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