Re: [PATCH v8 6/8] iommu/msi-iommu: iommu_msi_domain

[Alex Williamson]

On Thu, 28 Apr 2016 08:15:21 +0000
Eric Auger <eric.auger@xxxxxxxxxx> wrote:

> This function checks whether
> - the device belongs to a non default iommu domain
> - this iommu domain requires the MSI address to be mapped.
> 
> If those conditions are met, the function returns the iommu domain
> to be used for mapping the MSI doorbell; else it returns NULL.
> 
> Signed-off-by: Eric Auger <eric.auger@xxxxxxxxxx>
> 
> ---
> 
> v7 -> v8:
> - renamed  iommu_msi_mapping_desc_to_domain to iommu_msi_domain
> - the function now takes a struct device *
> - use DOMAIN_ATTR_MSI_GEOMETRY attribute
> ---
>  drivers/iommu/msi-iommu.c | 17 +++++++++++++++++
>  include/linux/msi-iommu.h | 14 ++++++++++++++
>  2 files changed, 31 insertions(+)
> 
> diff --git a/drivers/iommu/msi-iommu.c b/drivers/iommu/msi-iommu.c
> index 203e86e..023ff17 100644
> --- a/drivers/iommu/msi-iommu.c
> +++ b/drivers/iommu/msi-iommu.c
> @@ -243,3 +243,20 @@ unlock:
>  	}
>  }
>  EXPORT_SYMBOL_GPL(iommu_msi_put_doorbell_iova);
> +
> +struct iommu_domain *iommu_msi_domain(struct device *dev)
> +{
> +	struct iommu_domain *d = iommu_get_domain_for_dev(dev);
> +	struct iommu_domain_msi_geometry msi_geometry;
> +
> +	if (!d || (d->type == IOMMU_DOMAIN_DMA))
> +		return NULL;
> +
> +	iommu_domain_get_attr(d, DOMAIN_ATTR_MSI_GEOMETRY, &msi_geometry);
> +	if (!msi_geometry.programmable)

It feels like we're conflating ideas with using the "programmable" flag
in this way.  AIUI the IOMMU API consumer is supposed to see the
invalid MSI geometry with the programmable feature set and know to call
iommu_msi_set_aperture().  Looking for the msi_cookie would tell us if
that had been done, but that doesn't tell us if it should have been
done.  iommu_msi_msg_pa_to_va() handles this later, if we return
NULL here that function returns success otherwise it goes on to fail if
the iova or msi cookie is not set.  So really what we're trying to flag
is that the MSI geometry participates in the IOMMU-MSI API you've
created and we should pick a feature name that says that rather than
something as vague a "programmable".  Perhaps simply iommu_msi_api
rather than programmable.

BTW, I don't see that you ever set aperture_start/end once
iommu_msi_set_aperture() has been called.  It seems like doing so would
add some consistency to that MSI geometry attribute.

Nice series overall.  Thanks,

Alex

> +		return NULL;
> +
> +	return d;
> +}
> +EXPORT_SYMBOL_GPL(iommu_msi_domain);
> +
> diff --git a/include/linux/msi-iommu.h b/include/linux/msi-iommu.h
> index 1cd115f..114bd69 100644
> --- a/include/linux/msi-iommu.h
> +++ b/include/linux/msi-iommu.h
> @@ -81,6 +81,15 @@ int iommu_msi_get_doorbell_iova(struct iommu_domain *domain,
>   */
>  void iommu_msi_put_doorbell_iova(struct iommu_domain *domain, phys_addr_t addr);
>  
> +/**
> + * iommu_msi_domain: in case the device is upstream to an IOMMU and this IOMMU
> + * translates the MSI transaction, returns the iommu domain the MSI doorbell
> + * address must be mapped in; else returns NULL.
> + *
> + * @dev: device handle
> + */
> +struct iommu_domain *iommu_msi_domain(struct device *dev);
> +
>  #else
>  
>  static inline int
> @@ -100,5 +109,10 @@ static inline int iommu_msi_get_doorbell_iova(struct iommu_domain *domain,
>  static inline void iommu_msi_put_doorbell_iova(struct iommu_domain *domain,
>  					       phys_addr_t addr) {}
>  
> +static inline struct iommu_domain *iommu_msi_domain(struct device *dev)
> +{
> +	return NULL;
> +}
> +
>  #endif	/* CONFIG_IOMMU_MSI */
>  #endif	/* __MSI_IOMMU_H */