Re: [PATCH 3/4] vdpa: get_iova_range() is mandatory for device specific DMA translation

From: Jason Wang
Date: Wed Aug 05 2020 - 23:27:38 EST

On 2020/8/5 下午8:55, Michael S. Tsirkin wrote:
On Wed, Jun 17, 2020 at 11:29:46AM +0800, Jason Wang wrote:
In order to let userspace work correctly, get_iova_range() is a must
for the device that has its own DMA translation logic.
I guess you mean for a device.

However in absence of ths op, I don't see what is wrong with just
assuming device can access any address.

It's just for safe, if you want, we can assume any address without this op.

Signed-off-by: Jason Wang <jasowang@xxxxxxxxxx>
drivers/vdpa/vdpa.c | 4 ++++
1 file changed, 4 insertions(+)

diff --git a/drivers/vdpa/vdpa.c b/drivers/vdpa/vdpa.c
index de211ef3738c..ab7af978ef70 100644
--- a/drivers/vdpa/vdpa.c
+++ b/drivers/vdpa/vdpa.c
@@ -82,6 +82,10 @@ struct vdpa_device *__vdpa_alloc_device(struct device *parent,
if (!!config->dma_map != !!config->dma_unmap)
goto err;
+ if ((config->dma_map || config->set_map) &&
+ !config->get_iova_range)
+ goto err;
err = -ENOMEM;
vdev = kzalloc(size, GFP_KERNEL);
if (!vdev)
What about devices using an IOMMU for translation?
IOMMUs generally have a limited IOVA range too, right?

See patch 4 which query the IOMMU geometry in this case:

+        iommu_domain_get_attr(v->domain,
+                      DOMAIN_ATTR_GEOMETRY, &geo);
+        range.start = geo.aperture_start;
+        range.end = geo.aperture_end;