Fix: workqueue: futex wait: handle spurious futex wakeups
[urcu.git] / src / urcu-defer-impl.h
... / ...
CommitLineData
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 */
99struct 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
113void __attribute__((destructor)) rcu_defer_exit(void);
114
115extern void synchronize_rcu(void);
116
117/*
118 * rcu_defer_mutex nests inside defer_thread_mutex.
119 */
120static pthread_mutex_t rcu_defer_mutex = PTHREAD_MUTEX_INITIALIZER;
121static pthread_mutex_t defer_thread_mutex = PTHREAD_MUTEX_INITIALIZER;
122
123static int32_t defer_thread_futex;
124static 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 */
130static DEFINE_URCU_TLS(struct defer_queue, defer_queue);
131static CDS_LIST_HEAD(registry_defer);
132static pthread_t tid_defer;
133
134static 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 */
154static 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
164static 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 */
181static 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 if (uatomic_read(&defer_thread_futex) != -1)
198 return;
199 while (futex_noasync(&defer_thread_futex, FUTEX_WAIT, -1,
200 NULL, NULL, 0)) {
201 switch (errno) {
202 case EWOULDBLOCK:
203 /* Value already changed. */
204 return;
205 case EINTR:
206 /* Retry if interrupted by signal. */
207 break; /* Get out of switch. */
208 default:
209 /* Unexpected error. */
210 urcu_die(errno);
211 }
212 }
213 }
214}
215
216/*
217 * Must be called after Q.S. is reached.
218 */
219static void rcu_defer_barrier_queue(struct defer_queue *queue,
220 unsigned long head)
221{
222 unsigned long i;
223 void (*fct)(void *p);
224 void *p;
225
226 /*
227 * Tail is only modified when lock is held.
228 * Head is only modified by owner thread.
229 */
230
231 for (i = queue->tail; i != head;) {
232 cmm_smp_rmb(); /* read head before q[]. */
233 p = CMM_LOAD_SHARED(queue->q[i++ & DEFER_QUEUE_MASK]);
234 if (caa_unlikely(DQ_IS_FCT_BIT(p))) {
235 DQ_CLEAR_FCT_BIT(p);
236 queue->last_fct_out = p;
237 p = CMM_LOAD_SHARED(queue->q[i++ & DEFER_QUEUE_MASK]);
238 } else if (caa_unlikely(p == DQ_FCT_MARK)) {
239 p = CMM_LOAD_SHARED(queue->q[i++ & DEFER_QUEUE_MASK]);
240 queue->last_fct_out = p;
241 p = CMM_LOAD_SHARED(queue->q[i++ & DEFER_QUEUE_MASK]);
242 }
243 fct = queue->last_fct_out;
244 fct(p);
245 }
246 cmm_smp_mb(); /* push tail after having used q[] */
247 CMM_STORE_SHARED(queue->tail, i);
248}
249
250static void _rcu_defer_barrier_thread(void)
251{
252 unsigned long head, num_items;
253
254 head = URCU_TLS(defer_queue).head;
255 num_items = head - URCU_TLS(defer_queue).tail;
256 if (caa_unlikely(!num_items))
257 return;
258 synchronize_rcu();
259 rcu_defer_barrier_queue(&URCU_TLS(defer_queue), head);
260}
261
262void rcu_defer_barrier_thread(void)
263{
264 mutex_lock_defer(&rcu_defer_mutex);
265 _rcu_defer_barrier_thread();
266 mutex_unlock(&rcu_defer_mutex);
267}
268
269/*
270 * rcu_defer_barrier - Execute all queued rcu callbacks.
271 *
272 * Execute all RCU callbacks queued before rcu_defer_barrier() execution.
273 * All callbacks queued on the local thread prior to a rcu_defer_barrier() call
274 * are guaranteed to be executed.
275 * Callbacks queued by other threads concurrently with rcu_defer_barrier()
276 * execution are not guaranteed to be executed in the current batch (could
277 * be left for the next batch). These callbacks queued by other threads are only
278 * guaranteed to be executed if there is explicit synchronization between
279 * the thread adding to the queue and the thread issuing the defer_barrier call.
280 */
281
282void rcu_defer_barrier(void)
283{
284 struct defer_queue *index;
285 unsigned long num_items = 0;
286
287 if (cds_list_empty(&registry_defer))
288 return;
289
290 mutex_lock_defer(&rcu_defer_mutex);
291 cds_list_for_each_entry(index, &registry_defer, list) {
292 index->last_head = CMM_LOAD_SHARED(index->head);
293 num_items += index->last_head - index->tail;
294 }
295 if (caa_likely(!num_items)) {
296 /*
297 * We skip the grace period because there are no queued
298 * callbacks to execute.
299 */
300 goto end;
301 }
302 synchronize_rcu();
303 cds_list_for_each_entry(index, &registry_defer, list)
304 rcu_defer_barrier_queue(index, index->last_head);
305end:
306 mutex_unlock(&rcu_defer_mutex);
307}
308
309/*
310 * _defer_rcu - Queue a RCU callback.
311 */
312static void _defer_rcu(void (*fct)(void *p), void *p)
313{
314 unsigned long head, tail;
315
316 /*
317 * Head is only modified by ourself. Tail can be modified by reclamation
318 * thread.
319 */
320 head = URCU_TLS(defer_queue).head;
321 tail = CMM_LOAD_SHARED(URCU_TLS(defer_queue).tail);
322
323 /*
324 * If queue is full, or reached threshold. Empty queue ourself.
325 * Worse-case: must allow 2 supplementary entries for fct pointer.
