blob: 60a37f72c36cb45767ec7c71f28f8a8ed4e3a2e2 [file] [log] [blame]
Paolo Bonzini79117472013-05-13 13:29:47 +02001/*
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
Peter Maydellaafd7582016-01-29 17:49:55 +000029#include "qemu/osdep.h"
Paolo Bonzini79117472013-05-13 13:29:47 +020030#include "qemu/rcu.h"
31#include "qemu/atomic.h"
Paolo Bonzini26387f82013-05-13 17:49:24 +020032#include "qemu/thread.h"
Paolo Bonzinia4649822015-02-11 17:15:18 +010033#include "qemu/main-loop.h"
Daniel Brodsky6e8a3552020-04-03 21:21:08 -070034#include "qemu/lockable.h"
Yang Zhong5a22ab72017-12-20 21:16:46 +080035#if defined(CONFIG_MALLOC_TRIM)
36#include <malloc.h>
37#endif
Paolo Bonzini79117472013-05-13 13:29:47 +020038
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
46unsigned long rcu_gp_ctr = RCU_GP_LOCKED;
47
48QemuEvent rcu_gp_event;
Wen Congyangc097a602015-07-27 10:24:18 +080049static QemuMutex rcu_registry_lock;
50static QemuMutex rcu_sync_lock;
Paolo Bonzini79117472013-05-13 13:29:47 +020051
52/*
53 * Check whether a quiescent state was crossed between the beginning of
54 * update_counter_and_wait and now.
55 */
56static inline int rcu_gp_ongoing(unsigned long *ctr)
57{
58 unsigned long v;
59
60 v = atomic_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
Wen Congyangc097a602015-07-27 10:24:18 +080069/* Protected by rcu_registry_lock. */
Paolo Bonzini79117472013-05-13 13:29:47 +020070typedef QLIST_HEAD(, rcu_reader_data) ThreadList;
71static ThreadList registry = QLIST_HEAD_INITIALIZER(registry);
72
73/* Wait for previous parity/grace period to be empty of readers. */
74static 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 atomic_mb_set for index->waiting, and
86 * atomic_mb_read for index->ctr, memory barriers are placed
87 * manually since writes to different threads are independent.
Paolo Bonzinie11131b2016-09-19 11:27:46 +020088 * qemu_event_reset has acquire semantics, so no memory barrier
89 * is needed here.
Paolo Bonzini79117472013-05-13 13:29:47 +020090 */
Paolo Bonzini79117472013-05-13 13:29:47 +020091 QLIST_FOREACH(index, &registry, node) {
92 atomic_set(&index->waiting, true);
93 }
94
Paolo Bonzini77a8b842018-02-16 09:23:31 +010095 /* Here, order the stores to index->waiting before the loads of
Paolo Bonzinic8d38772018-02-16 10:04:18 +010096 * index->ctr. Pairs with smp_mb_placeholder() in rcu_read_unlock(),
Paolo Bonzini77a8b842018-02-16 09:23:31 +010097 * ensuring that the loads of index->ctr are sequentially consistent.
Paolo Bonzinie11131b2016-09-19 11:27:46 +020098 */
Paolo Bonzinic8d38772018-02-16 10:04:18 +010099 smp_mb_global();
Paolo Bonzini79117472013-05-13 13:29:47 +0200100
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 atomic_set(&index->waiting, false);
110 }
111 }
112
Paolo Bonzini79117472013-05-13 13:29:47 +0200113 if (QLIST_EMPTY(&registry)) {
114 break;
115 }
116
Wen Congyangc097a602015-07-27 10:24:18 +0800117 /* 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.
Paolo Bonzini79117472013-05-13 13:29:47 +0200133 */
Wen Congyangc097a602015-07-27 10:24:18 +0800134 qemu_mutex_unlock(&rcu_registry_lock);
Paolo Bonzini79117472013-05-13 13:29:47 +0200135 qemu_event_wait(&rcu_gp_event);
Wen Congyangc097a602015-07-27 10:24:18 +0800136 qemu_mutex_lock(&rcu_registry_lock);
Paolo Bonzini79117472013-05-13 13:29:47 +0200137 }
138
139 /* put back the reader list in the registry */
140 QLIST_SWAP(&registry, &qsreaders, node);
141}
142
143void synchronize_rcu(void)
144{
Daniel Brodsky6e8a3552020-04-03 21:21:08 -0700145 QEMU_LOCK_GUARD(&rcu_sync_lock);
Paolo Bonzini79117472013-05-13 13:29:47 +0200146
Paolo Bonzini77a8b842018-02-16 09:23:31 +0100147 /* Write RCU-protected pointers before reading p_rcu_reader->ctr.
Paolo Bonzinic8d38772018-02-16 10:04:18 +0100148 * Pairs with smp_mb_placeholder() in rcu_read_lock().
