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/rcu.h"
  31 #include "qemu/atomic.h"
  32 #include "qemu/thread.h"
  33 #include "qemu/main-loop.h"
  34 #if defined(CONFIG_MALLOC_TRIM)
  35 #include <malloc.h>
  36 #endif
  37
  38 /*
  39  * Global grace period counter.  Bit 0 is always one in rcu_gp_ctr.
  40  * Bits 1 and above are defined in synchronize_rcu.
  41  */
  42 #define RCU_GP_LOCKED           (1UL << 0)
  43 #define RCU_GP_CTR              (1UL << 1)
  44
  45 unsigned long rcu_gp_ctr = RCU_GP_LOCKED;
  46
  47 QemuEvent rcu_gp_event;
  48 static QemuMutex rcu_registry_lock;
  49 static QemuMutex rcu_sync_lock;
  50
  51 /*
  52  * Check whether a quiescent state was crossed between the beginning of
  53  * update_counter_and_wait and now.
  54  */
  55 static inline int rcu_gp_ongoing(unsigned long *ctr)
  56 {
  57     unsigned long v;
  58
  59     v = atomic_read(ctr);
  60     return v && (v != rcu_gp_ctr);
  61 }
  62
  63 /* Written to only by each individual reader. Read by both the reader and the
  64  * writers.
  65  */
  66 __thread struct rcu_reader_data rcu_reader;
  67
  68 /* Protected by rcu_registry_lock.  */
  69 typedef QLIST_HEAD(, rcu_reader_data) ThreadList;
  70 static ThreadList registry = QLIST_HEAD_INITIALIZER(registry);
  71
  72 /* Wait for previous parity/grace period to be empty of readers.  */
  73 static void wait_for_readers(void)
  74 {
  75     ThreadList qsreaders = QLIST_HEAD_INITIALIZER(qsreaders);
  76     struct rcu_reader_data *index, *tmp;
  77
  78     for (;;) {
  79         /* We want to be notified of changes made to rcu_gp_ongoing
  80          * while we walk the list.
  81          */
  82         qemu_event_reset(&rcu_gp_event);
  83
  84         /* Instead of using atomic_mb_set for index->waiting, and
  85          * atomic_mb_read for index->ctr, memory barriers are placed
  86          * manually since writes to different threads are independent.
  87          * qemu_event_reset has acquire semantics, so no memory barrier
  88          * is needed here.
  89          */
  90         QLIST_FOREACH(index, &registry, node) {
  91             atomic_set(&index->waiting, true);
  92         }
  93
  94         /* Here, order the stores to index->waiting before the loads of
  95          * index->ctr.  Pairs with smp_mb_placeholder() in rcu_read_unlock(),
  96          * ensuring that the loads of index->ctr are sequentially consistent.
  97          */
  98         smp_mb_global();
  99
 100         QLIST_FOREACH_SAFE(index, &registry, node, tmp) {
 101             if (!rcu_gp_ongoing(&index->ctr)) {
 102                 QLIST_REMOVE(index, node);
 103                 QLIST_INSERT_HEAD(&qsreaders, index, node);
 104
 105                 /* No need for mb_set here, worst of all we
 106                  * get some extra futex wakeups.
 107                  */
 108                 atomic_set(&index->waiting, false);
 109             }
 110         }
 111
 112         if (QLIST_EMPTY(&registry)) {
 113             break;
 114         }
 115
 116         /* Wait for one thread to report a quiescent state and try again.
 117          * Release rcu_registry_lock, so rcu_(un)register_thread() doesn't
 118          * wait too much time.
 119          *
 120          * rcu_register_thread() may add nodes to &registry; it will not
 121          * wake up synchronize_rcu, but that is okay because at least another
 122          * thread must exit its RCU read-side critical section before
 123          * synchronize_rcu is done.  The next iteration of the loop will
 124          * move the new thread's rcu_reader from &registry to &qsreaders,
 125          * because rcu_gp_ongoing() will return false.
 126          *
 127          * rcu_unregister_thread() may remove nodes from &qsreaders instead
 128          * of &registry if it runs during qemu_event_wait.  That's okay;
 129          * the node then will not be added back to &registry by QLIST_SWAP
 130          * below.  The invariant is that the node is part of one list when
 131          * rcu_registry_lock is released.
