util/rcu.c

   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-common.h"
  30 #include <stdio.h>
  31 #include <assert.h>
  32 #include <stdlib.h>
  33 #include <stdint.h>
  34 #include <errno.h>
  35 #include "qemu/rcu.h"
  36 #include "qemu/atomic.h"
  37 #include "qemu/thread.h"
  38 #include "qemu/main-loop.h"
  39
  40 /*
  41  * Global grace period counter.  Bit 0 is always one in rcu_gp_ctr.
  42  * Bits 1 and above are defined in synchronize_rcu.
  43  */
  44 #define RCU_GP_LOCKED           (1UL << 0)
  45 #define RCU_GP_CTR              (1UL << 1)
  46
  47 unsigned long rcu_gp_ctr = RCU_GP_LOCKED;
  48
  49 QemuEvent rcu_gp_event;
  50 static QemuMutex rcu_gp_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 = 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
  69 /* Protected by rcu_gp_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 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.
  88          * atomic_mb_set has a smp_wmb before...
  89          */
  90         smp_wmb();
  91         QLIST_FOREACH(index, &registry, node) {
  92             atomic_set(&index->waiting, true);
  93         }
  94
  95         /* ... and a smp_mb after.  */
  96         smp_mb();
  97
  98         QLIST_FOREACH_SAFE(index, &registry, node, tmp) {
  99             if (!rcu_gp_ongoing(&index->ctr)) {
 100                 QLIST_REMOVE(index, node);
 101                 QLIST_INSERT_HEAD(&qsreaders, index, node);
 102
 103                 /* No need for mb_set here, worst of all we
 104                  * get some extra futex wakeups.
 105                  */
 106                 atomic_set(&index->waiting, false);
 107             }
 108         }
 109
 110         /* atomic_mb_read has smp_rmb after.  */
 111         smp_rmb();
 112
 113         if (QLIST_EMPTY(&registry)) {
 114             break;
 115         }
 116
 117         /* Wait for one thread to report a quiescent state and
 118          * try again.
 119          */
 120         qemu_event_wait(&rcu_gp_event);
 121     }
 122
 123     /* put back the reader list in the registry */
 124     QLIST_SWAP(&registry, &qsreaders, node);
 125 }
 126
 127 void synchronize_rcu(void)
 128 {
 129     qemu_mutex_lock(&rcu_gp_lock);
 130
 131     if (!QLIST_EMPTY(&registry)) {
 132         /* In either case, the atomic_mb_set below blocks stores that free
 133          * old RCU-protected pointers.
 134          */
 135         if (sizeof(rcu_gp_ctr) < 8) {
 136             /* For architectures with 32-bit longs, a two-subphases algorithm
 137              * ensures we do not encounter overflow bugs.
 138              *
 139              * Switch parity: 0 -> 1, 1 -> 0.
 140              */
 141             atomic_mb_set(&rcu_gp_ctr, rcu_gp_ctr ^ RCU_GP_CTR);
 142             wait_for_readers();
 143             atomic_mb_set(&rcu_gp_ctr, rcu_gp_ctr ^ RCU_GP_CTR);
 144         } else {
 145             /* Increment current grace period.  */
 146             atomic_mb_set(&rcu_gp_ctr, rcu_gp_ctr + RCU_GP_CTR);
 147         }
 148
 149         wait_for_readers();
 150     }
 151
 152     qemu_mutex_unlock(&rcu_gp_lock);
 153 }
 154
 155
 156 #define RCU_CALL_MIN_SIZE        30
 157
 158 /* Multi-producer, single-consumer queue based on urcu/static/wfqueue.h
 159  * from liburcu.  Note that head is only used by the consumer.
 160  */
 161 static struct rcu_head dummy;
 162 static struct rcu_head *head = &dummy, **tail = &dummy.next;
 163 static int rcu_call_count;
 164 static QemuEvent rcu_call_ready_event;
 165
 166 static void enqueue(struct rcu_head *node)
 167 {
 168     struct rcu_head **old_tail;
 169
 170     node->next = NULL;
 171     old_tail = atomic_xchg(&tail, &node->next);
 172     atomic_mb_set(old_tail, node);
 173 }
 174
 175 static struct rcu_head *try_dequeue(void)
 176 {
 177     struct rcu_head *node, *next;
 178
 179 retry:
 180     /* Test for an empty list, which we do not expect.  Note that for
 181      * the consumer head and tail are always consistent.  The head
 182      * is consistent because only the consumer reads/writes it.
