Re: [PATCH 06/11] iommu: Expose group variants of dma ownership interfaces

From: Lu Baolu
Date: Tue Nov 16 2021 - 04:42:21 EST

Hi Christoph,

On 2021/11/15 21:27, Christoph Hellwig wrote:
On Mon, Nov 15, 2021 at 10:05:47AM +0800, Lu Baolu wrote:
The vfio needs to set DMA_OWNER_USER for the entire group when attaching

The vfio subsystem? driver?

"vfio subsystem"

it to a vfio container. So expose group variants of setting/releasing dma
ownership for this purpose.

This also exposes the helper iommu_group_dma_owner_unclaimed() for vfio
report to userspace if the group is viable to user assignment, for

.. for vfio to report .. ?


void iommu_device_release_dma_owner(struct device *dev, enum iommu_dma_owner owner);
+int iommu_group_set_dma_owner(struct iommu_group *group, enum iommu_dma_owner owner,
+ struct file *user_file);
+void iommu_group_release_dma_owner(struct iommu_group *group, enum iommu_dma_owner owner);

Pleae avoid all these overly long lines.

Sure. Thanks!

+static inline int iommu_group_set_dma_owner(struct iommu_group *group,
+ enum iommu_dma_owner owner,
+ struct file *user_file)
+ return -EINVAL;
+static inline void iommu_group_release_dma_owner(struct iommu_group *group,
+ enum iommu_dma_owner owner)
+static inline bool iommu_group_dma_owner_unclaimed(struct iommu_group *group)
+ return false;

Why do we need these stubs? All potential callers should already
require CONFIG_IOMMU_API? Same for the helpers added in patch 1, btw.

You are right. This helper is only for vfio which requires IOMMU_API. I
will remove this.

The helpers in patch 1 seem not the same. The driver core (or bus
dma_configure() callback as suggested) will also call them.

+ mutex_lock(&group->mutex);
+ ret = __iommu_group_set_dma_owner(group, owner, user_file);
+ mutex_unlock(&group->mutex);

+ mutex_lock(&group->mutex);
+ __iommu_group_release_dma_owner(group, owner);
+ mutex_unlock(&group->mutex);

Unless I'm missing something (just skipping over the patches),
the existing callers also take the lock just around these calls,
so we don't really need the __-prefixed lowlevel helpers.

Move mutex_lock/unlock will make the helper implementation easier. :-)
It seems to be common code style in iommu core. For example,
__iommu_attach_group(), __iommu_group_for_each_dev(), etc.

+ mutex_lock(&group->mutex);
+ owner = group->dma_owner;
+ mutex_unlock(&group->mutex);

No need for a lock to read a single scalar.

Adding the lock will make kcasn happy. Jason G also told me that

[citing from his review comment]
It is always incorrect to read concurrent data without an annotation
of some kind.

For instance it can cause mis-execution of logic where the compiler is
unaware that a value it loads is allowed to change - ie no READ_ONCE/WRITE_ONCE semantic.

+ return owner == DMA_OWNER_NONE;

Best regards,