On 26/01/17 09:53, Geert Uytterhoeven wrote: > Currently, the IPMMU/VMSA driver supports 32-bit I/O Virtual Addresses > only, and thus sets io_pgtable_cfg.ias = 32. However, it doesn't force > a 32-bit IOVA space through the IOMMU Domain Geometry. > > Hence if a device (e.g. SYS-DMAC) rightfully configures a 40-bit DMA > mask, it will still be handed out a 40-bit IOVA, outside the 32-bit IOVA > space, leading to out-of-bounds accesses of the PGD when mapping the > IOVA. > > Force a 32-bit IOMMU Domain Geometry to fix this. Reviewed-by: Robin Murphy <robin.murphy@xxxxxxx> > Signed-off-by: Geert Uytterhoeven <geert+renesas@xxxxxxxxx> > --- > Should the generic code restrict the geometry based on IAS instead? Which generic code? IAS is specific to the io-pgtable library (well, really it's an ARM-SMMU-ism, but "input address size" is a fairly portable concept), but io-pgtable is just factored-out driver helper code and doesn't know anything about iommu_domains and the IOMMU API. Conversely, the IOMMU API core and kernel code beyond also know nothing about io-pgtable internals - in fact the domain geometry *is* how the IOMMU driver communicates its configured address space size to the outside world. Robin. > --- > drivers/iommu/ipmmu-vmsa.c | 2 ++ > 1 file changed, 2 insertions(+) > > diff --git a/drivers/iommu/ipmmu-vmsa.c b/drivers/iommu/ipmmu-vmsa.c > index 798578f1676480c6..eb8b3df8733b15fb 100644 > --- a/drivers/iommu/ipmmu-vmsa.c > +++ b/drivers/iommu/ipmmu-vmsa.c > @@ -424,6 +424,8 @@ static int ipmmu_domain_init_context(struct ipmmu_vmsa_domain *domain) > domain->cfg.ias = 32; > domain->cfg.oas = 40; > domain->cfg.tlb = &ipmmu_gather_ops; > + domain->io_domain.geometry.aperture_end = DMA_BIT_MASK(32); > + domain->io_domain.geometry.force_aperture = true; > /* > * TODO: Add support for coherent walk through CCI with DVM and remove > * cache handling. For now, delegate it to the io-pgtable code. >