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/osdep.h"
  30 #include "qemu-common.h"
  31 #include "qemu/rcu.h"
  32 #include "qemu/atomic.h"
  33 #include "qemu/thread.h"
  34 #include "qemu/main-loop.h"
  35
  36 /*
  37  * Global grace period counter.  Bit 0 is always one in rcu_gp_ctr.
  38  * Bits 1 and above are defined in synchronize_rcu.
  39  */
  40 #define RCU_GP_LOCKED           (1UL << 0)
  41 #define RCU_GP_CTR              (1UL << 1)
  42
  43 unsigned long rcu_gp_ctr = RCU_GP_LOCKED;
  44
  45 QemuEvent rcu_gp_event;
  46 static QemuMutex rcu_registry_lock;
  47 static QemuMutex rcu_sync_lock;
  48
  49 /*
  50  * Check whether a quiescent state was crossed between the beginning of
  51  * update_counter_and_wait and now.
  52  */
  53 static inline int rcu_gp_ongoing(unsigned long *ctr)
  54 {
  55     unsigned long v;
  56
  57     v = atomic_read(ctr);
  58     return v && (v != rcu_gp_ctr);
  59 }
  60
  61 /* Written to only by each individual reader. Read by both the reader and the
  62  * writers.
  63  */
  64 __thread struct rcu_reader_data rcu_reader;
  65
  66 /* Protected by rcu_registry_lock.  */
  67 typedef QLIST_HEAD(, rcu_reader_data) ThreadList;
  68 static ThreadList registry = QLIST_HEAD_INITIALIZER(registry);
  69
  70 /* Wait for previous parity/grace period to be empty of readers.  */
  71 static void wait_for_readers(void)
  72 {
  73     ThreadList qsreaders = QLIST_HEAD_INITIALIZER(qsreaders);
  74     struct rcu_reader_data *index, *tmp;
  75
  76     for (;;) {
  77         /* We want to be notified of changes made to rcu_gp_ongoing
  78          * while we walk the list.
  79          */
  80         qemu_event_reset(&rcu_gp_event);
  81
  82         /* Instead of using atomic_mb_set for index->waiting, and
  83          * atomic_mb_read for index->ctr, memory barriers are placed
  84          * manually since writes to different threads are independent.
  85          * qemu_event_reset has acquire semantics, so no memory barrier
  86          * is needed here.
  87          */
  88         QLIST_FOREACH(index, &registry, node) {
  89             atomic_set(&index->waiting, true);
  90         }
  91
  92         /* Here, order the stores to index->waiting before the
  93          * loads of index->ctr.
  94          */
  95         smp_mb();
  96
  97         QLIST_FOREACH_SAFE(index, &registry, node, tmp) {
  98             if (!rcu_gp_ongoing(&index->ctr)) {
  99                 QLIST_REMOVE(index, node);
 100                 QLIST_INSERT_HEAD(&qsreaders, index, node);
 101
 102                 /* No need for mb_set here, worst of all we
 103                  * get some extra futex wakeups.
 104                  */
 105                 atomic_set(&index->waiting, false);
 106             }
 107         }
 108
 109         if (QLIST_EMPTY(&registry)) {
 110             break;
 111         }
 112
 113         /* Wait for one thread to report a quiescent state and try again.
 114          * Release rcu_registry_lock, so rcu_(un)register_thread() doesn't
 115          * wait too much time.
 116          *
 117          * rcu_register_thread() may add nodes to &registry; it will not
 118          * wake up synchronize_rcu, but that is okay because at least another
 119          * thread must exit its RCU read-side critical section before
 120          * synchronize_rcu is done.  The next iteration of the loop will
 121          * move the new thread's rcu_reader from &registry to &qsreaders,
 122          * because rcu_gp_ongoing() will return false.
 123          *
 124          * rcu_unregister_thread() may remove nodes from &qsreaders instead
 125          * of &registry if it runs during qemu_event_wait.  That's okay;
 126          * the node then will not be added back to &registry by QLIST_SWAP
 127          * below.  The invariant is that the node is part of one list when
 128          * rcu_registry_lock is released.
