Hi Olof, Thanks for the comments! >>Based on what I can see, it seems that you're trying to describe two pieces of hardware with only one device in the DT. That seems like an odd choice. T194 SOC HW is designed to use Two ARM SMMU's together like one SMMU. The IOVA accesses from the HW controllers clients is interleaved across these Two SMMU's by Memory controller HW. Even though, there are two ARM SMMU's, HW design requires SW programming these two ARM SMMU instances identically. >>Also, it seems like all three SMMUs share the same interrupt line? That's somewhat suboptimal IMHO, but harder to change. This is by HW design. SMMU interrupts are only expected if there is an illegal IOVA access, which shouldn't happen in practical. The SMMU interrupts are only used to log the faults and for debug purpose on T194. >>Why not instantiate both of them and create a reference between then such that the TLB and sync ops are done on both of them in the native driver? I.e. two arm_smmu structs and a pointer between then (i.e. add a "next shared smmu" pointer in the struct arm_smmu). DMA framework supports only one SMMU driver instance per device. SMMU driver setup iommu_ops pointer in the device struct during the boot when the platform device is added to the bus. There is no support to allow a platform device register with two different SMMU devices. Two different SMMU instances is 2x page table memory and 2x map/unmap overhead. With SMMU driver knowing about two smmu instances, page table memory usage, map/unmap Overhead would be efficient. >>As long as devices only references one of them, locking only that one should provide suitable protection as well. Earlier proposal on creating one Master SMMU node and slave SMMU nodes within isn't liked. Will has proposed creating a library and reusing the code. I have made the patches to share the code. https://www.spinics.net/lists/linux-tegra/msg35539.html https://www.spinics.net/lists/linux-tegra/msg35494.html >>Seems like a simpler approach than adding a new layer to the driver, but maybe >>I am missing some complexity here? Programming both SMMU instances identically with existing ARM SMMU driver is not possible with multiple DT nodes. In the following patch, by writel* macro override, both ARM SMMU instances are programmed identically in Tegra194 SMMU driver transparently. Only tlb sync and interrupt handling need to know about specific instance. https://lkml.org/lkml/2018/11/1/573 +#define writel_relaxed writel_relaxed_all +#define writeq_relaxed writeq_relaxed_all +#define writel writel_all -KR
