util/qemu-thread-win32.c

   1 /*
   2  * Win32 implementation for mutex/cond/thread functions
   3  *
   4  * Copyright Red Hat, Inc. 2010
   5  *
   6  * Author:
   7  *  Paolo Bonzini <pbonzini@redhat.com>
   8  *
   9  * This work is licensed under the terms of the GNU GPL, version 2 or later.
  10  * See the COPYING file in the top-level directory.
  11  *
  12  */
  13 #include "qemu/osdep.h"
  14 #include "qemu-common.h"
  15 #include "qemu/thread.h"
  16 #include "qemu/notify.h"
  17 #include <process.h>
  18
  19 static bool name_threads;
  20
  21 void qemu_thread_naming(bool enable)
  22 {
  23     /* But note we don't actually name them on Windows yet */
  24     name_threads = enable;
  25
  26     fprintf(stderr, "qemu: thread naming not supported on this host\n");
  27 }
  28
  29 static void error_exit(int err, const char *msg)
  30 {
  31     char *pstr;
  32
  33     FormatMessage(FORMAT_MESSAGE_FROM_SYSTEM | FORMAT_MESSAGE_ALLOCATE_BUFFER,
  34                   NULL, err, 0, (LPTSTR)&pstr, 2, NULL);
  35     fprintf(stderr, "qemu: %s: %s\n", msg, pstr);
  36     LocalFree(pstr);
  37     abort();
  38 }
  39
  40 void qemu_mutex_init(QemuMutex *mutex)
  41 {
  42     mutex->owner = 0;
  43     InitializeCriticalSection(&mutex->lock);
  44 }
  45
  46 void qemu_mutex_destroy(QemuMutex *mutex)
  47 {
  48     assert(mutex->owner == 0);
  49     DeleteCriticalSection(&mutex->lock);
  50 }
  51
  52 void qemu_mutex_lock(QemuMutex *mutex)
  53 {
  54     EnterCriticalSection(&mutex->lock);
  55
  56     /* Win32 CRITICAL_SECTIONs are recursive.  Assert that we're not
  57      * using them as such.
  58      */
  59     assert(mutex->owner == 0);
  60     mutex->owner = GetCurrentThreadId();
  61 }
  62
  63 int qemu_mutex_trylock(QemuMutex *mutex)
  64 {
  65     int owned;
  66
  67     owned = TryEnterCriticalSection(&mutex->lock);
  68     if (owned) {
  69         assert(mutex->owner == 0);
  70         mutex->owner = GetCurrentThreadId();
  71     }
  72     return !owned;
  73 }
  74
  75 void qemu_mutex_unlock(QemuMutex *mutex)
  76 {
  77     assert(mutex->owner == GetCurrentThreadId());
  78     mutex->owner = 0;
  79     LeaveCriticalSection(&mutex->lock);
  80 }
  81
  82 void qemu_rec_mutex_init(QemuRecMutex *mutex)
  83 {
  84     InitializeCriticalSection(&mutex->lock);
  85 }
  86
  87 void qemu_rec_mutex_destroy(QemuRecMutex *mutex)
  88 {
  89     DeleteCriticalSection(&mutex->lock);
  90 }
  91
  92 void qemu_rec_mutex_lock(QemuRecMutex *mutex)
  93 {
  94     EnterCriticalSection(&mutex->lock);
  95 }
  96
  97 int qemu_rec_mutex_trylock(QemuRecMutex *mutex)
  98 {
  99     return !TryEnterCriticalSection(&mutex->lock);
 100 }
 101
 102 void qemu_rec_mutex_unlock(QemuRecMutex *mutex)
 103 {
 104     LeaveCriticalSection(&mutex->lock);
 105 }
 106
 107 void qemu_cond_init(QemuCond *cond)
 108 {
 109     memset(cond, 0, sizeof(*cond));
 110
 111     cond->sema = CreateSemaphore(NULL, 0, LONG_MAX, NULL);
 112     if (!cond->sema) {
 113         error_exit(GetLastError(), __func__);
 114     }
 115     cond->continue_event = CreateEvent(NULL,    /* security */
 116                                        FALSE,   /* auto-reset */
 117                                        FALSE,   /* not signaled */
 118                                        NULL);   /* name */
 119     if (!cond->continue_event) {
 120         error_exit(GetLastError(), __func__);
 121     }
 122 }
 123
 124 void qemu_cond_destroy(QemuCond *cond)
 125 {
 126     BOOL result;
 127     result = CloseHandle(cond->continue_event);
 128     if (!result) {
 129         error_exit(GetLastError(), __func__);
 130     }
 131     cond->continue_event = 0;
 132     result = CloseHandle(cond->sema);
 133     if (!result) {
 134         error_exit(GetLastError(), __func__);
 135     }
 136     cond->sema = 0;
 137 }
 138
 139 void qemu_cond_signal(QemuCond *cond)
 140 {
 141     DWORD result;
 142
 143     /*
 144      * Signal only when there are waiters.  cond->waiters is
 145      * incremented by pthread_cond_wait under the external lock,
 146      * so we are safe about that.
