uml: use barrier() instead of mb()

signals_enabled and pending have requirements on the order in which they are
modified.  This used to be done by declaring them volatile and putting an mb()
where the ordering requirements were in effect.

After getting a better (I hope) understanding of how to do this correctly, the
volatile declarations are gone and the mb()'s replaced by barrier()'s.

One of the mb()'s was deleted because I see no problematic writes that could
be re-ordered past that point.

Signed-off-by: Jeff Dike <jdike@linux.intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

diff --git a/arch/um/os-Linux/signal.c b/arch/um/os-Linux/signal.c
index 7ff8f57b7150c56fcce0495466fb98e8580d5810..62a66f38a9135d4976c597850cf69b8afca1b449 100644 (file)
--- a/arch/um/os-Linux/signal.c
+++ b/arch/um/os-Linux/signal.c
@@ -15,6 +15,9 @@
 #include "sysdep/sigcontext.h"
 #include "user.h"
 
+/* Copied from linux/compiler-gcc.h since we can't include it directly */
+#define barrier() __asm__ __volatile__("": : :"memory")
+
 /*
  * These are the asynchronous signals.  SIGPROF is excluded because we want to
  * be able to profile all of UML, not just the non-critical sections.  If
@@ -27,13 +30,8 @@
 #define SIGVTALRM_BIT 1
 #define SIGVTALRM_MASK (1 << SIGVTALRM_BIT)
 
-/*
- * These are used by both the signal handlers and
- * block/unblock_signals.  I don't want modifications cached in a
- * register - they must go straight to memory.
- */
-static volatile int signals_enabled = 1;
-static volatile int pending = 0;
+static int signals_enabled;
+static unsigned int pending;
 
 void sig_handler(int sig, struct sigcontext *sc)
 {
@@ -198,7 +196,7 @@ void block_signals(void)
         * This might matter if gcc figures out how to inline this and
         * decides to shuffle this code into the caller.
         */
-       mb();
+       barrier();
 }
 
 void unblock_signals(void)
@@ -224,21 +222,11 @@ void unblock_signals(void)
                 * Setting signals_enabled and reading pending must
                 * happen in this order.
                 */
-               mb();
+               barrier();
 
                save_pending = pending;
-               if (save_pending == 0) {
-                       /*
-                        * This must return with signals enabled, so
-                        * this barrier ensures that writes are
-                        * flushed out before the return.  This might
-                        * matter if gcc figures out how to inline
-                        * this (unlikely, given its size) and decides
-                        * to shuffle this code into the caller.
-                        */
-                       mb();
+               if (save_pending == 0)
                        return;
-               }
 
                pending = 0;