Re: [PATCH net] vhost: Use kzalloc() to allocate vhost_msg_node

[Jason Wang]

On 2018å04æ28æ 09:51, Kevin Easton wrote:
> On Fri, Apr 27, 2018 at 09:07:56PM -0400, Kevin Easton wrote:
> > On Fri, Apr 27, 2018 at 07:05:45PM +0300, Michael S. Tsirkin wrote:
> > > On Fri, Apr 27, 2018 at 11:45:02AM -0400, Kevin Easton wrote:
> > > > The struct vhost_msg within struct vhost_msg_node is copied to userspace,
> > > > so it should be allocated with kzalloc() to ensure all structure padding
> > > > is zeroed.
> > > >
> > > > Signed-off-by: Kevin Easton <kevin@xxxxxxxxxxx>
> > > > Reported-by: syzbot+87cfa083e727a224754b@xxxxxxxxxxxxxxxxxxxxxxxxx
> > > Does it help if a patch naming the padding is applied,
> > > and then we init just the relevant field?
> > > Just curious.
> > No, I don't believe that is sufficient to fix the problem.
> Scratch that, somehow I missed the "..and then we init just the
> relevant field" part, sorry.
>
> There's still the padding after the vhost_iotlb_msg to consider.  It's
> named in the union but I don't think that's guaranteed to be initialised
> when the iotlb member of the union is used to initialise things.
>
> > I didn't name the padding in my original patch because I wasn't sure
> > if the padding actually exists on 32 bit architectures?
> This might still be a conce

Yes.

print &((struct vhost_msg *)0)->iotlb
$3 = (struct vhost_iotlb_msg *) 0x4


> At the end of the day, zeroing 96 bytes (the full size of vhost_msg_node)
> is pretty quick.
>
>      - Kevin

Right, and even if it may be used heavily in the data-path, zeroing is 
not the main delay in that path.

Thanks