 147      */
 148     if (cond->waiters == 0) {
 149         return;
 150     }
 151
 152     /*
 153      * Waiting threads decrement it outside the external lock, but
 154      * only if another thread is executing pthread_cond_broadcast and
 155      * has the mutex.  So, it also cannot be decremented concurrently
 156      * with this particular access.
 157      */
 158     cond->target = cond->waiters - 1;
 159     result = SignalObjectAndWait(cond->sema, cond->continue_event,
 160                                  INFINITE, FALSE);
 161     if (result == WAIT_ABANDONED || result == WAIT_FAILED) {
 162         error_exit(GetLastError(), __func__);
 163     }
 164 }
 165
 166 void qemu_cond_broadcast(QemuCond *cond)
 167 {
 168     BOOLEAN result;
 169     /*
 170      * As in pthread_cond_signal, access to cond->waiters and
 171      * cond->target is locked via the external mutex.
 172      */
 173     if (cond->waiters == 0) {
 174         return;
 175     }
 176
 177     cond->target = 0;
 178     result = ReleaseSemaphore(cond->sema, cond->waiters, NULL);
 179     if (!result) {
 180         error_exit(GetLastError(), __func__);
 181     }
 182
 183     /*
 184      * At this point all waiters continue. Each one takes its
 185      * slice of the semaphore. Now it's our turn to wait: Since
 186      * the external mutex is held, no thread can leave cond_wait,
 187      * yet. For this reason, we can be sure that no thread gets
 188      * a chance to eat *more* than one slice. OTOH, it means
 189      * that the last waiter must send us a wake-up.
 190      */
 191     WaitForSingleObject(cond->continue_event, INFINITE);
 192 }
 193
 194 void qemu_cond_wait(QemuCond *cond, QemuMutex *mutex)
 195 {
 196     /*
 197      * This access is protected under the mutex.
 198      */
 199     cond->waiters++;
 200
 201     /*
 202      * Unlock external mutex and wait for signal.
 203      * NOTE: we've held mutex locked long enough to increment
 204      * waiters count above, so there's no problem with
 205      * leaving mutex unlocked before we wait on semaphore.
 206      */
 207     qemu_mutex_unlock(mutex);
 208     WaitForSingleObject(cond->sema, INFINITE);
 209
 210     /* Now waiters must rendez-vous with the signaling thread and
 211      * let it continue.  For cond_broadcast this has heavy contention
 212      * and triggers thundering herd.  So goes life.
 213      *
 214      * Decrease waiters count.  The mutex is not taken, so we have
 215      * to do this atomically.
 216      *
 217      * All waiters contend for the mutex at the end of this function
 218      * until the signaling thread relinquishes it.  To ensure
 219      * each waiter consumes exactly one slice of the semaphore,
 220      * the signaling thread stops until it is told by the last
 221      * waiter that it can go on.
 222      */
 223     if (InterlockedDecrement(&cond->waiters) == cond->target) {
 224         SetEvent(cond->continue_event);
 225     }
 226
 227     qemu_mutex_lock(mutex);
 228 }
 229
 230 void qemu_sem_init(QemuSemaphore *sem, int init)
 231 {
 232     /* Manual reset.  */
 233     sem->sema = CreateSemaphore(NULL, init, LONG_MAX, NULL);
 234 }
 235
 236 void qemu_sem_destroy(QemuSemaphore *sem)
 237 {
 238     CloseHandle(sem->sema);
 239 }
 240
 241 void qemu_sem_post(QemuSemaphore *sem)
 242 {
 243     ReleaseSemaphore(sem->sema, 1, NULL);
 244 }
 245
 246 int qemu_sem_timedwait(QemuSemaphore *sem, int ms)
 247 {
 248     int rc = WaitForSingleObject(sem->sema, ms);
 249     if (rc == WAIT_OBJECT_0) {
 250         return 0;
 251     }
 252     if (rc != WAIT_TIMEOUT) {
 253         error_exit(GetLastError(), __func__);
 254     }
 255     return -1;
 256 }
 257
 258 void qemu_sem_wait(QemuSemaphore *sem)
 259 {
 260     if (WaitForSingleObject(sem->sema, INFINITE) != WAIT_OBJECT_0) {
 261         error_exit(GetLastError(), __func__);
 262     }
 263 }
 264
 265 /* Wrap a Win32 manual-reset event with a fast userspace path.  The idea
 266  * is to reset the Win32 event lazily, as part of a test-reset-test-wait
 267  * sequence.  Such a sequence is, indeed, how QemuEvents are used by
 268  * RCU and other subsystems!
