[IA64] Fix race in mm-context wrap-around logic.

The patch below should fix a race which could cause stale TLB entries.
Specifically, when 2 CPUs ended up racing for entrance to
wrap_mmu_context(). The losing CPU would find that by the time it
acquired ctx.lock, mm->context already had a valid value, but then it
failed to (re-)check the delayed TLB flushing logic and hence could
end up using a context number when there were still stale entries in
its TLB. The fix is to check for delayed TLB flushes only after
mm->context is valid (non-zero). The patch also makes GCC v4.x
happier by defining a non-volatile variant of mm_context_t called
nv_mm_context_t.

Signed-off-by: David Mosberger-Tang <[e-mail]>
Signed-off-by: Tony Luck <[e-mail]>

include/asm-ia64/mmu.h

···
 #define __MMU_H
 
 /*
- * Type for a context number.  We declare it volatile to ensure proper ordering when it's
- * accessed outside of spinlock'd critical sections (e.g., as done in activate_mm() and
- * init_new_context()).
+ * Type for a context number.  We declare it volatile to ensure proper
+ * ordering when it's accessed outside of spinlock'd critical sections
+ * (e.g., as done in activate_mm() and init_new_context()).
  */
 typedef volatile unsigned long mm_context_t;
+
+typedef unsigned long nv_mm_context_t;
 
 #endif

include/asm-ia64/mmu_context.h

···
 delayed_tlb_flush (void)
 {
 	extern void local_flush_tlb_all (void);
+	unsigned long flags;
 
 	if (unlikely(__ia64_per_cpu_var(ia64_need_tlb_flush))) {
-		local_flush_tlb_all();
-		__ia64_per_cpu_var(ia64_need_tlb_flush) = 0;
+		spin_lock_irqsave(&ia64_ctx.lock, flags);
+		{
+			if (__ia64_per_cpu_var(ia64_need_tlb_flush)) {
+				local_flush_tlb_all();
+				__ia64_per_cpu_var(ia64_need_tlb_flush) = 0;
+			}
+		}
+		spin_unlock_irqrestore(&ia64_ctx.lock, flags);
 	}
 }
 
-static inline mm_context_t
+static inline nv_mm_context_t
 get_mmu_context (struct mm_struct *mm)
 {
 	unsigned long flags;
-	mm_context_t context = mm->context;
+	nv_mm_context_t context = mm->context;
 
-	if (context)
-		return context;
-
-	spin_lock_irqsave(&ia64_ctx.lock, flags);
-	{
-		/* re-check, now that we've got the lock: */
-		context = mm->context;
-		if (context == 0) {
-			cpus_clear(mm->cpu_vm_mask);
-			if (ia64_ctx.next >= ia64_ctx.limit)
-				wrap_mmu_context(mm);
-			mm->context = context = ia64_ctx.next++;
+	if (unlikely(!context)) {
+		spin_lock_irqsave(&ia64_ctx.lock, flags);
+		{
+			/* re-check, now that we've got the lock: */
+			context = mm->context;
+			if (context == 0) {
+				cpus_clear(mm->cpu_vm_mask);
+				if (ia64_ctx.next >= ia64_ctx.limit)
+					wrap_mmu_context(mm);
+				mm->context = context = ia64_ctx.next++;
+			}
 		}
+		spin_unlock_irqrestore(&ia64_ctx.lock, flags);
 	}
-	spin_unlock_irqrestore(&ia64_ctx.lock, flags);
+	/*
+	 * Ensure we're not starting to use "context" before any old
+	 * uses of it are gone from our TLB.
+	 */
+	delayed_tlb_flush();
+
 	return context;
 }
 
···
 }
 
 static inline void
-reload_context (mm_context_t context)
+reload_context (nv_mm_context_t context)
 {
 	unsigned long rid;
 	unsigned long rid_incr = 0;
···
 static inline void
 activate_context (struct mm_struct *mm)
 {
-	mm_context_t context;
+	nv_mm_context_t context;
 
 	do {
 		context = get_mmu_context(mm);
···
 static inline void
 activate_mm (struct mm_struct *prev, struct mm_struct *next)
 {
-	delayed_tlb_flush();
-
 	/*
 	 * We may get interrupts here, but that's OK because interrupt handlers cannot
 	 * touch user-space.