Re: kvm [2087]: load/store instruction decoding not implemented

[Marc Zyngier <marc.zyngier@arm.com>]

On 24/02/15 12:47, Christoffer Dall wrote:
> On Tue, Feb 24, 2015 at 12:29:25PM +0000, Richard W.M. Jones wrote:
>> On Tue, Feb 24, 2015 at 09:15:18PM +0900, Peter Maydell wrote:
>>> On 24 February 2015 at 20:59, Richard W.M. Jones <rjones@redhat.com> wrote:
>>>>
>>>> https://bugzilla.redhat.com/show_bug.cgi?id=1194366
>>>>
>>>> Has anyone seen this KVM error?  Or have suggestions how to debug it
>>>> further?
>>>>
>>>>   kvm [2028]: load/store instruction decoding not implemented
>>>
>>> This is a fairly common thing to run into and google is bound
>>> to have references to past discussions. What has happened here
>>> is that the guest has attempted a "complex" load/store instruction
>>> to an area of RAM which is not mapped (ie not guest RAM).
>>> For this class of instructions the hardware doesn't provide
>>> syndrome information to allow us to figure out the address/size
>>> etc of the access, so we would have to actually decode the
>>> offending instruction and emulate it; this emulation isn't
>>> implemented.
>>>
>>> Complex insns are things like load-multiple (there's a complete
>>> list in the ARM ARM somewhere). Generally this indicates a guest
>>> bug because you really shouldn't be accessing devices with
>>> weird instructions like that (and you shouldn't be accessing
>>> unmapped memory at all).
>>
>> I'm not super-familiar with the aarch64 instruction set, but
>> according to qemu the instruction is:
>>
>> b8004403        str     w3, [x0],#4
>>
>> (in __copy_to_user).  My interpretation is this is storing the
>> lower 32 bits of x3 into the storage pointed to by x0 (+ 4 bytes?)
>> Is that one of the complicated ones?
> 
> Shouldn't be, I don't think aarch64 does any register write-back here.
> This is an aarch64 userspace process, right?
> 
> You can try adding some more debugging info to the print to get us the
> IPA it is failing on:
> 
> diff --git a/arch/arm/kvm/mmio.c b/arch/arm/kvm/mmio.c
> index 5d3bfc0..e468937 100644
> --- a/arch/arm/kvm/mmio.c
> +++ b/arch/arm/kvm/mmio.c
> @@ -182,7 +182,8 @@ int io_mem_abort(struct kvm_vcpu *vcpu, struct kvm_run *run,
>  		if (ret)
>  			return ret;
>  	} else {
> -		kvm_err("load/store instruction decoding not implemented\n");
> +		kvm_err("load/store instruction decoding not implemented (HSR: 0x%x, IPA: 0x%llx)\n",
> +			kvm_vcpu_get_hsr(vcpu), fault_ipa);
>  		return -ENOSYS;
>  	}
>  
> 
> I wonder why you're faulting on an address that your guest kernel is
> doing __copy_to_user() on in the first place though, hmmm.

Here's my theory: userspace is accessing something it should never
access (outside of RAM, basically), and doing so via a kernel interface.

Is this process accessing /dev/mem by any chance? dmidecode anyone?

	M.
-- 
Jazz is not dead. It just smells funny...