Re: [Qemu-devel] Ideas wiki for GSoC 2010

[Anthony Liguori <anthony@codemonkey.ws>]

On 03/15/2010 10:06 AM, Avi Kivity wrote:
> On 03/15/2010 03:23 PM, Anthony Liguori wrote:
>> On 03/15/2010 08:11 AM, Avi Kivity wrote:
>>> On 03/15/2010 03:03 PM, Joerg Roedel wrote:
>>>>
>>>>>> I will add another project - iommu emulation.  Could be very useful
>>>>>> for doing device assignment to nested guests, which could make
>>>>>> testing a lot easier.
>>>>> Our experiments show that nested device assignment is pretty much
>>>>> required for I/O performance in nested scenarios.
>>>> Really? I did a small test with virtio-blk in a nested guest (disk 
>>>> read
>>>> with dd, so not a real benchmark) and got a reasonable 
>>>> read-performance
>>>> of around 25MB/s from the disk in the l2-guest.
>>>>
>>>
>>> Your guest wasn't doing a zillion VMREADs and VMWRITEs every exit.
>>>
>>> I plan to reduce VMREAD/VMWRITE overhead for kvm, but not much we 
>>> can do for other guests.
>>
>> VMREAD/VMWRITEs are generally optimized by hypervisors as they tend 
>> to be costly.  KVM is a bit unusual in terms of how many times the 
>> instructions are executed per exit.
>
> Do you know offhand of any unnecessary read/writes?  There's 
> update_cr8_intercept(), but on normal exits, I don't see what else we 
> can remove.

Yeah, there are a number of examples.

vmcs_clear_bits() and vmcs_set_bits() read a field of the VMCS and then 
immediately writes it.  This is unnecessary as the same information 
could be kept in a shadow variable.  In vmx_fpu_activate, we call 
vmcs_clear_bits() followed immediately by vmcs_set_bits(). which means 
we're reading GUEST_CR0 twice and writing it twice.

vmx_get_rflags() reads from the VMCS and we frequently call get_rflags() 
followed by a set_rflags() to update a bit.  We also don't cache the 
value between calls and there's a few spots in the code that make 
multiple calls.

Regards,

Anthony Liguori