On Wed, Oct 16, 2024 at 08:53:20AM +0000, Peng Fan wrote: > > Subject: Re: [PATCH RFC 0/2] iommu/arm-smmu-v3: bypass streamid > > zero on i.MX95 > > > > On Wed, Oct 16, 2024 at 12:56:54AM +0000, Peng Fan wrote: > > > > Subject: Re: [PATCH RFC 0/2] iommu/arm-smmu-v3: bypass > > streamid zero > > > > on i.MX95 > > > > > > > > On 2024-10-15 4:14 am, Peng Fan (OSS) wrote: > > > > > i.MX95 eDMA3 connects to DSU ACP, supporting dma coherent > > > > memory to > > > > > memory operations. However TBU is in the path between eDMA3 > > > > and ACP, > > > > > need to bypass the default SID 0 to make eDMA3 work properly. > > > > > > > > I'm confused, why not just describe that the device owns this > > > > StreamID in the DT the normal way, i.e, "iommus = <&smmu 0>;"? > > > > > > Current SMMU-v3 driver not support streamID sharing, but we have > > > limited streamIDs(i.MX95 max supports 64 SIDs). However the eDMA > > > supports 64 channels, each channel may have a SID, but not > > supported > > > by current smmuv3 driver. > > > > > > We see smmu would degrade the performance in somecases, so we > > wanna to > > > give user an option to bypass SMMU for eDMA. > > > > If it's a system-wise global need, adding "iommu.passthrough=y" > > to the bootargs string should work for all trusted devices. And > > No. not system wide. > > > particular group can be changed to the IDENTITY domain too upon > > cmdline setting, e.g. > > echo identity | sudo tee > > /sys/class/pci_bus/0009\:01/device/iommu_group/type > > This means I still need to put iommus = <&smmu 0> under > the device tree node. > Yes, otherwise the bypassing an unassigned SID isn't something that the community wouldn't prefer as it smashes a hole in security for everyone. Does your hardware have a way all the devices can configure/program SIDs or does the hardware have multiple devices that generate SID 0? If it's programmable, I'd suggest reserving streamID 0 for bypass and let the system work with the other 63. Apart from that, if you can guarantee that the `eDMA3` would always access a fixed region of memory, then maybe use RMRs to install bypass. > Thanks, > Peng. > > > > > Thanks > > Nicolin > Thanks, Pranjal
