Hi Joerg On Mon, Mar 08, 2021 at 09:58:26AM +0800, Lu Baolu wrote: > Hi Joerg, > > On 3/4/21 8:26 PM, Joerg Roedel wrote: > >On Thu, Feb 25, 2021 at 02:26:51PM +0800, Lu Baolu wrote: > >>When the first level page table is used for IOVA translation, it only > >>supports Read-Only and Read-Write permissions. The Write-Only permission > >>is not supported as the PRESENT bit (implying Read permission) should > >>always set. When using second level, we still give separate permissions > >>that allows WriteOnly which seems inconsistent and awkward. There is no > >>use case we can think off, hence remove that configuration to make it > >>consistent. > > > >No use-case for WriteOnly mappings? How about DMA_FROM_DEVICE mappings? > > > > The statement of no use case is not correct. Sorry about it. > > As we have moved to use first level for IOVA translation, the first > level page table entry only provides RO and RW permissions. So if any > device driver specifies DMA_FROM_DEVICE attribution, it will get RW > permission in the page table. This patch aims to make the permissions > of second level and first level consistent. No impact on the use of > DMA_FROM_DEVICE attribution. >
That is the primary motivation, given that we have moved to 1st level for general IOVA, first level doesn't have a WO mapping. I didn't know enough about the history to determine if a WO without a READ is very useful. I guess the ZLR was invented to support those cases without a READ in PCIe. I Early Intel IOMMU's didn't handle ZLR properly, until we fixed it in the next generation. It just seemed opposite to the CPU page-tables, and we wanted to have consistent behavior. After moving to 1st level, we don't want things to work sometimes, and break if we use 2nd level for the same mappings. Hope this helps Cheers, Ashok
