Commit | Line | Data |
---|---|---|
0376e7b2 PM |
1 | #ifndef _URCU_DEFER_IMPL_H |
2 | #define _URCU_DEFER_IMPL_H | |
3 | ||
786ee85b | 4 | /* |
0376e7b2 PM |
5 | * urcu-defer-impl.h |
6 | * | |
7 | * Userspace RCU header - memory reclamation. | |
786ee85b | 8 | * |
0376e7b2 PM |
9 | * TO BE INCLUDED ONLY FROM URCU LIBRARY CODE. See urcu-defer.h for linking |
10 | * dynamically with the userspace rcu reclamation library. | |
786ee85b | 11 | * |
6982d6d7 | 12 | * Copyright (c) 2009 Mathieu Desnoyers <mathieu.desnoyers@efficios.com> |
0376e7b2 | 13 | * Copyright (c) 2009 Paul E. McKenney, IBM Corporation. |
786ee85b MD |
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 | |
0376e7b2 PM |
28 | * |
29 | * IBM's contributions to this file may be relicensed under LGPLv2 or later. | |
786ee85b MD |
30 | */ |
31 | ||
0376e7b2 | 32 | #include <stdlib.h> |
786ee85b | 33 | #include <pthread.h> |
0376e7b2 | 34 | #include <stdio.h> |
786ee85b MD |
35 | #include <signal.h> |
36 | #include <assert.h> | |
786ee85b MD |
37 | #include <string.h> |
38 | #include <errno.h> | |
39 | #include <poll.h> | |
4ce9e4f2 | 40 | #include <sys/time.h> |
4ce9e4f2 | 41 | #include <unistd.h> |
6d841bc2 | 42 | #include <stdint.h> |
786ee85b | 43 | |
41849996 | 44 | #include "urcu/futex.h" |
0376e7b2 PM |
45 | |
46 | #include <urcu/compiler.h> | |
47 | #include <urcu/arch.h> | |
a2e7bf9c | 48 | #include <urcu/uatomic.h> |
0376e7b2 PM |
49 | #include <urcu/list.h> |
50 | #include <urcu/system.h> | |
51 | ||
52 | /* | |
53 | * Number of entries in the per-thread defer queue. Must be power of 2. | |
54 | */ | |
55 | #define DEFER_QUEUE_SIZE (1 << 12) | |
56 | #define DEFER_QUEUE_MASK (DEFER_QUEUE_SIZE - 1) | |
57 | ||
58 | /* | |
59 | * Typically, data is aligned at least on the architecture size. | |
60 | * Use lowest bit to indicate that the current callback is changing. | |
61 | * Assumes that (void *)-2L is not used often. Used to encode non-aligned | |
62 | * functions and non-aligned data using extra space. | |
63 | * We encode the (void *)-2L fct as: -2L, fct, data. | |
64 | * We encode the (void *)-2L data as: -2L, fct, data. | |
65 | * Here, DQ_FCT_MARK == ~DQ_FCT_BIT. Required for the test order. | |
66 | */ | |
67 | #define DQ_FCT_BIT (1 << 0) | |
68 | #define DQ_IS_FCT_BIT(x) ((unsigned long)(x) & DQ_FCT_BIT) | |
69 | #define DQ_SET_FCT_BIT(x) \ | |
70 | (x = (void *)((unsigned long)(x) | DQ_FCT_BIT)) | |
71 | #define DQ_CLEAR_FCT_BIT(x) \ | |
72 | (x = (void *)((unsigned long)(x) & ~DQ_FCT_BIT)) | |
73 | #define DQ_FCT_MARK ((void *)(~DQ_FCT_BIT)) | |
74 | ||
75 | /* | |
76 | * This code section can only be included in LGPL 2.1 compatible source code. | |
77 | * See below for the function call wrappers which can be used in code meant to | |
78 | * be only linked with the Userspace RCU library. This comes with a small | |
79 | * performance degradation on the read-side due to the added function calls. | |
80 | * This is required to permit relinking with newer versions of the library. | |
81 | */ | |
82 | ||
83 | #ifdef DEBUG_RCU | |
84 | #define rcu_assert(args...) assert(args) | |
85 | #else | |
86 | #define rcu_assert(args...) | |
87 | #endif | |
88 | ||
89 | /* | |
90 | * defer queue. | |
91 | * Contains pointers. Encoded to save space when same callback is often used. | |
92 | * When looking up the next item: | |
93 | * - if DQ_FCT_BIT is set, set the current callback to DQ_CLEAR_FCT_BIT(ptr) | |
94 | * - next element contains pointer to data. | |
