s390/mm: optimize page table upgrade routine

There is a maximum of two new tables allocated on page table
upgrade. Because we know that a loop the current implementation
is based on could be unrolled with some improvements:

  * upgrade from 3 to 5 levels happens in one go - without an
    unnecessary re-take of page_table_lock in-between;

  * page tables initialization moved out of the atomic code;
Signed-off-by: Alexander Gordeev <agordeev@linux.ibm.com>
Reviewed-by: Heiko Carstens <heiko.carstens@de.ibm.com>
Signed-off-by: Heiko Carstens <heiko.carstens@de.ibm.com>
Signed-off-by: Vasily Gorbik <gor@linux.ibm.com>

arch/s390/mm/pgalloc.c

@@ -77,43 +77,65 @@ static void __crst_table_upgrade(void *arg)
 
 int crst_table_upgrade(struct mm_struct *mm, unsigned long end)
 {
-	unsigned long *table, *pgd;
-	int rc, notify;
+	unsigned long *pgd = NULL, *p4d = NULL, *__pgd;
+	unsigned long asce_limit = mm->context.asce_limit;
 
 	/* upgrade should only happen from 3 to 4, 3 to 5, or 4 to 5 levels */
-	VM_BUG_ON(mm->context.asce_limit < _REGION2_SIZE);
-	rc = 0;
-	notify = 0;
-	while (mm->context.asce_limit < end) {
-		table = crst_table_alloc(mm);
-		if (!table) {
-			rc = -ENOMEM;
-			break;
-		}
-		spin_lock_bh(&mm->page_table_lock);
-		pgd = (unsigned long *) mm->pgd;
-		if (mm->context.asce_limit == _REGION2_SIZE) {
-			crst_table_init(table, _REGION2_ENTRY_EMPTY);
-			p4d_populate(mm, (p4d_t *) table, (pud_t *) pgd);
-			mm->pgd = (pgd_t *) table;
-			mm->context.asce_limit = _REGION1_SIZE;
-			mm->context.asce = __pa(mm->pgd) | _ASCE_TABLE_LENGTH |
-				_ASCE_USER_BITS | _ASCE_TYPE_REGION2;
-			mm_inc_nr_puds(mm);
-		} else {
-			crst_table_init(table, _REGION1_ENTRY_EMPTY);
-			pgd_populate(mm, (pgd_t *) table, (p4d_t *) pgd);
-			mm->pgd = (pgd_t *) table;
-			mm->context.asce_limit = -PAGE_SIZE;
-			mm->context.asce = __pa(mm->pgd) | _ASCE_TABLE_LENGTH |
-				_ASCE_USER_BITS | _ASCE_TYPE_REGION1;
-		}
-		notify = 1;
-		spin_unlock_bh(&mm->page_table_lock);
+	VM_BUG_ON(asce_limit < _REGION2_SIZE);
+
+	if (end <= asce_limit)
+		return 0;
+
+	if (asce_limit == _REGION2_SIZE) {
+		p4d = crst_table_alloc(mm);
+		if (unlikely(!p4d))
+			goto err_p4d;
+		crst_table_init(p4d, _REGION2_ENTRY_EMPTY);
 	}
-	if (notify)
-		on_each_cpu(__crst_table_upgrade, mm, 0);
-	return rc;
+	if (end > _REGION1_SIZE) {
+		pgd = crst_table_alloc(mm);
+		if (unlikely(!pgd))
+			goto err_pgd;
+		crst_table_init(pgd, _REGION1_ENTRY_EMPTY);
+	}
+
+	spin_lock_bh(&mm->page_table_lock);
+
+	/*
+	 * This routine gets called with mmap_sem lock held and there is
+	 * no reason to optimize for the case of otherwise. However, if
+	 * that would ever change, the below check will let us know.
+	 */
+	VM_BUG_ON(asce_limit != mm->context.asce_limit);
+
+	if (p4d) {
+		__pgd = (unsigned long *) mm->pgd;
+		p4d_populate(mm, (p4d_t *) p4d, (pud_t *) __pgd);
+		mm->pgd = (pgd_t *) p4d;
+		mm->context.asce_limit = _REGION1_SIZE;
+		mm->context.asce = __pa(mm->pgd) | _ASCE_TABLE_LENGTH |
+			_ASCE_USER_BITS | _ASCE_TYPE_REGION2;
+		mm_inc_nr_puds(mm);
+	}
+	if (pgd) {
+		__pgd = (unsigned long *) mm->pgd;
+		pgd_populate(mm, (pgd_t *) pgd, (p4d_t *) __pgd);
+		mm->pgd = (pgd_t *) pgd;
+		mm->context.asce_limit = -PAGE_SIZE;
+		mm->context.asce = __pa(mm->pgd) | _ASCE_TABLE_LENGTH |
+			_ASCE_USER_BITS | _ASCE_TYPE_REGION1;
+	}
+
+	spin_unlock_bh(&mm->page_table_lock);
+
+	on_each_cpu(__crst_table_upgrade, mm, 0);
+
+	return 0;
+
+err_pgd:
+	crst_table_free(mm, p4d);
+err_p4d:
+	return -ENOMEM;
 }
 
 void crst_table_downgrade(struct mm_struct *mm)