[RFC] implement arm64_adjust_dma_zone as arm32 does?

[arnd@arndb.de (Arnd Bergmann)]

On Wednesday 03 December 2014 12:17:50 Catalin Marinas wrote:
> On Wed, Dec 03, 2014 at 11:02:20AM +0000, Jisheng Zhang wrote:
> > Dear Catalin,
> > 
> > On Wed, 3 Dec 2014 02:24:32 -0800
> > Catalin Marinas <catalin.marinas@arm.com> wrote:
> > 
> > > On Wed, Dec 03, 2014 at 09:58:55AM +0000, Jisheng Zhang wrote:
> > > > If one platform has one limitation only some banks are DMA-able, for
> > > > example 0-2GB. Under arm32, we can set the dma_zone_size, then
> > > > arm_adjust_dma_zone() will set the correct dma zone for us. But under
> > > > arm64, we can't do it.
> > > 
> > > Do you really have plans for such platform (and it won't have an IOMMU)
> > > or it's just theoretical? Currently ZONE_DMA is limited to the bottom
> > > 32-bit of RAM. The arm32 dma_zone_size relies on the machine descriptors
> > > which we don't have on arm64.
> > 
> > Thanks for your detailed clarification. Yes, there's plan for such platform and
> > there's no IOMMU.
> [...]
> > > What we could do is re-introduce ZONE_DMA32 for 32-bit dma masks and a
> > > ZONE_DMA for smaller masks. The problem is describing such ZONE_DMA mask
> > > via DT but very early before we set up the zones. A quick hack would be
> > > to limit ZONE_DMA to the minimum of 32-bit limit or the end of the first
> > > memblock, so you describe it as such in DT (but then we penalise
> > > platforms that can do DMA on all memblocks).
> > 
> > We can keep ZONE_DMA32 out from arm64, just describe the ZONE_DMA mask via.
> > DT as you pointed out, then setup the matching dma zone.
> > 
> > I'm wondering whether setting up ZONE_DMA mask via. DT is acceptable or not?
> 
> That would be better than machine descriptors, but I can't think of a
> nice way to describe this (dma-ranges in the top DT node or a new
> property for the memory node?).

Putting DMA ranges in the root DT node has the special meaning of defining
that *no* devices can do DMA to addresses outside of the top-level range,
which I assume is not what Jisheng wants (please clarify).

> We could try to get the minimum of all dma-ranges specified in the DT
> for those buses which are not hooked to an IOMMU but setting up ZONE_DMA
> needs to be done very early, before even unflattening the DT.
> 
> Cc'ing some arm-soc and DT people.

Another alternative would be to hardcode the size of the DMA zone at
compile time to some value that will work for everyone. It should still
work ok even if it's really small (e.g. 16MB), but that space may
get exhausted more easily. If we find a device that has its dma-ranges
smaller than ZONE_DMA, or ZONE_DMA is disabled, we would not allow that
device to perform any DMA, and that may still be ok in a lot of cases,
e.g. an SPI controller or a UART that can operate in polled mode as
a fallback.

Before we put too much infrastructure into the kernel to solve the
problem for all corner cases, let's see the extend of the brokenness
of this platform and assume that all future platforms are less broken
for the moment. If someone else comes up with a worse bug, it's up to
them to fix it.

Jisheng, can you elaborate on the specific errata?
Is this just for early pre-production silicon or do you expect this 
to ship in significant numbers to end-users?
Would anybody see this hardware on a machine that has upgradeable kernels?
Are there other reasons that already prevent you from running an unpatched
kernel indefinitely?
What DMA masters are affected by this, does it include any high-speed
devices, or could you fall back to polling MMIO for mainline kernels?

	Arnd