This patch fixes a race between when the disable cpu is marked online

and binding IPIs back to the CPU.  In some cases, an IPI would be sent
to CPU1 before it had allocated a new evtchn.  Moving smp_resume() call
before setting the cpu online fixes this race.  This fixes bug #228
(http://bugzilla.xensource.com/bugzilla/show_bug.cgi?id=228).

Signed-off-by: Ryan Harper <ryanh@us.ibm.com>

diff --git a/linux-2.6-xen-sparse/arch/xen/i386/kernel/process.c b/linux-2.6-xen-sparse/arch/xen/i386/kernel/process.c
index ca08c3bd2eafac7bc749a0546148b9b483585599..058fd3797cc542242b7f01b6aef8d21c722885b8 100644 (file)
--- a/linux-2.6-xen-sparse/arch/xen/i386/kernel/process.c
+++ b/linux-2.6-xen-sparse/arch/xen/i386/kernel/process.c
@@ -112,6 +112,10 @@ void xen_idle(void)
 
 #ifdef CONFIG_HOTPLUG_CPU
 #include <asm/nmi.h>
+#ifdef CONFIG_SMP
+extern void smp_suspend(void);
+extern void smp_resume(void);
+#endif
 /* We don't actually take CPU down, just spin without interrupts. */
 static inline void play_dead(void)
 {
@@ -120,6 +124,13 @@ static inline void play_dead(void)
                HYPERVISOR_yield();
 
        __flush_tlb_all();
+   /* 
+    * Restore IPI/IRQ mappings before marking online to prevent 
+    * race between pending interrupts and restoration of handler. 
+    */
+#ifdef CONFIG_SMP
+       smp_resume();
+#endif
        cpu_set(smp_processor_id(), cpu_online_map);
 }
 #else
@@ -135,10 +146,6 @@ static inline void play_dead(void)
  * low exit latency (ie sit in a loop waiting for
  * somebody to say that they'd like to reschedule)
  */
-#ifdef CONFIG_SMP
-extern void smp_suspend(void);
-extern void smp_resume(void);
-#endif
 void cpu_idle (void)
 {
        int cpu = _smp_processor_id();
@@ -166,9 +173,6 @@ void cpu_idle (void)
                                HYPERVISOR_vcpu_down(cpu);
 #endif
                                play_dead();
-#ifdef CONFIG_SMP
-                               smp_resume();
-#endif
                                local_irq_enable();
                        }