x86: allocate membarrier system call number
[urcu.git] / urcu.c
1 /*
2 * urcu.c
3 *
4 * Userspace RCU library
5 *
6 * Copyright (c) 2009 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
7 * Copyright (c) 2009 Paul E. McKenney, IBM Corporation.
8 *
9 * This library is free software; you can redistribute it and/or
10 * modify it under the terms of the GNU Lesser General Public
11 * License as published by the Free Software Foundation; either
12 * version 2.1 of the License, or (at your option) any later version.
13 *
14 * This library is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
17 * Lesser General Public License for more details.
18 *
19 * You should have received a copy of the GNU Lesser General Public
20 * License along with this library; if not, write to the Free Software
21 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
22 *
23 * IBM's contributions to this file may be relicensed under LGPLv2 or later.
24 */
25
26 #define _BSD_SOURCE
27 #define _GNU_SOURCE
28 #define _LGPL_SOURCE
29 #define _DEFAULT_SOURCE
30 #include <stdio.h>
31 #include <pthread.h>
32 #include <signal.h>
33 #include <assert.h>
34 #include <stdlib.h>
35 #include <stdint.h>
36 #include <string.h>
37 #include <errno.h>
38 #include <poll.h>
39
40 #include "urcu/arch.h"
41 #include "urcu/wfcqueue.h"
42 #include "urcu/map/urcu.h"
43 #include "urcu/static/urcu.h"
44 #include "urcu-pointer.h"
45 #include "urcu/tls-compat.h"
46
47 #include "urcu-die.h"
48 #include "urcu-wait.h"
49
50 /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */
51 #undef _LGPL_SOURCE
52 #include "urcu.h"
53 #define _LGPL_SOURCE
54
55 /*
56 * If a reader is really non-cooperative and refuses to commit its
57 * rcu_active_readers count to memory (there is no barrier in the reader
58 * per-se), kick it after 10 loops waiting for it.
59 */
60 #define KICK_READER_LOOPS 10
61
62 /*
63 * Active attempts to check for reader Q.S. before calling futex().
64 */
65 #define RCU_QS_ACTIVE_ATTEMPTS 100
66
67 /* If the headers do not support membarrier system call, fall back on RCU_MB */
68 #ifdef __NR_membarrier
69 # define membarrier(...) syscall(__NR_membarrier, __VA_ARGS__)
70 #else
71 # define membarrier(...) -ENOSYS
72 #endif
73
74 enum membarrier_cmd {
75 MEMBARRIER_CMD_QUERY = 0,
76 MEMBARRIER_CMD_SHARED = (1 << 0),
77 };
78
79 #ifdef RCU_MEMBARRIER
80 static int init_done;
81 int rcu_has_sys_membarrier;
82
83 void __attribute__((constructor)) rcu_init(void);
84 #endif
85
86 #ifdef RCU_MB
87 void rcu_init(void)
88 {
89 }
90 #endif
91
92 #ifdef RCU_SIGNAL
93 static int init_done;
94
95 void __attribute__((constructor)) rcu_init(void);
96 void __attribute__((destructor)) rcu_exit(void);
97 #endif
98
99 /*
100 * rcu_gp_lock ensures mutual exclusion between threads calling
101 * synchronize_rcu().
102 */
103 static pthread_mutex_t rcu_gp_lock = PTHREAD_MUTEX_INITIALIZER;
104 /*
105 * rcu_registry_lock ensures mutual exclusion between threads
106 * registering and unregistering themselves to/from the registry, and
107 * with threads reading that registry from synchronize_rcu(). However,
108 * this lock is not held all the way through the completion of awaiting
109 * for the grace period. It is sporadically released between iterations
110 * on the registry.
111 * rcu_registry_lock may nest inside rcu_gp_lock.
112 */
113 static pthread_mutex_t rcu_registry_lock = PTHREAD_MUTEX_INITIALIZER;
114 struct rcu_gp rcu_gp = { .ctr = RCU_GP_COUNT };
115
116 /*
117 * Written to only by each individual reader. Read by both the reader and the
118 * writers.
