On 10/22/26 2:54 PM, Anna Maniscalco wrote:
On 7/9/26 9:39 PM, Dmitry Baryshkov wrote:
On Thu, Jul 09, 2026 at 05:54:23PM +0300, Dmitry Baryshkov wrote:
On Thu, Jul 09, 2026 at 07:12:39PM +0530, Akhil P Oommen wrote:
On 7/7/2026 10:09 PM, Dmitry Baryshkov wrote:
On Tue, Jul 07, 2026 at 05:02:43PM +0200, Anna Maniscalco wrote:
On 7/6/26 5:04 PM, Dmitry Baryshkov wrote:
On Sun, Jul 05, 2026 at 09:13:37PM +0200, Anna Maniscalco wrote:
In order to associate the SID used by LPAC to it's own domain and
context bank we need a separate node on the DTS.
Can we use iommu-map instead of adding a new device using the same
addresses? You provide zero explanations, but it seems like you are
trying to separate a part of the GPU into a separate device, but
it's
unnecessary.
What I was trying to say is that we need to map sid0 and sid1 to
different
context banks.
From what I recall, so long as we use the same dt node, the
iommu drivert
provides no way of doing this.
What kgsl does is having a node for the iommu and then subnodes
for gfx lpac
and secure.
The address on the lpac device doesn't really matter. I have very
limted
knowledge of DTSs but I think we could just leave it out altogether?
I think, you were poined to the Glymur Iris patches. Another
option is
the QDA driver being submitted too. Both ot them use iommu-map to
describe the same usecase - a set of SIDs which are used by the
device
but which should not be a part of the defaut set. With it in place
you
don't need to have separate qcom,lpac devices, no additional
properties,
etc.
I'd recommend following the same approach.
iiuc, the conclusion in this recent thread is to use child nodes for a
similar case:
https://lore.kernel.org/all/[email protected]/
Let me check internally.
After a brief discussion: it can be either, depending on what suites
better (and if you can persuade DMA / IOMMU maintainers to let one extra
bus or not).
One thing sashiko is pointing out is that new dt should work with hold
kernel so, in general moving sid1 from there would break it since old
kernels don't know how to handle them separately.
Not sure how to address this.
Best regards,
Also, looking into iommu-map, it seems like I would ultimately still
need to add a child device.
Best regards,
--
Anna Maniscalco <[email protected]>
