X-Git-Url: https://git.draconx.ca/gitweb/cdecl99.git/blobdiff_plain/c11f668ec9c5b2f0016cf6920a111c52e895b24a..bf336dc52970daa394d119b11284a1b5016a74c0:/src/thread-w32.h

diff --git a/src/thread-w32.h b/src/thread-w32.h
index fb3bdc4..181ebfc 100644
--- a/src/thread-w32.h
+++ b/src/thread-w32.h
@@ -8,6 +8,19 @@
  * the Free Software Foundation, either version 3 of the License, or
  * (at your option) any later version.
  *
+ * The init_once implementation is adapted from the Pthreads-win32 library
+ * implementation based on MCS (Mellor-Crummy Scott) locks, originally
+ * distributed with the following copyright and permission notice:
+ *
+ *      Pthreads-win32 - POSIX Threads Library for Win32
+ *      Copyright(C) 1998 John E. Bossom
+ *      Copyright(C) 1999,2005 Pthreads-win32 contributors
+ *
+ *      This library is free software; you can redistribute it and/or
+ *      modify it under the terms of the GNU Lesser General Public
+ *      License as published by the Free Software Foundation; either
+ *      version 2 of the License, or (at your option) any later version.
+ *
  * This program is distributed in the hope that it will be useful,
  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
@@ -19,7 +32,6 @@
 
 #define WIN32_LEAN_AND_MEAN
 #include <windows.h>
-#include "windows-once.h"
 
 static DWORD tls_key;
 #define tls_key_valid (tls_key != TLS_OUT_OF_INDEXES)
@@ -36,16 +48,113 @@ static BOOL tls_set(void *p)
 	return TlsSetValue(tls_key, p);
 }
 
-static void init_once_with_tls(void)
+/*
+ * Synchronize with another thread's call to init_once_wait on the same
+ * HANDLE object, which must be initialized to null.  If init_once_wait
+ * got there first, we receive a handle to the event object which can
+ * then be signaled.
+ */
+static void init_once_signal(HANDLE *ep)
 {
-	tls_key = TlsAlloc();
-	init_once_cb();
+	/*
+	 * Note that INVALID_HANDLE_VALUE is distinct from a null pointer
+	 * and also distinct from any handle returned by CreateEvent.
+	 */
+	HANDLE e = INVALID_HANDLE_VALUE;
+
+	if ((e = InterlockedCompareExchangePointer(ep, e, NULL)))
+		SetEvent(e);
 }
 
+/*
+ * Synchronize with another thread's call to init_once_signal on the same
+ * HANDLE object.
+ */
+static void init_once_wait(HANDLE *ep)
+{
+#if _WIN64
+	if (!InterlockedAdd64((LONG64 *)ep, 0)) // load with memory barrier
+#else
+	if (!InterlockedAdd((LONG *)ep, 0))
+#endif
+	{
+		HANDLE e;
+
+		e = CreateEvent(NULL, FALSE, FALSE, NULL);
+		if (!InterlockedCompareExchangePointer(ep, e, NULL))
+			WaitForSingleObject(e, INFINITE);
+		CloseHandle(e);
+	}
+}
+
+/*
+ * This implementation using an MCS lock variation requires only a single
+ * pointer of shared global state initialized to null, and in the uncontended
+ * case does not require allocation of any Windows resources whatsoever.
+ *
+ * These locks are described in the paper:
+ *
+ *  "Algorithms for Scalable Synchronization on Shared-Memory Multiprocessors"
+ *   by John M. Mellor-Crummey and Michael L. Scott.
+ *   ACM Transactions on Computer Systems Volume 9, Issue 1 (Feb. 1991).
+ *
+ * The basic idea is that each thread has a local state, which includes a
+ * pointer to the next waiting thread.  The global state is a tail pointer to
+ * the last waiting thread.  The running thread holds the lock and also the
+ * pointer to the first waiter.
+ *
+ * On acquire, atomically swap our fresh local state with the global tail
+ * pointer, becoming the new last waiter.  We receive a pointer to the previous
+ * last waiter (or nothing, in the unlocked case).  At this point it is safe
+ * for a new thread to come along and update the tail pointer again.  If
+ * needed, we then update the last waiter to point to our thread, signal
+ * that this is completed, and then wait to be signaled.
+ *
+ * On release, if the tail pointer points to us there are no waiters, and this
+ * can be confirmed with an atomic compare and exchange to the done state,
+ * which is equivalent to the original state except that a subsequent acquirer
+ * will know that the initialization has been previously completed.
+ *
+ * If that didn't unlock the lock, we need to wait for the signal from the next
+ * thread (which may not have updated our next pointer yet), then signal the
+ * next thread to wake up.  Eventually the queue will empty and the lock is
+ * left in the done state, at which point a simple atomic load can determine
+ * that nothing else needs to happen.
+ */
 static int init_once(void)
 {
-	static glwthread_once_t ctrl = GLWTHREAD_ONCE_INIT;
+	struct once_state * const init_done = (void *)-1;
+
+	struct once_state {
+		struct once_state *next;
+		HANDLE ready_event, next_event;
+	};
+
+	static void * volatile tail;
+	struct once_state local = {0};
+	struct once_state *p;
+
+	/* fast path for the normal (init completed) case. */
+	if (tail == init_done)
+		return 1;
+
+	if (!(p = InterlockedExchangePointer(&tail, &local))) {
+		/* we're number one! */
+		tls_key = TlsAlloc();
+		init_once_cb();
+	} else if (p != init_done) {
+		/* contended, wait for predecessor */
+		p->next = &local;
+		init_once_signal(&p->next_event);
+		init_once_wait(&local.ready_event);
+	}
+
+	p = InterlockedCompareExchangePointer(&tail, init_done, &local);
+	if (p != &local) {
+		/* contended, wait for successor */
+		init_once_wait(&local.next_event);
+		init_once_signal(&local.next->ready_event);
+	}
 
-	glwthread_once(&ctrl, init_once_with_tls);
 	return 1;
 }