 129          */
 130         qemu_mutex_unlock(&rcu_registry_lock);
 131         qemu_event_wait(&rcu_gp_event);
 132         qemu_mutex_lock(&rcu_registry_lock);
 133     }
 134
 135     /* put back the reader list in the registry */
 136     QLIST_SWAP(&registry, &qsreaders, node);
 137 }
 138
 139 void synchronize_rcu(void)
 140 {
 141     qemu_mutex_lock(&rcu_sync_lock);
 142     qemu_mutex_lock(&rcu_registry_lock);
 143
 144     if (!QLIST_EMPTY(&registry)) {
 145         /* In either case, the atomic_mb_set below blocks stores that free
 146          * old RCU-protected pointers.
 147          */
 148         if (sizeof(rcu_gp_ctr) < 8) {
 149             /* For architectures with 32-bit longs, a two-subphases algorithm
 150              * ensures we do not encounter overflow bugs.
 151              *
 152              * Switch parity: 0 -> 1, 1 -> 0.
 153              */
 154             atomic_mb_set(&rcu_gp_ctr, rcu_gp_ctr ^ RCU_GP_CTR);
 155             wait_for_readers();
 156             atomic_mb_set(&rcu_gp_ctr, rcu_gp_ctr ^ RCU_GP_CTR);
 157         } else {
 158             /* Increment current grace period.  */
 159             atomic_mb_set(&rcu_gp_ctr, rcu_gp_ctr + RCU_GP_CTR);
 160         }
 161
 162         wait_for_readers();
 163     }
 164
 165     qemu_mutex_unlock(&rcu_registry_lock);
 166     qemu_mutex_unlock(&rcu_sync_lock);
 167 }
 168
 169
 170 #define RCU_CALL_MIN_SIZE        30
 171
 172 /* Multi-producer, single-consumer queue based on urcu/static/wfqueue.h
 173  * from liburcu.  Note that head is only used by the consumer.
 174  */
 175 static struct rcu_head dummy;
 176 static struct rcu_head *head = &dummy, **tail = &dummy.next;
 177 static int rcu_call_count;
 178 static QemuEvent rcu_call_ready_event;
 179
 180 static void enqueue(struct rcu_head *node)
 181 {
 182     struct rcu_head **old_tail;
 183
 184     node->next = NULL;
 185     old_tail = atomic_xchg(&tail, &node->next);
 186     atomic_mb_set(old_tail, node);
 187 }
 188
 189 static struct rcu_head *try_dequeue(void)
 190 {
 191     struct rcu_head *node, *next;
 192
 193 retry:
 194     /* Test for an empty list, which we do not expect.  Note that for
 195      * the consumer head and tail are always consistent.  The head
 196      * is consistent because only the consumer reads/writes it.
 197      * The tail, because it is the first step in the enqueuing.
 198      * It is only the next pointers that might be inconsistent.
 199      */
 200     if (head == &dummy && atomic_mb_read(&tail) == &dummy.next) {
 201         abort();
 202     }
 203
 204     /* If the head node has NULL in its next pointer, the value is
 205      * wrong and we need to wait until its enqueuer finishes the update.
 206      */
 207     node = head;
 208     next = atomic_mb_read(&head->next);
 209     if (!next) {
 210         return NULL;
 211     }
 212
 213     /* Since we are the sole consumer, and we excluded the empty case
 214      * above, the queue will always have at least two nodes: the
 215      * dummy node, and the one being removed.  So we do not need to update
 216      * the tail pointer.
 217      */
 218     head = next;
 219
 220     /* If we dequeued the dummy node, add it back at the end and retry.  */
 221     if (node == &dummy) {
 222         enqueue(node);
 223         goto retry;
 224     }
 225
 226     return node;
 227 }
 228
 229 static void *call_rcu_thread(void *opaque)
 230 {
 231     struct rcu_head *node;
 232
 233     rcu_register_thread();
 234
 235     for (;;) {
 236         int tries = 0;
 237         int n = atomic_read(&rcu_call_count);
 238
 239         /* Heuristically wait for a decent number of callbacks to pile up.
