Re: [RFC PATCH 0/4] generic loader FDT support (for direct Xen boot)

[Alex Bennée]

Alistair Francis <alistair23@gmail.com> writes:

> On Fri, Oct 9, 2020 at 10:07 AM Alex Bennée <alex.bennee@linaro.org> wrote:
>>
>> Hi,
>>
>> This series adds the ability to append FDT data for blobs loaded by
>> the generic loader. My principle use-case was to be able to directly
>> boot Xen with a kernel image which avoided having to:
>>
>>   - get the kernel image into my system image
>>   - deal with bugs in FDT munging by -bios firmware and/or grub
>>
>> as such this currently a developer hack that makes *my* life easy and
>> is thus presented as an RFC for discussion. While I've tested it with
>> Xen I'm sure the approach would be applicable to other hypervisors or
>> firmware which expect to consume FDT data pointing at various blobs.
>
> An interesting idea. I think this comes up enough that it's worth
> thinking about.
>
>>
>> An example command line that launches this is (magic from -kernel):
>>
>>   ./aarch64-softmmu/qemu-system-aarch64 -cpu cortex-a57 \
>>     -machine type=virt,virtualization=on -display none \
>>     -serial mon:stdio \
>>     -netdev user,id=unet,hostfwd=tcp::2222-:22 \
>>     -device virtio-net-pci,netdev=unet,id=virt-net,disable-legacy=on \
>>     -device virtio-scsi-pci,id=virt-scsi,disable-legacy=on \
>>     -blockdev 
>> driver=raw,node-name=hd,discard=unmap,file.driver=host_device,file.filename=/dev/zen-disk/debian-buster-arm64
>>  \
>>     -device scsi-hd,drive=hd,id=virt-scsi-hd \
>>     -smp 4 -m 4096 \
>>     -kernel ~/lsrc/xen.git/xen/xen \
>>     -append "dom0_mem=1G,max:1G loglvl=all guest_loglvl=all" \
>>     -device 
>> loader,addr=0x47000000,file=$HOME/lsrc/linux.git/builds/arm64/arch/arm64/boot/Image,len-fdt-compat=2,fdt-compat[0]='multiboot,,module',fdt-compat[1]='multiboot,,kernel',fdt-bootargs="root=/dev/sda2
>>  ro console=hvc0 earlyprintk=xen"
>
> You are loading the kernel `Image` file, but changing the device tree
> that was generated by QEMU and is being loaded in memory. As a user
> that is really confusing.

Well in this case the "kernel" is Xen which helpfully comes with a
kernel compatible header that the kernel loaded understand. The blob
could be any Dom0 kernel - it just happens to be a Linux kernel in this
case.

>
>>
>> So a couple of choices I've made doing this:
>>
>> Promoting *fdt to MachineState
>>
>> This seemed the simplest approach to making the fdt available to the
>> global state, especially as MachineState already has a *dtb pointer.
>> I've converted the ARM virt machine and a RISCV machine but if this is
>> acceptable I can convert the others.
>
> This seems fine to me.
>
>>
>> "Polluting" the generic loader arguments
>>
>> This was mainly out of convenience as the loader already knows the
>> size of the blob being loaded. However you could certainly argue it
>> makes sense to have a more generic "FDT expander" virtual device that
>> could for example query the QOM model somehow to find the details it
>> needs.
>
> That seems like a better option. Why not have a generic way to modify
> the device tree with a specific argument? It could either be -device
> loader,file=fdt,... or -fdt ...

Well it comes down to how much of a special case this is? Pretty much
all FDT (and ACPI for the matter) is basically down to the board level
models - and not all of them just the funky configurable ones. For other
board models we just expect the user to pass the FDT they got with their
kernel blob.

For modern VirtIO systems the only thing you really need to expose is
the root PCI bus because the rest of the devices are discover-able
there.

So the real question is are there any other -devices that we want to be
able to graft FDT entries on or is the generic loader enough of a
special case that we keep all the logic in there?

>
> Alistair
>
>>
>> FDT isn't the only way of passing system information up the boot
>> chain. We could reasonably want to do a similar thing with ACPI which
>> is what should be being used on SBSA like devices to communicate with
>> the hypervisor.
>>
>> Also relying on ,, in the QemuOpt parser is ugly. It might be worth
>> having better quoting support if I press on with this.
>>
>> So what do people think? Something worth developing?
>>
>>
>> Alex Bennée (4):
>>   hw/board: promote fdt from ARM VirtMachineState to MachineState
>>   hw/riscv: migrate fdt field to generic MachineState
>>   device_tree: add qemu_fdt_setprop_string_array helper
>>   generic_loader: allow the insertion of /chosen/module stanzas
>>
>>  include/hw/arm/virt.h            |   1 -
>>  include/hw/boards.h              |   1 +
>>  include/hw/core/generic-loader.h |   4 +
>>  include/hw/riscv/virt.h          |   1 -
>>  include/sysemu/device_tree.h     |  17 ++
>>  device_tree.c                    |  26 +++
>>  hw/arm/virt.c                    | 322 ++++++++++++++++---------------
>>  hw/core/generic-loader.c         |  42 ++++
>>  hw/riscv/virt.c                  |  18 +-
>>  9 files changed, 268 insertions(+), 164 deletions(-)
>>
>> --
>> 2.20.1
>>
>>


-- 
Alex Bennée