 132          */
 133         qemu_mutex_unlock(&rcu_registry_lock);
 134         qemu_event_wait(&rcu_gp_event);
 135         qemu_mutex_lock(&rcu_registry_lock);
 136     }
 137
 138     /* put back the reader list in the registry */
 139     QLIST_SWAP(&registry, &qsreaders, node);
 140 }
 141
 142 void synchronize_rcu(void)
 143 {
 144     qemu_mutex_lock(&rcu_sync_lock);
 145
 146     /* Write RCU-protected pointers before reading p_rcu_reader->ctr.
 147      * Pairs with smp_mb_placeholder() in rcu_read_lock().
 148      */
 149     smp_mb_global();
 150
 151     qemu_mutex_lock(&rcu_registry_lock);
 152     if (!QLIST_EMPTY(&registry)) {
 153         /* In either case, the atomic_mb_set below blocks stores that free
 154          * old RCU-protected pointers.
 155          */
 156         if (sizeof(rcu_gp_ctr) < 8) {
 157             /* For architectures with 32-bit longs, a two-subphases algorithm
 158              * ensures we do not encounter overflow bugs.
 159              *
 160              * Switch parity: 0 -> 1, 1 -> 0.
 161              */
 162             atomic_mb_set(&rcu_gp_ctr, rcu_gp_ctr ^ RCU_GP_CTR);
 163             wait_for_readers();
 164             atomic_mb_set(&rcu_gp_ctr, rcu_gp_ctr ^ RCU_GP_CTR);
 165         } else {
 166             /* Increment current grace period.  */
 167             atomic_mb_set(&rcu_gp_ctr, rcu_gp_ctr + RCU_GP_CTR);
 168         }
 169
 170         wait_for_readers();
 171     }
 172
 173     qemu_mutex_unlock(&rcu_registry_lock);
 174     qemu_mutex_unlock(&rcu_sync_lock);
 175 }
 176
 177
 178 #define RCU_CALL_MIN_SIZE        30
 179
 180 /* Multi-producer, single-consumer queue based on urcu/static/wfqueue.h
 181  * from liburcu.  Note that head is only used by the consumer.
 182  */
 183 static struct rcu_head dummy;
 184 static struct rcu_head *head = &dummy, **tail = &dummy.next;
 185 static int rcu_call_count;
 186 static QemuEvent rcu_call_ready_event;
 187
 188 static void enqueue(struct rcu_head *node)
 189 {
 190     struct rcu_head **old_tail;
 191
 192     node->next = NULL;
 193     old_tail = atomic_xchg(&tail, &node->next);
 194     atomic_mb_set(old_tail, node);
 195 }
 196
 197 static struct rcu_head *try_dequeue(void)
 198 {
 199     struct rcu_head *node, *next;
 200
 201 retry:
 202     /* Test for an empty list, which we do not expect.  Note that for
 203      * the consumer head and tail are always consistent.  The head
 204      * is consistent because only the consumer reads/writes it.
 205      * The tail, because it is the first step in the enqueuing.
 206      * It is only the next pointers that might be inconsistent.
 207      */
 208     if (head == &dummy && atomic_mb_read(&tail) == &dummy.next) {
 209         abort();
 210     }
 211
 212     /* If the head node has NULL in its next pointer, the value is
 213      * wrong and we need to wait until its enqueuer finishes the update.
 214      */
 215     node = head;
 216     next = atomic_mb_read(&head->next);
 217     if (!next) {
 218         return NULL;
 219     }
 220
 221     /* Since we are the sole consumer, and we excluded the empty case
 222      * above, the queue will always have at least two nodes: the
 223      * dummy node, and the one being removed.  So we do not need to update
 224      * the tail pointer.
 225      */
 226     head = next;
 227
 228     /* If we dequeued the dummy node, add it back at the end and retry.  */
 229     if (node == &dummy) {
 230         enqueue(node);
 231         goto retry;
 232     }
 233
 234     return node;
 235 }
 236
 237 static void *call_rcu_thread(void *opaque)
 238 {
 239     struct rcu_head *node;
 240
 241     rcu_register_thread();
 242
 243     for (;;) {
 244         int tries = 0;
 245         int n = atomic_read(&rcu_call_count);
 246
 247         /* Heuristically wait for a decent number of callbacks to pile up.
 248          * Fetch rcu_call_count now, we only must process elements that were
 249          * added before synchronize_rcu() starts.