119 */
120 DEFINE_URCU_TLS(struct rcu_reader, rcu_reader);
121
122 static CDS_LIST_HEAD(registry);
123
124 /*
125 * Queue keeping threads awaiting to wait for a grace period. Contains
126 * struct gp_waiters_thread objects.
127 */
128 static DEFINE_URCU_WAIT_QUEUE(gp_waiters);
129
130 static void mutex_lock(pthread_mutex_t *mutex)
131 {
132 int ret;
133
134 #ifndef DISTRUST_SIGNALS_EXTREME
135 ret = pthread_mutex_lock(mutex);
136 if (ret)
137 urcu_die(ret);
138 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
139 while ((ret = pthread_mutex_trylock(mutex)) != 0) {
140 if (ret != EBUSY && ret != EINTR)
141 urcu_die(ret);
142 if (CMM_LOAD_SHARED(URCU_TLS(rcu_reader).need_mb)) {
143 cmm_smp_mb();
144 _CMM_STORE_SHARED(URCU_TLS(rcu_reader).need_mb, 0);
145 cmm_smp_mb();
146 }
147 (void) poll(NULL, 0, 10);
148 }
149 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
150 }
151
152 static void mutex_unlock(pthread_mutex_t *mutex)
153 {
154 int ret;
155
156 ret = pthread_mutex_unlock(mutex);
157 if (ret)
158 urcu_die(ret);
159 }
160
161 #ifdef RCU_MEMBARRIER
162 static void smp_mb_master(void)
163 {
164 if (caa_likely(rcu_has_sys_membarrier))
165 (void) membarrier(MEMBARRIER_CMD_SHARED, 0);
166 else
167 cmm_smp_mb();
168 }
169 #endif
170
171 #ifdef RCU_MB
172 static void smp_mb_master(void)
173 {
174 cmm_smp_mb();
175 }
176 #endif
177
178 #ifdef RCU_SIGNAL
179 static void force_mb_all_readers(void)
180 {
181 struct rcu_reader *index;
182
183 /*
184 * Ask for each threads to execute a cmm_smp_mb() so we can consider the
185 * compiler barriers around rcu read lock as real memory barriers.
186 */
187 if (cds_list_empty(&registry))
188 return;
189 /*
190 * pthread_kill has a cmm_smp_mb(). But beware, we assume it performs
191 * a cache flush on architectures with non-coherent cache. Let's play
192 * safe and don't assume anything : we use cmm_smp_mc() to make sure the
193 * cache flush is enforced.
194 */
195 cds_list_for_each_entry(index, &registry, node) {
196 CMM_STORE_SHARED(index->need_mb, 1);
197 pthread_kill(index->tid, SIGRCU);
198 }
199 /*
200 * Wait for sighandler (and thus mb()) to execute on every thread.
201 *
202 * Note that the pthread_kill() will never be executed on systems
203 * that correctly deliver signals in a timely manner. However, it
204 * is not uncommon for kernels to have bugs that can result in
205 * lost or unduly delayed signals.
206 *
207 * If you are seeing the below pthread_kill() executing much at
208 * all, we suggest testing the underlying kernel and filing the
209 * relevant bug report. For Linux kernels, we recommend getting
210 * the Linux Test Project (LTP).
211 */
212 cds_list_for_each_entry(index, &registry, node) {
213 while (CMM_LOAD_SHARED(index->need_mb)) {
214 pthread_kill(index->tid, SIGRCU);
215 (void) poll(NULL, 0, 1);
216 }
217 }
218 cmm_smp_mb(); /* read ->need_mb before ending the barrier */
219 }
220
221 static void smp_mb_master(void)
222 {
223 force_mb_all_readers();
224 }
225 #endif /* #ifdef RCU_SIGNAL */
226
227 /*
228 * synchronize_rcu() waiting. Single thread.
