Spinlock helper function.
authordrepper <drepper>
Mon, 20 Apr 1998 18:13:00 +0000 (18:13 +0000)
committerdrepper <drepper>
Mon, 20 Apr 1998 18:13:00 +0000 (18:13 +0000)
linuxthreads/spinlock.c [new file with mode: 0644]

diff --git a/linuxthreads/spinlock.c b/linuxthreads/spinlock.c
new file mode 100644 (file)
index 0000000..170d9ae
--- /dev/null
@@ -0,0 +1,59 @@
+/* Linuxthreads - a simple clone()-based implementation of Posix        */
+/* threads for Linux.                                                   */
+/* Copyright (C) 1998 Xavier Leroy (Xavier.Leroy@inria.fr)              */
+/*                                                                      */
+/* This program is free software; you can redistribute it and/or        */
+/* modify it under the terms of the GNU Library 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        */
+/* GNU Library General Public License for more details.                 */
+
+/* Spin locks */
+
+#include <sched.h>
+#include <time.h>
+#include "pthread.h"
+#include "internals.h"
+#include "spinlock.h"
+
+/* This function is called if the inlined test-and-set in acquire() failed */
+
+/* The retry strategy is as follows:
+   - We test and set the spinlock MAX_SPIN_COUNT times, calling
+     sched_yield() each time.  This gives ample opportunity for other
+     threads with priority >= our priority to make progress and
+     release the spinlock.
+   - If a thread with priority < our priority owns the spinlock,
+     calling sched_yield() repeatedly is useless, since we're preventing
+     the owning thread from making progress and releasing the spinlock.
+     So, after MAX_SPIN_LOCK attemps, we suspend the calling thread
+     using nanosleep().  This again should give time to the owning thread
+     for releasing the spinlock.
+     Notice that the nanosleep() interval must not be too small,
+     since the kernel does busy-waiting for short intervals in a realtime
+     process (!).  The smallest duration that guarantees thread
+     suspension is currently 2ms.
+   - When nanosleep() returns, we try again, doing MAX_SPIN_COUNT
+     sched_yield(), then sleeping again if needed. */
+
+void __pthread_acquire(int * spinlock)
+{
+  int cnt = 0;
+  struct timespec tm;
+
+  while (testandset(spinlock)) {
+    if (cnt < MAX_SPIN_COUNT) {
+      sched_yield();
+      cnt++;
+    } else {
+      tm.tv_sec = 0;
+      tm.tv_nsec = SPIN_SLEEP_DURATION;
+      nanosleep(&tm, NULL);
+      cnt = 0;
+    }
+  }
+}