util/qemu-coroutine-lock.c

   1 /*
   2  * coroutine queues and locks
   3  *
   4  * Copyright (c) 2011 Kevin Wolf <kwolf@redhat.com>
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a copy
   7  * of this software and associated documentation files (the "Software"), to deal
   8  * in the Software without restriction, including without limitation the rights
   9  * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
  10  * copies of the Software, and to permit persons to whom the Software is
  11  * furnished to do so, subject to the following conditions:
  12  *
  13  * The above copyright notice and this permission notice shall be included in
  14  * all copies or substantial portions of the Software.
  15  *
  16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
  19  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  21  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
  22  * THE SOFTWARE.
  23  *
  24  * The lock-free mutex implementation is based on OSv
  25  * (core/lfmutex.cc, include/lockfree/mutex.hh).
  26  * Copyright (C) 2013 Cloudius Systems, Ltd.
  27  */
  28
  29 #include "qemu/osdep.h"
  30 #include "qemu/coroutine.h"
  31 #include "qemu/coroutine_int.h"
  32 #include "qemu/processor.h"
  33 #include "qemu/queue.h"
  34 #include "block/aio.h"
  35 #include "trace.h"
  36
  37 void qemu_co_queue_init(CoQueue *queue)
  38 {
  39     QSIMPLEQ_INIT(&queue->entries);
  40 }
  41
  42 void coroutine_fn qemu_co_queue_wait_impl(CoQueue *queue, QemuLockable *lock)
  43 {
  44     Coroutine *self = qemu_coroutine_self();
  45     QSIMPLEQ_INSERT_TAIL(&queue->entries, self, co_queue_next);
  46
  47     if (lock) {
  48         qemu_lockable_unlock(lock);
  49     }
  50
  51     /* There is no race condition here.  Other threads will call
  52      * aio_co_schedule on our AioContext, which can reenter this
  53      * coroutine but only after this yield and after the main loop
  54      * has gone through the next iteration.
  55      */
  56     qemu_coroutine_yield();
  57     assert(qemu_in_coroutine());
  58
  59     /* TODO: OSv implements wait morphing here, where the wakeup
  60      * primitive automatically places the woken coroutine on the
  61      * mutex's queue.  This avoids the thundering herd effect.
  62      * This could be implemented for CoMutexes, but not really for
  63      * other cases of QemuLockable.
  64      */
  65     if (lock) {
  66         qemu_lockable_lock(lock);
  67     }
  68 }
  69
  70 static bool qemu_co_queue_do_restart(CoQueue *queue, bool single)
  71 {
  72     Coroutine *next;
  73
  74     if (QSIMPLEQ_EMPTY(&queue->entries)) {
  75         return false;
  76     }
  77
  78     while ((next = QSIMPLEQ_FIRST(&queue->entries)) != NULL) {
  79         QSIMPLEQ_REMOVE_HEAD(&queue->entries, co_queue_next);
  80         aio_co_wake(next);
  81         if (single) {
  82             break;
  83         }
  84     }
  85     return true;
  86 }
  87
  88 bool qemu_co_queue_next(CoQueue *queue)
  89 {
  90     return qemu_co_queue_do_restart(queue, true);
  91 }
  92
  93 void qemu_co_queue_restart_all(CoQueue *queue)
  94 {
  95     qemu_co_queue_do_restart(queue, false);
  96 }
  97
  98 bool qemu_co_enter_next_impl(CoQueue *queue, QemuLockable *lock)
  99 {
 100     Coroutine *next;
 101
 102     next = QSIMPLEQ_FIRST(&queue->entries);
 103     if (!next) {
 104         return false;
 105     }
 106
 107     QSIMPLEQ_REMOVE_HEAD(&queue->entries, co_queue_next);
 108     if (lock) {
 109         qemu_lockable_unlock(lock);
 110     }
 111     aio_co_wake(next);
 112     if (lock) {
 113         qemu_lockable_lock(lock);
 114     }
 115     return true;
 116 }
 117
 118 bool qemu_co_queue_empty(CoQueue *queue)
 119 {
 120     return QSIMPLEQ_FIRST(&queue->entries) == NULL;
 121 }
 122
 123 /* The wait records are handled with a multiple-producer, single-consumer
 124  * lock-free queue.  There cannot be two concurrent pop_waiter() calls
 125  * because pop_waiter() can only be called while mutex->handoff is zero.
