On Mon, Jan 23, 2023 at 05:36:06PM +0100, Rob Herring wrote: > On Mon, Jan 23, 2023 at 10:02 AM Vincent Whitchurch > <vincent.whitchurch@xxxxxxxx> wrote: > dtc should complain about this... It probably does, the test framework currently doesn't report these to the test runner/writer; maybe it should. > > compatible = "virtio,uml"; > > Binding? There was some discussion earlier about whether a binding was needed here (you were on CC): https://lore.kernel.org/lkml/20211222103417.GB25135@xxxxxxxx/ > > > virtio-device-id = <1234>; > > ranges; > > > > pci { > > #address-cells = <3>; > > #size-cells = <2>; > > ranges = <0x0000000 0 0 0 0xf0000000 0 0x20000>; > > compatible = "virtio,device4d2", "pci"; > > "pci" is not a valid compatible string. I think it's there since I based this tree off from arch/x86/platform/ce4100/falconfalls.dts. I see that there is some code in arch/x86/kernel/devicetree.c to handle this compatible and register all platform devices under that. Do we need something like that for UML instead of this patch? > > > device_type = "pci"; > > bus-range = <0 0>; > > > > platform_parent: device@0,0 { > > compatible = "pci494f,dc8"; > > reg = <0x00000 0 0 0x0 0x10000>; > > ranges; > > > > uart@10000 { > > compatible = "google,goldfish-tty"; > > reg = <0x00000 0 0x10000 0 0x10000>; > > This is not a PCI device, so it shouldn't be using PCI addressing. > 'ranges' needs an entry (for each BAR) to translate to just a normal > MMIO bus with 1 or 2 address/size cells. Maybe we want a 'simple-bus' > node for each BAR. The FPGA series needs the same things, but that > aspect hasn't really been addressed as the first issue is populating > the PCI devices dynamically. Yes, this ranges stuff can be fixed in the Python code which generates these trees. In my cases the devicetree blob contains all the devices under the PCI devices, see my other email. > The DT address translation code should support all this > (MMIO->PCI->MMIO), but I don't think there's any existing examples. An > example (that I can test) would be great. If the unittest had that > example, I'd be thrilled. Anyone can run what I'm running since it uses UML and there is no real hardware, but the setup is a bit more complicated than an in-kernel unit test since there is a virtio backend in userspace which implements the "hardware". If you want to try it: git remote add vwax https://github.com/vwax/linux.git git fetch vwax git checkout vmax/roadtest/platform-wip make -C tools/testing/roadtest/ -j24 OPTS="-v -k platform" You should see a "PASSED roadtest/tests/base/test_platform.py::test_foo" if it works. See Documentation/dev-tools/roadtest.rst for more info. As mentioned in the other email, the only patches to the kernel proper in that tree are already posted ones and WIP fixes.