95 | * - else if item == DQ_FCT_MARK | |
96 | * - set the current callback to next element ptr | |
97 | * - following next element contains pointer to data. | |
98 | * - else current element contains data | |
99 | */ | |
100 | struct defer_queue { | |
101 | unsigned long head; /* add element at head */ | |
102 | void *last_fct_in; /* last fct pointer encoded */ | |
103 | unsigned long tail; /* next element to remove at tail */ | |
104 | void *last_fct_out; /* last fct pointer encoded */ | |
105 | void **q; | |
106 | /* registry information */ | |
107 | unsigned long last_head; | |
108 | struct cds_list_head list; /* list of thread queues */ | |
109 | }; | |
110 | ||
786ee85b MD |
111 | /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */ |
112 | #include "urcu-defer.h" | |
113 | ||
02be5561 | 114 | void __attribute__((destructor)) rcu_defer_exit(void); |
786ee85b MD |
115 | |
116 | extern void synchronize_rcu(void); | |
117 | ||
118 | /* | |
02be5561 | 119 | * rcu_defer_mutex nests inside defer_thread_mutex. |
786ee85b | 120 | */ |
02be5561 | 121 | static pthread_mutex_t rcu_defer_mutex = PTHREAD_MUTEX_INITIALIZER; |
786ee85b MD |
122 | static pthread_mutex_t defer_thread_mutex = PTHREAD_MUTEX_INITIALIZER; |
123 | ||
6d841bc2 | 124 | static int32_t defer_thread_futex; |
4ce9e4f2 | 125 | |
786ee85b MD |
126 | /* |
127 | * Written to only by each individual deferer. Read by both the deferer and | |
128 | * the reclamation tread. | |
129 | */ | |
3dce4bfa | 130 | static struct defer_queue __thread defer_queue; |
0376e7b2 | 131 | static CDS_LIST_HEAD(registry_defer); |
786ee85b | 132 | static pthread_t tid_defer; |
4ce9e4f2 | 133 | |
0376e7b2 | 134 | static void mutex_lock_defer(pthread_mutex_t *mutex) |
786ee85b MD |
135 | { |
136 | int ret; | |
137 | ||
138 | #ifndef DISTRUST_SIGNALS_EXTREME | |
139 | ret = pthread_mutex_lock(mutex); | |
140 | if (ret) { | |
141 | perror("Error in pthread mutex lock"); | |
142 | exit(-1); | |
143 | } | |
144 | #else /* #ifndef DISTRUST_SIGNALS_EXTREME */ | |
145 | while ((ret = pthread_mutex_trylock(mutex)) != 0) { | |
146 | if (ret != EBUSY && ret != EINTR) { | |
147 | printf("ret = %d, errno = %d\n", ret, errno); | |
148 | perror("Error in pthread mutex lock"); | |
149 | exit(-1); | |
150 | } | |
4ce9e4f2 | 151 | pthread_testcancel(); |
786ee85b MD |
152 | poll(NULL,0,10); |
153 | } | |
154 | #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */ | |
155 | } | |
156 | ||
04eb9c4f MD |
157 | /* |
158 | * Wake-up any waiting defer thread. Called from many concurrent threads. | |
159 | */ | |
160 | static void wake_up_defer(void) | |
161 | { | |
ec4e58a3 MD |
162 | if (unlikely(uatomic_read(&defer_thread_futex) == -1)) { |
163 | uatomic_set(&defer_thread_futex, 0); | |
0854ccff | 164 | futex_noasync(&defer_thread_futex, FUTEX_WAKE, 1, |
04eb9c4f MD |
165 | NULL, NULL, 0); |
166 | } | |
167 | } | |
168 | ||
169 | static unsigned long rcu_defer_num_callbacks(void) | |
170 | { | |
171 | unsigned long num_items = 0, head; | |
dbc6128f | 172 | struct defer_queue *index; |
04eb9c4f | 173 | |
0376e7b2 PM |
174 | mutex_lock_defer(&rcu_defer_mutex); |
175 | cds_list_for_each_entry(index, ®istry_defer, list) { | |
6cf3827c | 176 | head = CMM_LOAD_SHARED(index->head); |
dbc6128f | 177 | num_items += head - index->tail; |
04eb9c4f | 178 | } |
6abb4bd5 | 179 | mutex_unlock(&rcu_defer_mutex); |
04eb9c4f MD |
180 | return num_items; |
181 | } | |
182 | ||
183 | /* | |
