[PATCH v4] arm64: fix VTTBR_BADDR_MASK

[Christoffer Dall]

On Tue, Aug 19, 2014 at 09:53:51AM -0500, Joel Schopp wrote:
> 
> On 08/19/2014 09:37 AM, Christoffer Dall wrote:
> > On Tue, Aug 19, 2014 at 09:05:09AM -0500, Joel Schopp wrote:
> >> On 08/19/2014 07:22 AM, Christoffer Dall wrote:
> >>> On Mon, Aug 18, 2014 at 03:30:58PM -0500, Joel Schopp wrote:
> >>>>  #endif /* __ARM_KVM_MMU_H__ */
> >>>> diff --git a/arch/arm/kvm/mmu.c b/arch/arm/kvm/mmu.c
> >>>> index 16e7994..70f0f02 100644
> >>>> --- a/arch/arm/kvm/mmu.c
> >>>> +++ b/arch/arm/kvm/mmu.c
> >>>> @@ -521,6 +521,7 @@ int create_hyp_io_mappings(void *from, void *to, phys_addr_t phys_addr)
> >>>>   */
> >>>>  int kvm_alloc_stage2_pgd(struct kvm *kvm)
> >>>>  {
> >>>> +	unsigned int s2_pgds, s2_pgd_order;
> >>>>  	pgd_t *pgd;
> >>>>  
> >>>>  	if (kvm->arch.pgd != NULL) {
> >>>> @@ -528,10 +529,18 @@ int kvm_alloc_stage2_pgd(struct kvm *kvm)
> >>>>  		return -EINVAL;
> >>>>  	}
> >>>>  
> >>>> -	pgd = (pgd_t *)__get_free_pages(GFP_KERNEL, S2_PGD_ORDER);
> >>>> +	s2_pgds = (1 << (kvm_get_phys_addr_shift() - PGDIR_SHIFT));
> >>>> +	s2_pgd_order = get_order(s2_pgds * sizeof(pgd_t));
> >>>> +
> >>>> +	pgd = (pgd_t *)__get_free_pages(GFP_KERNEL, s2_pgd_order);
> >>>>  	if (!pgd)
> >>>>  		return -ENOMEM;
> >>>>  
> >>>> +	if ((unsigned long)pgd & ~vttbr_baddr_mask) {
> >>>> +		kvm_err("Stage-2 pgd not correctly aligned: %p\n", pgd);
> >>>> +		return -EFAULT;
> >>>> +	}
> >>>>
> >>>>
> >>>> There are two problems that I've found here.  The first problem is that
> >>>> vttbr_baddr_mask isn't allocated yet at this point in the code.
> >>> allocated? you mean assigned?
> >>> aren't you setting vttbr_baddr_mask in kvm_arch_init()?  that's
> >>> certainly called before kvm_arch_init_vm().
> >> Yes, I mean assigned, at least I got the first letter correct :)  All I
> >> know is that vttbr_baddr_mask was still zero and checking for zero and
> >> calling the set function gave it a value.
> > that sounds.... weird and wrong.  Hum.  Mind sticking a few prints in
> > there and figuring out what's causing this?
> >
> >>>
> >>>> The
> >>>> second problem is that pgd is a virtual address, ie pgd ==
> >>>> 0xfffffe03bbb40000 while the vttbr masks off the high bits for a
> >>>> physical address, ie vttbr_baddr_mask=0x00007ffffffe0000 .  Even
> >>>> correcting for those issues I haven't been able to make this check work
> >>>> properly.  I'll resend v5 the patch with all the other suggested changes.
> >>>>
> >>> What are the issues that you face?  Iow. what is the alignment of the
> >>> returned physical address?
> >>>
> >>> (You should be able to just to virt_to_phys(pgd) and use that to test
> >>> for the vttbr_baddr_mask).
> >> The addresses above are actually from my system, 64K page aligned on a
> >> 64K page kernel.  I did use virt_to_phys() and the kernel got a null
> >> dereference and paniced, I didn't trace down where the panic was occuring.
> >>
> > virt_to_phys() directly caused the null dereference?  That sounds bad!
> I don't think it was the virt_to_phys() directly causing the null
> dereference, but again I didn't trace it down.
> 
> >
> > Would you mind trying to trace this down?  I'll be happy to provide as
> > much help as I can along the way.
> I can break the kvm_alloc_stage2_pgd check into a separate patch on top
> of this one and circle back around to it after I finish another
> unrelated thing I'm working on.

that would be great, thanks.