On 31/01/18 09:37, Arnd Bergmann wrote:
On Wed, Jan 31, 2018 at 8:29 AM, Maxime Ripard
<maxime.ripard@xxxxxxxxxxxxxxxxxx> wrote:
Hi Thierry,
On Tue, Jan 30, 2018 at 11:01:50AM +0100, Thierry Reding wrote:
On Tue, Jan 30, 2018 at 10:59:16AM +0100, Thierry Reding wrote:
On Tue, Jan 30, 2018 at 10:24:48AM +0100, Arnd Bergmann wrote:
On Tue, Jan 30, 2018 at 8:54 AM, Maxime Ripard
<maxime.ripard@xxxxxxxxxxxxxxxxxx> wrote:
On Mon, Jan 29, 2018 at 03:34:02PM +0100, Arnd Bergmann wrote:
On Mon, Jan 29, 2018 at 10:25 AM, Linus Walleij
<linus.walleij@xxxxxxxxxx> wrote:
On Mon, Jan 29, 2018 at 9:25 AM, Maxime Ripard
<maxime.ripard@xxxxxxxxxxxxxxxxxx> wrote:
On Sat, Jan 27, 2018 at 05:14:26PM +0100, Linus Walleij wrote:
At one point we had discussed adding a 'dma-masters' property that
lists all the buses on which a device can be a dma master, and
the respective properties of those masters (iommu, coherency,
offset, ...).
IIRC at the time we decided that we could live without that complexity,
but perhaps we cannot.
Are you talking about this ?
https://elixir.free-electrons.com/linux/latest/source/Documentation/devicetree/bindings/dma/dma.txt#L41
It doesn't seem to be related to that issue to me. And in our
particular cases, all the devices are DMA masters, the RAM is just
mapped to another address.
No, that's not the one I was thinking of. The idea at the time was much
more generic, and not limited to dma engines. I don't recall the details,
but I think that Thierry was either involved or made the proposal at the
time.
Yeah, I vaguely remember discussing something like this before. A quick
search through my inbox yielded these two threads, mostly related to
IOMMU but I think there were some mentions about dma-ranges and so on as
well. I'll have to dig deeper into those threads to refresh my memories,
but I won't get around to it until later today.
If someone wants to read up on this in the meantime, here are the links:
https://lkml.org/lkml/2014/4/27/346
http://lists.infradead.org/pipermail/linux-arm-kernel/2014-May/257200.html
From a quick glance the issue of dma-ranges was something that we hand-
waved at the time.
Also found this, which seems to be relevant as well:
http://lists.infradead.org/pipermail/linux-arm-kernel/2014-May/252715.html
Adding Dave.
Thanks for the pointers, I started to read through it.
I guess we have to come up with two solutions here: a short term one
to address the users we already have in the kernel properly, and a
long term one where we could use that discussion as a starting point.
For the short term one, could we just set the device dma_pfn_offset to
PHYS_OFFSET at probe time, and use our dma_addr_t directly later on,
or would this also cause some issues?
That would certainly be an improvement over the current version,
it keeps the hack more localized. That's fine with me. Note that both
PHYS_OFFSET and dma_pfn_offset have architecture specific
meanings and they could in theory change, so ideally we'd do that
fixup somewhere in arch/arm/mach-sunxi/ at boot time before the
driver gets probed, but this wouldn't work on arm64 if we need it
there too.
dma_pfn_offset certainly *shouldn't* be architecture-specific; it's just
historically not been used consistently everywhere. That should now be
addressed by Christoph's big dma_direct_ops cleanup (now merged) which
fills in the places it was being missed in generic code. From quickly
skimming this thread, it sounds like that is really should be sufficient
for this particular hardware - if all its DMA goes through an
interconnect which makes RAM appear at 0, then you give it "dma-ranges =
<0 PHYS_OFFSET size>;" and things should 'just work' provided the DMA
API is otherwise being used correctly.
If different devices have differing DMA master address maps (such that
you can't just place a single dma-ranges on your "soc" node) then the
trick is to wrap them in intermediate "simple-bus" nodes with a
straight-though "ranges;" and the appropriate "dma-ranges = ...". Yes,
it's a bit of a bodge, but then pretending AXI/APB interconnects are
buses in the sense that DT assumes is already a bit of a bodge.
For the long term plan, from what I read from the discussion, it's
mostly centered around IOMMU indeed, and we don't have that. What we
would actually need is to break the assumption that the DMA "parent"
bus is the DT node's parent bus.
And I guess this could be done quite easily by adding a dma-parent
property with a phandle to the bus controller, that would have a
dma-ranges property of its own with the proper mapping described
there. It should be simple enough to support, but is there anything
that could prevent something like that to work properly (such as
preventing further IOMMU-related developments that were described in
those mail threads).
I've thought about it a little bit now. A dma-parent property would nicely
solve two problems:
- a device on a memory mapped control bus that is a bus master on
a different bus. This is the case we are talking about here AFAICT
- a device that is on a different kind of bus (i2c, spi, usb, ...) but also
happens to be a dma master on another bus. I suspect we have
some of these today and they work by accident because we set the
dma_mask and dma_map_ops quite liberally in the DT probe code,
but it really shouldn't work according to our bindings. We may also
have drivers that work around the issue by forcing the correct dma
mask and map_ops today, which makes them work but is rather
fragile.
Note that I recently started tightening this area up (see d89e2378a97f),
precisely because doing DMA configuration for things that shouldn't be
DMA masters turned out to be messing with legitimate devices in
conjunction with other unpleasant and fiddly circumstances elsewhere.
I can think of a couple of other problems that may or may not be
relevant in the future that would require a more complex solution:
- a device that is a bus master on more than one bus, e.g. a
DMA engine that can copy between the CPU address space and
another memory controller that is not visible to the CPU
Those definitely exist - e.g. as best I understand it the crypto SRAMs
on Marvell Armada XP parts which we describe as "mmio-sram" are not
actually CPU-accessible, and only the CESA DMA engine can move data
between them and main memory.
- a device that is connected to main memory both through an IOMMU
and directly through its parent bus, and the device itself is in
control over which of the two it uses (usually the IOMMU would
contol whether a device is bypassing translation)
*shudder*
- a device that has a single DMA address space with some form
of non-linear mapping to one or more parent buses. Some of these
can be expressed using the parent's dma-ranges properties, but
our code currently only looks at the first entry in dma-ranges.
That's definitely a thing, too. In fact the TI Keystone platform which
begat our current of_dma_get_range() implementation is like that; we
just only describe the offset upper half of the DMA address map where
dma-coherent RAM lives, not the bottom 2GB of peripherals which is the
same as the CPU view. See also the madness of the brcmstb PCIe patches
elsewhere on this list.
Another problem is the interaction with the dma-ranges and iommu
properties. I have not found any real problems here, but we certainly
need to be careful to define what happens in all combinations and
make sure that we document it well in the bindings and have those
reviewed by the affected parties, at least the ARM and PowerPC
architecture folks as well as the Nvidia and Renesas platform
maintainers, which in my experience have the most complex DMA
hardware.
I hold out hope that we never see DMA offsets mixed with IOMMUs (because
thankfully it doesn't make much sense from a hardware perspective), but
we certainly do want the size aspect of dma-ranges to describe cases
where there are fewer than expected address bits wired between the
device master interface and IOMMU slave interface. There's a fairly
long-standing issue with that being a bit broken which is hurting at
least one platform I know about.
Robin.