Paolo Bonzini77a8b842018-02-16 09:23:31 +0100149 */
Paolo Bonzinic8d38772018-02-16 10:04:18 +0100150 smp_mb_global();
Paolo Bonzini77a8b842018-02-16 09:23:31 +0100151
Daniel Brodsky6e8a3552020-04-03 21:21:08 -0700152 QEMU_LOCK_GUARD(&rcu_registry_lock);
Paolo Bonzini79117472013-05-13 13:29:47 +0200153 if (!QLIST_EMPTY(&registry)) {
154 /* In either case, the atomic_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 atomic_mb_set(&rcu_gp_ctr, rcu_gp_ctr ^ RCU_GP_CTR);
164 wait_for_readers();
165 atomic_mb_set(&rcu_gp_ctr, rcu_gp_ctr ^ RCU_GP_CTR);
166 } else {
167 /* Increment current grace period. */
168 atomic_mb_set(&rcu_gp_ctr, rcu_gp_ctr + RCU_GP_CTR);
169 }
170
171 wait_for_readers();
172 }
Paolo Bonzini79117472013-05-13 13:29:47 +0200173}
174
Paolo Bonzini26387f82013-05-13 17:49:24 +0200175
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 */
181static struct rcu_head dummy;
182static struct rcu_head *head = &dummy, **tail = &dummy.next;
183static int rcu_call_count;
184static QemuEvent rcu_call_ready_event;
185
186static void enqueue(struct rcu_head *node)
187{
188 struct rcu_head **old_tail;
189
190 node->next = NULL;
191 old_tail = atomic_xchg(&tail, &node->next);
192 atomic_mb_set(old_tail, node);
193}
194
195static struct rcu_head *try_dequeue(void)
196{
197 struct rcu_head *node, *next;
198
199retry:
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 && atomic_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 = atomic_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
235static void *call_rcu_thread(void *opaque)
236{
237 struct rcu_head *node;
238
Paolo Bonziniab28bd22015-07-09 08:55:38 +0200239 rcu_register_thread();
240
Paolo Bonzini26387f82013-05-13 17:49:24 +0200241 for (;;) {
242 int tries = 0;
243 int n = atomic_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 */
Paolo Bonzinia7d1d632015-02-11 15:51:54 +0100249 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);
Paolo Bonzini26387f82013-05-13 17:49:24 +0200253 n = atomic_read(&rcu_call_count);
Paolo Bonzinia7d1d632015-02-11 15:51:54 +0100254 if (n == 0) {
Yang Zhong5a22ab72017-12-20 21:16:46 +0800255#if defined(CONFIG_MALLOC_TRIM)
256 malloc_trim(4 * 1024 * 1024);
257#endif
Paolo Bonzinia7d1d632015-02-11 15:51:54 +0100258 qemu_event_wait(&rcu_call_ready_event);
259 }
Paolo Bonzini26387f82013-05-13 17:49:24 +0200260 }
Paolo Bonzinia7d1d632015-02-11 15:51:54 +0100261 n = atomic_read(&rcu_call_count);
Paolo Bonzini26387f82013-05-13 17:49:24 +0200262 }
263
264 atomic_sub(&rcu_call_count, n);
265 synchronize_rcu();
Paolo Bonzinia4649822015-02-11 17:15:18 +0100266 qemu_mutex_lock_iothread();
Paolo Bonzini26387f82013-05-13 17:49:24 +0200267 while (n > 0) {
268 node = try_dequeue();
269 while (!node) {
Paolo Bonzinia4649822015-02-11 17:15:18 +0100270 qemu_mutex_unlock_iothread();
Paolo Bonzini26387f82013-05-13 17:49:24 +0200271 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 }
Paolo Bonzinia4649822015-02-11 17:15:18 +0100277 qemu_mutex_lock_iothread();
Paolo Bonzini26387f82013-05-13 17:49:24 +0200278 }
279
280 n--;
281 node->func(node);
282 }
Paolo Bonzinia4649822015-02-11 17:15:18 +0100283 qemu_mutex_unlock_iothread();
Paolo Bonzini26387f82013-05-13 17:49:24 +0200284 }
285 abort();
286}
287
288void call_rcu1(struct rcu_head *node, void (*func)(struct rcu_head *node))
289{
290 node->func = func;
291 enqueue(node);
292 atomic_inc(&rcu_call_count);
293 qemu_event_set(&rcu_call_ready_event);
294}
295
Paolo Bonzini79117472013-05-13 13:29:47 +0200296void rcu_register_thread(void)
