On Fri, Jul 06, 2018 at 03:36:04PM +0100, Mark Rutland wrote: > On Fri, Jul 06, 2018 at 07:25:40AM -0700, Matthew Wilcox wrote: > > On Fri, Jul 06, 2018 at 02:41:44PM +0100, Mark Rutland wrote: > > > When idr_alloc() is called for the first time on an IDR (which has no > > > nodes in its radix tree), we end up with calculate_count() calling > > > get_slot_offset() with a NULL node, leading to a NULL pointer > > > dereference caught by UBSAN: > > > > Thanks! > > > > > The result of the load is passed into node_tag_get(), which ignores the > > > value when node is NULL. Typically, the compiler inlines both > > > get_slot_offset() and node_tag_get() into calculate_count(), optimizing > > > away the NULL-pointer dereference, and hence this doesn't typically > > > result in a boot failure. > > > > > > We can't rely on the compiler always doing this, and must avoid > > > dereferencing fields from node when it is potentially NULL. > > > > > > To do so, this patch folds the generation of offset into tag_get(), such > > > that this only happens when node is not NULL. Callers are updated to > > > pass the relevant slot, rather than the offset derived from it. > > > > I did think about this ... honest! My reasoning clearly glitched at some > > point. I think this is the right way to fix the problem. > > > > Acked-by: Matthew Wilcox <wi...@infradead.org> > > Cheers! > > I assume that Andrew will pick this up, if he's also happy with it.
I've just started fuzzing, and found this also applies with node_tag_set(). I'll spin a v2 with that fixed up, too. Thanks, Mark.
