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. | |
2ef1bfb2 LJ |
64 | * We encode the (void *)-2L data as either: |
65 | * fct | DQ_FCT_BIT, data (if fct is aligned), or | |
66 | * -2L, fct, data (if fct is not aligned). | |
0376e7b2 PM |
67 | * Here, DQ_FCT_MARK == ~DQ_FCT_BIT. Required for the test order. |
68 | */ | |
69 | #define DQ_FCT_BIT (1 << 0) | |
70 | #define DQ_IS_FCT_BIT(x) ((unsigned long)(x) & DQ_FCT_BIT) | |
71 | #define DQ_SET_FCT_BIT(x) \ | |
72 | (x = (void *)((unsigned long)(x) | DQ_FCT_BIT)) | |
73 | #define DQ_CLEAR_FCT_BIT(x) \ | |
74 | (x = (void *)((unsigned long)(x) & ~DQ_FCT_BIT)) | |
75 | #define DQ_FCT_MARK ((void *)(~DQ_FCT_BIT)) | |
76 | ||
77 | /* | |
78 | * This code section can only be included in LGPL 2.1 compatible source code. | |
79 | * See below for the function call wrappers which can be used in code meant to | |
80 | * be only linked with the Userspace RCU library. This comes with a small | |
81 | * performance degradation on the read-side due to the added function calls. | |
82 | * This is required to permit relinking with newer versions of the library. | |
83 | */ | |
84 | ||
85 | #ifdef DEBUG_RCU | |
86 | #define rcu_assert(args...) assert(args) | |
87 | #else | |
88 | #define rcu_assert(args...) | |
89 | #endif | |
90 | ||
91 | /* | |
92 | * defer queue. | |
93 | * Contains pointers. Encoded to save space when same callback is often used. | |
94 | * When looking up the next item: | |
95 | * - if DQ_FCT_BIT is set, set the current callback to DQ_CLEAR_FCT_BIT(ptr) | |
96 | * - next element contains pointer to data. | |
97 | * - else if item == DQ_FCT_MARK | |
98 | * - set the current callback to next element ptr | |
99 | * - following next element contains pointer to data. | |
100 | * - else current element contains data | |
101 | */ | |
102 | struct defer_queue { | |
103 | unsigned long head; /* add element at head */ | |
104 | void *last_fct_in; /* last fct pointer encoded */ | |
105 | unsigned long tail; /* next element to remove at tail */ | |
106 | void *last_fct_out; /* last fct pointer encoded */ | |
107 | void **q; | |
108 | /* registry information */ | |
109 | unsigned long last_head; | |
110 | struct cds_list_head list; /* list of thread queues */ | |
111 | }; | |
112 | ||
786ee85b MD |
113 | /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */ |
114 | #include "urcu-defer.h" | |
115 | ||
02be5561 | 116 | void __attribute__((destructor)) rcu_defer_exit(void); |
786ee85b MD |
117 | |
118 | extern void synchronize_rcu(void); | |
119 | ||
120 | /* | |
02be5561 | 121 | * rcu_defer_mutex nests inside defer_thread_mutex. |
786ee85b | 122 | */ |
02be5561 | 123 | static pthread_mutex_t rcu_defer_mutex = PTHREAD_MUTEX_INITIALIZER; |
786ee85b MD |
124 | static pthread_mutex_t defer_thread_mutex = PTHREAD_MUTEX_INITIALIZER; |
125 | ||
6d841bc2 | 126 | static int32_t defer_thread_futex; |
d7ff6cee | 127 | static int32_t defer_thread_stop; |
4ce9e4f2 | 128 | |
786ee85b MD |
129 | /* |
130 | * Written to only by each individual deferer. Read by both the deferer and | |
131 | * the reclamation tread. | |
132 | */ | |
3dce4bfa | 133 | static struct defer_queue __thread defer_queue; |
0376e7b2 | 134 | static CDS_LIST_HEAD(registry_defer); |
