xref: /qemu/util/rcu.c (revision 727385c4)
1 /*
2  * urcu-mb.c
3  *
4  * Userspace RCU library with explicit memory barriers
5  *
6  * Copyright (c) 2009 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
7  * Copyright (c) 2009 Paul E. McKenney, IBM Corporation.
8  * Copyright 2015 Red Hat, Inc.
9  *
10  * Ported to QEMU by Paolo Bonzini  <pbonzini@redhat.com>
11  *
12  * This library is free software; you can redistribute it and/or
13  * modify it under the terms of the GNU Lesser General Public
14  * License as published by the Free Software Foundation; either
15  * version 2.1 of the License, or (at your option) any later version.
16  *
17  * This library is distributed in the hope that it will be useful,
18  * but WITHOUT ANY WARRANTY; without even the implied warranty of
19  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
20  * Lesser General Public License for more details.
21  *
22  * You should have received a copy of the GNU Lesser General Public
23  * License along with this library; if not, write to the Free Software
24  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
25  *
26  * IBM's contributions to this file may be relicensed under LGPLv2 or later.
27  */
28 
29 #include "qemu/osdep.h"
30 #include "qemu/rcu.h"
31 #include "qemu/atomic.h"
32 #include "qemu/thread.h"
33 #include "qemu/main-loop.h"
34 #include "qemu/lockable.h"
35 #if defined(CONFIG_MALLOC_TRIM)
36 #include <malloc.h>
37 #endif
38 
39 /*
40  * Global grace period counter.  Bit 0 is always one in rcu_gp_ctr.
41  * Bits 1 and above are defined in synchronize_rcu.
42  */
43 #define RCU_GP_LOCKED           (1UL << 0)
44 #define RCU_GP_CTR              (1UL << 1)
45 
46 unsigned long rcu_gp_ctr = RCU_GP_LOCKED;
47 
48 QemuEvent rcu_gp_event;
49 static QemuMutex rcu_registry_lock;
50 static QemuMutex rcu_sync_lock;
51 
52 /*
53  * Check whether a quiescent state was crossed between the beginning of
54  * update_counter_and_wait and now.
55  */
56 static inline int rcu_gp_ongoing(unsigned long *ctr)
57 {
58     unsigned long v;
59 
60     v = qatomic_read(ctr);
61     return v && (v != rcu_gp_ctr);
62 }
63 
64 /* Written to only by each individual reader. Read by both the reader and the
65  * writers.
66  */
67 __thread struct rcu_reader_data rcu_reader;
68 
69 /* Protected by rcu_registry_lock.  */
70 typedef QLIST_HEAD(, rcu_reader_data) ThreadList;
71 static ThreadList registry = QLIST_HEAD_INITIALIZER(registry);
72 
73 /* Wait for previous parity/grace period to be empty of readers.  */
74 static void wait_for_readers(void)
75 {
76     ThreadList qsreaders = QLIST_HEAD_INITIALIZER(qsreaders);
77     struct rcu_reader_data *index, *tmp;
78 
79     for (;;) {
80         /* We want to be notified of changes made to rcu_gp_ongoing
81          * while we walk the list.
82          */
83         qemu_event_reset(&rcu_gp_event);
84 
85         /* Instead of using qatomic_mb_set for index->waiting, and
86          * qatomic_mb_read for index->ctr, memory barriers are placed
87          * manually since writes to different threads are independent.
88          * qemu_event_reset has acquire semantics, so no memory barrier
89          * is needed here.
90          */
91         QLIST_FOREACH(index, &registry, node) {
92             qatomic_set(&index->waiting, true);
93         }
94 
95         /* Here, order the stores to index->waiting before the loads of
96          * index->ctr.  Pairs with smp_mb_placeholder() in rcu_read_unlock(),
97          * ensuring that the loads of index->ctr are sequentially consistent.
98          */
99         smp_mb_global();
100 
101         QLIST_FOREACH_SAFE(index, &registry, node, tmp) {
102             if (!rcu_gp_ongoing(&index->ctr)) {
103                 QLIST_REMOVE(index, node);
104                 QLIST_INSERT_HEAD(&qsreaders, index, node);
105 
106                 /* No need for mb_set here, worst of all we
107                  * get some extra futex wakeups.
108                  */
109                 qatomic_set(&index->waiting, false);
110             }
111         }
112 
113         if (QLIST_EMPTY(&registry)) {
114             break;
115         }
116 
117         /* Wait for one thread to report a quiescent state and try again.
118          * Release rcu_registry_lock, so rcu_(un)register_thread() doesn't
119          * wait too much time.
120          *
121          * rcu_register_thread() may add nodes to &registry; it will not
122          * wake up synchronize_rcu, but that is okay because at least another
123          * thread must exit its RCU read-side critical section before
124          * synchronize_rcu is done.  The next iteration of the loop will
125          * move the new thread's rcu_reader from &registry to &qsreaders,
126          * because rcu_gp_ongoing() will return false.
