On Tue, Oct 9, 2018 at 1:27 PM Arnd Bergmann <arnd@xxxxxxxx> wrote:
>
> On Tue, Oct 9, 2018 at 9:03 AM Sunil Kovvuri <sunil.kovvuri@xxxxxxxxx> wrote:
> > On Mon, Oct 8, 2018 at 5:38 PM Arnd Bergmann <arnd@xxxxxxxx> wrote:
> > > On Sun, Oct 7, 2018 at 5:01 PM <sunil.kovvuri@xxxxxxxxx> wrote:
> > >
> > > >
> > > > + /* HW interprets RVU_AF_MSIXTR_BASE address as an IOVA, hence
> > > > + * create a IOMMU mapping for the physcial address configured by
> > > > + * firmware and reconfig RVU_AF_MSIXTR_BASE with IOVA.
> > > > + */
> > > > + cfg = rvu_read64(rvu, BLKADDR_RVUM, RVU_PRIV_CONST);
> > > > + max_msix = cfg & 0xFFFFF;
> > > > + phy_addr = rvu_read64(rvu, BLKADDR_RVUM, RVU_AF_MSIXTR_BASE);
> > > > + iova = dma_map_single(rvu->dev, (void *)phy_addr,
> > > > + max_msix * PCI_MSIX_ENTRY_SIZE,
> > > > + DMA_BIDIRECTIONAL);
> > > > + if (dma_mapping_error(rvu->dev, iova))
> > > > + return -ENOMEM;
> > > > +
> > > > + rvu_write64(rvu, BLKADDR_RVUM, RVU_AF_MSIXTR_BASE, (u64)iova);
> > > > + rvu->msix_base_iova = iova;
> > > > +
> > >
> > > I'm a bit puzzled by how this works. Does this rely on a specific iommu
> > > driver implementation? Normally a physical address makes no sense to the
> > > implementation backing dma_map_single() that tries to convert a
> > > linear kernel virtual address into a physical address.
> > >
> > > Arnd
> >
> > I understand what you are pointing at, but we did test this.
> > IOMMU on this silicon is standard ARM64 SMMUv3.
>
> I didn't doubt that you get the right results, I just couldn't see how ;-)
>
> > All dma_map_single does is virt_to_page and iommu_dma_map_page
> > converts it back i.e page_to_phys.
> > So the IOMMU driver gets the same physical address passed above and
> > creates a iova translation mapping.
> >
> > For reference below is the captured debug info for the same
> > =====
> > [ 19.435968] rvu_setup_msix_resources: phy_addr 0x3200000 iova 0xfff80000
> > [ 19.436967] rvu_setup_msix_resources: virt_to_page(phy_addr)
> > 0xffff7fe00000c800 page_to_phys(page) 0x3200000
> > offset_in_page(phy_addr) 0x00
> > =====
>
> I think if you enable CONFIG_DEBUG_VIRTUAL, the virt_to_page()
> above should trigger a warning in
>
> phys_addr_t __virt_to_phys(unsigned long x)
> {
> WARN(!__is_lm_address(x),
> "virt_to_phys used for non-linear address: %pK (%pS)\n",
> (void *)x,
> (void *)x);
>
> return __virt_to_phys_nodebug(x);
> }
>
> Can you verify that?
>
> Arnd
No, it isn't, as for 64bit systems CONFIG_SPARSEMEM_VMEMMAP is enabled.
But is there any alternative to what is being done ?
Thanks,
Sunil.
