arch/tile: fix up locking in pgtable.c slightly

We should be holding the init_mm.page_table_lock in shatter_huge_page()
since we are modifying the kernel page tables.  Then, only if we are
walking the other root page tables to update them, do we want to take
the pgd_lock.

Add a comment about taking the pgd_lock that we always do it with
interrupts disabled and therefore are not at risk from the tlbflush
IPI deadlock as is seen on x86.

Signed-off-by: Chris Metcalf <cmetcalf@tilera.com>

arch/tile/mm/pgtable.c

@@ -177,14 +177,10 @@ void shatter_huge_page(unsigned long addr)
 	if (!pmd_huge_page(*pmd))
 		return;
 
-	/*
-	 * Grab the pgd_lock, since we may need it to walk the pgd_list,
-	 * and since we need some kind of lock here to avoid races.
-	 */
-	spin_lock_irqsave(&pgd_lock, flags);
+	spin_lock_irqsave(&init_mm.page_table_lock, flags);
 	if (!pmd_huge_page(*pmd)) {
 		/* Lost the race to convert the huge page. */
-		spin_unlock_irqrestore(&pgd_lock, flags);
+		spin_unlock_irqrestore(&init_mm.page_table_lock, flags);
 		return;
 	}
 
@@ -194,21 +190,23 @@ void shatter_huge_page(unsigned long addr)
 
 #ifdef __PAGETABLE_PMD_FOLDED
 	/* Walk every pgd on the system and update the pmd there. */
+	spin_lock(&pgd_lock);
 	list_for_each(pos, &pgd_list) {
 		pmd_t *copy_pmd;
 		pgd = list_to_pgd(pos) + pgd_index(addr);
 		pud = pud_offset(pgd, addr);
 		copy_pmd = pmd_offset(pud, addr);
 		__set_pmd(copy_pmd, *pmd);
 	}
+	spin_unlock(&pgd_lock);
 #endif
 
 	/* Tell every cpu to notice the change. */
 	flush_remote(0, 0, NULL, addr, HPAGE_SIZE, HPAGE_SIZE,
 		     cpu_possible_mask, NULL, 0);
 
 	/* Hold the lock until the TLB flush is finished to avoid races. */
-	spin_unlock_irqrestore(&pgd_lock, flags);
+	spin_unlock_irqrestore(&init_mm.page_table_lock, flags);
 }
 
 /*
@@ -217,9 +215,13 @@ void shatter_huge_page(unsigned long addr)
  * against pageattr.c; it is the unique case in which a valid change
  * of kernel pagetables can't be lazily synchronized by vmalloc faults.
  * vmalloc faults work because attached pagetables are never freed.
- * The locking scheme was chosen on the basis of manfred's
- * recommendations and having no core impact whatsoever.
- * -- wli
+ *
+ * The lock is always taken with interrupts disabled, unlike on x86
+ * and other platforms, because we need to take the lock in
+ * shatter_huge_page(), which may be called from an interrupt context.
+ * We are not at risk from the tlbflush IPI deadlock that was seen on
+ * x86, since we use the flush_remote() API to have the hypervisor do
+ * the TLB flushes regardless of irq disabling.
  */
 DEFINE_SPINLOCK(pgd_lock);
 LIST_HEAD(pgd_list);