Re: [Xen-devel] [RFC PATCH v1 02/10] xen/arm: register mmio handler at runtime

[Julien Grall <julien.grall@xxxxxxxxxx>]

On 04/01/2014 10:34 AM, Vijay Kilari wrote:
> Hi Julien,

Hello Vijay,

> On Thu, Mar 27, 2014 at 8:32 PM, Julien Grall <julien.grall@xxxxxxxxxx> wrote:
>> On 03/27/2014 05:40 AM, Vijay Kilari wrote:
>>> On Wed, Mar 26, 2014 at 8:17 PM, Julien Grall <julien.grall@xxxxxxxxxx> 
>>> wrote:
>>>> On 03/26/2014 12:29 PM, Vijay Kilari wrote:
>>>>> Hi Julien,
>>>>>
>>>>>
>>>>> On Fri, Mar 21, 2014 at 10:53 PM, Julien Grall <julien.grall@xxxxxxxxxx> 
>>>>> wrote:
>>>>>> On 03/21/2014 05:17 PM, Ian Campbell wrote:
>>>>>>> On Wed, 2014-03-19 at 19:47 +0530, vijay.kilari@xxxxxxxxx wrote:
>>>>>>>> From: Vijaya Kumar K <Vijaya.Kumar@xxxxxxxxxxxxxxxxxx>
>>>>>>>>
>>>>>>>> mmio handlers are registers at compile time
>>>>>>>> for drivers like vuart and vgic.
>>>>>>>> Make mmio handler registered at runtime by
>>>>>>>> creating linked list of mmio handlers
>>>>>>>
>>>>>>> I'm not convinced of the need for this, certainly the vgic side can just
>>>>>>> demux into v2 or v3 as necessary.
>>>>>>
>>>>>> Demux the code just add an indirection. We could have a list of mmio
>>>>>> handler per domain and rely on it to call the right handler. A bit like 
>>>>>> x86.
>>>>>>
>>>>>  Until Andrii adds IOMMU  handling should keep this patch? and adopt
>>>>> to it later?
>>>>
>>>> I'm not sure to understand. IHMO, it doesn't sound right to upstream a
>>>> patch that we know it will be superseded in a couple of months. Why
>>>> can't you rework your patch to have it in good shape now?
>>>>
>>>  I assume that Andii's patch will be required to make mmio handlers
>>> per domain. If not let me know references in x86 that I can make it
>>> per domain specific mmio handlers.
>>
>> I have already pointed out to the x86 code few mails before.
>>
>> You can look at xen/arch/x86/intercept.c.
> 
>    From the x86/hvm/intercept.c file, the mmio handling of x86 is
> similar to existing arm mmio handling in arm/io.c file, where the
> handlers are registered statically.

Not really, x86 use an hybrid approach:
    - static handler for IO region that are handled by every guest (hpet
...)
     - dynamic handler registered by register_io_handler

>   I understand that only requirement is to make vgic as domain specific
> similar to vuart. Is that what you expect. Can you please make it clear?

vuart and vgic handler won't be the only dynamic handlers. On some
platform (e.g the cubieboard) we will have to use a similar solution to
be able to passthrough UARTs to dom0. Currently they are blacklisted
because the UART used by Xen shares the page with the other UARTs.

Using the x86 approach will allow us to register the right vGIC
emulation for every guest. We might want to start a shiny FreeBSD guest
with GICv2 emulation and a Linux guest with a GICv3.

> 
>  IMO, the io handling in x86/hvm/intercept.c is not required for arm.

IHMO, as MMIO is often trapped, the performance is critical. Your
solution won't improve the current implementation because you still have
to browse unused MMIO handler each time the guest will trap into Xen.

Regards,

-- 
Julien Grall

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel