On Thu, Aug 14, 2025 at 07:32:36PM -0700, Shyam Saini wrote: > On Thu, Aug 14, 2025 at 09:39:58PM -0300, Jason Gunthorpe wrote: > > On Thu, Aug 14, 2025 at 04:30:18PM -0700, Shyam Saini wrote: > > > or were you referring to [2]? > > > > > > In that case, the PCI child node data needs to be parsed, which is > > > currently handled individually by each host controller driver. > > > > Yes, this looks like it may be what I was thinking of, the pci@1,0 > > specifes the BDF effectively > > In that case, we'll need to parse the child DTS nodes properly > within of_iommu_get_resv_regions(). I'll include this in v4.
Kinda surprised this isn't happening already? It would be good to refer to the original specs and describe how whatetever you propose is aligned there. > > Presumably this is a fixed issue of the platform. You never did > > explain how your system has such werdio behavior, or how something > > like iommu=pt can function on it... > Yes, this issue is platform-specific. On this platform, the default MSI IOVA > range overlaps with an address region that is reserved for another purpose, > Other than that we haven't observed any obvious issues for DMA operations Usually DRAM is at the default MSI IOVA address, so if you run iommu=pt then presumably your DRAM map has a nice hole in it.. But maybe the memory map told the OS about that. Jason
