Use Interlocked calls in win32 once() implementation.

This is simpler than the previous scheme, which tried to allocate
the CRITICAL_SECTION struct in a thread-safe manner before it
could use it to run the wrapped function in a thread-safe manner.

Change-Id: I172e5544e5f16403a3a0e5e2b9104b1292a0d786

vpx_ports/vpx_once.h

@@ -1,5 +1,5 @@
 /*
- *  Copyright (c) 2011 The WebM project authors. All Rights Reserved.
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
  *
  *  Use of this source code is governed by a BSD-style license
  *  that can be found in the LICENSE file in the root of the source
@@ -13,63 +13,83 @@
 
 #include "vpx_config.h"
 
+/* Implement a function wrapper to guarantee initialization
+ * thread-safety for library singletons.
+ *
+ * NOTE: These functions use static locks, and can only be
+ * used with one common argument per compilation unit. So
+ *
+ * file1.c:
+ *   vpx_once(foo);
+ *   ...
+ *   vpx_once(foo);
+ *
+ *   file2.c:
+ *     vpx_once(bar);
+ *
+ * will ensure foo() and bar() are each called only once, but in
+ *
+ * file1.c:
+ *   vpx_once(foo);
+ *   vpx_once(bar):
+ *
+ * bar() will never be called because the lock is used up
+ * by the call to foo().
+ */
+
 #if CONFIG_MULTITHREAD && defined(_WIN32)
 #include <windows.h>
 #include <stdlib.h>
+/* Declare a per-compilation-unit state variable to track the progress
+ * of calling func() only once. This must be at global scope because
+ * local initializers are not thread-safe in MSVC prior to Visual
+ * Studio 2015.
+ *
+ * As a static, once_state will be zero-initialized as program start.
+ */
+static LONG once_state;
 static void once(void (*func)(void))
 {
-    static CRITICAL_SECTION *lock;
+    /* Try to advance once_state from its initial value of 0 to 1.
-    static LONG waiters;
+     * Only one thread can succeed in doing so.
-    static int done;
-    void *lock_ptr = &lock;
-
-    /* If the initialization is complete, return early. This isn't just an
-     * optimization, it prevents races on the destruction of the global
-     * lock.
      */
-    if(done)
+    if (InterlockedCompareExchange(&once_state, 1, 0) == 0) {
-        return;
+        /* We're the winning thread, having set once_state to 1.
-
+         * Call our function. */
-    InterlockedIncrement(&waiters);
-
-    /* Get a lock. We create one and try to make it the one-true-lock,
-     * throwing it away if we lost the race.
-     */
-
-    {
-        /* Scope to protect access to new_lock */
-        CRITICAL_SECTION *new_lock = malloc(sizeof(CRITICAL_SECTION));
-        InitializeCriticalSection(new_lock);
-        if (InterlockedCompareExchangePointer(lock_ptr, new_lock, NULL) != NULL)
-        {
-            DeleteCriticalSection(new_lock);
-            free(new_lock);
-        }
-    }
-
-    /* At this point, we have a lock that can be synchronized on. We don't
-     * care which thread actually performed the allocation.
-     */
-
-    EnterCriticalSection(lock);
-
-    if (!done)
-    {
         func();
-        done = 1;
+        /* Now advance once_state to 2, unblocking any other threads. */
+        InterlockedIncrement(&once_state);
+        return;
     }
 
-    LeaveCriticalSection(lock);
+    /* We weren't the winning thread, but we want to block on
-
+     * the state variable so we don't return before func()
-    /* Last one out should free resources. The destructed objects are
+     * has finished executing elsewhere.
-     * protected by checking if(done) above.
+     *
+     * Try to advance once_state from 2 to 2, which is only possible
+     * after the winning thead advances it from 1 to 2.
      */
-    if(!InterlockedDecrement(&waiters))
+    while (InterlockedCompareExchange(&once_state, 2, 2) != 2) {
-    {
+        /* State isn't yet 2. Try again.
-        DeleteCriticalSection(lock);
+         *
-        free(lock);
+         * We are used for singleton initialization functions,
-        lock = NULL;
+         * which should complete quickly. Contention will likewise
+         * be rare, so it's worthwhile to use a simple but cpu-
+         * intensive busy-wait instead of successive backoff,
+         * waiting on a kernel object, or another heavier-weight scheme.
+         *
+         * We can at least yield our timeslice.
+         */
+        Sleep(0);
     }
+
+    /* We've seen once_state advance to 2, so we know func()
+     * has been called. And we've left once_state as we found it,
+     * so other threads will have the same experience.
+     *
+     * It's safe to return now.
+     */
+    return;
 }