned long head)
229 {
230 unsigned long i;
231 void (*fct)(void *p);
232 void *p;
233
234 /*
235 * Tail is only modified when lock is held.
236 * Head is only modified by owner thread.
237 */
238
239 for (i = queue->tail; i != head;) {
240 cmm_smp_rmb(); /* read head before q[]. */
241 p = CMM_LOAD_SHARED(queue->q[i++ & DEFER_QUEUE_MASK]);
242 if (caa_unlikely(DQ_IS_FCT_BIT(p))) {
243 DQ_CLEAR_FCT_BIT(p);
244 queue->last_fct_out = p;
245 p = CMM_LOAD_SHARED(queue->q[i++ & DEFER_QUEUE_MASK]);
246 } else if (caa_unlikely(p == DQ_FCT_MARK)) {
247 p = CMM_LOAD_SHARED(queue->q[i++ & DEFER_QUEUE_MASK]);
248 queue->last_fct_out = p;
249 p = CMM_LOAD_SHARED(queue->q[i++ & DEFER_QUEUE_MASK]);
250 }
251 fct = queue->last_fct_out;
252 fct(p);
253 }
254 cmm_smp_mb(); /* push tail after having used q[] */
255 CMM_STORE_SHARED(queue->tail, i);
256 }
257
258 static void _rcu_defer_barrier_thread(void)
259 {
260 unsigned long head, num_items;
261
262 head = URCU_TLS(defer_queue).head;
263 num_items = head - URCU_TLS(defer_queue).tail;
264 if (caa_unlikely(!num_items))
265 return;
266 synchronize_rcu();
267 rcu_defer_barrier_queue(&URCU_TLS(defer_queue), head);
268 }
269
270 void rcu_defer_barrier_thread(void)
271 {
272 mutex_lock_defer(&rcu_defer_mutex);
273 _rcu_defer_barrier_thread();
274 mutex_unlock(&rcu_defer_mutex);
275 }
276
277 /*
278 * rcu_defer_barrier - Execute all queued rcu callbacks.
279 *
280 * Execute all RCU callbacks queued before rcu_defer_barrier() execution.
281 * All callbacks queued on the local thread prior to a rcu_defer_barrier() call
282 * are guaranteed to be executed.
283 * Callbacks queued by other threads concurrently with rcu_defer_barrier()
284 * execution are not guaranteed to be executed in the current batch (could
285 * be left for the next batch). These callbacks queued by other threads are only
286 * guaranteed to be executed if there is explicit synchronization between
287 * the thread adding to the queue and the thread issuing the defer_barrier call.
288 */
289
290 void rcu_defer_barrier(void)
291 {
292 struct defer_queue *index;
293 unsigned long num_items = 0;
294
295 if (cds_list_empty(&registry_defer))
296 return;
297
298 mutex_lock_defer(&rcu_defer_mutex);
299 cds_list_for_each_entry(index, &registry_defer, list) {
300 index->last_head = CMM_LOAD_SHARED(index->head);
301 num_items += index->last_head - index->tail;
302 }
303 if (caa_likely(!num_items)) {
304 /*
305 * We skip the grace period because there are no queued
306 * callbacks to execute.
307 */
308 goto end;
309 }
310 synchronize_rcu();
311 cds_list_for_each_entry(index, &registry_defer, list)
312 rcu_defer_barrier_queue(index, index->last_head);
313 end:
314 mutex_unlock(&rcu_defer_mutex);
315 }
316
317 /*
318 * _defer_rcu - Queue a RCU callback.
319 */
320 static void _defer_rcu(void (*fct)(void *p), void *p)
321 {
322 unsigned long head, tail;
323
324 /*
325 * Head is only modified by ourself. Tail can be modified by reclamation
326 * thread.
327 */
328 head = URCU_TLS(defer_queue).head;
329 tail = CMM_LOAD_SHARED(URCU_TLS(defer_queue).tail);
330
331 /*
332 * If queue is full, or reached threshold. Empty queue ourself.
333 * Worse-case: must allow 2 supplementary entries for fct pointer.
334 */
335 if (caa_unlikely(head - tail >= DEFER_QUEUE_SIZE - 2)) {
336 urcu_posix_assert(head - tail <= DEFER_QUEUE_SIZE);
337 rcu_defer_barrier_thread();
338 urcu_posix_assert(head - CMM_LOAD_SHARED(URCU_TLS(defer_queue).tail) == 0);
339 }
340
341 /*
342 * Encode:
343 * if the function is not changed and the data is aligned and it is
344 * not the marker:
345 * store the data
346 * otherwise if the function is aligned and its not the marker:
347 * store the function with DQ_FCT_BIT
348 * store the data
349 * otherwise:
350 * store the marker (DQ_FCT_MARK)
351 * store the function
352 * store the data
353 *
354 * Decode: see the comments before 'struct defer_queue'
355 * or the code in rcu_defer_barrier_queue().