326 */
327 if (caa_unlikely(head - tail >= DEFER_QUEUE_SIZE - 2)) {
328 urcu_posix_assert(head - tail <= DEFER_QUEUE_SIZE);
329 rcu_defer_barrier_thread();
330 urcu_posix_assert(head - CMM_LOAD_SHARED(URCU_TLS(defer_queue).tail) == 0);
331 }
332
333 /*
334 * Encode:
335 * if the function is not changed and the data is aligned and it is
336 * not the marker:
337 * store the data
338 * otherwise if the function is aligned and its not the marker:
339 * store the function with DQ_FCT_BIT
340 * store the data
341 * otherwise:
342 * store the marker (DQ_FCT_MARK)
343 * store the function
344 * store the data
345 *
346 * Decode: see the comments before 'struct defer_queue'
347 * or the code in rcu_defer_barrier_queue().
348 */
349 if (caa_unlikely(URCU_TLS(defer_queue).last_fct_in != fct
350 || DQ_IS_FCT_BIT(p)
351 || p == DQ_FCT_MARK)) {
352 URCU_TLS(defer_queue).last_fct_in = fct;
353 if (caa_unlikely(DQ_IS_FCT_BIT(fct) || fct == DQ_FCT_MARK)) {
354 _CMM_STORE_SHARED(URCU_TLS(defer_queue).q[head++ & DEFER_QUEUE_MASK],
355 DQ_FCT_MARK);
356 _CMM_STORE_SHARED(URCU_TLS(defer_queue).q[head++ & DEFER_QUEUE_MASK],
357 fct);
358 } else {
359 DQ_SET_FCT_BIT(fct);
360 _CMM_STORE_SHARED(URCU_TLS(defer_queue).q[head++ & DEFER_QUEUE_MASK],
361 fct);
362 }
363 }
364 _CMM_STORE_SHARED(URCU_TLS(defer_queue).q[head++ & DEFER_QUEUE_MASK], p);
365 cmm_smp_wmb(); /* Publish new pointer before head */
366 /* Write q[] before head. */
367 CMM_STORE_SHARED(URCU_TLS(defer_queue).head, head);
368 cmm_smp_mb(); /* Write queue head before read futex */
369 /*
370 * Wake-up any waiting defer thread.
371 */
372 wake_up_defer();
373}
374
375static void *thr_defer(void *args __attribute__((unused)))
376{
377 for (;;) {
378 /*
379 * "Be green". Don't wake up the CPU if there is no RCU work
380 * to perform whatsoever. Aims at saving laptop battery life by
381 * leaving the processor in sleep state when idle.
382 */
383 wait_defer();
384 /* Sleeping after wait_defer to let many callbacks enqueue */
385 (void) poll(NULL,0,100); /* wait for 100ms */
386 rcu_defer_barrier();
387 }
388
389 return NULL;
390}
391
392/*
393 * library wrappers to be used by non-LGPL compatible source code.
394 */
395
396void defer_rcu(void (*fct)(void *p), void *p)
397{
398 _defer_rcu(fct, p);
399}
400
401static void start_defer_thread(void)
402{
403 int ret;
404
405 ret = pthread_create(&tid_defer, NULL, thr_defer, NULL);
406 urcu_posix_assert(!ret);
407}
408
409static void stop_defer_thread(void)
410{
411 int ret;
412 void *tret;
413
414 _CMM_STORE_SHARED(defer_thread_stop, 1);
415 /* Store defer_thread_stop before testing futex */
416 cmm_smp_mb();
417 wake_up_defer();
418
419 ret = pthread_join(tid_defer, &tret);
420 urcu_posix_assert(!ret);
421
422 CMM_STORE_SHARED(defer_thread_stop, 0);
423 /* defer thread should always exit when futex value is 0 */
424 urcu_posix_assert(uatomic_read(&defer_thread_futex) == 0);
425}
426
427int rcu_defer_register_thread(void)
428{
429 int was_empty;
430
431 urcu_posix_assert(URCU_TLS(defer_queue).last_head == 0);
432 urcu_posix_assert(URCU_TLS(defer_queue).q == NULL);
433 URCU_TLS(defer_queue).q = malloc(sizeof(void *) * DEFER_QUEUE_SIZE);
434 if (!URCU_TLS(defer_queue).q)
435 return -ENOMEM;
436
437 mutex_lock_defer(&defer_thread_mutex);
438 mutex_lock_defer(&rcu_defer_mutex);
439 was_empty = cds_list_empty(&registry_defer);
440 cds_list_add(&URCU_TLS(defer_queue).list, &registry_defer);
441 mutex_unlock(&rcu_defer_mutex);
442
443 if (was_empty)
444 start_defer_thread();
445 mutex_unlock(&defer_thread_mutex);
446 return 0;
447}
448
449void rcu_defer_unregister_thread(void)
450{
451 int is_empty;
452
453 mutex_lock_defer(&defer_thread_mutex);
454 mutex_lock_defer(&rcu_defer_mutex);
455 cds_list_del(&URCU_TLS(defer_queue).list);
456 _rcu_defer_barrier_thread();
457 free(URCU_TLS(defer_queue).q);
458 URCU_TLS(defer_queue).q = NULL;
459 is_empty = cds_list_empty(&registry_defer);
460 mutex_unlock(&rcu_defer_mutex);
461
462 if (is_empty)
463 stop_defer_thread();
464 mutex_unlock(&defer_thread_mutex);
465}
466
467void rcu_defer_exit(void)
468{
469 urcu_posix_assert(cds_list_empty(&registry_defer));
470}
471
472#endif /* _URCU_DEFER_IMPL_H */
This page took 0.027817 seconds and 4 git commands to generate.