Re: KVM guest sometimes failed to boot because of kernel stack overflow if KPTI is enabled on a hisilicon ARM64 platform.

[Will Deacon]

Hi Wei,

On Thu, Jun 21, 2018 at 11:14:28AM +0100, Wei Xu wrote:
> On 2018/6/21 10:18, Will Deacon wrote:
> > On Thu, Jun 21, 2018 at 09:38:53AM +0100, James Morse wrote:
> >> On 20/06/18 17:25, Wei Xu wrote:
> >>>     [    0.042421] Insufficient stack space to handle exception!
> >>>     [    0.042423] ESR: 0x96000046 -- DABT (current EL)
> >>>     [    0.043730] FAR: 0xffff0000093a80e0
> >>>     [    0.044714] Task stack: [0xffff0000093a8000..0xffff0000093ac000]
> >>
> >> This was a level 2 translation fault on a write, to an address that is within
> >> the stack....
> >>
> >>
> >>>     [    0.051113] IRQ stack: [0xffff000008000000..0xffff000008004000]
> >>>     [    0.057610] Overflow stack: [0xffff80003efce2f0..0xffff80003efcf2f0]
> >>>     [    0.064003] CPU: 0 PID: 12 Comm: migration/0 Not tainted
> >>> 4.17.0-45865-g2b31fe7-dirty #10
> >>>     [    0.072201] Hardware name: linux,dummy-virt (DT)
> >>
> >>>     [    0.076797] pstate: 604003c5 (nZCv DAIF +PAN -UAO)
> >>>     [    0.081727] pc : el1_sync+0x0/0xb0
> >>
> >> ... from the vectors.
> >>
> >>
> >>>     [    0.085217] lr : kpti_install_ng_mappings+0x120/0x214
> >>
> >> What I think is happening is: we come out of the kpti idmap with the stack
> >> unmapped. Shortly after we access the stack, which faults. el1_sync faults as
> >> well when it tries to push the registers to the stack, and we keep going until
> >> we overflow the stack.
> >>
> >> I can't reproduce this with kvmtool or qemu in the model.
> > 
> > Hmm, one thing that occurs to me is that the kpti_install_ng_mappings()
> > code leaves the nG bit set in table entries, which is actually IGNORED in
> > the architecture.
> > 
> > Wei -- does the diff below help at all? Make sure you disable CONFIG_KASAN,
> > otherwise your kernel will take an age to boot.
> 
> Yes, amazing! This patch resolved the issue.

Great...

> I have tested 50 times and can not reproduce the issue any more.
> Could you please tell more why this patch works?

You might need to ask your CPU design team ;)

Without this patch, the code in idmap_kpti_install_ng_mappings() sets
bit 11 in table descriptors so that we can keep track of which parts of
the page table we've visited. With this patch, we don't bother tracking
and potentially rewalk parts of the page table (which takes a very long
time if KASAN is enabled).

The architecture documents I've looked at are clear that bit 11 is IGNORED
by the CPU, which:

  "Indicates that the architecture guarantees that the bit or field is not
   interpreted or modified by hardware."

Please can you double-check that your CPU is indeed ignoring bit 11 in
non-leaf (table) descriptors?

Thanks,

Will