356 */
357 if (caa_unlikely(URCU_TLS(defer_queue).last_fct_in != fct
358 || DQ_IS_FCT_BIT(p)
359 || p == DQ_FCT_MARK)) {
360 URCU_TLS(defer_queue).last_fct_in = fct;
361 if (caa_unlikely(DQ_IS_FCT_BIT(fct) || fct == DQ_FCT_MARK)) {
362 _CMM_STORE_SHARED(URCU_TLS(defer_queue).q[head++ & DEFER_QUEUE_MASK],
363 DQ_FCT_MARK);
364 _CMM_STORE_SHARED(URCU_TLS(defer_queue).q[head++ & DEFER_QUEUE_MASK],
365 fct);
366 } else {
367 DQ_SET_FCT_BIT(fct);
368 _CMM_STORE_SHARED(URCU_TLS(defer_queue).q[head++ & DEFER_QUEUE_MASK],
369 fct);
370 }
371 }
372 _CMM_STORE_SHARED(URCU_TLS(defer_queue).q[head++ & DEFER_QUEUE_MASK], p);
373 cmm_smp_wmb(); /* Publish new pointer before head */
374 /* Write q[] before head. */
375 CMM_STORE_SHARED(URCU_TLS(defer_queue).head, head);
376 cmm_smp_mb(); /* Write queue head before read futex */
377 /*
378 * Wake-up any waiting defer thread.
379 */
380 wake_up_defer();
381 }
382
383 static void *thr_defer(void *args __attribute__((unused)))
384 {
385 for (;;) {
386 /*
387 * "Be green". Don't wake up the CPU if there is no RCU work
388 * to perform whatsoever. Aims at saving laptop battery life by
389 * leaving the processor in sleep state when idle.
390 */
391 wait_defer();
392 /* Sleeping after wait_defer to let many callbacks enqueue */
393 (void) poll(NULL,0,100); /* wait for 100ms */
394 rcu_defer_barrier();
395 }
396
397 return NULL;
398 }
399
400 /*
401 * library wrappers to be used by non-LGPL compatible source code.
402 */
403
404 void defer_rcu(void (*fct)(void *p), void *p)
405 {
406 _defer_rcu(fct, p);
407 }
408
409 static void start_defer_thread(void)
410 {
411 int ret;
412 sigset_t newmask, oldmask;
413
414 ret = sigfillset(&newmask);
415 urcu_posix_assert(!ret);
416 ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask);
417 urcu_posix_assert(!ret);
418
419 ret = pthread_create(&tid_defer, NULL, thr_defer, NULL);
420 if (ret)
421 urcu_die(ret);
422
423 ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
424 urcu_posix_assert(!ret);
425 }
426
427 static void stop_defer_thread(void)
428 {
429 int ret;
430 void *tret;
431
432 _CMM_STORE_SHARED(defer_thread_stop, 1);
433 /* Store defer_thread_stop before testing futex */
434 cmm_smp_mb();
435 wake_up_defer();
436
437 ret = pthread_join(tid_defer, &tret);
438 urcu_posix_assert(!ret);
439
440 CMM_STORE_SHARED(defer_thread_stop, 0);
441 /* defer thread should always exit when futex value is 0 */
442 urcu_posix_assert(uatomic_read(&defer_thread_futex) == 0);
443 }
444
445 int rcu_defer_register_thread(void)
446 {
447 int was_empty;
448
449 urcu_posix_assert(URCU_TLS(defer_queue).last_head == 0);
450 urcu_posix_assert(URCU_TLS(defer_queue).q == NULL);
451 URCU_TLS(defer_queue).q = malloc(sizeof(void *) * DEFER_QUEUE_SIZE);
452 if (!URCU_TLS(defer_queue).q)
453 return -ENOMEM;
454
455 mutex_lock_defer(&defer_thread_mutex);
456 mutex_lock_defer(&rcu_defer_mutex);
457 was_empty = cds_list_empty(&registry_defer);
458 cds_list_add(&URCU_TLS(defer_queue).list, &registry_defer);
459 mutex_unlock(&rcu_defer_mutex);
460
461 if (was_empty)
462 start_defer_thread();
463 mutex_unlock(&defer_thread_mutex);
464 return 0;
465 }
466
467 void rcu_defer_unregister_thread(void)
468 {
469 int is_empty;
470
471 mutex_lock_defer(&defer_thread_mutex);
472 mutex_lock_defer(&rcu_defer_mutex);
473 cds_list_del(&URCU_TLS(defer_queue).list);
474 _rcu_defer_barrier_thread();
475 free(URCU_TLS(defer_queue).q);
476 URCU_TLS(defer_queue).q = NULL;
477 is_empty = cds_list_empty(&registry_defer);
478 mutex_unlock(&rcu_defer_mutex);
479
480 if (is_empty)
481 stop_defer_thread();
482 mutex_unlock(&defer_thread_mutex);
483 }
484
485 void rcu_defer_exit(void)
486 {
487 urcu_posix_assert(cds_list_empty(&registry_defer));
488 }
489
490 #endif /* _URCU_DEFER_IMPL_H */
This page took 0.048117 seconds and 4 git commands to generate.