Re: [Xen-devel] [PATCH v4 07/17] x86/hvm: add length to mmio check op

[Andrew Cooper <andrew.cooper3@xxxxxxxxxx>]

On 25/06/15 14:36, Paul Durrant wrote:
>> -----Original Message-----
>> From: Andrew Cooper [mailto:andrew.cooper3@xxxxxxxxxx]
>> Sent: 25 June 2015 14:34
>> To: Jan Beulich
>> Cc: Paul Durrant; xen-devel@xxxxxxxxxxxxxxxxxxxx; Keir (Xen.org)
>> Subject: Re: [PATCH v4 07/17] x86/hvm: add length to mmio check op
>>
>> On 25/06/15 13:46, Jan Beulich wrote:
>>>>>> On 25.06.15 at 14:21, <andrew.cooper3@xxxxxxxxxx> wrote:
>>>> On 24/06/15 12:24, Paul Durrant wrote:
>>>>> When memory mapped I/O is range checked by internal handlers, the
>> length
>>>>> of the access should be taken into account.
>>>>>
>>>>> Signed-off-by: Paul Durrant <paul.durrant@xxxxxxxxxx>
>>>>> Cc: Keir Fraser <keir@xxxxxxx>
>>>>> Cc: Jan Beulich <jbeulich@xxxxxxxx>
>>>>> Cc: Andrew Cooper <andrew.cooper3@xxxxxxxxxx>
>>>>>
>>>> For what purpose?  The length of the access doesn't affect which handler
>>>> should accept the IO.
>>>>
>>>> This length check now causes an MMIO handler to not claim an access
>>>> which straddles the upper boundary.
>>>>
>>>> It is probably fine to terminate such an access early, but it isn't fine
>>>> to pass such a straddled access to the default ioreq server.
>>> No, without involving the length in the check we can end up with
>>> check() saying "Yes, mine" but read() or write() saying "Not me".
>>> What I would agree with is for the generic handler to split the
>>> access if the first byte fits, but the final byte doesn't.
>> I discussed this with Paul over lunch.  I had not considered how IO gets
>> forwarded to the device model for shared implementations.
>>
>> Is it reasonable to split a straddled access and direct the halves at
>> different handlers? This is not in line with how other hardware behaves
>> (PCIe will reject any straddled access).  Furthermore, given small MMIO
>> regions and larger registers, there is no guarantee that a single split
>> will suffice.
>>
>> I see in the other thread going on that a domain_crash() is deemed ok
>> for now, which is fine my me.
>>
> I think that also allows me to simplfy the patch since I don't have to modify 
> the mmio_check op any more. I simply call it once for the first byte of the 
> access and, if it accepts, verify that it also accepts the last byte of the 
> access.

At that point, I would say it would be easier to modify the claim check
to return "yes/straddled/no" rather than calling it twice.

~Andrew

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