Re: [PATCH v5 04/12] PCI: brcmstb: add dma-range mapping for inbound traffic

From: Robin Murphy
Date: Wed Sep 26 2018 - 06:57:03 EST

On 26/09/18 03:59, Florian Fainelli wrote:


On 9/24/2018 8:01 AM, Jim Quinlan wrote:
On Mon, Sep 24, 2018 at 4:25 AM Ard Biesheuvel
<ard.biesheuvel@xxxxxxxxxx> wrote:

On Fri, 21 Sep 2018 at 19:41, Jim Quinlan <jim2101024@xxxxxxxxx> wrote:

On Thu, Sep 20, 2018 at 5:39 PM Florian Fainelli <f.fainelli@xxxxxxxxx> wrote:

On 09/20/2018 02:33 PM, Ard Biesheuvel wrote:
On 20 September 2018 at 14:31, Florian Fainelli <f.fainelli@xxxxxxxxx> wrote:
On 09/20/2018 02:04 PM, Ard Biesheuvel wrote:
On 20 September 2018 at 13:55, Florian Fainelli <f.fainelli@xxxxxxxxx> wrote:
On 09/19/2018 07:19 PM, Ard Biesheuvel wrote:
On 19 September 2018 at 07:31, Jim Quinlan <jim2101024@xxxxxxxxx> wrote:
The Broadcom STB PCIe host controller is intimately related to the
memory subsystem. This close relationship adds complexity to how cpu
system memory is mapped to PCIe memory. Ideally, this mapping is an
identity mapping, or an identity mapping off by a constant. Not so in
this case.

Consider the Broadcom reference board BCM97445LCC_4X8 which has 6 GB
of system memory. Here is how the PCIe controller maps the
system memory to PCIe memory:

ÂÂ memc0-a@[ÂÂÂÂÂÂÂ 0....3fffffff] <=> pci@[ 0....3fffffff]
ÂÂ memc0-b@[100000000...13fffffff] <=> pci@[ 40000000....7fffffff]
ÂÂ memc1-a@[ 40000000....7fffffff] <=> pci@[ 80000000....bfffffff]
ÂÂ memc1-b@[300000000...33fffffff] <=> pci@[ c0000000....ffffffff]
ÂÂ memc2-a@[ 80000000....bfffffff] <=> pci@[100000000...13fffffff]
ÂÂ memc2-b@[c00000000...c3fffffff] <=> pci@[140000000...17fffffff]


So is describing this as

dma-ranges = <0x0 0x0 0x0 0x0 0x0 0x40000000>,
ÂÂÂÂÂÂÂÂÂÂÂÂÂ <0x0 0x40000000 0x1 0x0 0x0 0x40000000>,
ÂÂÂÂÂÂÂÂÂÂÂÂÂ <0x0 0x80000000 0x0 0x40000000 0x0 0x40000000>,
ÂÂÂÂÂÂÂÂÂÂÂÂÂ <0x0 0xc0000000 0x3 0x0 0x0 0x40000000>,
ÂÂÂÂÂÂÂÂÂÂÂÂÂ <0x1 0x0 0x0 0x80000000 0x0 0x40000000>,
ÂÂÂÂÂÂÂÂÂÂÂÂÂ <0x1 0x40000000 0x0 0xc0000000 0x0 0x40000000>;

not working for you? I haven't tried this myself, but since DT permits
describing the inbound mappings this way, we should fix the code if it
doesn't work at the moment.

You mean encoding the memory controller index in the first cell? If that
works, that's indeed a much cleaner solution, though is it standard
compliant in any form?

No those are just memory addresses (although I may have screwed up the
order). From Documentation/devicetree/booting-without-of.txt:

"""
Optional property:
- dma-ranges: <prop-encoded-array> encoded as arbitrary number of triplets of
ÂÂÂÂÂÂÂÂ (child-bus-address, parent-bus-address, length). Each triplet specified
ÂÂÂÂÂÂÂÂ describes a contiguous DMA address range.
"""


Then I am confused by your comment, that's what this patch does, it adds
support for reading "dma-ranges" from Device Tree and setting up inbound
windows using that. The only caveat is that because the PCIe root
complex has some ties with the memory bus architecture it is connected
to (SCB in our case) there is still a requirement to know the
translation between a given physical address and its backing memory
controller/aperture.


Ah ok, apologies for the noise then.

I was hoping that having working support for dma-ranges would remove
the need for the special phys<->dma conversion routines.

What you describe definitively works with platform devices, but I am not
sure this is working for PCIe devices, although, conceptually it should,
yes.
Sorry for my delay in responding. One problem is that
of_dma_configure() only looks at the first dma-range given and then
converts it to dev->dma_pfn_offset which is respected by the DMA API.
However, we often have multiple dma-ranges, not just one. This is the
big issue.


Given the recent attention to getting these APIs in shape, this may be
something Robin or Christoph may care to look into?

It looks like this has been brought up before in the "[RFC PATCH] of:
Fix DMA configuration for non-DT masters" thread aka

https://lists.linuxfoundation.org/pipermail/iommu/2017-April/021325.html

In the thread "Oza Oza", a Broadcom coworker probably dealing with the
same exact problem as I, enumerates three problems. #1 and #2 are
the exact same ones I've just given: the "dma-ranges" prop of the RC
DT node is "skipped", and of_dma_get_range() only considers the first
entry in any "dma-ranges".

Robin, is that something that is expected or should the "dma-ranges" somehow propagate from host bridge down the PCIe end-point drivers?

Nope, the code is most definitely incomplete - it's sufficient to support the system it was originally needed for (i.e. platform devices with a single range), but can by no means even pretend to support the binding fully. Furthermore, the way that PCI support was later grafted into of_dma_configure() was *only* in support of dma-coherent without consideration for dma-ranges. Hence the current mess.


Thanks, Jim


In any case, the description of dma-ranges should be in sync with the
way Linux interprets it, so this is either a documentation bug or a
DMA layer bug.

There is another issue with of_dma_configure() being invoked by the EP
driver on "bridge->parent->of_node", which is our RC device,
Of_dma_configure() calls of_dma_range() on the of_get_next_parent() of
our RC's device node and this misses the dma-ranges property which is
contained within the RC. I think I could workaround this but there is
no getting around the first problem.


IIUC dma-ranges should be added to the parent bus of a device, which I
guess is slightly ambiguous for a root complex that incorporates a
host bridge.

Humm, why is that ambiguous for a host bridge/root complex?

The real problem is that FDT machines don't describe the PCI hierarchy in DT as proper OF does, so we have this awkward crossing between the DT model and the Linux device model where the devices have no DT representation and the "parent bus" is a DT leaf node, which cocks up the way the current code is expecting to work.

Robin.