 183      * The tail, because it is the first step in the enqueuing.
 184      * It is only the next pointers that might be inconsistent.
 185      */
 186     if (head == &dummy && atomic_mb_read(&tail) == &dummy.next) {
 187         abort();
 188     }
 189
 190     /* If the head node has NULL in its next pointer, the value is
 191      * wrong and we need to wait until its enqueuer finishes the update.
 192      */
 193     node = head;
 194     next = atomic_mb_read(&head->next);
 195     if (!next) {
 196         return NULL;
 197     }
 198
 199     /* Since we are the sole consumer, and we excluded the empty case
 200      * above, the queue will always have at least two nodes: the
 201      * dummy node, and the one being removed.  So we do not need to update
 202      * the tail pointer.
 203      */
 204     head = next;
 205
 206     /* If we dequeued the dummy node, add it back at the end and retry.  */
 207     if (node == &dummy) {
 208         enqueue(node);
 209         goto retry;
 210     }
 211
 212     return node;
 213 }
 214
 215 static void *call_rcu_thread(void *opaque)
 216 {
 217     struct rcu_head *node;
 218
 219     for (;;) {
 220         int tries = 0;
 221         int n = atomic_read(&rcu_call_count);
 222
 223         /* Heuristically wait for a decent number of callbacks to pile up.
 224          * Fetch rcu_call_count now, we only must process elements that were
 225          * added before synchronize_rcu() starts.
 226          */
 227         while (n == 0 || (n < RCU_CALL_MIN_SIZE && ++tries <= 5)) {
 228             g_usleep(10000);
 229             if (n == 0) {
 230                 qemu_event_reset(&rcu_call_ready_event);
 231                 n = atomic_read(&rcu_call_count);
 232                 if (n == 0) {
 233                     qemu_event_wait(&rcu_call_ready_event);
 234                 }
 235             }
 236             n = atomic_read(&rcu_call_count);
 237         }
 238
 239         atomic_sub(&rcu_call_count, n);
 240         synchronize_rcu();
 241         qemu_mutex_lock_iothread();
 242         while (n > 0) {
 243             node = try_dequeue();
 244             while (!node) {
 245                 qemu_mutex_unlock_iothread();
 246                 qemu_event_reset(&rcu_call_ready_event);
 247                 node = try_dequeue();
 248                 if (!node) {
 249                     qemu_event_wait(&rcu_call_ready_event);
 250                     node = try_dequeue();
 251                 }
 252                 qemu_mutex_lock_iothread();
 253             }
 254
 255             n--;
 256             node->func(node);
 257         }
 258         qemu_mutex_unlock_iothread();
 259     }
 260     abort();
 261 }
 262
 263 void call_rcu1(struct rcu_head *node, void (*func)(struct rcu_head *node))
 264 {
 265     node->func = func;
 266     enqueue(node);
 267     atomic_inc(&rcu_call_count);
 268     qemu_event_set(&rcu_call_ready_event);
 269 }
 270
 271 void rcu_register_thread(void)
 272 {
 273     assert(rcu_reader.ctr == 0);
 274     qemu_mutex_lock(&rcu_gp_lock);
 275     QLIST_INSERT_HEAD(&registry, &rcu_reader, node);
 276     qemu_mutex_unlock(&rcu_gp_lock);
 277 }
 278
 279 void rcu_unregister_thread(void)
 280 {
 281     qemu_mutex_lock(&rcu_gp_lock);
 282     QLIST_REMOVE(&rcu_reader, node);
 283     qemu_mutex_unlock(&rcu_gp_lock);
 284 }
 285
 286 static void __attribute__((__constructor__)) rcu_init(void)
 287 {
 288     QemuThread thread;
 289
 290     qemu_mutex_init(&rcu_gp_lock);
 291     qemu_event_init(&rcu_gp_event, true);
 292
 293     qemu_event_init(&rcu_call_ready_event, false);
 294     qemu_thread_create(&thread, "call_rcu", call_rcu_thread,
 295                        NULL, QEMU_THREAD_DETACHED);
 296
 297     rcu_register_thread();
 298 }