cpus-common: simplify locking for start_exclusive/end_exclusive

It is not necessary to hold qemu_cpu_list_mutex throughout the
exclusive section, because no other exclusive section can run
while pending_cpus != 0.

exclusive_idle() is called in cpu_exec_start(), and that prevents
any CPUs created after start_exclusive() from entering cpu_exec()
during an exclusive section.

Reviewed-by: Richard Henderson <rth@twiddle.net>
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

cpus-common.c

@@ -171,8 +171,7 @@ static inline void exclusive_idle(void)
 }
 
 /* Start an exclusive operation.
-   Must only be called from outside cpu_exec, takes
-   qemu_cpu_list_lock.   */
+   Must only be called from outside cpu_exec.  */
 void start_exclusive(void)
 {
     CPUState *other_cpu;
@@ -191,11 +190,17 @@ void start_exclusive(void)
     while (pending_cpus > 1) {
         qemu_cond_wait(&exclusive_cond, &qemu_cpu_list_lock);
     }
+
+    /* Can release mutex, no one will enter another exclusive
+     * section until end_exclusive resets pending_cpus to 0.
+     */
+    qemu_mutex_unlock(&qemu_cpu_list_lock);
 }
 
-/* Finish an exclusive operation.  Releases qemu_cpu_list_lock.  */
+/* Finish an exclusive operation.  */
 void end_exclusive(void)
 {
+    qemu_mutex_lock(&qemu_cpu_list_lock);
     pending_cpus = 0;
     qemu_cond_broadcast(&exclusive_resume);
     qemu_mutex_unlock(&qemu_cpu_list_lock);

docs/tcg-exclusive.promela

@@ -98,9 +98,11 @@ byte has_waiter[N_CPUS];
     do                                                            \
       :: pending_cpus > 1 -> COND_WAIT(exclusive_cond, mutex);    \
       :: else             -> break;                               \
-    od
+    od;                                                           \
+    MUTEX_UNLOCK(mutex);
 
 #define end_exclusive()                                           \
+    MUTEX_LOCK(mutex);                                            \
     pending_cpus = 0;                                             \
     COND_BROADCAST(exclusive_resume);                             \
     MUTEX_UNLOCK(mutex);

include/qom/cpu.h

@@ -846,17 +846,13 @@ void cpu_exec_end(CPUState *cpu);
  * cpu_exec are exited immediately.  CPUs that call cpu_exec_start
  * during the exclusive section go to sleep until this CPU calls
  * end_exclusive.
- *
- * Returns with the CPU list lock taken (which nests outside all
- * other locks except the BQL).
  */
 void start_exclusive(void);
 
 /**
  * end_exclusive:
  *
  * Concludes an exclusive execution section started by start_exclusive.
- * Releases the CPU list lock.
  */
 void end_exclusive(void);