[RFC PATCH v1 0/6] Resolve unwanted DMA backing with IOMMU

[Dmitry Osipenko]

On Friday, 27 July 2018 21:31:34 MSK Joerg Roedel wrote:
> On Fri, Jul 27, 2018 at 11:13:31AM -0600, Rob Herring wrote:
> > I don't follow why we need a property rather than being implied by the
> > device's (the GPU) compatible string.
> 
> There might be devices where either setup works, with or without IOMMU
> translation, and the firmware can set the property depending on whether
> the user wants more performance or more security.
> 
> If we have a whitelist in the kernel this gets more complicated, we
> probably need additional kernel-parameters to overwrite those whitelist
> entries. Having a property in the device-tree seems to be a better way
> here, imho.

IIUC, device-tree should be considered to be "written in stone" for a consumer 
device and hence firmware property isn't something that could be easily 
changed. The kernel-parameter will be much more universal. Anyway the global 
whitelisting should be a different topic for discussion, right now we need a 
kind of private whitelisting that is internal to kernel.