RE: [RFC v2 1/2] vfio/pci: export common symbols in vfio-pci

From: Liu, Yi L
Date: Tue Mar 26 2019 - 08:37:42 EST


> From: Alex Williamson [mailto:alex.williamson@xxxxxxxxxx]
> Sent: Tuesday, March 26, 2019 2:17 AM
> To: Liu, Yi L <yi.l.liu@xxxxxxxxx>
> Subject: Re: [RFC v2 1/2] vfio/pci: export common symbols in vfio-pci
>
> On Sat, 23 Mar 2019 11:06:44 +0000
> "Liu, Yi L" <yi.l.liu@xxxxxxxxx> wrote:
> > Hi Alex,

[...]

> >
> > I tried to get a common file which includes the definitions of the module
> > options and the common interfaces and get it linked separately with each
> > module. It works well when linked separately by config the
> > CONFIG_VFIO_PCI=m and CONFIG_VFIO_PCI_MDEV=m in kernel
> > configuration file. CONFIG_VFIO_PCI_MDEV is a new Kconfig macro
> > for the mdev wrapped version driver. However, if building the vfio-pci
> > and the mdev wrapped version into kernel image (config the
> > CONFIG_VFIO_PCI=y and CONFIG_VFIO_PCI_MDEV=y), then the symbols
> > defined in the common file will be shared thus doesn't allow dissimilar
> > user settings.
> >
> > Per my understanding, I think we expect to allow simultaneous usage of
> > the two drivers. So I think the way above doesn't meet our expectation.
>
> I agree. They should be related in implementation only, from a user
> perspective they should be entirely separate.
>
> > I considered a possible proposal as below. May listen to your opinion
> > on it before heading to cook. Also, better idea is welcomed. :-)
> >
> > - get a common file includes interfaces which are common and have
> > input parameters to differentiate the calling from vfio-pci and the
> > wrapped version. e.g. vfio_pci_rw(). may call it as vfio_pci_common.c.
> >
> > - get another common file includes the definitions of the module options,
> > and the functions which referred the options. Define all of them as static.
> > may call it as common.c
> >
> > - get vfio_pci.c which includes the module_init/exit interfaces and driver
> > registration operations of vfio-pci.ko. This file should include the common.c
> > above to have same module options with the mdev wrapped version.
> >
> > - get vfio_pci_mdev.c which includes the module_init/exit interfaces and
> > driver registration operations of vfio-pci-mdev.ko. It should also include
> > the common.c above to have same module options with vfio-pci.ko.
> >
> > - Makefile:
> > vfio-pci-y := vfio_pci.o vfio_pci_common.o vfio_pci_intrs.o vfio_pci_rdwr.o
> vfio_pci_config.o
> > vfio-pci-$(CONFIG_VFIO_PCI_IGD) += vfio_pci_igd.o
> > vfio-pci-$(CONFIG_VFIO_PCI_NVLINK2) += vfio_pci_nvlink2.o
> >
> > vfio-pci-mdev-y := vfio_pci_mdev.o vfio_pci_common.o vfio_pci_intrs.o
> vfio_pci_rdwr.o vfio_pci_config.o
> > vfio-pci-mdev-$(CONFIG_VFIO_PCI_IGD) += vfio_pci_igd.o
> > vfio-pci-mdev-$(CONFIG_VFIO_PCI_NVLINK2) += vfio_pci_nvlink2.o
> >
> > obj-$(CONFIG_VFIO_PCI) += vfio-pci.o
> > obj-$(CONFIG_VFIO_PCI_MDEV) += vfio-pci-mdev.o
>
> Each module needs it's own module_init/exit and will register its own
> struct pci_driver, which gives us separate control of the probe and

Agreed.

> remove callbacks. I think we want the drivers to have the same module
> parameters initially, but we don't necessarily want to require it for
> any future options, so we can duplicate the parameter declarations.
> Then to support the shared code, I think we can easily push nointxmask,
> disable_vga, and disable_idle_d3 into bools on the struct
> vfio_pci_device, which would be allocated and set by each module's
> probe function before calling the shared probe function.

sounds good to me.

> vfio_fill_ids() could take a pointer to the array to keep them separate
> between modules.

Agreed.

> I think that just leaves the config permission bits,
> vfio_pci_{un}init_perm_bits(). Could we use a simple atomic reference
> counter on those to potentially share them so they get initialized by
> the first caller and freed by the last user, at least in the case of
> both drivers being compiled statically into the kernel? Thanks,

Sure, I can add it. The two modules will still share the cap_perms and
ecap_perms config bits when built statically in kernel. However, I think
such share is reasonable. I'll check if any other similar bits in other files.

> Alex

Thanks for the suggestions, Alex. Let me prepare another RFC.

Regards,
Yi Liu