786ee85b | 135 | static pthread_t tid_defer; |
4ce9e4f2 | 136 | |
0376e7b2 | 137 | static void mutex_lock_defer(pthread_mutex_t *mutex) |
786ee85b MD |
138 | { |
139 | int ret; | |
140 | ||
141 | #ifndef DISTRUST_SIGNALS_EXTREME | |
142 | ret = pthread_mutex_lock(mutex); | |
143 | if (ret) { | |
144 | perror("Error in pthread mutex lock"); | |
145 | exit(-1); | |
146 | } | |
147 | #else /* #ifndef DISTRUST_SIGNALS_EXTREME */ | |
148 | while ((ret = pthread_mutex_trylock(mutex)) != 0) { | |
149 | if (ret != EBUSY && ret != EINTR) { | |
150 | printf("ret = %d, errno = %d\n", ret, errno); | |
151 | perror("Error in pthread mutex lock"); | |
152 | exit(-1); | |
153 | } | |
154 | poll(NULL,0,10); | |
155 | } | |
156 | #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */ | |
157 | } | |
158 | ||
04eb9c4f MD |
159 | /* |
160 | * Wake-up any waiting defer thread. Called from many concurrent threads. | |
161 | */ | |
162 | static void wake_up_defer(void) | |
163 | { | |
a0b7f7ea | 164 | if (caa_unlikely(uatomic_read(&defer_thread_futex) == -1)) { |
ec4e58a3 | 165 | uatomic_set(&defer_thread_futex, 0); |
0854ccff | 166 | futex_noasync(&defer_thread_futex, FUTEX_WAKE, 1, |
04eb9c4f MD |
167 | NULL, NULL, 0); |
168 | } | |
169 | } | |
170 | ||
171 | static unsigned long rcu_defer_num_callbacks(void) | |
172 | { | |
173 | unsigned long num_items = 0, head; | |
dbc6128f | 174 | struct defer_queue *index; |
04eb9c4f | 175 | |
0376e7b2 PM |
176 | mutex_lock_defer(&rcu_defer_mutex); |
177 | cds_list_for_each_entry(index, ®istry_defer, list) { | |
6cf3827c | 178 | head = CMM_LOAD_SHARED(index->head); |
dbc6128f | 179 | num_items += head - index->tail; |
04eb9c4f | 180 | } |
6abb4bd5 | 181 | mutex_unlock(&rcu_defer_mutex); |
04eb9c4f MD |
182 | return num_items; |
183 | } | |
184 | ||
185 | /* | |
186 | * Defer thread waiting. Single thread. | |
187 | */ | |
188 | static void wait_defer(void) | |
189 | { | |
ec4e58a3 | 190 | uatomic_dec(&defer_thread_futex); |
d7ff6cee MD |
191 | /* Write futex before read queue */ |
192 | /* Write futex before read defer_thread_stop */ | |
193 | cmm_smp_mb(); | |
194 | if (_CMM_LOAD_SHARED(defer_thread_stop)) { | |
195 | uatomic_set(&defer_thread_futex, 0); | |
196 | pthread_exit(0); | |
197 | } | |
04eb9c4f | 198 | if (rcu_defer_num_callbacks()) { |
5481ddb3 | 199 | cmm_smp_mb(); /* Read queue before write futex */ |
04eb9c4f | 200 | /* Callbacks are queued, don't wait. */ |
ec4e58a3 | 201 | uatomic_set(&defer_thread_futex, 0); |
04eb9c4f | 202 | } else { |
5481ddb3 | 203 | cmm_smp_rmb(); /* Read queue before read futex */ |
ec4e58a3 | 204 | if (uatomic_read(&defer_thread_futex) == -1) |
0854ccff | 205 | futex_noasync(&defer_thread_futex, FUTEX_WAIT, -1, |
04eb9c4f MD |
206 | NULL, NULL, 0); |
207 | } | |
208 | } | |
209 | ||
786ee85b MD |
210 | /* |
211 | * Must be called after Q.S. is reached. | |
212 | */ | |
213 | static void rcu_defer_barrier_queue(struct defer_queue *queue, | |
804b4375 | 214 | unsigned long head) |
786ee85b MD |
215 | { |
216 | unsigned long i; | |
804b4375 MD |
217 | void (*fct)(void *p); |
218 | void *p; | |
786ee85b MD |
219 | |
220 | /* | |
221 | * Tail is only modified when lock is held. | |