184 | * Defer thread waiting. Single thread. | |
185 | */ | |
186 | static void wait_defer(void) | |
187 | { | |
ec4e58a3 | 188 | uatomic_dec(&defer_thread_futex); |
5481ddb3 | 189 | cmm_smp_mb(); /* Write futex before read queue */ |
04eb9c4f | 190 | if (rcu_defer_num_callbacks()) { |
5481ddb3 | 191 | cmm_smp_mb(); /* Read queue before write futex */ |
04eb9c4f | 192 | /* Callbacks are queued, don't wait. */ |
ec4e58a3 | 193 | uatomic_set(&defer_thread_futex, 0); |
04eb9c4f | 194 | } else { |
5481ddb3 | 195 | cmm_smp_rmb(); /* Read queue before read futex */ |
ec4e58a3 | 196 | if (uatomic_read(&defer_thread_futex) == -1) |
0854ccff | 197 | futex_noasync(&defer_thread_futex, FUTEX_WAIT, -1, |
04eb9c4f MD |
198 | NULL, NULL, 0); |
199 | } | |
200 | } | |
201 | ||
786ee85b MD |
202 | /* |
203 | * Must be called after Q.S. is reached. | |
204 | */ | |
205 | static void rcu_defer_barrier_queue(struct defer_queue *queue, | |
804b4375 | 206 | unsigned long head) |
786ee85b MD |
207 | { |
208 | unsigned long i; | |
804b4375 MD |
209 | void (*fct)(void *p); |
210 | void *p; | |
786ee85b MD |
211 | |
212 | /* | |
213 | * Tail is only modified when lock is held. | |
214 | * Head is only modified by owner thread. | |
215 | */ | |
216 | ||
804b4375 | 217 | for (i = queue->tail; i != head;) { |
5481ddb3 | 218 | cmm_smp_rmb(); /* read head before q[]. */ |
6cf3827c | 219 | p = CMM_LOAD_SHARED(queue->q[i++ & DEFER_QUEUE_MASK]); |
804b4375 | 220 | if (unlikely(DQ_IS_FCT_BIT(p))) { |
804b4375 MD |
221 | DQ_CLEAR_FCT_BIT(p); |
222 | queue->last_fct_out = p; | |
6cf3827c | 223 | p = CMM_LOAD_SHARED(queue->q[i++ & DEFER_QUEUE_MASK]); |
804b4375 | 224 | } else if (unlikely(p == DQ_FCT_MARK)) { |
6cf3827c | 225 | p = CMM_LOAD_SHARED(queue->q[i++ & DEFER_QUEUE_MASK]); |
804b4375 | 226 | queue->last_fct_out = p; |
6cf3827c | 227 | p = CMM_LOAD_SHARED(queue->q[i++ & DEFER_QUEUE_MASK]); |
29cdb8d8 | 228 | } |
804b4375 | 229 | fct = queue->last_fct_out; |
804b4375 | 230 | fct(p); |
786ee85b | 231 | } |
5481ddb3 | 232 | cmm_smp_mb(); /* push tail after having used q[] */ |
6cf3827c | 233 | CMM_STORE_SHARED(queue->tail, i); |
786ee85b MD |
234 | } |
235 | ||
236 | static void _rcu_defer_barrier_thread(void) | |
237 | { | |
0d0e6c21 | 238 | unsigned long head, num_items; |
786ee85b MD |
239 | |
240 | head = defer_queue.head; | |
0d0e6c21 MD |
241 | num_items = head - defer_queue.tail; |
242 | if (unlikely(!num_items)) | |
243 | return; | |
786ee85b MD |
244 | synchronize_rcu(); |
245 | rcu_defer_barrier_queue(&defer_queue, head); | |
246 | } | |
247 | ||
786ee85b MD |
248 | void rcu_defer_barrier_thread(void) |
249 | { | |
0376e7b2 | 250 | mutex_lock_defer(&rcu_defer_mutex); |
786ee85b | 251 | _rcu_defer_barrier_thread(); |
6abb4bd5 | 252 | mutex_unlock(&rcu_defer_mutex); |
786ee85b MD |
253 | } |
254 | ||
0d0e6c21 MD |
255 | /* |
256 | * rcu_defer_barrier - Execute all queued rcu callbacks. | |
257 | * | |
258 | * Execute all RCU callbacks queued before rcu_defer_barrier() execution. | |
259 | * All callbacks queued on the local thread prior to a rcu_defer_barrier() call | |
260 | * are guaranteed to be executed. | |
261 | * Callbacks queued by other threads concurrently with rcu_defer_barrier() | |
262 | * execution are not guaranteed to be executed in the current batch (could | |
263 | * be left for the next batch). These callbacks queued by other threads are only | |
264 | * guaranteed to be executed if there is explicit synchronization between | |