 126  * This can happen in three cases:
 127  * - in qemu_co_mutex_unlock, before the hand-off protocol has started.
 128  *   In this case, qemu_co_mutex_lock will see mutex->handoff == 0 and
 129  *   not take part in the handoff.
 130  * - in qemu_co_mutex_lock, if it steals the hand-off responsibility from
 131  *   qemu_co_mutex_unlock.  In this case, qemu_co_mutex_unlock will fail
 132  *   the cmpxchg (it will see either 0 or the next sequence value) and
 133  *   exit.  The next hand-off cannot begin until qemu_co_mutex_lock has
 134  *   woken up someone.
 135  * - in qemu_co_mutex_unlock, if it takes the hand-off token itself.
 136  *   In this case another iteration starts with mutex->handoff == 0;
 137  *   a concurrent qemu_co_mutex_lock will fail the cmpxchg, and
 138  *   qemu_co_mutex_unlock will go back to case (1).
 139  *
 140  * The following functions manage this queue.
 141  */
 142 typedef struct CoWaitRecord {
 143     Coroutine *co;
 144     QSLIST_ENTRY(CoWaitRecord) next;
 145 } CoWaitRecord;
 146
 147 static void push_waiter(CoMutex *mutex, CoWaitRecord *w)
 148 {
 149     w->co = qemu_coroutine_self();
 150     QSLIST_INSERT_HEAD_ATOMIC(&mutex->from_push, w, next);
 151 }
 152
 153 static void move_waiters(CoMutex *mutex)
 154 {
 155     QSLIST_HEAD(, CoWaitRecord) reversed;
 156     QSLIST_MOVE_ATOMIC(&reversed, &mutex->from_push);
 157     while (!QSLIST_EMPTY(&reversed)) {
 158         CoWaitRecord *w = QSLIST_FIRST(&reversed);
 159         QSLIST_REMOVE_HEAD(&reversed, next);
 160         QSLIST_INSERT_HEAD(&mutex->to_pop, w, next);
 161     }
 162 }
 163
 164 static CoWaitRecord *pop_waiter(CoMutex *mutex)
 165 {
 166     CoWaitRecord *w;
 167
 168     if (QSLIST_EMPTY(&mutex->to_pop)) {
 169         move_waiters(mutex);
 170         if (QSLIST_EMPTY(&mutex->to_pop)) {
 171             return NULL;
 172         }
 173     }
 174     w = QSLIST_FIRST(&mutex->to_pop);
 175     QSLIST_REMOVE_HEAD(&mutex->to_pop, next);
 176     return w;
 177 }
 178
 179 static bool has_waiters(CoMutex *mutex)
 180 {
 181     return QSLIST_EMPTY(&mutex->to_pop) || QSLIST_EMPTY(&mutex->from_push);
 182 }
 183
 184 void qemu_co_mutex_init(CoMutex *mutex)
 185 {
 186     memset(mutex, 0, sizeof(*mutex));
 187 }
 188
 189 static void coroutine_fn qemu_co_mutex_wake(CoMutex *mutex, Coroutine *co)
 190 {
 191     /* Read co before co->ctx; pairs with smp_wmb() in
 192      * qemu_coroutine_enter().
 193      */
 194     smp_read_barrier_depends();
 195     mutex->ctx = co->ctx;
 196     aio_co_wake(co);
 197 }
 198
 199 static void coroutine_fn qemu_co_mutex_lock_slowpath(AioContext *ctx,
 200                                                      CoMutex *mutex)
 201 {
 202     Coroutine *self = qemu_coroutine_self();
 203     CoWaitRecord w;
 204     unsigned old_handoff;
 205
 206     trace_qemu_co_mutex_lock_entry(mutex, self);
 207     w.co = self;
 208     push_waiter(mutex, &w);
 209
 210     /* This is the "Responsibility Hand-Off" protocol; a lock() picks from
 211      * a concurrent unlock() the responsibility of waking somebody up.
