On Tue, Jan 12, 2021 at 7:28 PM Ray Jui <ray.jui@xxxxxxxxxxxx> wrote:
> On 2020-12-15 7:49 a.m., Arnd Bergmann wrote:
> > On Tue, Dec 15, 2020 at 4:41 PM Florian Fainelli <f.fainelli@xxxxxxxxx> wrote:
> >>
> >> On 12/15/2020 5:19 AM, Bharat Gooty wrote:
> >>> Since the IOMMU is disabled and DMA engine is on 32-bit bus, We can not
> >>> give the complete DDR for the USB DMA.
> >>> So restricting the usable DAM size to 4GB.
> >>
> >> Thanks, can you make this a proper patch submission along with a Fixes:
> >> tag that is:
> >>
> >> Fixes: 2013a4b684b6 ("arm64: dts: broadcom: clear the warnings caused by
> >> empty dma-ranges")
> >
> > Yes, that would be helpful, though I would appreciate a better description
> > that explains what is actually going on: is it the device or the bus that
> > has the 32-bit limitation, and if it is indeed a bug in the device, why do
> > you pretend that this is a 64-bit device on a 32-bit bus instead (this
> > could also use a comment in the dts file)?
> >
> > Arnd
> >
>
> Sorry for the delay in reply. Bharat finally got time to do some
> investigation to confirm the following:
>
> These USB controllers indeed can address 64-bit. However, on the bus
> internally, only 40-bits are connected to the interconnect of CCN. As a
> result, the 'dma-ranges' should be modified to address 40-bit in size.
a 40-bit range is effectively the same as a 64-bit range though, so I
think you'll still need a driver quirk if the device cannot use the whole 40
bit addressing but is limited to 32 bits because of a bug in its dma engine.
> We also have a somewhat related question, is it true that since v5.10,
> defining of 'dma-ranges' on the bus node where its child device node has
> implication of IOMMU usage (through 'iommus' or 'iommu-map') is now
> mandatory? My understanding is that the 'dma-ranges' in this scheme will
> define the IOVA address to system address mapping required by all
> devices on that bus. Please help to confirm if my understanding is correct
I have not actually seen that change. In general, you should always
have a dma-ranges property in the parent of a DMA master, I think
we just never enforced that and fall back to a 32-bit mask by default.
Arnd