265 | * the thread adding to the queue and the thread issuing the defer_barrier call. | |
266 | */ | |
267 | ||
786ee85b MD |
268 | void rcu_defer_barrier(void) |
269 | { | |
dbc6128f | 270 | struct defer_queue *index; |
0d0e6c21 | 271 | unsigned long num_items = 0; |
786ee85b | 272 | |
0376e7b2 | 273 | if (cds_list_empty(®istry_defer)) |
786ee85b MD |
274 | return; |
275 | ||
0376e7b2 PM |
276 | mutex_lock_defer(&rcu_defer_mutex); |
277 | cds_list_for_each_entry(index, ®istry_defer, list) { | |
6cf3827c | 278 | index->last_head = CMM_LOAD_SHARED(index->head); |
dbc6128f | 279 | num_items += index->last_head - index->tail; |
0d0e6c21 MD |
280 | } |
281 | if (likely(!num_items)) { | |
282 | /* | |
283 | * We skip the grace period because there are no queued | |
284 | * callbacks to execute. | |
285 | */ | |
286 | goto end; | |
287 | } | |
786ee85b | 288 | synchronize_rcu(); |
0376e7b2 | 289 | cds_list_for_each_entry(index, ®istry_defer, list) |
dbc6128f | 290 | rcu_defer_barrier_queue(index, index->last_head); |
0d0e6c21 | 291 | end: |
6abb4bd5 | 292 | mutex_unlock(&rcu_defer_mutex); |
786ee85b MD |
293 | } |
294 | ||
2c22932b | 295 | /* |
b4f313b7 | 296 | * _defer_rcu - Queue a RCU callback. |
2c22932b | 297 | */ |
3614f13c | 298 | void _defer_rcu(void (*fct)(void *p), void *p) |
2c22932b MD |
299 | { |
300 | unsigned long head, tail; | |
301 | ||
302 | /* | |
303 | * Head is only modified by ourself. Tail can be modified by reclamation | |
304 | * thread. | |
305 | */ | |
306 | head = defer_queue.head; | |
6cf3827c | 307 | tail = CMM_LOAD_SHARED(defer_queue.tail); |
2c22932b MD |
308 | |
309 | /* | |
ec8e44cf | 310 | * If queue is full, or reached threshold. Empty queue ourself. |
2c22932b MD |
311 | * Worse-case: must allow 2 supplementary entries for fct pointer. |
312 | */ | |
62f71961 | 313 | if (unlikely(head - tail >= DEFER_QUEUE_SIZE - 2)) { |
2c22932b MD |
314 | assert(head - tail <= DEFER_QUEUE_SIZE); |
315 | rcu_defer_barrier_thread(); | |
6cf3827c | 316 | assert(head - CMM_LOAD_SHARED(defer_queue.tail) == 0); |
2c22932b MD |
317 | } |
318 | ||
319 | if (unlikely(defer_queue.last_fct_in != fct)) { | |
320 | defer_queue.last_fct_in = fct; | |
321 | if (unlikely(DQ_IS_FCT_BIT(fct) || fct == DQ_FCT_MARK)) { | |
322 | /* | |
323 | * If the function to encode is not aligned or the | |
324 | * marker, write DQ_FCT_MARK followed by the function | |
325 | * pointer. | |
326 | */ | |
6cf3827c | 327 | _CMM_STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK], |
2c22932b | 328 | DQ_FCT_MARK); |
6cf3827c | 329 | _CMM_STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK], |
2c22932b MD |
330 | fct); |
331 | } else { | |
332 | DQ_SET_FCT_BIT(fct); | |
6cf3827c | 333 | _CMM_STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK], |
2c22932b MD |
334 | fct); |
335 | } | |
336 | } else { | |
337 | if (unlikely(DQ_IS_FCT_BIT(p) || p == DQ_FCT_MARK)) { | |
338 | /* | |
339 | * If the data to encode is not aligned or the marker, | |
340 | * write DQ_FCT_MARK followed by the function pointer. | |
341 | */ | |
6cf3827c | 342 | _CMM_STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK], |
2c22932b | 343 | DQ_FCT_MARK); |
6cf3827c | 344 | _CMM_STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK], |
2c22932b MD |
345 | fct); |
346 | } | |
347 | } | |
6cf3827c | 348 | _CMM_STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK], p); |
5481ddb3 | 349 | cmm_smp_wmb(); /* Publish new pointer before head */ |