 269  *
 270  * Valid transitions:
 271  * - free->set, when setting the event
 272  * - busy->set, when setting the event, followed by futex_wake
 273  * - set->free, when resetting the event
 274  * - free->busy, when waiting
 275  *
 276  * set->busy does not happen (it can be observed from the outside but
 277  * it really is set->free->busy).
 278  *
 279  * busy->free provably cannot happen; to enforce it, the set->free transition
 280  * is done with an OR, which becomes a no-op if the event has concurrently
 281  * transitioned to free or busy (and is faster than cmpxchg).
 282  */
 283
 284 #define EV_SET         0
 285 #define EV_FREE        1
 286 #define EV_BUSY       -1
 287
 288 void qemu_event_init(QemuEvent *ev, bool init)
 289 {
 290     /* Manual reset.  */
 291     ev->event = CreateEvent(NULL, TRUE, TRUE, NULL);
 292     ev->value = (init ? EV_SET : EV_FREE);
 293 }
 294
 295 void qemu_event_destroy(QemuEvent *ev)
 296 {
 297     CloseHandle(ev->event);
 298 }
 299
 300 void qemu_event_set(QemuEvent *ev)
 301 {
 302     /* qemu_event_set has release semantics, but because it *loads*
 303      * ev->value we need a full memory barrier here.
 304      */
 305     smp_mb();
 306     if (atomic_read(&ev->value) != EV_SET) {
 307         if (atomic_xchg(&ev->value, EV_SET) == EV_BUSY) {
 308             /* There were waiters, wake them up.  */
 309             SetEvent(ev->event);
 310         }
 311     }
 312 }
 313
 314 void qemu_event_reset(QemuEvent *ev)
 315 {
 316     unsigned value;
 317
 318     value = atomic_read(&ev->value);
 319     smp_mb_acquire();
 320     if (value == EV_SET) {
 321         /* If there was a concurrent reset (or even reset+wait),
 322          * do nothing.  Otherwise change EV_SET->EV_FREE.
 323          */
 324         atomic_or(&ev->value, EV_FREE);
 325     }
 326 }
 327
 328 void qemu_event_wait(QemuEvent *ev)
 329 {
 330     unsigned value;
 331
 332     value = atomic_read(&ev->value);
 333     smp_mb_acquire();
 334     if (value != EV_SET) {
 335         if (value == EV_FREE) {
 336             /* qemu_event_set is not yet going to call SetEvent, but we are
 337              * going to do another check for EV_SET below when setting EV_BUSY.
 338              * At that point it is safe to call WaitForSingleObject.
 339              */
 340             ResetEvent(ev->event);
 341
 342             /* Tell qemu_event_set that there are waiters.  No need to retry
 343              * because there cannot be a concurent busy->free transition.
 344              * After the CAS, the event will be either set or busy.
