Re: [PATCH v5 3/3] PCI: Set dma-can-stall for HiSilicon chips

[Will Deacon]

On Thu, Aug 26, 2021 at 08:12:12PM +0100, Robin Murphy wrote:
> On 2021-08-26 19:26, Bjorn Helgaas wrote:
> > [+cc Will, Robin, Joerg, hoping for an ack]
> > 
> > On Tue, Jul 13, 2021 at 10:54:36AM +0800, Zhangfei Gao wrote:
> > > HiSilicon KunPeng920 and KunPeng930 have devices appear as PCI but are
> > > actually on the AMBA bus. These fake PCI devices can support SVA via
> > > SMMU stall feature, by setting dma-can-stall for ACPI platforms.
> > > 
> > > Signed-off-by: Zhangfei Gao <zhangfei.gao@xxxxxxxxxx>
> > > Signed-off-by: Jean-Philippe Brucker <jean-philippe@xxxxxxxxxx>
> > > Signed-off-by: Zhou Wang <wangzhou1@xxxxxxxxxxxxx>
> > > ---
> > >   drivers/pci/quirks.c | 13 +++++++++++++
> > >   1 file changed, 13 insertions(+)
> > > 
> > > diff --git a/drivers/pci/quirks.c b/drivers/pci/quirks.c
> > > index 5d46ac6..03b0f98 100644
> > > --- a/drivers/pci/quirks.c
> > > +++ b/drivers/pci/quirks.c
> > > @@ -1823,10 +1823,23 @@ DECLARE_PCI_FIXUP_CLASS_FINAL(PCI_VENDOR_ID_HUAWEI, 0x1610, PCI_CLASS_BRIDGE_PCI
> > >   static void quirk_huawei_pcie_sva(struct pci_dev *pdev)
> > >   {
> > > +	struct property_entry properties[] = {
> > > +		PROPERTY_ENTRY_BOOL("dma-can-stall"),
> > 
> > "dma-can-stall" is used in arm_smmu_probe_device() to help set
> > master->stall_enabled.
> > 
> > I don't know the implications, so it'd be nice to get an ack from a
> > maintainer of that code.
> 
> If it helps,
> 
> Acked-by: Robin Murphy <robin.murphy@xxxxxxx>
> 
> Normally stalling must not be enabled for PCI devices, since it would break
> the PCI requirement for free-flowing writes and may lead to deadlock. We
> expect PCI devices to support ATS and PRI if they want to be fault-tolerant,
> so there's no ACPI binding to describe anything else, even when a "PCI"
> device turns out to be a regular old SoC device dressed up as a RCiEP and
> normal rules don't apply.
> 
> I'm taking it on trust that stalling really is safe for all possible
> matching devices here (in general, deadlock may still be possible in the SoC
> interconnect depending on topology, hence why it's an explicit opt-in even
> for platform devices), but TBH either way I think I'd rather have this as a
> quirk in the kernel under our control, than have vendors attempt to play
> tricks with _DSD properties out in the field :)

I think a comment next to the quirk echoing the commit message and your
first paragraph above would be really helpful here.

Will