2c22932b | 350 | /* Write q[] before head. */ |
6cf3827c | 351 | CMM_STORE_SHARED(defer_queue.head, head); |
5481ddb3 | 352 | cmm_smp_mb(); /* Write queue head before read futex */ |
2c22932b MD |
353 | /* |
354 | * Wake-up any waiting defer thread. | |
355 | */ | |
356 | wake_up_defer(); | |
357 | } | |
358 | ||
786ee85b MD |
359 | void *thr_defer(void *args) |
360 | { | |
361 | for (;;) { | |
4ce9e4f2 | 362 | pthread_testcancel(); |
4ce9e4f2 MD |
363 | /* |
364 | * "Be green". Don't wake up the CPU if there is no RCU work | |
365 | * to perform whatsoever. Aims at saving laptop battery life by | |
366 | * leaving the processor in sleep state when idle. | |
367 | */ | |
4ce9e4f2 | 368 | wait_defer(); |
4ce9e4f2 | 369 | /* Sleeping after wait_defer to let many callbacks enqueue */ |
71df5ef4 | 370 | poll(NULL,0,100); /* wait for 100ms */ |
786ee85b MD |
371 | rcu_defer_barrier(); |
372 | } | |
373 | ||
374 | return NULL; | |
375 | } | |
376 | ||
377 | /* | |
378 | * library wrappers to be used by non-LGPL compatible source code. | |
379 | */ | |
380 | ||
3614f13c | 381 | void defer_rcu(void (*fct)(void *p), void *p) |
786ee85b | 382 | { |
3614f13c | 383 | _defer_rcu(fct, p); |
786ee85b MD |
384 | } |
385 | ||
786ee85b MD |
386 | static void start_defer_thread(void) |
387 | { | |
388 | int ret; | |
389 | ||
dbc6128f | 390 | ret = pthread_create(&tid_defer, NULL, thr_defer, NULL); |
786ee85b MD |
391 | assert(!ret); |
392 | } | |
393 | ||
394 | static void stop_defer_thread(void) | |
395 | { | |
396 | int ret; | |
397 | void *tret; | |
398 | ||
4ce9e4f2 MD |
399 | pthread_cancel(tid_defer); |
400 | wake_up_defer(); | |
786ee85b MD |
401 | ret = pthread_join(tid_defer, &tret); |
402 | assert(!ret); | |
403 | } | |
404 | ||
7fdbbd61 | 405 | int rcu_defer_register_thread(void) |
786ee85b | 406 | { |
dbc6128f MD |
407 | int was_empty; |
408 | ||
409 | assert(defer_queue.last_head == 0); | |
410 | assert(defer_queue.q == NULL); | |
411 | defer_queue.q = malloc(sizeof(void *) * DEFER_QUEUE_SIZE); | |
7fdbbd61 MD |
412 | if (!defer_queue.q) |
413 | return -ENOMEM; | |
786ee85b | 414 | |
0376e7b2 PM |
415 | mutex_lock_defer(&defer_thread_mutex); |
416 | mutex_lock_defer(&rcu_defer_mutex); | |
417 | was_empty = cds_list_empty(®istry_defer); | |
418 | cds_list_add(&defer_queue.list, ®istry_defer); | |
6abb4bd5 | 419 | mutex_unlock(&rcu_defer_mutex); |
786ee85b | 420 | |
dbc6128f | 421 | if (was_empty) |
786ee85b | 422 | start_defer_thread(); |
6abb4bd5 | 423 | mutex_unlock(&defer_thread_mutex); |
7fdbbd61 | 424 | return 0; |
786ee85b MD |
425 | } |
426 | ||
427 | void rcu_defer_unregister_thread(void) | |
428 | { | |
dbc6128f | 429 | int is_empty; |
786ee85b | 430 | |
0376e7b2 PM |
431 | mutex_lock_defer(&defer_thread_mutex); |
432 | mutex_lock_defer(&rcu_defer_mutex); | |
16aa9ee8 | 433 | cds_list_del(&defer_queue.list); |
786ee85b MD |
434 | _rcu_defer_barrier_thread(); |
435 | free(defer_queue.q); | |
436 | defer_queue.q = NULL; | |
0376e7b2 | 437 | is_empty = cds_list_empty(®istry_defer); |
6abb4bd5 | 438 | mutex_unlock(&rcu_defer_mutex); |
786ee85b | 439 | |
dbc6128f | 440 | if (is_empty) |
786ee85b | 441 | stop_defer_thread(); |
6abb4bd5 | 442 | mutex_unlock(&defer_thread_mutex); |
786ee85b MD |
443 | } |
444 | ||
02be5561 | 445 | void rcu_defer_exit(void) |
786ee85b | 446 | { |
0376e7b2 | 447 | assert(cds_list_empty(®istry_defer)); |
786ee85b | 448 | } |
0376e7b2 PM |
449 | |
450 | #endif /* _URCU_DEFER_IMPL_H */ |