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 #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 = 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_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 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          * qemu_event_reset has acquire semantics, so no memory barrier
  89          * is needed here.
  90          */
  91         QLIST_FOREACH(index, &registry, node) {
  92             atomic_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                 atomic_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 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     }
 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 = atomic_xchg(&tail, &node->next);
 192     atomic_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 && 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
 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 = 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          */
 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 = atomic_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 = atomic_read(&rcu_call_count);
 262         }
 263
 264         atomic_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     atomic_inc(&rcu_call_count);
 293     qemu_event_set(&rcu_call_ready_event);
 294 }
 295
 296 void rcu_register_thread(void)
 297 {
 298     assert(rcu_reader.ctr == 0);
 299     qemu_mutex_lock(&rcu_registry_lock);
 300     QLIST_INSERT_HEAD(&registry, &rcu_reader, node);
 301     qemu_mutex_unlock(&rcu_registry_lock);
 302 }
 303
 304 void rcu_unregister_thread(void)
 305 {
 306     qemu_mutex_lock(&rcu_registry_lock);
 307     QLIST_REMOVE(&rcu_reader, node);
 308     qemu_mutex_unlock(&rcu_registry_lock);
 309 }
 310
 311 static void rcu_init_complete(void)
 312 {
 313     QemuThread thread;
 314
 315     qemu_mutex_init(&rcu_registry_lock);
 316     qemu_mutex_init(&rcu_sync_lock);
 317     qemu_event_init(&rcu_gp_event, true);
 318
 319     qemu_event_init(&rcu_call_ready_event, false);
 320
 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      */
 324     qemu_thread_create(&thread, "call_rcu", call_rcu_thread,
 325                        NULL, QEMU_THREAD_DETACHED);
 326
 327     rcu_register_thread();
 328 }
 329
 330 static int atfork_depth = 1;
 331
 332 void rcu_enable_atfork(void)
 333 {
 334     atfork_depth++;
 335 }
 336
 337 void rcu_disable_atfork(void)
 338 {
 339     atfork_depth--;
 340 }
 341
 342 #ifdef CONFIG_POSIX
 343 static void rcu_init_lock(void)
 344 {
 345     if (atfork_depth < 1) {
 346         return;
 347     }
 348
 349     qemu_mutex_lock(&rcu_sync_lock);
 350     qemu_mutex_lock(&rcu_registry_lock);
 351 }
 352
 353 static void rcu_init_unlock(void)
 354 {
 355     if (atfork_depth < 1) {
 356         return;
 357     }
 358
 359     qemu_mutex_unlock(&rcu_registry_lock);
 360     qemu_mutex_unlock(&rcu_sync_lock);
 361 }
 362
 363 static void rcu_init_child(void)
 364 {
 365     if (atfork_depth < 1) {
 366         return;
 367     }
 368
 369     memset(&registry, 0, sizeof(registry));
 370     rcu_init_complete();
 371 }
 372 #endif
 373
 374 static void __attribute__((__constructor__)) rcu_init(void)
 375 {
 376     smp_mb_global_init();
 377 #ifdef CONFIG_POSIX
 378     pthread_atfork(rcu_init_lock, rcu_init_unlock, rcu_init_child);
 379 #endif
 380     rcu_init_complete();
 381 }