Re: [PATCH v7 3/6] vfio/type1: Update iova list on detach

[Alex Williamson <alex.williamson@redhat.com>]

On Wed, 26 Jun 2019 16:12:45 +0100
Shameer Kolothum <shameerali.kolothum.thodi@huawei.com> wrote:

> Get a copy of iova list on _group_detach and try to update the list.
> On success replace the current one with the copy. Leave the list as
> it is if update fails.
> 
> Signed-off-by: Shameer Kolothum <shameerali.kolothum.thodi@huawei.com>
> ---
>  drivers/vfio/vfio_iommu_type1.c | 91 +++++++++++++++++++++++++++++++++
>  1 file changed, 91 insertions(+)
> 
> diff --git a/drivers/vfio/vfio_iommu_type1.c b/drivers/vfio/vfio_iommu_type1.c
> index b6bfdfa16c33..e872fb3a0f39 100644
> --- a/drivers/vfio/vfio_iommu_type1.c
> +++ b/drivers/vfio/vfio_iommu_type1.c
> @@ -1873,12 +1873,88 @@ static void vfio_sanity_check_pfn_list(struct vfio_iommu *iommu)
>  	WARN_ON(iommu->notifier.head);
>  }
>  
> +/*
> + * Called when a domain is removed in detach. It is possible that
> + * the removed domain decided the iova aperture window. Modify the
> + * iova aperture with the smallest window among existing domains.
> + */
> +static void vfio_iommu_aper_expand(struct vfio_iommu *iommu,
> +				   struct list_head *iova_copy)
> +{
> +	struct vfio_domain *domain;
> +	struct iommu_domain_geometry geo;
> +	struct vfio_iova *node;
> +	dma_addr_t start = 0;
> +	dma_addr_t end = (dma_addr_t)~0;
> +
> +	list_for_each_entry(domain, &iommu->domain_list, next) {
> +		iommu_domain_get_attr(domain->domain, DOMAIN_ATTR_GEOMETRY,
> +				      &geo);
> +		if (geo.aperture_start > start)
> +			start = geo.aperture_start;
> +		if (geo.aperture_end < end)
> +			end = geo.aperture_end;
> +	}
> +
> +	/* Modify aperture limits. The new aper is either same or bigger */
> +	node = list_first_entry(iova_copy, struct vfio_iova, list);
> +	node->start = start;
> +	node = list_last_entry(iova_copy, struct vfio_iova, list);
> +	node->end = end;
> +}
> +
> +/*
> + * Called when a group is detached. The reserved regions for that
> + * group can be part of valid iova now. But since reserved regions
> + * may be duplicated among groups, populate the iova valid regions
> + * list again.
> + */
> +static int vfio_iommu_resv_refresh(struct vfio_iommu *iommu,
> +				   struct list_head *iova_copy)
> +{
> +	struct vfio_domain *d;
> +	struct vfio_group *g;
> +	struct vfio_iova *node;
> +	dma_addr_t start, end;
> +	LIST_HEAD(resv_regions);
> +	int ret;
> +
> +	list_for_each_entry(d, &iommu->domain_list, next) {
> +		list_for_each_entry(g, &d->group_list, next)
> +			iommu_get_group_resv_regions(g->iommu_group,
> +						     &resv_regions);

Need to account for failure case here too.

> +	}
> +
> +	if (list_empty(&resv_regions))
> +		return 0;
> +
> +	node = list_first_entry(iova_copy, struct vfio_iova, list);
> +	start = node->start;
> +	node = list_last_entry(iova_copy, struct vfio_iova, list);
> +	end = node->end;
> +
> +	/* purge the iova list and create new one */
> +	vfio_iommu_iova_free(iova_copy);
> +
> +	ret = vfio_iommu_aper_resize(iova_copy, start, end);
> +	if (ret)
> +		goto done;
> +
> +	/* Exclude current reserved regions from iova ranges */
> +	ret = vfio_iommu_resv_exclude(iova_copy, &resv_regions);
> +done:
> +	vfio_iommu_resv_free(&resv_regions);
> +	return ret;
> +}
> +
>  static void vfio_iommu_type1_detach_group(void *iommu_data,
>  					  struct iommu_group *iommu_group)
>  {
>  	struct vfio_iommu *iommu = iommu_data;
>  	struct vfio_domain *domain;
>  	struct vfio_group *group;
> +	bool iova_copy_fail;
> +	LIST_HEAD(iova_copy);
>  
>  	mutex_lock(&iommu->lock);
>  
> @@ -1901,6 +1977,12 @@ static void vfio_iommu_type1_detach_group(void *iommu_data,
>  		}
>  	}
>  
> +	/*
> +	 * Get a copy of iova list. If success, use copy to update the
> +	 * list and to replace the current one.
> +	 */
> +	iova_copy_fail = !!vfio_iommu_iova_get_copy(iommu, &iova_copy);
> +
>  	list_for_each_entry(domain, &iommu->domain_list, next) {
>  		group = find_iommu_group(domain, iommu_group);
>  		if (!group)
> @@ -1926,10 +2008,19 @@ static void vfio_iommu_type1_detach_group(void *iommu_data,
>  			iommu_domain_free(domain->domain);
>  			list_del(&domain->next);
>  			kfree(domain);
> +			if (!iova_copy_fail && !list_empty(&iommu->domain_list))
> +				vfio_iommu_aper_expand(iommu, &iova_copy);
>  		}
>  		break;
>  	}
>  
> +	if (!iova_copy_fail && !list_empty(&iommu->domain_list)) {
> +		if (!vfio_iommu_resv_refresh(iommu, &iova_copy))
> +			vfio_iommu_iova_insert_copy(iommu, &iova_copy);
> +		else
> +			vfio_iommu_iova_free(&iova_copy);
> +	}

The iova_copy_fail and list_empty tests are rather ugly, could we avoid
them by pushing the tests to the expand and refresh functions?  ie. it
looks like vfio_iommu_aper_expand() could test list_empty(iova_copy),
the list_for_each on domain_list doesn't need special handling.  Same
for vfio_iommu_resv_refresh().  This would also fix the bug above that
I think we don't free iova_copy if domain_list becomes empty during
this operation.  Thanks,

Alex