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