ARM PMD disallow half-written entries crashing things

On ARM with the 2-level page table format, a PMD entry is represented by
two consecutive section entries covering 2MB of virtual space.

However, static mappings always were allowed to use separate 1MB section
entries.  This means in practice that a static mapping may create half
populated PMDs via create_mapping().

Since commit 0536bdf33f (ARM: move iotable mappings within the vmalloc
region) those static mappings are located in the vmalloc area. We must
ensure no such half populated PMDs are accessible once vmalloc() or
ioremap() start looking at the vmalloc area for nearby free virtual
address ranges, or various things leading to a kernel crash will happen.

Signed-off-by: Nicolas Pitre <npitre@linaro.org>
Reported-by: Santosh Shilimkar
Tested-by: R, Sricharan
Reviewed-by: Catalin Marinas
Cc: stable@vger.kernel.org

1 files changed, 74 insertions, 0 deletions

diff --git a/arch/arm/mm/mmu.c b/arch/arm/mm/mmu.c
index 8d10225e0ff0..d9085a644feb 100644
--- a/arch/arm/mm/mmu.c
+++ b/arch/arm/mm/mmu.c
@@ -814,6 +814,79 @@ void __init iotable_init(struct map_desc *io_desc, int nr)
 	}
 }
 
+#ifndef CONFIG_ARM_LPAE
+
+/*
+ * The Linux PMD is made of two consecutive section entries covering 2MB
+ * (see definition in include/asm/pgtable-2level.h).  However a call to
+ * create_mapping() may optimize static mappings by using individual
+ * 1MB section mappings.  This leaves the actual PMD potentially half
+ * initialized if the top or bottom section entry isn't used, leaving it
+ * open to problems if a subsequent ioremap() or vmalloc() tries to use
+ * the virtual space left free by that unused section entry.
+ *
+ * Let's avoid the issue by inserting dummy vm entries covering the unused
+ * PMD halves once the static mappings are in place.
+ */
+
+static void __init pmd_empty_section_gap(unsigned long addr)
+{
+	struct vm_struct *vm;
+
+	vm = early_alloc_aligned(sizeof(*vm), __alignof__(*vm));
+	vm->addr = (void *)addr;
+	vm->size = SECTION_SIZE;
+	vm->flags = VM_IOREMAP | VM_ARM_STATIC_MAPPING;
+	vm->caller = pmd_empty_section_gap;
+	vm_area_add_early(vm);
+}
+
+static void __init fill_pmd_gaps(void)
+{
+	struct vm_struct *vm;
+	unsigned long addr, next = 0;
+	pmd_t *pmd;
+
+	/* we're still single threaded hence no lock needed here */
+	for (vm = vmlist; vm; vm = vm->next) {
+		if (!(vm->flags & VM_ARM_STATIC_MAPPING))
+			continue;
+		addr = (unsigned long)vm->addr;
+		if (addr < next)
+			continue;
+
+		/*
+		 * Check if this vm starts on an odd section boundary.
+		 * If so and the first section entry for this PMD is free
+		 * then we block the corresponding virtual address.
+		 */
+		if ((addr & ~PMD_MASK) == SECTION_SIZE) {
+			pmd = pmd_off_k(addr);
+			if (pmd_none(*pmd))
+				pmd_empty_section_gap(addr & PMD_MASK);
+		}
+
+		/*
+		 * Then check if this vm ends on an odd section boundary.
+		 * If so and the second section entry for this PMD is empty
+		 * then we block the corresponding virtual address.
+		 */
+		addr += vm->size;
+		if ((addr & ~PMD_MASK) == SECTION_SIZE) {
+			pmd = pmd_off_k(addr) + 1;
+			if (pmd_none(*pmd))
+				pmd_empty_section_gap(addr);
+		}
+
+		/* no need to look at any vm entry until we hit the next PMD */
+		next = (addr + PMD_SIZE - 1) & PMD_MASK;
+	}
+}
+
+#else
+#define fill_pmd_gaps() do { } while (0)
+#endif
+
 static void * __initdata vmalloc_min =
 	(void *)(VMALLOC_END - (240 << 20) - VMALLOC_OFFSET);
 
@@ -1095,6 +1168,7 @@ static void __init devicemaps_init(struct machine_desc *mdesc)
 	 */
 	if (mdesc->map_io)
 		mdesc->map_io();
+	fill_pmd_gaps();
 
 	/*
 	 * Finally flush the caches and tlb to ensure that we're in a