Re: [PATCH] of: change fixup of dma-ranges size to error

From: Rob Herring
Date: Thu Apr 06 2017 - 18:42:43 EST


On Thu, Apr 6, 2017 at 1:37 PM, Frank Rowand <frowand.list@xxxxxxxxx> wrote:
> On 04/06/17 07:03, Rob Herring wrote:
>> On Thu, Apr 6, 2017 at 1:18 AM, <frowand.list@xxxxxxxxx> wrote:
>>> From: Frank Rowand <frank.rowand@xxxxxxxx>
>>>
>>> of_dma_get_range() has workaround code to fixup a device tree that
>>> incorrectly specified a mask instead of a size for property
>>> dma-ranges. That device tree was fixed a year ago in v4.6, so
>>> the workaround is no longer needed. Leave a data validation
>>> check in place, but no longer do the fixup. Move the check
>>> one level deeper in the call stack so that other possible users
>>> of dma-ranges will also be protected.
>>>
>>> The fix to the device tree was in
>>> commit c91cb9123cdd ("dtb: amd: Fix DMA ranges in device tree").
>>
>> NACK.
>> This was by design. You can't represent a size of 2^64 or 2^32.
>
> I agree that being unable to represent a size of 2^32 in a u32 and
> a size of 2^64 in a u64 is the underlying issue.
>
> But the code to convert a mask to a size is _not_ design, it is a
> hack that temporarily worked around a device tree that did not follow
> the dma-ranges binding in the ePAPR.

Since when is (2^64 - 1) not a size. It's a perfectly valid size in
DT. And there's probably not a system in the world that needs access
to that last byte. Is it completely accurate description if we
subtract off 1? No, but it is still a valid range (so would be
subtracting 12345).

> That device tree was corrected a year ago to provide a size instead of
> a mask.

You are letting Linux implementation details influence your DT
thinking. DT is much more flexible in that it supports a base address
and size (and multiple of them) while Linux can only deal with a
single address mask. If Linux dealt with base + size, then we wouldn't
be having this conversation. As long as Linux only deals with masks,
we're going to have to have some sort of work-around to deal with
them.

>> Well, technically you can for the latter, but then you have to grow
>> #size-cells to 2 for an otherwise all 32-bit system which seems kind
>> of pointless and wasteful. You could further restrict this to only
>> allow ~0 and not just any case with bit 0 set.
>>
>> I'm pretty sure AMD is not the only system. There were 32-bit systems too.
>
> I examined all instances of property dma-ranges in in tree dts files in
> Linux 4.11-rc1. There are none that incorrectly specify mask instead of
> size.

Okay, but there are ones for ranges at least. See ecx-2000.dts.

> #size-cells only changes to 2 for the dma-ranges property and the ranges
> property when size is 2^32, so that is a very small amount of space.
>
> The patch does not allow for a size of 2^64. If a system requires a
> size of 2^64 then the type of size needs to increase to be larger
> than a u64. If you would like for the code to be defensive and
> detect a device tree providing a size of 2^64 then I can add a
> check to of_dma_get_range() to return -EINVAL if #size-cells > 2.
> When that error triggers, the type of size can be changed.

#size-cells > 2 is completely broken for anything but PCI. I doubt it
is easily fixed without some special casing (i.e. a different hack)
until we have 128-bit support. I hope to retire before we need to
support that.

Rob