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 #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
  96          * loads of index->ctr.
  97          */
  98         smp_mb();
  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     qemu_mutex_lock(&rcu_registry_lock);
 146
 147     if (!QLIST_EMPTY(&registry)) {
 148         /* In either case, the atomic_mb_set below blocks stores that free
 149          * old RCU-protected pointers.
 150          */
 151         if (sizeof(rcu_gp_ctr) < 8) {
 152             /* For architectures with 32-bit longs, a two-subphases algorithm
 153              * ensures we do not encounter overflow bugs.
 154              *
 155              * Switch parity: 0 -> 1, 1 -> 0.
 156              */
 157             atomic_mb_set(&rcu_gp_ctr, rcu_gp_ctr ^ RCU_GP_CTR);
 158             wait_for_readers();
 159             atomic_mb_set(&rcu_gp_ctr, rcu_gp_ctr ^ RCU_GP_CTR);
 160         } else {
 161             /* Increment current grace period.  */
 162             atomic_mb_set(&rcu_gp_ctr, rcu_gp_ctr + RCU_GP_CTR);
 163         }
 164
 165         wait_for_readers();
 166     }
 167
 168     qemu_mutex_unlock(&rcu_registry_lock);
 169     qemu_mutex_unlock(&rcu_sync_lock);
 170 }
 171
 172
 173 #define RCU_CALL_MIN_SIZE        30
 174
 175 /* Multi-producer, single-consumer queue based on urcu/static/wfqueue.h
 176  * from liburcu.  Note that head is only used by the consumer.
 177  */
 178 static struct rcu_head dummy;
 179 static struct rcu_head *head = &dummy, **tail = &dummy.next;
 180 static int rcu_call_count;
 181 static QemuEvent rcu_call_ready_event;
 182
 183 static void enqueue(struct rcu_head *node)
 184 {
 185     struct rcu_head **old_tail;
 186
 187     node->next = NULL;
 188     old_tail = atomic_xchg(&tail, &node->next);
 189     atomic_mb_set(old_tail, node);
 190 }
 191
 192 static struct rcu_head *try_dequeue(void)
 193 {
 194     struct rcu_head *node, *next;
 195
 196 retry:
 197     /* Test for an empty list, which we do not expect.  Note that for
 198      * the consumer head and tail are always consistent.  The head
 199      * is consistent because only the consumer reads/writes it.
 200      * The tail, because it is the first step in the enqueuing.
 201      * It is only the next pointers that might be inconsistent.
 202      */
 203     if (head == &dummy && atomic_mb_read(&tail) == &dummy.next) {
 204         abort();
 205     }
 206
 207     /* If the head node has NULL in its next pointer, the value is
 208      * wrong and we need to wait until its enqueuer finishes the update.
 209      */
 210     node = head;
 211     next = atomic_mb_read(&head->next);
 212     if (!next) {
 213         return NULL;
 214     }
 215
 216     /* Since we are the sole consumer, and we excluded the empty case
 217      * above, the queue will always have at least two nodes: the
 218      * dummy node, and the one being removed.  So we do not need to update
 219      * the tail pointer.
 220      */
 221     head = next;
 222
 223     /* If we dequeued the dummy node, add it back at the end and retry.  */
 224     if (node == &dummy) {
 225         enqueue(node);
 226         goto retry;
 227     }
 228
 229     return node;
 230 }
 231
 232 static void *call_rcu_thread(void *opaque)
 233 {
 234     struct rcu_head *node;
 235
 236     rcu_register_thread();
 237
 238     for (;;) {
 239         int tries = 0;
 240         int n = atomic_read(&rcu_call_count);
 241
 242         /* Heuristically wait for a decent number of callbacks to pile up.
 243          * Fetch rcu_call_count now, we only must process elements that were
 244          * added before synchronize_rcu() starts.