 212      */
 213     old_handoff = qatomic_mb_read(&mutex->handoff);
 214     if (old_handoff &&
 215         has_waiters(mutex) &&
 216         qatomic_cmpxchg(&mutex->handoff, old_handoff, 0) == old_handoff) {
 217         /* There can be no concurrent pops, because there can be only
 218          * one active handoff at a time.
 219          */
 220         CoWaitRecord *to_wake = pop_waiter(mutex);
 221         Coroutine *co = to_wake->co;
 222         if (co == self) {
 223             /* We got the lock ourselves!  */
 224             assert(to_wake == &w);
 225             mutex->ctx = ctx;
 226             return;
 227         }
 228
 229         qemu_co_mutex_wake(mutex, co);
 230     }
 231
 232     qemu_coroutine_yield();
 233     trace_qemu_co_mutex_lock_return(mutex, self);
 234 }
 235
 236 void coroutine_fn qemu_co_mutex_lock(CoMutex *mutex)
 237 {
 238     AioContext *ctx = qemu_get_current_aio_context();
 239     Coroutine *self = qemu_coroutine_self();
 240     int waiters, i;
 241
 242     /* Running a very small critical section on pthread_mutex_t and CoMutex
 243      * shows that pthread_mutex_t is much faster because it doesn't actually
 244      * go to sleep.  What happens is that the critical section is shorter
 245      * than the latency of entering the kernel and thus FUTEX_WAIT always
 246      * fails.  With CoMutex there is no such latency but you still want to
 247      * avoid wait and wakeup.  So introduce it artificially.
 248      */
 249     i = 0;
 250 retry_fast_path:
 251     waiters = qatomic_cmpxchg(&mutex->locked, 0, 1);
 252     if (waiters != 0) {
 253         while (waiters == 1 && ++i < 1000) {
 254             if (qatomic_read(&mutex->ctx) == ctx) {
 255                 break;
 256             }
 257             if (qatomic_read(&mutex->locked) == 0) {
 258                 goto retry_fast_path;
 259             }
 260             cpu_relax();
 261         }
 262         waiters = qatomic_fetch_inc(&mutex->locked);
 263     }
 264
 265     if (waiters == 0) {
 266         /* Uncontended.  */
 267         trace_qemu_co_mutex_lock_uncontended(mutex, self);
 268         mutex->ctx = ctx;
 269     } else {
 270         qemu_co_mutex_lock_slowpath(ctx, mutex);
 271     }
 272     mutex->holder = self;
 273     self->locks_held++;
 274 }
 275
 276 void coroutine_fn qemu_co_mutex_unlock(CoMutex *mutex)
 277 {
 278     Coroutine *self = qemu_coroutine_self();
 279
 280     trace_qemu_co_mutex_unlock_entry(mutex, self);
 281
 282     assert(mutex->locked);
 283     assert(mutex->holder == self);
 284     assert(qemu_in_coroutine());
 285
 286     mutex->ctx = NULL;
 287     mutex->holder = NULL;
 288     self->locks_held--;
 289     if (qatomic_fetch_dec(&mutex->locked) == 1) {
 290         /* No waiting qemu_co_mutex_lock().  Pfew, that was easy!  */
 291         return;
 292     }
 293
 294     for (;;) {
 295         CoWaitRecord *to_wake = pop_waiter(mutex);
 296         unsigned our_handoff;
 297
 298         if (to_wake) {
 299             qemu_co_mutex_wake(mutex, to_wake->co);
 300             break;
 301         }
 302
 303         /* Some concurrent lock() is in progress (we know this because
 304          * mutex->locked was >1) but it hasn't yet put itself on the wait
 305          * queue.  Pick a sequence number for the handoff protocol (not 0).
 306          */
 307         if (++mutex->sequence == 0) {
 308             mutex->sequence = 1;
 309         }
 310
 311         our_handoff = mutex->sequence;
 312         qatomic_mb_set(&mutex->handoff, our_handoff);
 313         if (!has_waiters(mutex)) {
 314             /* The concurrent lock has not added itself yet, so it
 315              * will be able to pick our handoff.
