On Thu, 24 Sep 2020 at 15:45, Ard Biesheuvel <ardb@xxxxxxxxxx> wrote:
>
> On Thu, 24 Sep 2020 at 15:28, Rob Herring <robh@xxxxxxxxxx> wrote:
> >
> > On Thu, Sep 24, 2020 at 5:00 AM Ard Biesheuvel <ardb@xxxxxxxxxx> wrote:
> > >
> > > On Wed, 23 Sep 2020 at 08:28, Jisheng Zhang <Jisheng.Zhang@xxxxxxxxxxxxx> wrote:
> > > >
> > > > Currently, dw_pcie_msi_init() allocates and maps page for msi, then
> > > > program the PCIE_MSI_ADDR_LO and PCIE_MSI_ADDR_HI. The Root Complex
> > > > may lose power during suspend-to-RAM, so when we resume, we want to
> > > > redo the latter but not the former. If designware based driver (for
> > > > example, pcie-tegra194.c) calls dw_pcie_msi_init() in resume path, the
> > > > previous msi page will be leaked.
> > > >
> > > > Move the allocate and map msi page from dw_pcie_msi_init() to
> > > > dw_pcie_host_init() to fix this problem.
> > > >
> > > > Fixes: 56e15a238d92 ("PCI: tegra: Add Tegra194 PCIe support")
> > > > Signed-off-by: Jisheng Zhang <Jisheng.Zhang@xxxxxxxxxxxxx>
> > >
> > > Why do you allocate a page for this in the first place? Isn't
> > > PCIE_MSI_ADDR_HI:PCIE_MSI_ADDR_LO simply a magic DMA address that
> > > never gets forwarded across to the CPU side of the host bridge, and
> > > triggers a SPI instead, which gets handled by reading
> > > PCIE_MSI_INTR0_STATUS ?
> >
> > My question too after digging into this some more. I've asked the
> > question on the thread that further complicated all this changing from
> > virt_to_phys() to dma_map_page()[1].
> >
> > > Couldn't you just map the zero page instead?
> >
> > Why a page even? You could use PCIE_MSI_ADDR_LO address itself even.
> > Or just an address in the driver data which is what some other drivers
> > do.
> >
>
> PCIE_MSI_ADDR_LO itself could collide with an actual DRAM address if
> any translation is applied on inbound transactions. Using an actual
> DRAM address avoids that.
... although the MSI doorbell register on the GIC, for instance, needs
to be DMA addressable as well, of course.