 345              */
 346             if (atomic_cmpxchg(&ev->value, EV_FREE, EV_BUSY) == EV_SET) {
 347                 value = EV_SET;
 348             } else {
 349                 value = EV_BUSY;
 350             }
 351         }
 352         if (value == EV_BUSY) {
 353             WaitForSingleObject(ev->event, INFINITE);
 354         }
 355     }
 356 }
 357
 358 struct QemuThreadData {
 359     /* Passed to win32_start_routine.  */
 360     void             *(*start_routine)(void *);
 361     void             *arg;
 362     short             mode;
 363     NotifierList      exit;
 364
 365     /* Only used for joinable threads. */
 366     bool              exited;
 367     void             *ret;
 368     CRITICAL_SECTION  cs;
 369 };
 370
 371 static bool atexit_registered;
 372 static NotifierList main_thread_exit;
 373
 374 static __thread QemuThreadData *qemu_thread_data;
 375
 376 static void run_main_thread_exit(void)
 377 {
 378     notifier_list_notify(&main_thread_exit, NULL);
 379 }
 380
 381 void qemu_thread_atexit_add(Notifier *notifier)
 382 {
 383     if (!qemu_thread_data) {
 384         if (!atexit_registered) {
 385             atexit_registered = true;
 386             atexit(run_main_thread_exit);
 387         }
 388         notifier_list_add(&main_thread_exit, notifier);
 389     } else {
 390         notifier_list_add(&qemu_thread_data->exit, notifier);
 391     }
 392 }
 393
 394 void qemu_thread_atexit_remove(Notifier *notifier)
 395 {
 396     notifier_remove(notifier);
 397 }
 398
 399 static unsigned __stdcall win32_start_routine(void *arg)
 400 {
 401     QemuThreadData *data = (QemuThreadData *) arg;
 402     void *(*start_routine)(void *) = data->start_routine;
 403     void *thread_arg = data->arg;
 404
 405     qemu_thread_data = data;
 406     qemu_thread_exit(start_routine(thread_arg));
 407     abort();
 408 }
 409
 410 void qemu_thread_exit(void *arg)
 411 {
 412     QemuThreadData *data = qemu_thread_data;
 413
 414     notifier_list_notify(&data->exit, NULL);
 415     if (data->mode == QEMU_THREAD_JOINABLE) {
 416         data->ret = arg;
 417         EnterCriticalSection(&data->cs);
 418         data->exited = true;
 419         LeaveCriticalSection(&data->cs);
 420     } else {
 421         g_free(data);
 422     }
 423     _endthreadex(0);
 424 }
 425
 426 void *qemu_thread_join(QemuThread *thread)
 427 {
 428     QemuThreadData *data;
 429     void *ret;
 430     HANDLE handle;
 431
 432     data = thread->data;
 433     if (data->mode == QEMU_THREAD_DETACHED) {
 434         return NULL;
 435     }
 436
 437     /*
 438      * Because multiple copies of the QemuThread can exist via
 439      * qemu_thread_get_self, we need to store a value that cannot
 440      * leak there.  The simplest, non racy way is to store the TID,
 441      * discard the handle that _beginthreadex gives back, and
 442      * get another copy of the handle here.
 443      */
 444     handle = qemu_thread_get_handle(thread);
 445     if (handle) {
 446         WaitForSingleObject(handle, INFINITE);
 447         CloseHandle(handle);
 448     }
 449     ret = data->ret;
 450     DeleteCriticalSection(&data->cs);
 451     g_free(data);
 452     return ret;
 453 }
 454
 455 void qemu_thread_create(QemuThread *thread, const char *name,
 456                        void *(*start_routine)(void *),
 457                        void *arg, int mode)
 458 {
 459     HANDLE hThread;
 460     struct QemuThreadData *data;
 461
 462     data = g_malloc(sizeof *data);
 463     data->start_routine = start_routine;
 464     data->arg = arg;
 465     data->mode = mode;
 466     data->exited = false;
 467     notifier_list_init(&data->exit);
 468
 469     if (data->mode != QEMU_THREAD_DETACHED) {
 470         InitializeCriticalSection(&data->cs);
 471     }
 472
 473     hThread = (HANDLE) _beginthreadex(NULL, 0, win32_start_routine,
 474                                       data, 0, &thread->tid);
 475     if (!hThread) {
 476         error_exit(GetLastError(), __func__);
 477     }
 478     CloseHandle(hThread);
 479     thread->data = data;
 480 }
 481
 482 void qemu_thread_get_self(QemuThread *thread)
 483 {
 484     thread->data = qemu_thread_data;
 485     thread->tid = GetCurrentThreadId();
 486 }
 487
 488 HANDLE qemu_thread_get_handle(QemuThread *thread)
 489 {
 490     QemuThreadData *data;
 491     HANDLE handle;
 492
 493     data = thread->data;
 494     if (data->mode == QEMU_THREAD_DETACHED) {
 495         return NULL;
 496     }
 497
 498     EnterCriticalSection(&data->cs);
 499     if (!data->exited) {
 500         handle = OpenThread(SYNCHRONIZE | THREAD_SUSPEND_RESUME, FALSE,
 501                             thread->tid);
 502     } else {
 503         handle = NULL;
 504     }
 505     LeaveCriticalSection(&data->cs);
 506     return handle;
 507 }
 508
 509 bool qemu_thread_is_self(QemuThread *thread)
 510 {
 511     return GetCurrentThreadId() == thread->tid;
 512 }