Re: [PATCH v4 1/3] x86: Introduce a new constant KERNEL_MAPPING_SIZE

[Baoquan He <bhe@redhat.com>]

On 03/03/17 at 11:07pm, Baoquan He wrote:
> On 03/03/17 at 03:28pm, Borislav Petkov wrote:
> > On Fri, Mar 03, 2017 at 09:11:52PM +0800, Baoquan He wrote:
> > > And another meaning of defining kernel iamge size and mapping size
> > > differently is we can randomize the limited kernel image in the mapping
> > > area. If they are the same or kernel image can be very large, the
> > > position will be fixed or very few, kernel text KASLR will be
> > > meaningless.
> > 
> > This is simply not true:
> > 
> > @@ -408,9 +408,9 @@ static unsigned long find_random_virt_addr(unsigned long minimum,
> >         /*
> >          * There are how many CONFIG_PHYSICAL_ALIGN-sized slots
> >          * that can hold image_size within the range of minimum to
> > -        * KERNEL_IMAGE_SIZE?
> > +        * KERNEL_MAPPING_SIZE?
> >          */
> > -       slots = (KERNEL_IMAGE_SIZE - minimum - image_size) /
> > +       slots = (KERNEL_MAPPING_SIZE - minimum - image_size) /
> >                  CONFIG_PHYSICAL_ALIGN + 1;
> > 
> > *With* kaslr, KERNEL_IMAGE_SIZE = 1G and KERNEL_MAPPING_SIZE = 1G.
> > Before your patch KERNEL_IMAGE_SIZE = 1G too with kaslr enabled.
> 
> 512M and 1G is the first case, just an example. Usually kernel image size
			~ worst, sorry, typo
> is only about 20M, from my laptop.
> 
> Yes, before KERNEL_IMAGE_SIZE is 1G with kaslr enabled. when you
> suggested taking a fixed size for the KERNEL_IMAGE_SIZE, but not changed
> back and forth with the kaslr set or not, I started to consider this.
> 
> See the 1G, hard-constrainted because of level2_kernel_pgt. In the
> future we could put kernel mapping area in another place to remove the
> 1G limitation, could be 10G or 512G since virtual address are so
> redundent, just an assumption, kernel KASLR can benefit from this
> actually, but we can't make upper value of kernel image size also be
> that big. That will make linker script checking lose meaning.
> 
> Thanks
> Baoquan
>