On Tue, 2022-09-27 at 13:56 -0300, Jason Gunthorpe wrote: > On Tue, Sep 27, 2022 at 06:33:48PM +0200, Niklas Schnelle wrote: > > > Not sure what the non-MSI reservation is for? It does seem like x86_64 > > also uses this for quite large ranges. > > There are lots of things that are unsuitable for DMA on x86 platforms, > unfortunately.. But yeah, I'm not sure either. > > > This is because I'm getting a map request for an IOVA in the reserved > > region. > > How come? iova_reserve_iommu_regions() reads the reserved regions and > loads them as reserved into the iovad which should cause > iommu_dma_alloc_iova() and alloc_iova_fast() to not return values in > those ranges. > > It all looks like it is supposed to work > > Did something go wrong in the initialization order perhaps? > > Jason It was of course a classic off-by-one, the table size is a number of entries but geometry.aperture_end seems to be the largest allowed IOVA. So we need: s390_domain->domain.geometry.force_aperture = true; s390_domain->domain.geometry.aperture_start = 0; s390_domain->domain.geometry.aperture_end = ZPCI_TABLE_SIZE_RT - 1; Otherwise the first IOVA allocated is ZPCI_TABLE_SIZE_RT itself. Similarly we need the second reserved region if (zdev->end_dma < ZPCI_TABLE_SIZE_RT - 1). In your patch I think you had the MAX_DMA_TABLE_ADDR name right but would have also calculated the number of entries. On the other hand with the dma-iommu.c conversion it no longer makes sense to lower zdev->end_dma artificially, so at least on current machine LPARs we would end up with just a lower reserved region 0x0000000000000000 to 0x00000000ffffffff and can use IOVAs up to aperture_end.