297{
298 assert(rcu_reader.ctr == 0);
Wen Congyangc097a602015-07-27 10:24:18 +0800299 qemu_mutex_lock(&rcu_registry_lock);
Paolo Bonzini79117472013-05-13 13:29:47 +0200300 QLIST_INSERT_HEAD(&registry, &rcu_reader, node);
Wen Congyangc097a602015-07-27 10:24:18 +0800301 qemu_mutex_unlock(&rcu_registry_lock);
Paolo Bonzini79117472013-05-13 13:29:47 +0200302}
303
304void rcu_unregister_thread(void)
305{
Wen Congyangc097a602015-07-27 10:24:18 +0800306 qemu_mutex_lock(&rcu_registry_lock);
Paolo Bonzini79117472013-05-13 13:29:47 +0200307 QLIST_REMOVE(&rcu_reader, node);
Wen Congyangc097a602015-07-27 10:24:18 +0800308 qemu_mutex_unlock(&rcu_registry_lock);
Paolo Bonzini79117472013-05-13 13:29:47 +0200309}
310
Paolo Bonzini21b7cf92015-03-05 16:53:48 +0100311static void rcu_init_complete(void)
Paolo Bonzini79117472013-05-13 13:29:47 +0200312{
Paolo Bonzini26387f82013-05-13 17:49:24 +0200313 QemuThread thread;
314
Wen Congyangc097a602015-07-27 10:24:18 +0800315 qemu_mutex_init(&rcu_registry_lock);
316 qemu_mutex_init(&rcu_sync_lock);
Paolo Bonzini79117472013-05-13 13:29:47 +0200317 qemu_event_init(&rcu_gp_event, true);
Paolo Bonzini26387f82013-05-13 17:49:24 +0200318
319 qemu_event_init(&rcu_call_ready_event, false);
Paolo Bonzini21b7cf92015-03-05 16:53:48 +0100320
321 /* The caller is assumed to have iothread lock, so the call_rcu thread
322 * must have been quiescent even after forking, just recreate it.
323 */
Paolo Bonzini26387f82013-05-13 17:49:24 +0200324 qemu_thread_create(&thread, "call_rcu", call_rcu_thread,
325 NULL, QEMU_THREAD_DETACHED);
326
Paolo Bonzini79117472013-05-13 13:29:47 +0200327 rcu_register_thread();
328}
Paolo Bonzini21b7cf92015-03-05 16:53:48 +0100329
Paolo Bonzini73c6e402016-01-27 08:49:21 +0100330static int atfork_depth = 1;
331
332void rcu_enable_atfork(void)
333{
334 atfork_depth++;
335}
336
337void rcu_disable_atfork(void)
338{
339 atfork_depth--;
340}
341
Paolo Bonzini21b7cf92015-03-05 16:53:48 +0100342#ifdef CONFIG_POSIX
343static void rcu_init_lock(void)
344{
Paolo Bonzini73c6e402016-01-27 08:49:21 +0100345 if (atfork_depth < 1) {
346 return;
347 }
348
Wen Congyangc097a602015-07-27 10:24:18 +0800349 qemu_mutex_lock(&rcu_sync_lock);
350 qemu_mutex_lock(&rcu_registry_lock);
Paolo Bonzini21b7cf92015-03-05 16:53:48 +0100351}
352
353static void rcu_init_unlock(void)
354{
Paolo Bonzini73c6e402016-01-27 08:49:21 +0100355 if (atfork_depth < 1) {
356 return;
357 }
358
Wen Congyangc097a602015-07-27 10:24:18 +0800359 qemu_mutex_unlock(&rcu_registry_lock);
360 qemu_mutex_unlock(&rcu_sync_lock);
Paolo Bonzini21b7cf92015-03-05 16:53:48 +0100361}
362
Paolo Bonzini2a96a552016-03-25 14:00:51 +0100363static void rcu_init_child(void)
Paolo Bonzini21b7cf92015-03-05 16:53:48 +0100364{
Paolo Bonzini2a96a552016-03-25 14:00:51 +0100365 if (atfork_depth < 1) {
366 return;
367 }
368
Paolo Bonzini21b7cf92015-03-05 16:53:48 +0100369 memset(&registry, 0, sizeof(registry));
370 rcu_init_complete();
371}
Paolo Bonzini2a96a552016-03-25 14:00:51 +0100372#endif
Paolo Bonzini21b7cf92015-03-05 16:53:48 +0100373
374static void __attribute__((__constructor__)) rcu_init(void)
375{
Paolo Bonzinic8d38772018-02-16 10:04:18 +0100376 smp_mb_global_init();
Paolo Bonzini21b7cf92015-03-05 16:53:48 +0100377#ifdef CONFIG_POSIX
Paolo Bonzini2a96a552016-03-25 14:00:51 +0100378 pthread_atfork(rcu_init_lock, rcu_init_unlock, rcu_init_child);
Paolo Bonzini21b7cf92015-03-05 16:53:48 +0100379#endif
380 rcu_init_complete();
381}