229 */
230 static void wait_gp(void)
231 {
232 /* Read reader_gp before read futex */
233 smp_mb_master();
234 if (uatomic_read(&rcu_gp.futex) != -1)
235 return;
236 while (futex_async(&rcu_gp.futex, FUTEX_WAIT, -1,
237 NULL, NULL, 0)) {
238 switch (errno) {
239 case EWOULDBLOCK:
240 /* Value already changed. */
241 return;
242 case EINTR:
243 /* Retry if interrupted by signal. */
244 break; /* Get out of switch. */
245 default:
246 /* Unexpected error. */
247 urcu_die(errno);
248 }
249 }
250 }
251
252 /*
253 * Always called with rcu_registry lock held. Releases this lock between
254 * iterations and grabs it again. Holds the lock when it returns.
255 */
256 static void wait_for_readers(struct cds_list_head *input_readers,
257 struct cds_list_head *cur_snap_readers,
258 struct cds_list_head *qsreaders)
259 {
260 unsigned int wait_loops = 0;
261 struct rcu_reader *index, *tmp;
262 #ifdef HAS_INCOHERENT_CACHES
263 unsigned int wait_gp_loops = 0;
264 #endif /* HAS_INCOHERENT_CACHES */
265
266 /*
267 * Wait for each thread URCU_TLS(rcu_reader).ctr to either
268 * indicate quiescence (not nested), or observe the current
269 * rcu_gp.ctr value.
270 */
271 for (;;) {
272 if (wait_loops < RCU_QS_ACTIVE_ATTEMPTS)
273 wait_loops++;
274 if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
275 uatomic_dec(&rcu_gp.futex);
276 /* Write futex before read reader_gp */
277 smp_mb_master();
278 }
279
280 cds_list_for_each_entry_safe(index, tmp, input_readers, node) {
281 switch (rcu_reader_state(&index->ctr)) {
282 case RCU_READER_ACTIVE_CURRENT:
283 if (cur_snap_readers) {
284 cds_list_move(&index->node,
285 cur_snap_readers);
286 break;
287 }
288 /* Fall-through */
289 case RCU_READER_INACTIVE:
290 cds_list_move(&index->node, qsreaders);
291 break;
292 case RCU_READER_ACTIVE_OLD:
293 /*
294 * Old snapshot. Leaving node in
295 * input_readers will make us busy-loop
296 * until the snapshot becomes current or
297 * the reader becomes inactive.
298 */
299 break;
300 }
301 }
302
303 #ifndef HAS_INCOHERENT_CACHES
304 if (cds_list_empty(input_readers)) {
305 if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
306 /* Read reader_gp before write futex */
307 smp_mb_master();
308 uatomic_set(&rcu_gp.futex, 0);
309 }
310 break;
311 } else {
312 /* Temporarily unlock the registry lock. */
313 mutex_unlock(&rcu_registry_lock);
314 if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS)
315 wait_gp();
316 else
317 caa_cpu_relax();
318 /* Re-lock the registry lock before the next loop. */
319 mutex_lock(&rcu_registry_lock);
320 }
321 #else /* #ifndef HAS_INCOHERENT_CACHES */
322 /*
323 * BUSY-LOOP. Force the reader thread to commit its
324 * URCU_TLS(rcu_reader).ctr update to memory if we wait
325 * for too long.
326 */
327 if (cds_list_empty(input_readers)) {
328 if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
329 /* Read reader_gp before write futex */
330 smp_mb_master();
331 uatomic_set(&rcu_gp.futex, 0);
332 }
333 break;
334 } else {
335 if (wait_gp_loops == KICK_READER_LOOPS) {
336 smp_mb_master();
337 wait_gp_loops = 0;
338 }
339 /* Temporarily unlock the registry lock. */
340 mutex_unlock(&rcu_registry_lock);
341 if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
342 wait_gp();
343 wait_gp_loops++;
344 } else {
345 caa_cpu_relax();
346 }
347 /* Re-lock the registry lock before the next loop. */
348 mutex_lock(&rcu_registry_lock);
349 }
350 #endif /* #else #ifndef HAS_INCOHERENT_CACHES */
351 }
352 }
353
354 void synchronize_rcu(void)
355 {
356 CDS_LIST_HEAD(cur_snap_readers);
357 CDS_LIST_HEAD(qsreaders);
358 DEFINE_URCU_WAIT_NODE(wait, URCU_WAIT_WAITING);
359 struct urcu_waiters waiters;
360
361 /*
362 * Add ourself to gp_waiters queue of threads awaiting to wait
363 * for a grace period. Proceed to perform the grace period only
364 * if we are the first thread added into the queue.
