Re: [PATCH 02/10] swiotlb-xen: remove start_dma_addr

[Stefano Stabellini <sstabellini@xxxxxxxxxx>]

On Fri, 22 May 2020, Julien Grall wrote:
> On 22/05/2020 04:55, Stefano Stabellini wrote:
> > On Thu, 21 May 2020, Julien Grall wrote:
> > > Hi,
> > > 
> > > On 21/05/2020 00:45, Stefano Stabellini wrote:
> > > > From: Stefano Stabellini <stefano.stabellini@xxxxxxxxxx>
> > > > 
> > > > It is not strictly needed. Call virt_to_phys on xen_io_tlb_start
> > > > instead. It will be useful not to have a start_dma_addr around with the
> > > > next patches.
> > > > 
> > > > Signed-off-by: Stefano Stabellini <stefano.stabellini@xxxxxxxxxx>
> > > > ---
> > > >    drivers/xen/swiotlb-xen.c | 5 +----
> > > >    1 file changed, 1 insertion(+), 4 deletions(-)
> > > > 
> > > > diff --git a/drivers/xen/swiotlb-xen.c b/drivers/xen/swiotlb-xen.c
> > > > index a42129cba36e..b5e0492b07b9 100644
> > > > --- a/drivers/xen/swiotlb-xen.c
> > > > +++ b/drivers/xen/swiotlb-xen.c
> > > > @@ -52,8 +52,6 @@ static unsigned long xen_io_tlb_nslabs;
> > > >     * Quick lookup value of the bus address of the IOTLB.
> > > >     */
> > > >    -static u64 start_dma_addr;
> > > > -
> > > >    /*
> > > >     * Both of these functions should avoid XEN_PFN_PHYS because
> > > > phys_addr_t
> > > >     * can be 32bit when dma_addr_t is 64bit leading to a loss in
> > > > @@ -241,7 +239,6 @@ int __ref xen_swiotlb_init(int verbose, bool early)
> > > >                 m_ret = XEN_SWIOTLB_EFIXUP;
> > > >                 goto error;
> > > >         }
> > > > -       start_dma_addr = xen_virt_to_bus(xen_io_tlb_start);
> > > >         if (early) {
> > > >                 if (swiotlb_init_with_tbl(xen_io_tlb_start,
> > > > xen_io_tlb_nslabs,
> > > >                          verbose))
> > > > @@ -389,7 +386,7 @@ static dma_addr_t xen_swiotlb_map_page(struct device
> > > > *dev, struct page *page,
> > > >          */
> > > >         trace_swiotlb_bounced(dev, dev_addr, size, swiotlb_force);
> > > >    -    map = swiotlb_tbl_map_single(dev, start_dma_addr, phys,
> > > > +       map = swiotlb_tbl_map_single(dev, 
> > > > virt_to_phys(xen_io_tlb_start),
> > > > phys,
> > > 
> > > xen_virt_to_bus() is implemented as xen_phys_to_bus(virt_to_phys()). Can
> > > you
> > > explain how the two are equivalent?
> > 
> > They are not equivalent. Looking at what swiotlb_tbl_map_single expects,
> > and also the implementation of swiotlb_init_with_tbl, I think
> > virt_to_phys is actually the one we want.
> > 
> > swiotlb_tbl_map_single compares the argument with __pa(tlb) which is
> > __pa(xen_io_tlb_start) which is virt_to_phys(xen_io_tlb_start).
> 
> I can't find such check in master. What is your baseline? Could you point to
> the exact line of code?

My base is b85051e755b0e9d6dd8f17ef1da083851b83287d, which is master
from a couple of days back.


xen_swiotlb_init calls swiotlb_init_with_tbl which takes a virt address
as a parameter (xen_io_tlb_start), it gets converted to phys and stored
in io_tlb_start as a physical address.

Later, xen_swiotlb_map_page calls swiotlb_tbl_map_single passing a
dma addr as parameter (tbl_dma_addr). tbl_dma_addr is used to calculate
the right slot in the swiotlb buffer to use. (Strangely tbl_dma_addr is
a dma_addr_t and it is not converted to phys_addr_t before doing
operations... I think tbl_dma_addr should be a phys addr.)

The comparison with io_tlb_start is done here:

        do {
                while (iommu_is_span_boundary(index, nslots, offset_slots,
                                              max_slots)) {
                        index += stride;
                        if (index >= io_tlb_nslabs)
                                index = 0;
                        if (index == wrap)
                                goto not_found;
                }

index is io_tlb_start and offset_slots is derived by tbl_dma_addr.