On 18/10/2019 7:44 pm, Marek Vasut wrote:
[...]
Say you have a single hardware window, and this DT property (1-cell
numbers for simplicity:
dma-ranges = <0x00000000 0x00000000 0x80000000>;
Driver reads one entry and programs the window to 2GB@0, DMA setup
parses the first entry and sets device masks to 0x7fffffff, and
everything's fine.
Now say we describe the exact same address range this way instead:
dma-ranges = <0x00000000 0x00000000 0x40000000,
0x40000000 0x40000000 0x40000000>;
Driver reads one entry and programs the window to 1GB@0, DMA setup
parses the first entry and sets device masks to 0x3fffffff, and *today*,
things are suboptimal but happen to work.
Now say we finally get round to fixing the of_dma code to properly
generate DMA masks that actually include all usable address bits, a user
upgrades their kernel package, and reboots with that same DT...
Driver reads one entry and programs the window to 1GB@0, DMA setup
parses all entries and sets device masks to 0x7fffffff, devices start
randomly failing or throwing DMA errors half the time, angry user looks
at the changelog to find that somebody decided their now-corrupted
filesystem is less important than the fact that hey, at least the
machine didn't refuse to boot because the DT was obviously wrong. Are
you sure that shouldn't be a problem?
I think you picked a rather special case here and arrived as a DMA mask
which just fails in this special case. Such special case doesn't happen
here, and even if it did, I would expect Linux to merge those two ranges
or do something sane ? If the DMA mask is set incorrectly, that's a bug
of the DMA code I would think.
The mask is not set incorrectly - DMA masks represent the number of
address bits the device (or intermediate interconnect in the case of the
bus mask) is capable of driving. Thus when DMA is limited to a specific
address range, the masks should be wide enough to cover the topmost
address of that range (unless the device's own capability is inherently
narrower).
Then the mask should be 0x7fffffff in both cases I'd say.
Yes, *that's my whole point*. It *should* be that for both cases, but
today it works out not to be because the current code is a bit crap.
What DMA mask would you get if those two entries had a gap inbetween
them ? E.g.:
dma-ranges = <0x00000000 0x00000000 0x20000000,
0x40000000 0x40000000 0x20000000>;
The gap itself is immaterial - the highest address described by those
ranges is 0x5fffffff, which requires 31 bits to drive, thus the
appropriate mask would be 0x7fffffff.
OK, here's an real non-simplified example
I would really like an answer to the simple example above before we
start inventing convoluted ones.
Sigh... I did say "real". However convoluted it may look to you, this
window configuration was "invented" ~6 years ago when the Arm Juno SoCs
were first brought up and has been shipping ever since.
(note that these windows are fixed and not programmed by Linux):
dma-ranges = <0x02000000 0x0 0x2c1c0000 0x0 0x2c1c0000 0x0 0x00040000>,
<0x02000000 0x0 0x80000000 0x0 0x80000000 0x0 0x80000000>,
<0x43000000 0x8 0x80000000 0x8 0x80000000 0x2 0x00000000>;
The DMA masks for the devices behind this bridge *should* be 35 bits,
because that's the size of the largest usable address. Currently,
however, because of the of_dma code's deficiency they would end up being
an utterly useless 30 bits, which isn't even enough to reach the start
of RAM. Thus I can't actually have this property in my DT, and as a
result I can't enable the IOMMU, because *that* also needs to know the
ranges in order to reserve the unusable gaps between the windows once
address translation is in play.
How is this related to this particular patch ? This looks more like some
DMA internal topic.
It's trying to get across the point that "dma-ranges" is consumed by
multiple separate parts of the kernel, and implicitly relying on the way
they currently interact is a mistake which will come back to bite you.
Changing one consumer (your driver) to allow devices to exist despite
describing unusable DMA ranges might happen to seem OK today due to
shortcomings in another consumer (the DMA API layer), but once the DMA
API layer gets improved (necessarily, as above), your changes will cause
subtle DMA breakage on R-Car systems with more populated DRAM banks than
available PCI windows, which is arguably even worse than having PCI
simply refuse to probe in that situation.
Robin.
