Re: [PATCH v8 13/14] IB/core: use MAP_DIRECT to fix / enable RDMA to DAX mappings

On Wed, Oct 11, 2017 at 4:54 AM, Joerg Roedel [off-list ref] wrote:

On Tue, Oct 10, 2017 at 07:50:12AM -0700, Dan Williams wrote:

quoted

+static void ib_umem_lease_break(void *__umem)
+{
+     struct ib_umem *umem = umem;
+     struct ib_device *idev = umem->context->device;
+     struct device *dev = idev->dma_device;
+     struct scatterlist *sgl = umem->sg_head.sgl;
+
+     iommu_unmap(umem->iommu, sg_dma_address(sgl) & PAGE_MASK,
+                     iommu_sg_num_pages(dev, sgl, umem->npages));
+}

This looks like an invitation to break your code by random iommu-driver
changes. There is no guarantee that an iommu-backed dma-api
implemenation will map exactly iommu_sg_num_pages() pages for a given
sg-list. In other words, you are mixing the use of the IOMMU-API and the
DMA-API in an incompatible way that only works because you know the
internals of the iommu-drivers.

I've seen in another patch that your changes strictly require an IOMMU,
so you what you should do instead is to switch from the DMA-API to the
IOMMU-API and do the address-space management yourself.

Ok, I'll switch over completely to the iommu api for this. It will
also address Robin's concern.