Re: ce56a86e2a ("x86/mm: Limit mmap() of /dev/mem to valid physical addresses"): kernel BUG at arch/x86/mm/physaddr.c:79!

[Sander Eikelenboom]

On 26/10/17 19:49, Craig Bergstrom wrote:
> Sander, thanks for the details, they've been very useful.
> 
> I suspect that your host system's mem=2048M parameter is causing the
> problem.  Any chance you can confirm by removing the parameter and
> running the guest code path?

I removed it, but kept the hypervisor limiting dom0 memory to 2046M intact (in grub using the xen bootcmd: 
"multiboot       /xen-4.10.gz  dom0_mem=2048M,max:2048M ....."

Unfortunately that doesn't change anything, the guest still fails to start with the same errors.

> More specifically, since you're telling the kernel that it's high
> memory address is at 2048M and your device is at 0xfe1fe000 (~4G), the
> new mmap() limits are preventing you from mapping addresses that are
> explicitly disallowed by the parameter.
> 

Which would probably mean the current patch prohibits hard limiting the dom0 memory to a certain value (below 4G)
at least in combination with PCI-passthrough. So the only thing left would be to have no hard memory restriction on dom0
and rely on auto-ballooning, but I'm not a great fan of that.

I don't know how KVM handles setting memory limits for the host system, but perhaps it suffers from the same issue.

I also tried the patch from one of your last mails to make the check "less strict", 
but still get the same errors (when using the hard memory limits).

--
Sander

 
> 
> On Thu, Oct 26, 2017 at 10:39 AM, Ingo Molnar <mingo@xxxxxxxxxx> wrote:
>>
>> * Craig Bergstrom <craigb@xxxxxxxxxx> wrote:
>>
>>> Yes, not much time left for 4.14, it might be reasonable to pull the
>>> change out since it's causing problems. [...]
>>
>> Ok, I'll queue up a revert tomorrow morning and send it to Linus ASAP if there's
>> no good fix by then. In hindsight I should have queued it for v4.15 ...
>>
>> Thanks,
>>
>>         Ingo