Re: 2.6.31.5 x86-64 KVM: 'emulation failed (pagetable) rip 7fb49335da7b 66 0f 7f 07'

[Avi Kivity <avi@redhat.com>]

On 11/29/2009 04:23 PM, Avi Kivity wrote:
> On 11/29/2009 03:48 PM, Nix wrote:
>> On 29 Nov 2009, Avi Kivity uttered the following:
>>> 66 0f 7f 07              movdqa %xmm0,(%rdi)
>>>
>>> which we don't emulate.
>> x86-64 glibc 2.10 memset(), perhaps? On SSE-capable platforms that does
>> a whole bunch of
>>
>> L(SSE0QB):  movdqa %xmm0,-0xb0(%rdi)
>> L(SSE0QA):  movdqa %xmm0,-0xa0(%rdi)
>> L(SSE0Q9):  movdqa %xmm0,-0x90(%rdi)
>> L(SSE0Q8):  movdqa %xmm0,-0x80(%rdi)
>> L(SSE0Q7):  movdqa %xmm0,-0x70(%rdi)
>> L(SSE0Q6):  movdqa %xmm0,-0x60(%rdi)
>> L(SSE0Q5):  movdqa %xmm0,-0x50(%rdi)
>> L(SSE0Q4):  movdqa %xmm0,-0x40(%rdi)
>> L(SSE0Q3):  movdqa %xmm0,-0x30(%rdi)
>> L(SSE0Q2):  movdqa %xmm0,-0x20(%rdi)
>> L(SSE0Q1):  movdqa %xmm0,-0x10(%rdi)
>> L(SSE0Q0):  retq
>>
>> (multiple blocks of this, catering for alignment, I guess)
>>
>> and x86-64 is always SSE-capable.
>
> Most likely, either this or something similar is called on a userspace 
> device driver.  Can you check if this is triggered by starting X?
>
> If so, we'll have to emulate this instruction, which will be a bitch.
>

Not just in the emulator, also in the userspace interface.  We only 
support up to 8 bytes mmio.

Our options in fixing this are:

- extend mmio support to 16 (32? more?) bytes, require new kernel and 
new userspace
- use coalesced_mmio for this
    - if the mmio happens for a non coalesced mmio region, we flush 
immediately
    - what if userspace doesn't support coalesced mmio? (unlikely these 
days)
    - doesn't work for reads (unlikely for 16 byte accesses?)

-- 
error compiling committee.c: too many arguments to function