1 #ifndef _URCU_DEFER_IMPL_H
2 #define _URCU_DEFER_IMPL_H
7 * Userspace RCU header - memory reclamation.
9 * TO BE INCLUDED ONLY FROM URCU LIBRARY CODE. See urcu-defer.h for linking
10 * dynamically with the userspace rcu reclamation library.
12 * Copyright (c) 2009 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
13 * Copyright (c) 2009 Paul E. McKenney, IBM Corporation.
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.
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.
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
29 * IBM's contributions to this file may be relicensed under LGPLv2 or later.
44 #include "urcu/futex.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>
55 * Number of entries in the per-thread defer queue. Must be power of 2.
57 #define DEFER_QUEUE_SIZE (1 << 12)
58 #define DEFER_QUEUE_MASK (DEFER_QUEUE_SIZE - 1)
61 * Typically, data is aligned at least on the architecture size.
62 * Use lowest bit to indicate that the current callback is changing.
63 * Assumes that (void *)-2L is not used often. Used to encode non-aligned
64 * functions and non-aligned data using extra space.
65 * We encode the (void *)-2L fct as: -2L, fct, data.
66 * We encode the (void *)-2L data as either:
67 * fct | DQ_FCT_BIT, data (if fct is aligned), or
68 * -2L, fct, data (if fct is not aligned).
69 * Here, DQ_FCT_MARK == ~DQ_FCT_BIT. Required for the test order.
71 #define DQ_FCT_BIT (1 << 0)
72 #define DQ_IS_FCT_BIT(x) ((unsigned long)(x) & DQ_FCT_BIT)
73 #define DQ_SET_FCT_BIT(x) \
74 (x = (void *)((unsigned long)(x) | DQ_FCT_BIT))
75 #define DQ_CLEAR_FCT_BIT(x) \
76 (x = (void *)((unsigned long)(x) & ~DQ_FCT_BIT))
77 #define DQ_FCT_MARK ((void *)(~DQ_FCT_BIT))
80 * This code section can only be included in LGPL 2.1 compatible source code.
81 * See below for the function call wrappers which can be used in code meant to
82 * be only linked with the Userspace RCU library. This comes with a small
83 * performance degradation on the read-side due to the added function calls.
84 * This is required to permit relinking with newer versions of the library.
88 #define rcu_assert(args...) assert(args)
90 #define rcu_assert(args...)
95 * Contains pointers. Encoded to save space when same callback is often used.
96 * When looking up the next item:
97 * - if DQ_FCT_BIT is set, set the current callback to DQ_CLEAR_FCT_BIT(ptr)
98 * - next element contains pointer to data.
99 * - else if item == DQ_FCT_MARK
100 * - set the current callback to next element ptr
101 * - following next element contains pointer to data.
102 * - else current element contains data
105 unsigned long head
; /* add element at head */
106 void *last_fct_in
; /* last fct pointer encoded */
107 unsigned long tail
; /* next element to remove at tail */
108 void *last_fct_out
; /* last fct pointer encoded */
110 /* registry information */
111 unsigned long last_head
;
112 struct cds_list_head list
; /* list of thread queues */
115 /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */
116 #include "urcu-defer.h"
118 void __attribute__((destructor
)) rcu_defer_exit(void);
120 extern void synchronize_rcu(void);
123 * rcu_defer_mutex nests inside defer_thread_mutex.
125 static pthread_mutex_t rcu_defer_mutex
= PTHREAD_MUTEX_INITIALIZER
;
126 static pthread_mutex_t defer_thread_mutex
= PTHREAD_MUTEX_INITIALIZER
;
128 static int32_t defer_thread_futex
;
129 static int32_t defer_thread_stop
;
132 * Written to only by each individual deferer. Read by both the deferer and
133 * the reclamation tread.
135 static DEFINE_URCU_TLS(struct defer_queue
, defer_queue
);
136 static CDS_LIST_HEAD(registry_defer
);
137 static pthread_t tid_defer
;
139 static void mutex_lock_defer(pthread_mutex_t
*mutex
)
143 #ifndef DISTRUST_SIGNALS_EXTREME
144 ret
= pthread_mutex_lock(mutex
);
147 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
148 while ((ret
= pthread_mutex_trylock(mutex
)) != 0) {
149 if (ret
!= EBUSY
&& ret
!= EINTR
)
153 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
157 * Wake-up any waiting defer thread. Called from many concurrent threads.
