s390/mm: four page table levels vs. fork

The fork of a process with four page table levels is broken since
git commit 6252d702 "[S390] dynamic page tables."

All new mm contexts are created with three page table levels and
an asce limit of 4TB. If the parent has four levels dup_mmap will
add vmas to the new context which are outside of the asce limit.
The subsequent call to copy_page_range will walk the three level
page table structure of the new process with non-zero pgd and pud
indexes. This leads to memory clobbers as the pgd_index *and* the
pud_index is added to the mm->pgd pointer without a pgd_deref
in between.

The init_new_context() function is selecting the number of page
table levels for a new context. The function is used by mm_init()
which in turn is called by dup_mm() and mm_alloc(). These two are
used by fork() and exec(). The init_new_context() function can
distinguish the two cases by looking at mm->context.asce_limit,
for fork() the mm struct has been copied and the number of page
table levels may not change. For exec() the mm_alloc() function
set the new mm structure to zero, in this case a three-level page
table is created as the temporary stack space is located at
STACK_TOP_MAX = 4TB.

This fixes CVE-2016-2143.
Reported-by: Marcin Kościelnicki <koriakin@0x04.net>
Reviewed-by: Heiko Carstens <heiko.carstens@de.ibm.com>
Cc: stable@vger.kernel.org
Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com>

arch/s390/include/asm/mmu_context.h

@@ -15,17 +15,25 @@
 static inline int init_new_context(struct task_struct *tsk,
 				   struct mm_struct *mm)
 {
+	spin_lock_init(&mm->context.list_lock);
+	INIT_LIST_HEAD(&mm->context.pgtable_list);
+	INIT_LIST_HEAD(&mm->context.gmap_list);
 	cpumask_clear(&mm->context.cpu_attach_mask);
 	atomic_set(&mm->context.attach_count, 0);
 	mm->context.flush_mm = 0;
-	mm->context.asce_bits = _ASCE_TABLE_LENGTH | _ASCE_USER_BITS;
-	mm->context.asce_bits |= _ASCE_TYPE_REGION3;
 #ifdef CONFIG_PGSTE
 	mm->context.alloc_pgste = page_table_allocate_pgste;
 	mm->context.has_pgste = 0;
 	mm->context.use_skey = 0;
 #endif
-	mm->context.asce_limit = STACK_TOP_MAX;
+	if (mm->context.asce_limit == 0) {
+		/* context created by exec, set asce limit to 4TB */
+		mm->context.asce_bits = _ASCE_TABLE_LENGTH |
+			_ASCE_USER_BITS | _ASCE_TYPE_REGION3;
+		mm->context.asce_limit = STACK_TOP_MAX;
+	} else if (mm->context.asce_limit == (1UL << 31)) {
+		mm_inc_nr_pmds(mm);
+	}
 	crst_table_init((unsigned long *) mm->pgd, pgd_entry_type(mm));
 	return 0;
 }
@@ -111,8 +119,6 @@ static inline void activate_mm(struct mm_struct *prev,
 static inline void arch_dup_mmap(struct mm_struct *oldmm,
 				 struct mm_struct *mm)
 {
-	if (oldmm->context.asce_limit < mm->context.asce_limit)
-		crst_table_downgrade(mm, oldmm->context.asce_limit);
 }
 
 static inline void arch_exit_mmap(struct mm_struct *mm)

arch/s390/include/asm/pgalloc.h

@@ -100,12 +100,26 @@ static inline void pud_populate(struct mm_struct *mm, pud_t *pud, pmd_t *pmd)
 
 static inline pgd_t *pgd_alloc(struct mm_struct *mm)
 {
-	spin_lock_init(&mm->context.list_lock);
-	INIT_LIST_HEAD(&mm->context.pgtable_list);
-	INIT_LIST_HEAD(&mm->context.gmap_list);
-	return (pgd_t *) crst_table_alloc(mm);
+	unsigned long *table = crst_table_alloc(mm);
+
+	if (!table)
+		return NULL;
+	if (mm->context.asce_limit == (1UL << 31)) {
+		/* Forking a compat process with 2 page table levels */
+		if (!pgtable_pmd_page_ctor(virt_to_page(table))) {
+			crst_table_free(mm, table);
+			return NULL;
+		}
+	}
+	return (pgd_t *) table;
+}
+
+static inline void pgd_free(struct mm_struct *mm, pgd_t *pgd)
+{
+	if (mm->context.asce_limit == (1UL << 31))
+		pgtable_pmd_page_dtor(virt_to_page(pgd));
+	crst_table_free(mm, (unsigned long *) pgd);
 }
-#define pgd_free(mm, pgd) crst_table_free(mm, (unsigned long *) pgd)
 
 static inline void pmd_populate(struct mm_struct *mm,
 				pmd_t *pmd, pgtable_t pte)