 316              */
 317             break;
 318         }
 319
 320         /* Try to do the handoff protocol ourselves; if somebody else has
 321          * already taken it, however, we're done and they're responsible.
 322          */
 323         if (qatomic_cmpxchg(&mutex->handoff, our_handoff, 0) != our_handoff) {
 324             break;
 325         }
 326     }
 327
 328     trace_qemu_co_mutex_unlock_return(mutex, self);
 329 }
 330
 331 void qemu_co_rwlock_init(CoRwlock *lock)
 332 {
 333     memset(lock, 0, sizeof(*lock));
 334     qemu_co_queue_init(&lock->queue);
 335     qemu_co_mutex_init(&lock->mutex);
 336 }
 337
 338 void qemu_co_rwlock_rdlock(CoRwlock *lock)
 339 {
 340     Coroutine *self = qemu_coroutine_self();
 341
 342     qemu_co_mutex_lock(&lock->mutex);
 343     /* For fairness, wait if a writer is in line.  */
 344     while (lock->pending_writer) {
 345         qemu_co_queue_wait(&lock->queue, &lock->mutex);
 346     }
 347     lock->reader++;
 348     qemu_co_mutex_unlock(&lock->mutex);
 349
 350     /* The rest of the read-side critical section is run without the mutex.  */
 351     self->locks_held++;
 352 }
 353
 354 void qemu_co_rwlock_unlock(CoRwlock *lock)
 355 {
 356     Coroutine *self = qemu_coroutine_self();
 357
 358     assert(qemu_in_coroutine());
 359     if (!lock->reader) {
 360         /* The critical section started in qemu_co_rwlock_wrlock.  */
 361         qemu_co_queue_restart_all(&lock->queue);
 362     } else {
 363         self->locks_held--;
 364
 365         qemu_co_mutex_lock(&lock->mutex);
 366         lock->reader--;
 367         assert(lock->reader >= 0);
 368         /* Wakeup only one waiting writer */
 369         if (!lock->reader) {
 370             qemu_co_queue_next(&lock->queue);
 371         }
 372     }
 373     qemu_co_mutex_unlock(&lock->mutex);
 374 }
 375
 376 void qemu_co_rwlock_downgrade(CoRwlock *lock)
 377 {
 378     Coroutine *self = qemu_coroutine_self();
 379
 380     /* lock->mutex critical section started in qemu_co_rwlock_wrlock or
 381      * qemu_co_rwlock_upgrade.
 382      */
 383     assert(lock->reader == 0);
 384     lock->reader++;
 385     qemu_co_mutex_unlock(&lock->mutex);
 386
 387     /* The rest of the read-side critical section is run without the mutex.  */
 388     self->locks_held++;
 389 }
 390
 391 void qemu_co_rwlock_wrlock(CoRwlock *lock)
 392 {
 393     qemu_co_mutex_lock(&lock->mutex);
 394     lock->pending_writer++;
 395     while (lock->reader) {
 396         qemu_co_queue_wait(&lock->queue, &lock->mutex);
 397     }
 398     lock->pending_writer--;
 399
 400     /* The rest of the write-side critical section is run with
 401      * the mutex taken, so that lock->reader remains zero.
 402      * There is no need to update self->locks_held.
 403      */
 404 }
 405
 406 void qemu_co_rwlock_upgrade(CoRwlock *lock)
 407 {
 408     Coroutine *self = qemu_coroutine_self();
 409
 410     qemu_co_mutex_lock(&lock->mutex);
 411     assert(lock->reader > 0);
 412     lock->reader--;
 413     lock->pending_writer++;
 414     while (lock->reader) {
 415         qemu_co_queue_wait(&lock->queue, &lock->mutex);
 416     }
 417     lock->pending_writer--;
 418
 419     /* The rest of the write-side critical section is run with
 420      * the mutex taken, similar to qemu_co_rwlock_wrlock.  Do
 421      * not account for the lock twice in self->locks_held.
 422      */
 423     self->locks_held--;
 424 }