159 static void wake_up_defer(void)
161 if (caa_unlikely(uatomic_read(&defer_thread_futex
) == -1)) {
162 uatomic_set(&defer_thread_futex
, 0);
163 futex_noasync(&defer_thread_futex
, FUTEX_WAKE
, 1,
168 static unsigned long rcu_defer_num_callbacks(void)
170 unsigned long num_items
= 0, head
;
171 struct defer_queue
*index
;
173 mutex_lock_defer(&rcu_defer_mutex
);
174 cds_list_for_each_entry(index
, ®istry_defer
, list
) {
175 head
= CMM_LOAD_SHARED(index
->head
);
176 num_items
+= head
- index
->tail
;
178 mutex_unlock(&rcu_defer_mutex
);
183 * Defer thread waiting. Single thread.
185 static void wait_defer(void)
187 uatomic_dec(&defer_thread_futex
);
188 /* Write futex before read queue */
189 /* Write futex before read defer_thread_stop */
191 if (_CMM_LOAD_SHARED(defer_thread_stop
)) {
192 uatomic_set(&defer_thread_futex
, 0);
195 if (rcu_defer_num_callbacks()) {
196 cmm_smp_mb(); /* Read queue before write futex */
197 /* Callbacks are queued, don't wait. */
198 uatomic_set(&defer_thread_futex
, 0);
200 cmm_smp_rmb(); /* Read queue before read futex */
201 if (uatomic_read(&defer_thread_futex
) == -1)
202 futex_noasync(&defer_thread_futex
, FUTEX_WAIT
, -1,
208 * Must be called after Q.S. is reached.
210 static void rcu_defer_barrier_queue(struct defer_queue
*queue
,
214 void (*fct
)(void *p
);
218 * Tail is only modified when lock is held.
219 * Head is only modified by owner thread.
222 for (i
= queue
->tail
; i
!= head
;) {
223 cmm_smp_rmb(); /* read head before q[]. */
224 p
= CMM_LOAD_SHARED(queue
->q
[i
++ & DEFER_QUEUE_MASK
]);
225 if (caa_unlikely(DQ_IS_FCT_BIT(p
))) {
227 queue
->last_fct_out
= p
;
228 p
= CMM_LOAD_SHARED(queue
->q
[i
++ & DEFER_QUEUE_MASK
]);
229 } else if (caa_unlikely(p
== DQ_FCT_MARK
)) {
230 p
= CMM_LOAD_SHARED(queue
->q
[i
++ & DEFER_QUEUE_MASK
]);
231 queue
->last_fct_out
= p
;
232 p
= CMM_LOAD_SHARED(queue
->q
[i
++ & DEFER_QUEUE_MASK
]);
234 fct
= queue
->last_fct_out
;
237 cmm_smp_mb(); /* push tail after having used q[] */
238 CMM_STORE_SHARED(queue
->tail
, i
);
241 static void _rcu_defer_barrier_thread(void)
243 unsigned long head
, num_items
;
245 head
= URCU_TLS(defer_queue
).head
;
246 num_items
= head
- URCU_TLS(defer_queue
).tail
;
247 if (caa_unlikely(!num_items
))
250 rcu_defer_barrier_queue(&URCU_TLS(defer_queue
), head
);
253 void rcu_defer_barrier_thread(void)
255 mutex_lock_defer(&rcu_defer_mutex
);
256 _rcu_defer_barrier_thread();
257 mutex_unlock(&rcu_defer_mutex
);
261 * rcu_defer_barrier - Execute all queued rcu callbacks.
263 * Execute all RCU callbacks queued before rcu_defer_barrier() execution.
264 * All callbacks queued on the local thread prior to a rcu_defer_barrier() call
265 * are guaranteed to be executed.
266 * Callbacks queued by other threads concurrently with rcu_defer_barrier()
267 * execution are not guaranteed to be executed in the current batch (could
268 * be left for the next batch). These callbacks queued by other threads are only
269 * guaranteed to be executed if there is explicit synchronization between
270 * the thread adding to the queue and the thread issuing the defer_barrier call.
273 void rcu_defer_barrier(void)
275 struct defer_queue
*index
;
276 unsigned long num_items
= 0;
278 if (cds_list_empty(®istry_defer
))
281 mutex_lock_defer(&rcu_defer_mutex
);
282 cds_list_for_each_entry(index
, ®istry_defer
, list
) {
283 index
->last_head
= CMM_LOAD_SHARED(index
->head
);
284 num_items
+= index
->last_head
- index
->tail
;
286 if (caa_likely(!num_items
)) {
288 * We skip the grace period because there are no queued
289 * callbacks to execute.
294 cds_list_for_each_entry(index
, ®istry_defer
, list
)
295 rcu_defer_barrier_queue(index
, index
->last_head
);
297 mutex_unlock(&rcu_defer_mutex
);
301 * _defer_rcu - Queue a RCU callback.
303 static void _defer_rcu(void (*fct
)(void *p
), void *p
)
305 unsigned long head
, tail
;
308 * Head is only modified by ourself. Tail can be modified by reclamation
311 head
= URCU_TLS(defer_queue
).head
;
312 tail
= CMM_LOAD_SHARED(URCU_TLS(defer_queue
).tail
);
315 * If queue is full, or reached threshold. Empty queue ourself.
