Le jeudi 19 novembre 2026 à 11:08 -0400, Jason Gunthorpe a écrit : > On Thu, Nov 19, 2026 at 09:52:59AM -0500, Nicolas Dufresne wrote: > > Le jeudi 19 novembre 2026 à 10:31 -0400, Jason Gunthorpe a écrit : > > > With the additional twist that alot of embedded HW has apparently > > > choosen to not implement SSID but instead has a pool of SIDs that one > > > device can use, functionally similar to SSID. > > > > > > We'd need also API support for unbundling the DT iommus=[] SID list > > > into something that the SIDs can be attached more like a PASID.  That > > > need, I think, has become clear with all the DT tricks and fake busses > > > to try to work around this issue. :| > > > > The word "fake" never flies with the DT maintainers. On that subject, I still > > don't know what you mean with iommus=[]. > > When you write your DT you can list all the iommu streams in one > iommus property. Like this randomly picked out of the DT files: > > iommus = <&sysmmu_decon0x>, <&sysmmu_decon1x>; > > If you do that then Linux sets the struct device containing it to > put the same paging domain on all of those iommus. > > It doesn't make any sense to open code this in drivers when you can > just have the core code do it by writing your DT to match how the > driver wants to work... The will be rejected, as it does not map the hardware. If you check multi-core codecs, each pair of core+iommu are independent from each other, at least enough that they could be driven by different device drivers. And so they have been described separately in the DTs. I don't see a fault here. Its a Linux design decision to manage them as a group, so Linux can schedule them fairly between processes. It would also conflict with the power management, which arguably should happen independently. On this last bit, I'm actually a bit concern of how careless the multi-core patchset are regarding PM of the device that owns the shared domain, it only works by accident as they schedule core0 first, and other cores are only used if core0 is busy (there is no affinity). Collecting all cores using the component mechanism, and picking one master and sharing its domain to all other is pretty small effort, I don't want to go too crazy here. But I'm open for better DT mapping if something would be accepted by the DT maintainer. Nicolas