OMAP and MSM IOMMU driver misbehavior

OMAP and MSM IOMMU driver misbehavior

[Joerg Roedel]

Hi,

while reviewing another IOMMU driver again I came across a problem in
the IOMMU drivers for OMAP and MSM platforms. In both drivers the
'domain_destroy with devices attached' case isn't handled correctly.

OMAP driver seems not to track the devices attached to a domain at all.
So when a domain is destroyed it can happen that the hardware still
references old (and already freed) page-table pointers.

MSM tracks devices in a domain, but does not automatically remove the
devices from a domain that is about to be destroyed.

Please tell me when I mis-read the code, otherwise please fix this in
your drivers so that we can get consistent behavior for IOMMU-API
users :-)

Thanks,

	Joerg

-- 
AMD Operating System Research Center

Advanced Micro Devices GmbH Einsteinring 24 85609 Dornach
General Managers: Alberto Bozzo
Registration: Dornach, Landkr. Muenchen; Registerger. Muenchen, HRB Nr. 43632

Re: OMAP and MSM IOMMU driver misbehavior

[Ohad Ben-Cohen]

Hi Joerg,

On Mon, Jan 23, 2012 at 4:03 PM, Joerg Roedel <joerg.roedel@amd.com> wrote:
> Please tell me when I mis-read the code, otherwise please fix this in
> your drivers so that we can get consistent behavior for IOMMU-API
> users :-)

I'm a bit tied up with some schedule I have to meet this week, but
I'll definitely take a look and make sure it is fixed.

Somehow I trust you didn't mis-read anything :)

Thanks for reporting this,
Ohad.

Re: OMAP and MSM IOMMU driver misbehavior

[Stepan Moskovchenko]

On 1/23/2012 6:03 AM, Joerg Roedel wrote:
> Hi,
>
> while reviewing another IOMMU driver again I came across a problem in
> the IOMMU drivers for OMAP and MSM platforms. In both drivers the
> 'domain_destroy with devices attached' case isn't handled correctly.
>
> OMAP driver seems not to track the devices attached to a domain at all.
> So when a domain is destroyed it can happen that the hardware still
> references old (and already freed) page-table pointers.
>
> MSM tracks devices in a domain, but does not automatically remove the
> devices from a domain that is about to be destroyed.
>
> Please tell me when I mis-read the code, otherwise please fix this in
> your drivers so that we can get consistent behavior for IOMMU-API
> users :-)
>
> Thanks,
>
> 	Joerg

Hello

I believe your analysis is correct, and it is a legitimate problem. The 
driver does keep a list of devices attached to a domain, so it should 
not be too hard to detach them. However, I have been quite occupied with 
other things lately, but I can try to get to it when I have some free 
time. Calling detach_dev on each element is what needs to happen in 
theory, but I feel like the main detach_dev code will need to be broken 
out to handle the locking properly. Still, it does not sound 
particularly difficult.

Steve