Re: [PATCH 4/4] vfio: vfio_iommu_type1: implement VFIO_IOMMU_INFO_CAPABILITIES

[Pierre Morel]

On 16/05/2019 20:40, Alex Williamson wrote:
> On Fri, 10 May 2019 10:22:35 +0200
> Pierre Morel <pmorel@xxxxxxxxxxxxx> wrote:
>
> > We implement a capability intercafe for VFIO_IOMMU_GET_INFO and add the
> > first capability: VFIO_IOMMU_INFO_CAPABILITIES.
> >
> > When calling the ioctl, the user must specify
> > VFIO_IOMMU_INFO_CAPABILITIES to retrieve the capabilities and must check
> > in the answer if capabilities are supported.
> > Older kernel will not check nor set the VFIO_IOMMU_INFO_CAPABILITIES in
> > the flags of vfio_iommu_type1_info.
> >
> > The iommu get_attr callback will be called to retrieve the specific
> > attributes and fill the capabilities, VFIO_IOMMU_INFO_CAP_QFN for the
> > PCI query function attributes and VFIO_IOMMU_INFO_CAP_QGRP for the
> > PCI query function group.
> >
> > Signed-off-by: Pierre Morel <pmorel@xxxxxxxxxxxxx>
> > ---
> >   drivers/vfio/vfio_iommu_type1.c | 95 ++++++++++++++++++++++++++++++++++++++++-
> >   1 file changed, 94 insertions(+), 1 deletion(-)
> >
> > diff --git a/drivers/vfio/vfio_iommu_type1.c b/drivers/vfio/vfio_iommu_type1.c
> > index d0f731c..f7f8120 100644
> > --- a/drivers/vfio/vfio_iommu_type1.c
> > +++ b/drivers/vfio/vfio_iommu_type1.c
> > @@ -1658,6 +1658,70 @@ static int vfio_domains_have_iommu_cache(struct vfio_iommu *iommu)
> >   	return ret;
> >   }
> >   
> > +int vfio_iommu_type1_caps(struct vfio_iommu *iommu, struct vfio_info_cap *caps,
> > +			  size_t size)
> > +{
> > +	struct vfio_domain *d;
> > +	struct vfio_iommu_type1_info_block *info_fn;
> > +	struct vfio_iommu_type1_info_block *info_grp;
> > +	unsigned long total_size, fn_size, grp_size;
> > +	int ret;
> > +
> > +	d = list_first_entry(&iommu->domain_list, struct vfio_domain, next);
> > +	if (!d)
> > +		return -ENODEV;
> > +	/* The size of these capabilities are device dependent */
> > +	fn_size = iommu_domain_get_attr(d->domain,
> > +					DOMAIN_ATTR_ZPCI_FN_SIZE, NULL);
> > +	if (fn_size < 0)
> > +		return fn_size;
>
> What if non-Z archs want to use this?  The function is architected
> specifically for this one use case, fail if any component is not there
> which means it requires a re-write to add further support.  If
> ZPCI_FN_SIZE isn't support, move on to the next thing.

yes, clear.

> > +	fn_size +=  sizeof(struct vfio_info_cap_header);
> > +	total_size = fn_size;
>
> Here too, total_size should be initialized to zero and each section +=
> the size they'd like to add.

thanks, clear too.

> > +
> > +	grp_size = iommu_domain_get_attr(d->domain,
> > +					 DOMAIN_ATTR_ZPCI_GRP_SIZE, NULL);
> > +	if (grp_size < 0)
> > +		return grp_size;
> > +	grp_size +=  sizeof(struct vfio_info_cap_header);
> > +	total_size += grp_size;
> > +
> > +	/* Tell caller to call us with a greater buffer */
> > +	if (total_size > size) {
> > +		caps->size = total_size;
> > +		return 0;
> > +	}
> > +
> > +	info_fn = kzalloc(fn_size, GFP_KERNEL);
> > +	if (!info_fn)
> > +		return -ENOMEM;
>
> Maybe fn_size was zero because we're not on Z.
>
> > +	ret = iommu_domain_get_attr(d->domain,
> > +				    DOMAIN_ATTR_ZPCI_FN, &info_fn->data);
>
> Kernel internal structures != user api.  Thanks,
>
> Alex

Thanks a lot Alex,
I understand the concerns, I was too focussed on Z, I will rework this 
as you said:
- definition of the user API and
- take care that another architecture may want to use the interface.

Regards,
Pierre



--
Pierre Morel
Linux/KVM/QEMU in BÃblingen - Germany