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