Re: [PATCH] LDT improvements

[Ingo Molnar]

* Andy Lutomirski <luto@xxxxxxxxxxxxxx> wrote:

> 
> 
> > On Dec 7, 2017, at 9:23 AM, Thomas Gleixner <tglx@xxxxxxxxxxxxx> wrote:
> > 
> >> On Thu, 7 Dec 2017, Andy Lutomirski wrote:
> >> 
> >>> On Thu, Dec 7, 2017 at 4:43 AM, Borislav Petkov <bp@xxxxxxxxx> wrote:
> >>>> On Wed, Dec 06, 2017 at 11:22:21PM -0800, Andy Lutomirski wrote:
> >>>> I think I like this approach.  I also think it might be nice to move the
> >>>> whole cpu_entry_area into this new pgd range so that we can stop mucking
> >>>> around with the fixmap.
> >>> 
> >>> Yeah, and also, I don't like the idea of sacrificing a whole PGD
> >>> only for the LDT crap which is optional, even. Frankly - and this
> >>> is just me - I'd make CONFIG_KERNEL_PAGE_TABLE_ISOLATION xor
> >>> CONFIG_MODIFY_LDT_SYSCALL and don't give a rat's *ss about the LDT.
> >> 
> >> The PGD sacrifice doesn't bother me.  Putting a writable LDT map at a
> >> constant address does bother me.  We could probably get away with RO
> >> if we trapped and handled the nasty faults, but that could be very
> >> problematic.
> > 
> > Where is the problem? You can map it RO into user space with the USER bit
> > cleared. The kernel knows how to access the real stuff.
> 
> Blows up when the CPU tries to set the accessed bit.

BTW., could we force the accessed bit to be always set, without breaking the ABI?

> > The approach I've taken is to create a VMA and map it into user space with
> > the USER bit cleared. A little bit more effort code wise, but that avoids
> > all the page table muck and keeps it straight attached to the process.
> > 
> > Will post once in a bit.
> 
> I don't love mucking with user address space.  I'm also quite nervous about 
> putting it in our near anything that could pass an access_ok check, since we're 
> totally screwed if the bad guys can figure out how to write to it.

Hm, robustness of the LDT address wrt. access_ok() is a valid concern.

Can we have vmas with high addresses, in the vmalloc space for example?
IIRC the GPU code has precedents in that area.

Since this is x86-64, limitation of the vmalloc() space is not an issue.

I like Thomas's solution:

 - have the LDT in a regular mmap space vma (hence per process ASLR randomized), 
   but with the system bit set.

 - That would be an advantage even for non-PTI kernels, because mmap() is probably 
   more randomized than kmalloc().

 - It would also be a cleaner approach all around, and would avoid the fixmap
   complications and the scheduler muckery.

Thanks,

	Ingo