 245          */
 246         while (n == 0 || (n < RCU_CALL_MIN_SIZE && ++tries <= 5)) {
 247             g_usleep(10000);
 248             if (n == 0) {
 249                 qemu_event_reset(&rcu_call_ready_event);
 250                 n = atomic_read(&rcu_call_count);
 251                 if (n == 0) {
 252 #if defined(CONFIG_MALLOC_TRIM)
 253                     malloc_trim(4 * 1024 * 1024);
 254 #endif
 255                     qemu_event_wait(&rcu_call_ready_event);
 256                 }
 257             }
 258             n = atomic_read(&rcu_call_count);
 259         }
 260
 261         atomic_sub(&rcu_call_count, n);
 262         synchronize_rcu();
 263         qemu_mutex_lock_iothread();
 264         while (n > 0) {
 265             node = try_dequeue();
 266             while (!node) {
 267                 qemu_mutex_unlock_iothread();
 268                 qemu_event_reset(&rcu_call_ready_event);
 269                 node = try_dequeue();
 270                 if (!node) {
 271                     qemu_event_wait(&rcu_call_ready_event);
 272                     node = try_dequeue();
 273                 }
 274                 qemu_mutex_lock_iothread();
 275             }
 276
 277             n--;
 278             node->func(node);
 279         }
 280         qemu_mutex_unlock_iothread();
 281     }
 282     abort();
 283 }
 284
 285 void call_rcu1(struct rcu_head *node, void (*func)(struct rcu_head *node))
 286 {
 287     node->func = func;
 288     enqueue(node);
 289     atomic_inc(&rcu_call_count);
 290     qemu_event_set(&rcu_call_ready_event);
 291 }
 292
 293 void rcu_register_thread(void)
 294 {
 295     assert(rcu_reader.ctr == 0);
 296     qemu_mutex_lock(&rcu_registry_lock);
 297     QLIST_INSERT_HEAD(&registry, &rcu_reader, node);
 298     qemu_mutex_unlock(&rcu_registry_lock);
 299 }
 300
 301 void rcu_unregister_thread(void)
 302 {
 303     qemu_mutex_lock(&rcu_registry_lock);
 304     QLIST_REMOVE(&rcu_reader, node);
 305     qemu_mutex_unlock(&rcu_registry_lock);
 306 }
 307
 308 static void rcu_init_complete(void)
 309 {
 310     QemuThread thread;
 311
 312     qemu_mutex_init(&rcu_registry_lock);
 313     qemu_mutex_init(&rcu_sync_lock);
 314     qemu_event_init(&rcu_gp_event, true);
 315
 316     qemu_event_init(&rcu_call_ready_event, false);
 317
 318     /* The caller is assumed to have iothread lock, so the call_rcu thread
 319      * must have been quiescent even after forking, just recreate it.
 320      */
 321     qemu_thread_create(&thread, "call_rcu", call_rcu_thread,
 322                        NULL, QEMU_THREAD_DETACHED);
 323
 324     rcu_register_thread();
 325 }
 326
 327 static int atfork_depth = 1;
 328
 329 void rcu_enable_atfork(void)
 330 {
 331     atfork_depth++;
 332 }
 333
 334 void rcu_disable_atfork(void)
 335 {
 336     atfork_depth--;
 337 }
 338
 339 #ifdef CONFIG_POSIX
 340 static void rcu_init_lock(void)
 341 {
 342     if (atfork_depth < 1) {
 343         return;
 344     }
 345
 346     qemu_mutex_lock(&rcu_sync_lock);
 347     qemu_mutex_lock(&rcu_registry_lock);
 348 }
 349
 350 static void rcu_init_unlock(void)
 351 {
 352     if (atfork_depth < 1) {
 353         return;
 354     }
 355
 356     qemu_mutex_unlock(&rcu_registry_lock);
 357     qemu_mutex_unlock(&rcu_sync_lock);
 358 }
 359
 360 static void rcu_init_child(void)
 361 {
 362     if (atfork_depth < 1) {
 363         return;
 364     }
 365
 366     memset(&registry, 0, sizeof(registry));
 367     rcu_init_complete();
 368 }
 369 #endif
 370
 371 static void __attribute__((__constructor__)) rcu_init(void)
 372 {
 373 #ifdef CONFIG_POSIX
 374     pthread_atfork(rcu_init_lock, rcu_init_unlock, rcu_init_child);
 375 #endif
 376     rcu_init_complete();
 377 }