Re: [Qemu-devel] How to generate custom fw paths for IDE devices?

[Laszlo Ersek <lersek@redhat.com>]

On 07/19/18 19:19, Mark Cave-Ayland wrote:
> On 19/07/18 09:29, Laszlo Ersek wrote:
>
>> (updating Marcel's address to his GMail one)
>>
>> On 07/18/18 23:13, Mark Cave-Ayland wrote:
>>> Hi all,
>>>
>>> Following on from a couple of patches I've previously posted to the
>>> mailing list at
>>> https://lists.gnu.org/archive/html/qemu-devel/2018-06/msg08836.html
>>> I've made some good progress with trying to add bootindex support to
>>> OpenBIOS but I'm stuck with generating the IDE device paths from
>>> QEMU.
>>>
>>> According to OpenBIOS the device path for a cdrom on a sun4u machine
>>> should be:
>>>
>>>    /pci@1fe,0/pci@1,1/ide@3/ide1@8100/cdrom@0
>>>
>>> whereas with my working patchset I'm currently generating:
>>>
>>>    /pci@1fe,0/pci@1,1/ide@3/drive@1
>>>
>>> The issue is that the drive@1 part is being generated by the IDE
>>> drive device attached to the IDE bus in hw/ide/qdev.c, and so I
>>> think I need to override idebus_get_fw_dev_path() to manually
>>> generate the remainder of the path including both the controller and
>>> the correctly named drive node.
>>>
>>> One option may be to consider subclassing IDEBus and overriding
>>> idebus_get_fw_dev_path() there, but the cmd646 device is a child of
>>> TYPE_PCI_IDE which has its own internal IDEBus and so it seems
>>> overriding it is impossible.
>>>
>>> Can anyone point me in the right direction as to how to generate the
>>> correct fw path for IDE devices in the above format for sun4u
>>> machines?
>>
>> What prevents you from recognizing, in the guest firmware, the
>> OpenFirmware device path that is currently generated by QEMU?
>
> In a word: compatibility. With the older SPARC/PPC machines the OSs
> are *very* picky about device names and often make assumptions about
> the DT layout instead of parsing it properly, simply because at that
> time no-one really thought of running the software on anything but a
> real machine.
>
>> I mean, the device path generated by QEMU looks technically correct;
>> it reflects how the IDE controller sits in a PCI B/D/F, and how the
>> IDE drive sits on an IDE controller. Or do you actually have an
>> intermediate IDE controller (at "address 0x8100" on the top IDE
>> controller) in the sun4u machine type? Is the address 0x8100 actually
>> needed by the firmware?
>>
>> If so, perhaps you could turn that intermediate IDE controller
>> ("internal IDEBus") into its own class, and chain the instance of
>> that class like the rest of the bus controllers are chained. (Just
>> speculating...)
>
> In the case of the IDE device, the devices represent the
> primary/secondary interfaces and not the separate controllers: the
> primary lives at 0x8000 and the secondary at 0x8100.
>
> The other problem with changing this is that currently as there is no
> bootindex support, these map to the legacy -hda and -cdrom options
> correctly so even if I could do this, and even if all the OSs would
> still parse the device path, it would have to be a complete cutover
> which would likely involve me being on the receiving end of some angry
> emails.
>
> For PPC I've been playing with the macio device because MacOS refused
> to boot unless the nodes are named "ata-3" and "ata-4". So far I've
> only been able to do this by implementing a "dummy" macio Bus,
> reworking and attaching the macio-ide device to it, implementing
> minial Bus support, and then overriding the fw path function to
> rewrite the device name based upon its QOM type. That's a whole lot of
> work just to rename a device from "ide" to "ata-3" in the DT.
>
> If you take a look at the function Thomas mentioned in his email
> (https://git.qemu.org/?p=qemu.git;a=blob;f=hw/ppc/spapr.c;h=421b2dd09b515502cd11ccdd26420a8117f80cda;hb=e1ea55668ffe6ce558a063f3a9621b761738e1f2#l2866)
> suddenly it makes sense why I had to suggest patches for naming PCI
> devices: the SLOF/SPAPR authors decided to replace the entire
> FwPathProvider so then all that is necessary is to rewrite the paths
> for the relevant devices in one place; this also nicely handles the
> difference between IDE vs. SCSI vs. USB vs. virtio DT nodes.
>
> I certainly think it's worth keeping the PCI/sysbus patches I
> submitted to simplify the required logic, however it's clear to me
> that the above solution from SPAPR is going to be the best way forward
> for PPC and SPARC. Presumably this also explains why the patches
> didn't exist in the first place because the SLOF/SPAPR folks ignored
> the existing infrastructure and went ahead and did their own thing.
>
> From a design perspective I can completely understand why someone
> would come up with a design with a 1:1 correspondence between qdev and
> fw paths, but in reality it's the details that mean this just doesn't
> quite work in real life. In particular ISTM it's a big red flag that
> both IEEE-1275-based BIOSes, OpenBIOS and SLOF (upon which the design
> is heavily influenced) have to ignore the infrastructure based upon
> qdev and provide their own implementations.
>
> Are there any other similar issues around other BIOSes, e.g. s390,
> SeaBIOS at all?

I guess I could call the OVMF situation a "similar issue" :) UEFI is
totally independent of OpenFirmware (device paths and anything else); it
uses UEFI device paths. QEMU still populates the "bootorder" fw_cfg file
with OFW devpaths, so in OVMF I had to write a parser+translator, from
OFW to UEFI. It's one of the hairiest places in OVMF.

Obviously the particulars of the OFW devpaths exported by QEMU (i.e.,
the driver-name parts) didn't matter much, as long as the OFW devpaths
contained all the information necessary for the translation (including
structure / nesting), and as long as they were *stable*.

In practice they are stable, so that's great. Structurally /
nesting-wise, they are possible to translate too. Regarding the
"remaining" information content / expressive power, the OFW devpaths
generated by QEMU are not expressive enough -- in general -- to
*uniquely* map to UEFI boot options. So OVMF can only use them as
*prefixes*, after translation. These translated prefixes are used for
connecting (binding) sub-trees of bootable devices, and for filtering
and reordering UEFI boot options that core edk2 code generates for those
connected devices.

Nowadays we rarely have to touch this code; basically only when a new
kind of device (esp. bus) is added to QEMU (or needs to be enabled for
this kind of boot order matching in OVMF).

Oh and the PXB (PCI expander bus, a kind of extra root bridge/bus) was
particular fun. For recognizing *that*, we had to modify QEMU indeed. It
took a three-sided design discussion between SeaBIOS, QEMU, and OVMF. I
think you remember "explicit_ofw_unit_address", and
pxb_host_ofw_unit_address() :)

https://git.qemu.org/?p=qemu.git;a=commitdiff;h=48ea3dedc54dbcb3c738ddef02a336739910ecfd
https://github.com/tianocore/edk2/commit/5eb0b80afc4185f11379ab317f0b4d1b5520ef96
https://github.com/tianocore/edk2/commit/4fc18df9139bf5942249c77132d033a298b11c29

Laszlo