Re: [RFC 4/7] gpu: nova-core: populate GSP_VF_INFO when vGPU is enabled

[Joel Fernandes <joelagnelf@nvidia.com>]

Hi Zhi,

On Tue, Dec 09, 2025 at 03:41:14PM +0200, Zhi Wang wrote:
> On Sat, 6 Dec 2025 21:32:51 -0500
> Joel Fernandes <joelagnelf@nvidia.com> wrote:
[..]
> > > 0x00000004); +
> > > +            let val = bar.read32(0x88000 + 0xbfc);
> > > +            info.b64bitBar2 = u8::from((val & 0x00000006) ==
> > > 0x00000004);
> > 
> > Please no magic numbers, please use proper named constants with
> > documentation comments explaining the values.
> > 
> > Also BAR reads here need proper register macro definitions/access.
> > 
> 
> That is true. :) But this is because there is no register definition in
> the OpenRM code/non OpenRM code as well. I have no idea about the name
> and bit definitions of this register.
> 
> Suppose I will have to find some clues from some folks then document
> them here when going to patches request for merged. :)


I think these magic numbers are PCIe config space related. I found a couple of references [1] [2] [3]

[1]
In Open GPU docs, I see 0x00088000 is NV_PCFG but this is on Turing, lets
confirm what it is on other architectures (if not common, should it go
through a HAL?).

https://github.com/NVIDIA/open-gpu-kernel-modules/blob/a5bfb10e75a4046c5d991c65f49b5d29151e68cf/src/common/inc/swref/published/turing/tu102/dev_nv_xve.h#L4

and 0xbf4 is SRIOV capability headers, per the same header file:
NV_XVE_SRIOV_CAP_HDR10

Also the bit definition is not documented in that public header, but I find
from internal sources that what you're trying to do with the "& 0x6" is
determine whether the VF BAR is capable of 64-bit addressing:

 Bits [2:1] is VF_BAR1_ADR_TYPE and = 2 means the BAR is capable of 64-bit
 addressing, and = 0 means 32-bit.

I wonder if the format of these capability headers are present in the PCI
specification? It is worth checking, I find some very similar mentions of the
value 2 being 64-bit in https://wiki.osdev.org/PCI as well.

[2]
In Nouveau I found the 0x88000
  drivers/gpu/drm/nouveau/nouveau_reg.h +684

With a bunch of ids and such which is typical of what is in config space:

#    define NV50_PBUS_PCI_ID                                0x00088000
#        define NV50_PBUS_PCI_ID_VENDOR_ID                  0x0000ffff
#        define NV50_PBUS_PCI_ID_VENDOR_ID__SHIFT                    0
#        define NV50_PBUS_PCI_ID_DEVICE_ID                  0xffff0000
#        define NV50_PBUS_PCI_ID_DEVICE_ID__SHIFT                   16

Perhaps this is something pdev.config_read_dword() should be giving?

[3] This one I am not sure off, but the link
https://envytools.readthedocs.io/en/latest/hw/bus/pci.html says that on NV40+
cards, all 0x1000 bytes of PCIE config space are mapped to MMIO register
space at addresses 0x88000-0x88fff. This matches exactly the magic number in
your patch.

Also, I wonder if we need to determine if the BARs can be 64-bit addressed, do
we have requirements for BAR sizes > 4GB for vGPU and if not, do we need to
determine the BAR size addressability?

Also, shouldn't the PCI core subsystem be automatically determining if the
BARs are 64-bit addressable? Not sure if that belongs in the driver. It would
be good to understand how this is supposed to work.

thanks,

 - Joel