222 | * Head is only modified by owner thread. | |
223 | */ | |
224 | ||
804b4375 | 225 | for (i = queue->tail; i != head;) { |
5481ddb3 | 226 | cmm_smp_rmb(); /* read head before q[]. */ |
6cf3827c | 227 | p = CMM_LOAD_SHARED(queue->q[i++ & DEFER_QUEUE_MASK]); |
a0b7f7ea | 228 | if (caa_unlikely(DQ_IS_FCT_BIT(p))) { |
804b4375 MD |
229 | DQ_CLEAR_FCT_BIT(p); |
230 | queue->last_fct_out = p; | |
6cf3827c | 231 | p = CMM_LOAD_SHARED(queue->q[i++ & DEFER_QUEUE_MASK]); |
a0b7f7ea | 232 | } else if (caa_unlikely(p == DQ_FCT_MARK)) { |
6cf3827c | 233 | p = CMM_LOAD_SHARED(queue->q[i++ & DEFER_QUEUE_MASK]); |
804b4375 | 234 | queue->last_fct_out = p; |
6cf3827c | 235 | p = CMM_LOAD_SHARED(queue->q[i++ & DEFER_QUEUE_MASK]); |
29cdb8d8 | 236 | } |
804b4375 | 237 | fct = queue->last_fct_out; |
804b4375 | 238 | fct(p); |
786ee85b | 239 | } |
5481ddb3 | 240 | cmm_smp_mb(); /* push tail after having used q[] */ |
6cf3827c | 241 | CMM_STORE_SHARED(queue->tail, i); |
786ee85b MD |
242 | } |
243 | ||
244 | static void _rcu_defer_barrier_thread(void) | |
245 | { | |
0d0e6c21 | 246 | unsigned long head, num_items; |
786ee85b MD |
247 | |
248 | head = defer_queue.head; | |
0d0e6c21 | 249 | num_items = head - defer_queue.tail; |
a0b7f7ea | 250 | if (caa_unlikely(!num_items)) |
0d0e6c21 | 251 | return; |
786ee85b MD |
252 | synchronize_rcu(); |
253 | rcu_defer_barrier_queue(&defer_queue, head); | |
254 | } | |
255 | ||
786ee85b MD |
256 | void rcu_defer_barrier_thread(void) |
257 | { | |
0376e7b2 | 258 | mutex_lock_defer(&rcu_defer_mutex); |
786ee85b | 259 | _rcu_defer_barrier_thread(); |
6abb4bd5 | 260 | mutex_unlock(&rcu_defer_mutex); |
786ee85b MD |
261 | } |
262 | ||
0d0e6c21 MD |
263 | /* |
264 | * rcu_defer_barrier - Execute all queued rcu callbacks. | |
265 | * | |
266 | * Execute all RCU callbacks queued before rcu_defer_barrier() execution. | |
267 | * All callbacks queued on the local thread prior to a rcu_defer_barrier() call | |
268 | * are guaranteed to be executed. | |
269 | * Callbacks queued by other threads concurrently with rcu_defer_barrier() | |
270 | * execution are not guaranteed to be executed in the current batch (could | |
271 | * be left for the next batch). These callbacks queued by other threads are only | |
272 | * guaranteed to be executed if there is explicit synchronization between | |
273 | * the thread adding to the queue and the thread issuing the defer_barrier call. | |
274 | */ | |
275 | ||
786ee85b MD |
276 | void rcu_defer_barrier(void) |
277 | { | |
dbc6128f | 278 | struct defer_queue *index; |
0d0e6c21 | 279 | unsigned long num_items = 0; |
786ee85b | 280 | |
0376e7b2 | 281 | if (cds_list_empty(®istry_defer)) |
786ee85b MD |
282 | return; |
283 | ||
0376e7b2 PM |
284 | mutex_lock_defer(&rcu_defer_mutex); |
285 | cds_list_for_each_entry(index, ®istry_defer, list) { | |
6cf3827c | 286 | index->last_head = CMM_LOAD_SHARED(index->head); |
dbc6128f | 287 | num_items += index->last_head - index->tail; |
0d0e6c21 | 288 | } |
a0b7f7ea | 289 | if (caa_likely(!num_items)) { |
0d0e6c21 MD |
290 | /* |
291 | * We skip the grace period because there are no queued | |
292 | * callbacks to execute. | |
293 | */ | |
294 | goto end; | |
295 | } | |