365 * The implicit memory barrier before urcu_wait_add()
366 * orders prior memory accesses of threads put into the wait
367 * queue before their insertion into the wait queue.
368 */
369 if (urcu_wait_add(&gp_waiters, &wait) != 0) {
370 /* Not first in queue: will be awakened by another thread. */
371 urcu_adaptative_busy_wait(&wait);
372 /* Order following memory accesses after grace period. */
373 cmm_smp_mb();
374 return;
375 }
376 /* We won't need to wake ourself up */
377 urcu_wait_set_state(&wait, URCU_WAIT_RUNNING);
378
379 mutex_lock(&rcu_gp_lock);
380
381 /*
382 * Move all waiters into our local queue.
383 */
384 urcu_move_waiters(&waiters, &gp_waiters);
385
386 mutex_lock(&rcu_registry_lock);
387
388 if (cds_list_empty(&registry))
389 goto out;
390
391 /*
392 * All threads should read qparity before accessing data structure
393 * where new ptr points to. Must be done within rcu_registry_lock
394 * because it iterates on reader threads.
395 */
396 /* Write new ptr before changing the qparity */
397 smp_mb_master();
398
399 /*
400 * Wait for readers to observe original parity or be quiescent.
401 * wait_for_readers() can release and grab again rcu_registry_lock
402 * interally.
403 */
404 wait_for_readers(&registry, &cur_snap_readers, &qsreaders);
405
406 /*
407 * Must finish waiting for quiescent state for original parity before
408 * committing next rcu_gp.ctr update to memory. Failure to do so could
409 * result in the writer waiting forever while new readers are always
410 * accessing data (no progress). Enforce compiler-order of load
411 * URCU_TLS(rcu_reader).ctr before store to rcu_gp.ctr.
412 */
413 cmm_barrier();
414
415 /*
416 * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
417 * model easier to understand. It does not have a big performance impact
418 * anyway, given this is the write-side.
419 */
420 cmm_smp_mb();
421
422 /* Switch parity: 0 -> 1, 1 -> 0 */
423 CMM_STORE_SHARED(rcu_gp.ctr, rcu_gp.ctr ^ RCU_GP_CTR_PHASE);
424
425 /*
426 * Must commit rcu_gp.ctr update to memory before waiting for quiescent
427 * state. Failure to do so could result in the writer waiting forever
428 * while new readers are always accessing data (no progress). Enforce
429 * compiler-order of store to rcu_gp.ctr before load rcu_reader ctr.
430 */
431 cmm_barrier();
432
433 /*
434 *
435 * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
436 * model easier to understand. It does not have a big performance impact
437 * anyway, given this is the write-side.
438 */
439 cmm_smp_mb();
440
441 /*
442 * Wait for readers to observe new parity or be quiescent.
443 * wait_for_readers() can release and grab again rcu_registry_lock
444 * interally.
445 */
446 wait_for_readers(&cur_snap_readers, NULL, &qsreaders);
447
448 /*
449 * Put quiescent reader list back into registry.
450 */
451 cds_list_splice(&qsreaders, &registry);
452
453 /*
454 * Finish waiting for reader threads before letting the old ptr
455 * being freed. Must be done within rcu_registry_lock because it
456 * iterates on reader threads.
457 */
458 smp_mb_master();
459 out:
460 mutex_unlock(&rcu_registry_lock);
461 mutex_unlock(&rcu_gp_lock);
462
463 /*
464 * Wakeup waiters only after we have completed the grace period
465 * and have ensured the memory barriers at the end of the grace
466 * period have been issued.