 250          */
 251         while (n == 0 || (n < RCU_CALL_MIN_SIZE && ++tries <= 5)) {
 252             g_usleep(10000);
 253             if (n == 0) {
 254                 qemu_event_reset(&rcu_call_ready_event);
 255                 n = atomic_read(&rcu_call_count);
 256                 if (n == 0) {
 257 #if defined(CONFIG_MALLOC_TRIM)
 258                     malloc_trim(4 * 1024 * 1024);
 259 #endif
 260                     qemu_event_wait(&rcu_call_ready_event);
 261                 }
 262             }
 263             n = atomic_read(&rcu_call_count);
 264         }
 265
 266         atomic_sub(&rcu_call_count, n);
 267         synchronize_rcu();
 268         qemu_mutex_lock_iothread();
 269         while (n > 0) {
 270             node = try_dequeue();
 271             while (!node) {
 272                 qemu_mutex_unlock_iothread();
 273                 qemu_event_reset(&rcu_call_ready_event);
 274                 node = try_dequeue();
 275                 if (!node) {
 276                     qemu_event_wait(&rcu_call_ready_event);
 277                     node = try_dequeue();
 278                 }
 279                 qemu_mutex_lock_iothread();
 280             }
 281
 282             n--;
 283             node->func(node);
 284         }
 285         qemu_mutex_unlock_iothread();
 286     }
 287     abort();
 288 }
 289
 290 void call_rcu1(struct rcu_head *node, void (*func)(struct rcu_head *node))
 291 {
 292     node->func = func;
 293     enqueue(node);
 294     atomic_inc(&rcu_call_count);
 295     qemu_event_set(&rcu_call_ready_event);
 296 }
 297
 298 void rcu_register_thread(void)
 299 {
 300     assert(rcu_reader.ctr == 0);
 301     qemu_mutex_lock(&rcu_registry_lock);
 302     QLIST_INSERT_HEAD(&registry, &rcu_reader, node);
 303     qemu_mutex_unlock(&rcu_registry_lock);
 304 }
 305
 306 void rcu_unregister_thread(void)
 307 {
 308     qemu_mutex_lock(&rcu_registry_lock);
 309     QLIST_REMOVE(&rcu_reader, node);
 310     qemu_mutex_unlock(&rcu_registry_lock);
 311 }
 312
 313 static void rcu_init_complete(void)
 314 {
 315     QemuThread thread;
 316
 317     qemu_mutex_init(&rcu_registry_lock);
 318     qemu_mutex_init(&rcu_sync_lock);
 319     qemu_event_init(&rcu_gp_event, true);
 320
 321     qemu_event_init(&rcu_call_ready_event, false);
 322
 323     /* The caller is assumed to have iothread lock, so the call_rcu thread
 324      * must have been quiescent even after forking, just recreate it.
 325      */
 326     qemu_thread_create(&thread, "call_rcu", call_rcu_thread,
 327                        NULL, QEMU_THREAD_DETACHED);
 328
 329     rcu_register_thread();
 330 }
 331
 332 static int atfork_depth = 1;
 333
 334 void rcu_enable_atfork(void)
 335 {
 336     atfork_depth++;
 337 }
 338
 339 void rcu_disable_atfork(void)
 340 {
 341     atfork_depth--;
 342 }
 343
 344 #ifdef CONFIG_POSIX
 345 static void rcu_init_lock(void)
 346 {
 347     if (atfork_depth < 1) {
 348         return;
 349     }
 350
 351     qemu_mutex_lock(&rcu_sync_lock);
 352     qemu_mutex_lock(&rcu_registry_lock);
 353 }
 354
 355 static void rcu_init_unlock(void)
 356 {
 357     if (atfork_depth < 1) {
 358         return;
 359     }
 360
 361     qemu_mutex_unlock(&rcu_registry_lock);
 362     qemu_mutex_unlock(&rcu_sync_lock);
 363 }
 364
 365 static void rcu_init_child(void)
 366 {
 367     if (atfork_depth < 1) {
 368         return;
 369     }
 370
 371     memset(&registry, 0, sizeof(registry));
 372     rcu_init_complete();
 373 }
 374 #endif
 375
 376 static void __attribute__((__constructor__)) rcu_init(void)
 377 {
 378     smp_mb_global_init();
 379 #ifdef CONFIG_POSIX
 380     pthread_atfork(rcu_init_lock, rcu_init_unlock, rcu_init_child);
 381 #endif
 382     rcu_init_complete();
 383 }