127          *
128          * rcu_unregister_thread() may remove nodes from &qsreaders instead
129          * of &registry if it runs during qemu_event_wait.  That's okay;
130          * the node then will not be added back to &registry by QLIST_SWAP
131          * below.  The invariant is that the node is part of one list when
132          * rcu_registry_lock is released.
133          */
134         qemu_mutex_unlock(&rcu_registry_lock);
135         qemu_event_wait(&rcu_gp_event);
136         qemu_mutex_lock(&rcu_registry_lock);
137     }
138 
139     /* put back the reader list in the registry */
140     QLIST_SWAP(&registry, &qsreaders, node);
141 }
142 
143 void synchronize_rcu(void)
144 {
145     QEMU_LOCK_GUARD(&rcu_sync_lock);
146 
147     /* Write RCU-protected pointers before reading p_rcu_reader->ctr.
148      * Pairs with smp_mb_placeholder() in rcu_read_lock().
149      */
150     smp_mb_global();
151 
152     QEMU_LOCK_GUARD(&rcu_registry_lock);
153     if (!QLIST_EMPTY(&registry)) {
154         /* In either case, the qatomic_mb_set below blocks stores that free
155          * old RCU-protected pointers.
156          */
157         if (sizeof(rcu_gp_ctr) < 8) {
158             /* For architectures with 32-bit longs, a two-subphases algorithm
159              * ensures we do not encounter overflow bugs.
160              *
161              * Switch parity: 0 -> 1, 1 -> 0.
162              */
163             qatomic_mb_set(&rcu_gp_ctr, rcu_gp_ctr ^ RCU_GP_CTR);
164             wait_for_readers();
165             qatomic_mb_set(&rcu_gp_ctr, rcu_gp_ctr ^ RCU_GP_CTR);
166         } else {
167             /* Increment current grace period.  */
168             qatomic_mb_set(&rcu_gp_ctr, rcu_gp_ctr + RCU_GP_CTR);
169         }
170 
171         wait_for_readers();
172     }
173 }
174 
175 
176 #define RCU_CALL_MIN_SIZE        30
177 
178 /* Multi-producer, single-consumer queue based on urcu/static/wfqueue.h
179  * from liburcu.  Note that head is only used by the consumer.
180  */
181 static struct rcu_head dummy;
182 static struct rcu_head *head = &dummy, **tail = &dummy.next;
183 static int rcu_call_count;
184 static QemuEvent rcu_call_ready_event;
185 
186 static void enqueue(struct rcu_head *node)
187 {
188     struct rcu_head **old_tail;
189 
190     node->next = NULL;
191     old_tail = qatomic_xchg(&tail, &node->next);
192     qatomic_mb_set(old_tail, node);
193 }
194 
195 static struct rcu_head *try_dequeue(void)
196 {
197     struct rcu_head *node, *next;
198 
199 retry:
200     /* Test for an empty list, which we do not expect.  Note that for
201      * the consumer head and tail are always consistent.  The head
202      * is consistent because only the consumer reads/writes it.
203      * The tail, because it is the first step in the enqueuing.
204      * It is only the next pointers that might be inconsistent.
205      */
206     if (head == &dummy && qatomic_mb_read(&tail) == &dummy.next) {
207         abort();
208     }
209 
210     /* If the head node has NULL in its next pointer, the value is
211      * wrong and we need to wait until its enqueuer finishes the update.
212      */
213     node = head;
214     next = qatomic_mb_read(&head->next);
215     if (!next) {
216         return NULL;
217     }
218 
219     /* Since we are the sole consumer, and we excluded the empty case
220      * above, the queue will always have at least two nodes: the
221      * dummy node, and the one being removed.  So we do not need to update
222      * the tail pointer.
223      */
224     head = next;
225 
226     /* If we dequeued the dummy node, add it back at the end and retry.  */
227     if (node == &dummy) {
228         enqueue(node);
229         goto retry;
230     }
231 
232     return node;
233 }
234 
235 static void *call_rcu_thread(void *opaque)
236 {
237     struct rcu_head *node;
238 
239     rcu_register_thread();
240 
241     for (;;) {
242         int tries = 0;
243         int n = qatomic_read(&rcu_call_count);
244 
245         /* Heuristically wait for a decent number of callbacks to pile up.
246          * Fetch rcu_call_count now, we only must process elements that were
247          * added before synchronize_rcu() starts.