 240          * Fetch rcu_call_count now, we only must process elements that were
 241          * added before synchronize_rcu() starts.
 242          */
 243         while (n == 0 || (n < RCU_CALL_MIN_SIZE && ++tries <= 5)) {
 244             g_usleep(10000);
 245             if (n == 0) {
 246                 qemu_event_reset(&rcu_call_ready_event);
 247                 n = atomic_read(&rcu_call_count);
 248                 if (n == 0) {
 249                     qemu_event_wait(&rcu_call_ready_event);
 250                 }
 251             }
 252             n = atomic_read(&rcu_call_count);
 253         }
 254
 255         atomic_sub(&rcu_call_count, n);
 256         synchronize_rcu();
 257         qemu_mutex_lock_iothread();
 258         while (n > 0) {
 259             node = try_dequeue();
 260             while (!node) {
 261                 qemu_mutex_unlock_iothread();
 262                 qemu_event_reset(&rcu_call_ready_event);
 263                 node = try_dequeue();
 264                 if (!node) {
 265                     qemu_event_wait(&rcu_call_ready_event);
 266                     node = try_dequeue();
 267                 }
 268                 qemu_mutex_lock_iothread();
 269             }
 270
 271             n--;
 272             node->func(node);
 273         }
 274         qemu_mutex_unlock_iothread();
 275     }
 276     abort();
 277 }
 278
 279 void call_rcu1(struct rcu_head *node, void (*func)(struct rcu_head *node))
 280 {
 281     node->func = func;
 282     enqueue(node);
 283     atomic_inc(&rcu_call_count);
 284     qemu_event_set(&rcu_call_ready_event);
 285 }
 286
 287 void rcu_register_thread(void)
 288 {
 289     assert(rcu_reader.ctr == 0);
 290     qemu_mutex_lock(&rcu_registry_lock);
 291     QLIST_INSERT_HEAD(&registry, &rcu_reader, node);
 292     qemu_mutex_unlock(&rcu_registry_lock);
 293 }
 294
 295 void rcu_unregister_thread(void)
 296 {
 297     qemu_mutex_lock(&rcu_registry_lock);
 298     QLIST_REMOVE(&rcu_reader, node);
 299     qemu_mutex_unlock(&rcu_registry_lock);
 300 }
 301
 302 static void rcu_init_complete(void)
 303 {
 304     QemuThread thread;
 305
 306     qemu_mutex_init(&rcu_registry_lock);
 307     qemu_mutex_init(&rcu_sync_lock);
 308     qemu_event_init(&rcu_gp_event, true);
 309
 310     qemu_event_init(&rcu_call_ready_event, false);
 311
 312     /* The caller is assumed to have iothread lock, so the call_rcu thread
 313      * must have been quiescent even after forking, just recreate it.
 314      */
 315     qemu_thread_create(&thread, "call_rcu", call_rcu_thread,
 316                        NULL, QEMU_THREAD_DETACHED);
 317
 318     rcu_register_thread();
 319 }
 320
 321 #ifdef CONFIG_POSIX
 322 static void rcu_init_lock(void)
 323 {
 324     qemu_mutex_lock(&rcu_sync_lock);
 325     qemu_mutex_lock(&rcu_registry_lock);
 326 }
 327
 328 static void rcu_init_unlock(void)
 329 {
 330     qemu_mutex_unlock(&rcu_registry_lock);
 331     qemu_mutex_unlock(&rcu_sync_lock);
 332 }
 333 #endif
 334
 335 void rcu_after_fork(void)
 336 {
 337     memset(&registry, 0, sizeof(registry));
 338     rcu_init_complete();
 339 }
 340
 341 static void __attribute__((__constructor__)) rcu_init(void)
 342 {
 343 #ifdef CONFIG_POSIX
 344     pthread_atfork(rcu_init_lock, rcu_init_unlock, rcu_init_unlock);
 345 #endif
 346     rcu_init_complete();
 347 }