786ee85b | 296 | synchronize_rcu(); |
0376e7b2 | 297 | cds_list_for_each_entry(index, ®istry_defer, list) |
dbc6128f | 298 | rcu_defer_barrier_queue(index, index->last_head); |
0d0e6c21 | 299 | end: |
6abb4bd5 | 300 | mutex_unlock(&rcu_defer_mutex); |
786ee85b MD |
301 | } |
302 | ||
2c22932b | 303 | /* |
b4f313b7 | 304 | * _defer_rcu - Queue a RCU callback. |
2c22932b | 305 | */ |
3614f13c | 306 | void _defer_rcu(void (*fct)(void *p), void *p) |
2c22932b MD |
307 | { |
308 | unsigned long head, tail; | |
309 | ||
310 | /* | |
311 | * Head is only modified by ourself. Tail can be modified by reclamation | |
312 | * thread. | |
313 | */ | |
314 | head = defer_queue.head; | |
6cf3827c | 315 | tail = CMM_LOAD_SHARED(defer_queue.tail); |
2c22932b MD |
316 | |
317 | /* | |
ec8e44cf | 318 | * If queue is full, or reached threshold. Empty queue ourself. |
2c22932b MD |
319 | * Worse-case: must allow 2 supplementary entries for fct pointer. |
320 | */ | |
a0b7f7ea | 321 | if (caa_unlikely(head - tail >= DEFER_QUEUE_SIZE - 2)) { |
2c22932b MD |
322 | assert(head - tail <= DEFER_QUEUE_SIZE); |
323 | rcu_defer_barrier_thread(); | |
6cf3827c | 324 | assert(head - CMM_LOAD_SHARED(defer_queue.tail) == 0); |
2c22932b MD |
325 | } |
326 | ||
2ef1bfb2 LJ |
327 | /* |
328 | * Encode: | |
329 | * if the function is not changed and the data is aligned and it is | |
330 | * not the marker: | |
331 | * store the data | |
332 | * otherwise if the function is aligned and its not the marker: | |
333 | * store the function with DQ_FCT_BIT | |
334 | * store the data | |
335 | * otherwise: | |
336 | * store the marker (DQ_FCT_MARK) | |
337 | * store the function | |
338 | * store the data | |
339 | * | |
340 | * Decode: see the comments before 'struct defer_queue' | |
341 | * or the code in rcu_defer_barrier_queue(). | |
342 | */ | |
a0b7f7ea | 343 | if (caa_unlikely(defer_queue.last_fct_in != fct |
2ef1bfb2 LJ |
344 | || DQ_IS_FCT_BIT(p) |
345 | || p == DQ_FCT_MARK)) { | |
2c22932b | 346 | defer_queue.last_fct_in = fct; |
a0b7f7ea | 347 | if (caa_unlikely(DQ_IS_FCT_BIT(fct) || fct == DQ_FCT_MARK)) { |
6cf3827c | 348 | _CMM_STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK], |
2c22932b | 349 | DQ_FCT_MARK); |
6cf3827c | 350 | _CMM_STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK], |
2c22932b MD |
351 | fct); |
352 | } else { | |
353 | DQ_SET_FCT_BIT(fct); | |
6cf3827c | 354 | _CMM_STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK], |
2c22932b MD |
355 | fct); |
356 | } | |
2c22932b | 357 | } |
6cf3827c | 358 | _CMM_STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK], p); |
5481ddb3 | 359 | cmm_smp_wmb(); /* Publish new pointer before head */ |
2c22932b | 360 | /* Write q[] before head. */ |
6cf3827c | 361 | CMM_STORE_SHARED(defer_queue.head, head); |
5481ddb3 | 362 | cmm_smp_mb(); /* Write queue head before read futex */ |
2c22932b MD |
363 | /* |
364 | * Wake-up any waiting defer thread. | |
365 | */ | |
366 | wake_up_defer(); | |
367 | } | |
368 | ||
786ee85b MD |
369 | void *thr_defer(void *args) |
370 | { | |
371 | for (;;) { | |
4ce9e4f2 MD |
372 | /* |
373 | * "Be green". Don't wake up the CPU if there is no RCU work | |
374 | * to perform whatsoever. Aims at saving laptop battery life by | |
