On 27-Oct-23 10:54 AM, Christoph Hellwig wrote:
On Wed, Oct 25, 2023 at 12:26:58PM +0530, jeshwank wrote:
+ tee_buf->vaddr = dma_alloc_coherent(psp->dev, size, &tee_buf->dma, gfp);
+ if (!tee_buf->vaddr || !tee_buf->dma) {
+ kfree(tee_buf);
+ return NULL;
+ }
+
+ tee_buf->size = size;
+
+ /* Check whether IOMMU is present. If present, translate IOVA
+ * to physical address, else the dma handle is the physical
+ * address.
+ */
+ dom = iommu_get_domain_for_dev(psp->dev);
+ if (dom)
+ tee_buf->paddr = iommu_iova_to_phys(dom, tee_buf->dma);
+ else
No, you can't mix the DMA API and iommu API. You need to stick to one
or the other.
Can you please elaborate a bit more? Is it because in the presence of IOMMU,
a contiguous DMA or bus-address-space "buffer" may be mapped non-contiguously
into the physical address space? As a result, for buffers larger than a page
size, when PSP tries to map the physical address into it's address space, the
PSP Trusted OS will not be able to read back the entire buffer data.
CPU CPU Bus
Virtual Physical Address
Address Address Space
Space Space
+-------+ +------+ +------+
| | |MMIO | Offset | |
| | Virtual |Space | applied | |
C +-------+ --------> B +------+ ----------> +------+ A
| | mapping | | by host | |
+-----+ | | | | bridge | | +--------+
| | | | +------+ | | | |
| CPU | | | | RAM | | | | Device |
| | | | | | | | | |
+-----+ +-------+ +------+ +------+ +--------+
| | Virtual |Buffer| Mapping | |
X +-------+ --------> Y +------+ <---------- +------+ Z
| | mapping | RAM | by IOMMU
| | | |
| | | |
+-------+ +------+
Reference diagram from:
https://www.kernel.org/doc/Documentation/DMA-API-HOWTO.txt
Regards,
Jeshwanth