Re: [PATCH 1/2] kvm: Batch writes to MMIO

[Avi Kivity <avi@qumranet.com>]

Laurent Vivier wrote:
>>>   
>>>       
>> This breaks ordering on smp guests. batch_data needs to be a kvm thing, 
>> not a vcpu thing, and locked, of course.
>>     
>
> - is ordering between vcpu important when we already delay operations ?
>   

Yes.  Ordering is very different from delaying.  Also, we will only 
delay writes which don't matter (those that simply latch something into 
a register, which is later read).

> - using vcpu avoids the lock
>   

Correctness beats performance

> - Why PIO (pio_data) are vcpu things and not kvm things, then ?
>   

pio batching is used for just a single instruction that can work with 
batches (ins/outs), so a guest can't expect to see partial output.

With mmio, you can have

1. mmio
2. raise flag in shared memory
3. other vcpu sees flag, does mmio
4. spin-wait for ack from other vcpu
5. mmio

You can't do that with ins/outs.

-- 
Do not meddle in the internals of kernels, for they are subtle and quick to panic.


-------------------------------------------------------------------------
This SF.net email is sponsored by the 2008 JavaOne(SM) Conference 
Don't miss this year's exciting event. There's still time to save $100. 
Use priority code J8TL2D2. 
http://ad.doubleclick.net/clk;198757673;13503038;p?http://java.sun.com/javaone