375 | * leaving the processor in sleep state when idle. | |
376 | */ | |
4ce9e4f2 | 377 | wait_defer(); |
4ce9e4f2 | 378 | /* Sleeping after wait_defer to let many callbacks enqueue */ |
71df5ef4 | 379 | poll(NULL,0,100); /* wait for 100ms */ |
786ee85b MD |
380 | rcu_defer_barrier(); |
381 | } | |
382 | ||
383 | return NULL; | |
384 | } | |
385 | ||
386 | /* | |
387 | * library wrappers to be used by non-LGPL compatible source code. | |
388 | */ | |
389 | ||
3614f13c | 390 | void defer_rcu(void (*fct)(void *p), void *p) |
786ee85b | 391 | { |
3614f13c | 392 | _defer_rcu(fct, p); |
786ee85b MD |
393 | } |
394 | ||
786ee85b MD |
395 | static void start_defer_thread(void) |
396 | { | |
397 | int ret; | |
398 | ||
dbc6128f | 399 | ret = pthread_create(&tid_defer, NULL, thr_defer, NULL); |
786ee85b MD |
400 | assert(!ret); |
401 | } | |
402 | ||
403 | static void stop_defer_thread(void) | |
404 | { | |
405 | int ret; | |
406 | void *tret; | |
407 | ||
d7ff6cee MD |
408 | _CMM_STORE_SHARED(defer_thread_stop, 1); |
409 | /* Store defer_thread_stop before testing futex */ | |
410 | cmm_smp_mb(); | |
4ce9e4f2 | 411 | wake_up_defer(); |
d7ff6cee | 412 | |
786ee85b MD |
413 | ret = pthread_join(tid_defer, &tret); |
414 | assert(!ret); | |
d7ff6cee MD |
415 | |
416 | CMM_STORE_SHARED(defer_thread_stop, 0); | |
417 | /* defer thread should always exit when futex value is 0 */ | |
418 | assert(uatomic_read(&defer_thread_futex) == 0); | |
786ee85b MD |
419 | } |
420 | ||
7fdbbd61 | 421 | int rcu_defer_register_thread(void) |
786ee85b | 422 | { |
dbc6128f MD |
423 | int was_empty; |
424 | ||
425 | assert(defer_queue.last_head == 0); | |
426 | assert(defer_queue.q == NULL); | |
427 | defer_queue.q = malloc(sizeof(void *) * DEFER_QUEUE_SIZE); | |
7fdbbd61 MD |
428 | if (!defer_queue.q) |
429 | return -ENOMEM; | |
786ee85b | 430 | |
0376e7b2 PM |
431 | mutex_lock_defer(&defer_thread_mutex); |
432 | mutex_lock_defer(&rcu_defer_mutex); | |
433 | was_empty = cds_list_empty(®istry_defer); | |
434 | cds_list_add(&defer_queue.list, ®istry_defer); | |
6abb4bd5 | 435 | mutex_unlock(&rcu_defer_mutex); |
786ee85b | 436 | |
dbc6128f | 437 | if (was_empty) |
786ee85b | 438 | start_defer_thread(); |
6abb4bd5 | 439 | mutex_unlock(&defer_thread_mutex); |
7fdbbd61 | 440 | return 0; |
786ee85b MD |
441 | } |
442 | ||
443 | void rcu_defer_unregister_thread(void) | |
444 | { | |
dbc6128f | 445 | int is_empty; |
786ee85b | 446 | |
0376e7b2 PM |
447 | mutex_lock_defer(&defer_thread_mutex); |
448 | mutex_lock_defer(&rcu_defer_mutex); | |
16aa9ee8 | 449 | cds_list_del(&defer_queue.list); |
786ee85b MD |
450 | _rcu_defer_barrier_thread(); |
451 | free(defer_queue.q); | |
452 | defer_queue.q = NULL; | |
0376e7b2 | 453 | is_empty = cds_list_empty(®istry_defer); |
6abb4bd5 | 454 | mutex_unlock(&rcu_defer_mutex); |
786ee85b | 455 | |
dbc6128f | 456 | if (is_empty) |
786ee85b | 457 | stop_defer_thread(); |
6abb4bd5 | 458 | mutex_unlock(&defer_thread_mutex); |
786ee85b MD |
459 | } |
460 | ||
02be5561 | 461 | void rcu_defer_exit(void) |
786ee85b | 462 | { |
0376e7b2 | 463 | assert(cds_list_empty(®istry_defer)); |
786ee85b | 464 | } |
0376e7b2 PM |
465 | |
466 | #endif /* _URCU_DEFER_IMPL_H */ |