316 * Worse-case: must allow 2 supplementary entries for fct pointer.
318 if (caa_unlikely(head
- tail
>= DEFER_QUEUE_SIZE
- 2)) {
319 assert(head
- tail
<= DEFER_QUEUE_SIZE
);
320 rcu_defer_barrier_thread();
321 assert(head
- CMM_LOAD_SHARED(URCU_TLS(defer_queue
).tail
) == 0);
326 * if the function is not changed and the data is aligned and it is
329 * otherwise if the function is aligned and its not the marker:
330 * store the function with DQ_FCT_BIT
333 * store the marker (DQ_FCT_MARK)
337 * Decode: see the comments before 'struct defer_queue'
338 * or the code in rcu_defer_barrier_queue().
340 if (caa_unlikely(URCU_TLS(defer_queue
).last_fct_in
!= fct
342 || p
== DQ_FCT_MARK
)) {
343 URCU_TLS(defer_queue
).last_fct_in
= fct
;
344 if (caa_unlikely(DQ_IS_FCT_BIT(fct
) || fct
== DQ_FCT_MARK
)) {
345 _CMM_STORE_SHARED(URCU_TLS(defer_queue
).q
[head
++ & DEFER_QUEUE_MASK
],
347 _CMM_STORE_SHARED(URCU_TLS(defer_queue
).q
[head
++ & DEFER_QUEUE_MASK
],
351 _CMM_STORE_SHARED(URCU_TLS(defer_queue
).q
[head
++ & DEFER_QUEUE_MASK
],
355 _CMM_STORE_SHARED(URCU_TLS(defer_queue
).q
[head
++ & DEFER_QUEUE_MASK
], p
);
356 cmm_smp_wmb(); /* Publish new pointer before head */
357 /* Write q[] before head. */
358 CMM_STORE_SHARED(URCU_TLS(defer_queue
).head
, head
);
359 cmm_smp_mb(); /* Write queue head before read futex */
361 * Wake-up any waiting defer thread.
366 static void *thr_defer(void *args
)
370 * "Be green". Don't wake up the CPU if there is no RCU work
371 * to perform whatsoever. Aims at saving laptop battery life by
372 * leaving the processor in sleep state when idle.
375 /* Sleeping after wait_defer to let many callbacks enqueue */
376 poll(NULL
,0,100); /* wait for 100ms */
384 * library wrappers to be used by non-LGPL compatible source code.
387 void defer_rcu(void (*fct
)(void *p
), void *p
)
392 static void start_defer_thread(void)
396 ret
= pthread_create(&tid_defer
, NULL
, thr_defer
, NULL
);
400 static void stop_defer_thread(void)
405 _CMM_STORE_SHARED(defer_thread_stop
, 1);
406 /* Store defer_thread_stop before testing futex */
410 ret
= pthread_join(tid_defer
, &tret
);
413 CMM_STORE_SHARED(defer_thread_stop
, 0);
414 /* defer thread should always exit when futex value is 0 */
415 assert(uatomic_read(&defer_thread_futex
) == 0);
418 int rcu_defer_register_thread(void)
422 assert(URCU_TLS(defer_queue
).last_head
== 0);
423 assert(URCU_TLS(defer_queue
).q
== NULL
);
424 URCU_TLS(defer_queue
).q
= malloc(sizeof(void *) * DEFER_QUEUE_SIZE
);
425 if (!URCU_TLS(defer_queue
).q
)
428 mutex_lock_defer(&defer_thread_mutex
);
429 mutex_lock_defer(&rcu_defer_mutex
);
430 was_empty
= cds_list_empty(®istry_defer
);
431 cds_list_add(&URCU_TLS(defer_queue
).list
, ®istry_defer
);
432 mutex_unlock(&rcu_defer_mutex
);
435 start_defer_thread();
436 mutex_unlock(&defer_thread_mutex
);
440 void rcu_defer_unregister_thread(void)
444 mutex_lock_defer(&defer_thread_mutex
);
445 mutex_lock_defer(&rcu_defer_mutex
);
446 cds_list_del(&URCU_TLS(defer_queue
).list
);
447 _rcu_defer_barrier_thread();
448 free(URCU_TLS(defer_queue
).q
);
449 URCU_TLS(defer_queue
).q
= NULL
;
450 is_empty
= cds_list_empty(®istry_defer
);
451 mutex_unlock(&rcu_defer_mutex
);
455 mutex_unlock(&defer_thread_mutex
);
458 void rcu_defer_exit(void)
460 assert(cds_list_empty(®istry_defer
));
463 #endif /* _URCU_DEFER_IMPL_H */