Ensure good ordering of memory instruction in cpu_exec

The IO thread, when it senses cpu_single_env == 0, expects exit_request
to be checked later on. A compiler scheduling constraint is not strong
enough to ensure this on modern architecture. A memory fence is needed
as well.

Signed-off-by: Olivier Hainque <hainque@adacore.com>
Signed-off-by: Fabien Chouteau <chouteau@adacore.com>
Reviewed-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Stefan Weil <sw@weilnetz.de>

diff --git a/cpu-exec.c b/cpu-exec.c
index e74e55656a4028579562615b010472d999d3176a..aa8fa893d9cafffbc13fcc1e5dce713d15e23648 100644 (file)
--- a/cpu-exec.c
+++ b/cpu-exec.c
@@ -217,6 +217,14 @@ int cpu_exec(CPUArchState *env)
 
     cpu_single_env = env;
 
+    /* As long as cpu_single_env is null, up to the assignment just above,
+     * requests by other threads to exit the execution loop are expected to
+     * be issued using the exit_request global. We must make sure that our
+     * evaluation of the global value is performed past the cpu_single_env
+     * value transition point, which requires a memory barrier as well as
+     * an instruction scheduling constraint on modern architectures.  */
+    smp_mb();
+
     if (unlikely(exit_request)) {
         cpu->exit_request = 1;
     }