[SeaBIOS] [RFC v3] pciinit: setup mcfg for pxb-pcie to support multiple pci domains

[Zihan Yang]

HI Laszlo
Laszlo Ersek <lersek at redhat.com> 于2018年9月26日周三 上午1:17写道：
>
> On 09/25/18 17:38, Kevin O'Connor wrote:
> > On Mon, Sep 17, 2018 at 11:02:59PM +0800, Zihan Yang wrote:
> >> To support multiple pci domains of pxb-pcie device in qemu, we need to setup
> >> mcfg range in seabios. We use [0x80000000, 0xb0000000) to hold new domain mcfg
> >> table for now, and we need to retrieve the desired mcfg size of each pxb-pcie
> >> from a hidden bar because they may not need the whole 256 busses, which also
> >> enables us to support more domains within a limited range (768MB)
> >
> > At a highlevel, this looks okay to me.  I'd like to see additional
> > reviews from others more familiar with the QEMU PCI code, though.
> >
> > Is the plan to do the same thing for OVMF?
>
> I remain entirely unconvinced that this feature is useful. (I've stated
> so before.)
>
> I believe the latest QEMU RFC posting (v5) is here:
>
> [Qemu-devel] [RFC v5 0/6] pci_expander_brdige: support separate pci
> domain for pxb-pcie
>
> http://mid.mail-archive.com/1537196258-12581-1-git-send-email-whois.zihan.yang@gmail.com
>
> First, I fail to see the use case where ~256 PCI bus numbers aren't
> enough. If I strain myself, perhaps I can imagine using ~200 PCIe root
> ports on Q35 (each of which requires a separate bus number), so that we
> can independently hot-plug 200 devices then. And that's supposedly not
> enough, because we want... 300? 400? A thousand? Doesn't sound realistic
> to me. (This is not meant to be a strawman argument, I really have no
> idea what the feature would be useful for.)

It might not be very intuitive, but it indeed exists. The very
beginning discussion
about 4 months ago has mentioned a possible use case, and I paste it here

- We have Ray from Intel trying to use 1000 virtio-net devices
- We may have a VM managing some backups (tapes), we may have a lot of these.
- We may want indeed to create a nested solution as Michael mentioned.

The thread can be found in
https://lists.gnu.org/archive/html/qemu-devel/2018-05/msg04667.html

Also, a later post from a person in Dell stated in the list that he would need
this feature for Intel VMD In Dell EMC? I have no idea about the details,
but since they went here for help, I guess they do can benefit from it somehow.

> Second, the v5 RFC doesn't actually address the alleged bus number
> shortage. IIUC, it supports a low number of ECAM ranges under 4GB, but
> those are (individually) limited in the bus number ranges they can
> accommodate (due to 32-bit address space shortage). So more or less the
> current approach just fragments the bus number space we already have, to
> multiple domains.
>
> Third, should a subsequent iteration of the QEMU series put those extra
> ECAMs above 4GB, with the intent to leave the enumeration of those
> hierarchies to the "guest OS", it would present an incredible
> implementation mess for OVMF. If people gained the ability to attach
> storage or network to those domains, on the QEMU command line, they
> would expect to boot off of them, using UEFI. Then OVMF would have to
> make sure the controllers could be bound by their respective UEFI
> drivers. That in turn would require functional config space access
> (ECAM) at semi-random 64-bit addresses.

I'm not familiar with OVMF, so I'm afraid I don't know how to make it easier
for OVMF, the division of 64bit space in OVMF is out of my purpose. There
is no plan to implement it in OVMF for now, we just want to make the
seabios/qemu patch a proof of concept.

As for the seabios, it access devices through port 0xcf8/0xcfc, which is binded
to q35 host in qemu. If we want to change the mmconfig size of pxb-pcie
(instead of using whole 256MB), we must know its desired size, which is passed
as a hidden bar. Unfortunately the configuration space of pxb-pcie device
cannot be accessed with 0xcf8/0xcfc because they are in different host bridge.
At this time the ECAM is not configured yet, so no able to use MMIO too.
In previous version, I tried to bind pxb host to other ports in qemu, so that we
can use port io to access the config space of pxb-pcie, but it seems a
little dirty,

Another issue is how seabios initialize things. It will first do pci_setup when
things like RSDP is not loaded. It is inconvenient to retrieve MCFG table and
other information, so we cannot infer the mmconfig addr and size from MCFG
table in seabios.

Therefore we fall back to an alternative that we support 4x of devices as the
first step, and let the guest os do the initialization. The inability to boot
from devices in another domain is indeed an issue, and we don't have very
good solution to it yet.

Things might change in the future if we can figure out a better solution, and I
hope we can have an easier and more elegant solution in OVMF. But now
we are just trying to give a possible solution as a poc.

Thanks
Zihan