[edk2-devel] [PATCH 0/9] ArmVirtPkg: support two PL011 UARTs

[Laszlo Ersek]

On 10/26/23 16:55, Ard Biesheuvel wrote:
> On Thu, 26 Oct 2023 at 16:46, Julien Grall <julien at xen.org> wrote:
>>
>> Hi,
>>
>> On 26/10/2023 15:21, Peter Maydell wrote:
>>> On Tue, 10 Oct 2023 at 16:33, Laszlo Ersek <lersek at redhat.com> wrote:
>>>> On 10/10/23 09:43, Ard Biesheuvel wrote:
>>>>> Thanks for looking into this - a cleanup was overdue here.
>>>>>
>>>>> I will take a look in more detail later, but one thing that occurred
>>>>> to me when reading this overview is that having a separate DEBUG
>>>>> serial port would permit us to
>>>>>
>>>>> a) remove it from the DT
>>>>
>>>> ... as in, hide it from Linux, I assume?
>>>>
>>>>> b) add a runtime mapping for it
>>>>> c) keep using it after ExitBootServices
>>>>>
>>>>> This could be useful for debugging issues with the variable store etc.
>>>>>
>>>>> Not saying this is something to address in this series, but I'd like
>>>>> to hear your take on this.
>>>>>
>>>>
>>>> Sounds like a useful feature.
>>>>
>>>> I see four challenges:
>>>>
>>>>
>>>> (1) We'd have to coordinate it with Peter. If we hide any one of the
>>>> serial ports from Linux, that may not be what QEMU intends for Linux to
>>>> happen. Linux currently ties getties to all serial ports -- via the
>>>> serial* aliases, IIUC. Thus, some "positive identification" in the DT
>>>> could be necessary (i.e., that edk2 was welcome to hide that port from
>>>> Linux).
>>>
>>> The potential awkwardness here is that what the guest thinks about
>>> the serial ports depends on the ACPI table fragments which QEMU
>>> provides. EDK2 would need to edit the table fragment to remove any
>>> mention of the second UART if it wanted to hide it from the kernel.
>>> I don't know how hard that would be in EDK2.
>>
>> I am not sure if it would help EDK2 in this case. But we had a similar
>> problem when adding support for ACPI in Xen. It was not trivial to
>> remove the UART from the ACPI tables provided by the host. So we ended
>> up to introduce the STAO table [1]. This is used to describe which
>> device will be hidden to the OS.
>>
> 
> I'd much rather have an implementation of the _STA method on those
> devices where the underlying AML queries the fwcfg MMIO device

Yes, this is possible too; the AML generated by QEMU need not be
constant, it can interact at OS runtime with QEMU. An example for this
is the VCPU hotplug register block. The hot(un)plug AML is super
sophisticated, it connects the guest kernel, QEMU (via the register
block) and the firmware (via raising SMIs).

One word of caution against accessing fw_cfg specifically, from AML:
fw_cfg is already exposed via guest sysfs (IIRC -- see
"drivers/firmware/qemu_fw_cfg.c"), because fw_cfg ended up being
"abused" (IMO) as a generic info channel from the host-side user to
guest OS + applications. (I call this "abuse" because "fw_cfg" literally
says "firmware configuration" in the name, and this particular use case
is anything *but* firmware configuration.) Accessing the same device
from AML may not be without risks.

But "non-fw_cfg platform registers" might work well with AML.

Laszlo



-=-=-=-=-=-=-=-=-=-=-=-
Groups.io Links: You receive all messages sent to this group.
View/Reply Online (#110123): https://edk2.groups.io/g/devel/message/110123
Mute This Topic: https://groups.io/mt/101834880/1813853
Group Owner: devel+owner at edk2.groups.io
Unsubscribe: https://edk2.groups.io/g/devel/leave/3943202/1813853/130120423/xyzzy [edk2-devel-archive at redhat.com]
-=-=-=-=-=-=-=-=-=-=-=-