467 */
468 urcu_wake_all_waiters(&waiters);
469 }
470
471 /*
472 * library wrappers to be used by non-LGPL compatible source code.
473 */
474
475 void rcu_read_lock(void)
476 {
477 _rcu_read_lock();
478 }
479
480 void rcu_read_unlock(void)
481 {
482 _rcu_read_unlock();
483 }
484
485 int rcu_read_ongoing(void)
486 {
487 return _rcu_read_ongoing();
488 }
489
490 void rcu_register_thread(void)
491 {
492 URCU_TLS(rcu_reader).tid = pthread_self();
493 assert(URCU_TLS(rcu_reader).need_mb == 0);
494 assert(!(URCU_TLS(rcu_reader).ctr & RCU_GP_CTR_NEST_MASK));
495
496 mutex_lock(&rcu_registry_lock);
497 assert(!URCU_TLS(rcu_reader).registered);
498 URCU_TLS(rcu_reader).registered = 1;
499 rcu_init(); /* In case gcc does not support constructor attribute */
500 cds_list_add(&URCU_TLS(rcu_reader).node, &registry);
501 mutex_unlock(&rcu_registry_lock);
502 }
503
504 void rcu_unregister_thread(void)
505 {
506 mutex_lock(&rcu_registry_lock);
507 assert(URCU_TLS(rcu_reader).registered);
508 URCU_TLS(rcu_reader).registered = 0;
509 cds_list_del(&URCU_TLS(rcu_reader).node);
510 mutex_unlock(&rcu_registry_lock);
511 }
512
513 #ifdef RCU_MEMBARRIER
514 void rcu_init(void)
515 {
516 int ret;
517
518 if (init_done)
519 return;
520 init_done = 1;
521 ret = membarrier(MEMBARRIER_CMD_QUERY, 0);
522 if (ret >= 0 && (ret & MEMBARRIER_CMD_SHARED)) {
523 rcu_has_sys_membarrier = 1;
524 }
525 }
526 #endif
527
528 #ifdef RCU_SIGNAL
529 static void sigrcu_handler(int signo, siginfo_t *siginfo, void *context)
530 {
531 /*
532 * Executing this cmm_smp_mb() is the only purpose of this signal handler.
533 * It punctually promotes cmm_barrier() into cmm_smp_mb() on every thread it is
534 * executed on.
535 */
536 cmm_smp_mb();
537 _CMM_STORE_SHARED(URCU_TLS(rcu_reader).need_mb, 0);
538 cmm_smp_mb();
539 }
540
541 /*
542 * rcu_init constructor. Called when the library is linked, but also when
543 * reader threads are calling rcu_register_thread().
544 * Should only be called by a single thread at a given time. This is ensured by
545 * holing the rcu_registry_lock from rcu_register_thread() or by running
546 * at library load time, which should not be executed by multiple
547 * threads nor concurrently with rcu_register_thread() anyway.
548 */
549 void rcu_init(void)
550 {
551 struct sigaction act;
552 int ret;
553
554 if (init_done)
555 return;
556 init_done = 1;
557
558 act.sa_sigaction = sigrcu_handler;
559 act.sa_flags = SA_SIGINFO | SA_RESTART;
560 sigemptyset(&act.sa_mask);
561 ret = sigaction(SIGRCU, &act, NULL);
562 if (ret)
563 urcu_die(errno);
564 }
565
566 void rcu_exit(void)
567 {
568 /*
569 * Don't unregister the SIGRCU signal handler anymore, because
570 * call_rcu threads could still be using it shortly before the
571 * application exits.
572 * Assertion disabled because call_rcu threads are now rcu
573 * readers, and left running at exit.
574 * assert(cds_list_empty(&registry));
575 */
576 }
577
578 #endif /* #ifdef RCU_SIGNAL */
579
580 DEFINE_RCU_FLAVOR(rcu_flavor);
581
582 #include "urcu-call-rcu-impl.h"
583 #include "urcu-defer-impl.h"
This page took 0.061101 seconds and 4 git commands to generate.