248          */
249         while (n == 0 || (n < RCU_CALL_MIN_SIZE && ++tries <= 5)) {
250             g_usleep(10000);
251             if (n == 0) {
252                 qemu_event_reset(&rcu_call_ready_event);
253                 n = qatomic_read(&rcu_call_count);
254                 if (n == 0) {
255 #if defined(CONFIG_MALLOC_TRIM)
256                     malloc_trim(4 * 1024 * 1024);
257 #endif
258                     qemu_event_wait(&rcu_call_ready_event);
259                 }
260             }
261             n = qatomic_read(&rcu_call_count);
262         }
263 
264         qatomic_sub(&rcu_call_count, n);
265         synchronize_rcu();
266         qemu_mutex_lock_iothread();
267         while (n > 0) {
268             node = try_dequeue();
269             while (!node) {
270                 qemu_mutex_unlock_iothread();
271                 qemu_event_reset(&rcu_call_ready_event);
272                 node = try_dequeue();
273                 if (!node) {
274                     qemu_event_wait(&rcu_call_ready_event);
275                     node = try_dequeue();
276                 }
277                 qemu_mutex_lock_iothread();
278             }
279 
280             n--;
281             node->func(node);
282         }
283         qemu_mutex_unlock_iothread();
284     }
285     abort();
286 }
287 
288 void call_rcu1(struct rcu_head *node, void (*func)(struct rcu_head *node))
289 {
290     node->func = func;
291     enqueue(node);
292     qatomic_inc(&rcu_call_count);
293     qemu_event_set(&rcu_call_ready_event);
294 }
295 
296 
297 struct rcu_drain {
298     struct rcu_head rcu;
299     QemuEvent drain_complete_event;
300 };
301 
302 static void drain_rcu_callback(struct rcu_head *node)
303 {
304     struct rcu_drain *event = (struct rcu_drain *)node;
305     qemu_event_set(&event->drain_complete_event);
306 }
307 
308 /*
309  * This function ensures that all pending RCU callbacks
310  * on the current thread are done executing
311 
312  * drops big qemu lock during the wait to allow RCU thread
313  * to process the callbacks
314  *
315  */
316 
317 void drain_call_rcu(void)
318 {
319     struct rcu_drain rcu_drain;
320     bool locked = qemu_mutex_iothread_locked();
321 
322     memset(&rcu_drain, 0, sizeof(struct rcu_drain));
323     qemu_event_init(&rcu_drain.drain_complete_event, false);
324 
325     if (locked) {
326         qemu_mutex_unlock_iothread();
327     }
328 
329 
330     /*
331      * RCU callbacks are invoked in the same order as in which they
332      * are registered, thus we can be sure that when 'drain_rcu_callback'
333      * is called, all RCU callbacks that were registered on this thread
334      * prior to calling this function are completed.
335      *
336      * Note that since we have only one global queue of the RCU callbacks,
337      * we also end up waiting for most of RCU callbacks that were registered
338      * on the other threads, but this is a side effect that shoudn't be
339      * assumed.
340      */
341 
342     call_rcu1(&rcu_drain.rcu, drain_rcu_callback);
343     qemu_event_wait(&rcu_drain.drain_complete_event);
344 
345     if (locked) {
346         qemu_mutex_lock_iothread();
347     }
348 
349 }
350 
351 void rcu_register_thread(void)
352 {
353     assert(rcu_reader.ctr == 0);
354     qemu_mutex_lock(&rcu_registry_lock);
355     QLIST_INSERT_HEAD(&registry, &rcu_reader, node);
356     qemu_mutex_unlock(&rcu_registry_lock);
357 }
358 
359 void rcu_unregister_thread(void)
360 {
361     qemu_mutex_lock(&rcu_registry_lock);
362     QLIST_REMOVE(&rcu_reader, node);
363     qemu_mutex_unlock(&rcu_registry_lock);
364 }
365 
366 static void rcu_init_complete(void)
367 {
368     QemuThread thread;
369 
370     qemu_mutex_init(&rcu_registry_lock);
371     qemu_mutex_init(&rcu_sync_lock);
372     qemu_event_init(&rcu_gp_event, true);
373 
374     qemu_event_init(&rcu_call_ready_event, false);
375 
376     /* The caller is assumed to have iothread lock, so the call_rcu thread
377      * must have been quiescent even after forking, just recreate it.
378      */
379     qemu_thread_create(&thread, "call_rcu", call_rcu_thread,
380                        NULL, QEMU_THREAD_DETACHED);
381 
382     rcu_register_thread();
383 }
384 
385 static int atfork_depth = 1;
386 
387 void rcu_enable_atfork(void)
388 {
389     atfork_depth++;
390 }
391 
392 void rcu_disable_atfork(void)
393 {
394     atfork_depth--;
395 }
396 
397 #ifdef CONFIG_POSIX
398 static void rcu_init_lock(void)
399 {
400     if (atfork_depth < 1) {
401         return;
402     }
403 
404     qemu_mutex_lock(&rcu_sync_lock);
405     qemu_mutex_lock(&rcu_registry_lock);
406 }
407 
408 static void rcu_init_unlock(void)
409 {
410     if (atfork_depth < 1) {
411         return;
412     }
413 
414     qemu_mutex_unlock(&rcu_registry_lock);
415     qemu_mutex_unlock(&rcu_sync_lock);
416 }
417 
418 static void rcu_init_child(void)
419 {
420     if (atfork_depth < 1) {
421         return;
422     }
423 
424     memset(&registry, 0, sizeof(registry));
425     rcu_init_complete();
426 }
427 #endif
428 
429 static void __attribute__((__constructor__)) rcu_init(void)
430 {
431     smp_mb_global_init();
432 #ifdef CONFIG_POSIX
433     pthread_atfork(rcu_init_lock, rcu_init_unlock, rcu_init_child);
434 #endif
435     rcu_init_complete();
436 }
437