mm: vmap fix overflow
Nick Piggin [Fri, 27 Feb 2009 22:03:03 +0000 (14:03 -0800)]
The new vmap allocator can wrap the address and get confused in the case
of large allocations or VMALLOC_END near the end of address space.

Problem reported by Christoph Hellwig on a 32-bit XFS workload.

Signed-off-by: Nick Piggin <npiggin@suse.de>
Reported-by: Christoph Hellwig <hch@lst.de>
Cc: <stable@kernel.org> [2.6.28.x]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm/vmalloc.c

index 903cad4..ed3705e 100644 (file)
@@ -323,6 +323,7 @@ static struct vmap_area *alloc_vmap_area(unsigned long size,
        unsigned long addr;
        int purged = 0;
 
+       BUG_ON(!size);
        BUG_ON(size & ~PAGE_MASK);
 
        va = kmalloc_node(sizeof(struct vmap_area),
@@ -334,6 +335,9 @@ retry:
        addr = ALIGN(vstart, align);
 
        spin_lock(&vmap_area_lock);
+       if (addr + size - 1 < addr)
+               goto overflow;
+
        /* XXX: could have a last_hole cache */
        n = vmap_area_root.rb_node;
        if (n) {
@@ -365,6 +369,8 @@ retry:
 
                while (addr + size > first->va_start && addr + size <= vend) {
                        addr = ALIGN(first->va_end + PAGE_SIZE, align);
+                       if (addr + size - 1 < addr)
+                               goto overflow;
 
                        n = rb_next(&first->rb_node);
                        if (n)
@@ -375,6 +381,7 @@ retry:
        }
 found:
        if (addr + size > vend) {
+overflow:
                spin_unlock(&vmap_area_lock);
                if (!purged) {
                        purge_vmap_area_lazy();