Re: [PATCH 1/2] kernel/resource: Add new flag IORESOURCE_SHARED

From: Ricardo Ribalda Delgado
Date: Tue Jun 09 2015 - 08:34:32 EST


Hello Grant



On Tue, Jun 9, 2015 at 1:13 PM, Grant Likely <grant.likely@xxxxxxxxxx> wrote:
>
> On Mon, 8 Jun 2015 22:02:06 +0200
> , Ricardo Ribalda Delgado <ricardo.ribalda@xxxxxxxxx>
> wrote:
> > Hello Grant
> >
> > On Mon, Jun 8, 2015 at 8:23 PM, Grant Likely <grant.likely@xxxxxxxxxx> wrote:
> > > On Fri, 5 Jun 2015 12:51:17 +0200
> > > , Ricardo Ribalda Delgado <ricardo.ribalda@xxxxxxxxx>
> > > wrote:
> > >> Some device tree platforms have not defined correctly their memory
> > >> resources (i.e. Overlapping or duplication of resources).
> > >> To avoid this issue we have historically avoided to add their resources to
> > >> the resource tree. This leads to code duplication and oops when trying to
> > >> unload dynamically a device tree (feature introduced recently).
> > >>
> > >> This new flag tells the resource system that a resource can be shared by
> > >> multiple owners, so we can support device trees with problems at the
> > >> same time that we do not duplicate code or crash when unloading the
> > >> device tree.
> > >>
> > >> Signed-off-by: Ricardo Ribalda Delgado <ricardo.ribalda@xxxxxxxxx>
> > >> ---
> > >
> > > I'm really not comfortable with this change. The resource tree code is
> > > complicated enough as is. Adding this exception case quite probably adds
> > > corner cases that aren't property dealt with. If two regions overlay,
> > > and then request_region is called? Which region does it become a child
> > > of? And that's just off the top of my head. I don't want to hack in
> > > changes to the resource code for what is a corner case.
> >
> > I see your concern, perhaps you could provide a testcase and we can
> > find out if it fails or not. So far I have tested a device tree with
> > two devices on the same memory region, each device managed by a
> > driver.
>
> Actually, you need to provide the test case. You need to show that
> you've thought through all the implications and corner cases on the
> resource code. This is a non-trivial change to the how the resource code
> works, and you need to demonstrate that your really understand the
> implications of what you are doing.



On non broken platforms:

it will work exactly as it works today.

On broken platforms:

I have tried with duplicated devices: Both requesting the region via
devm_ioremap_resource

a0000000-dfffffff : PCI Bus 0000:00
b0000000-cfffffff : PCI Bus 0000:01
b0030000-b003ffff : /axi1/pcie_bridge_1
b0030000-b003ffff : /axi1/pcie_bridge
b0030000-b003ffff : /axi1/pcie_bridge
b0030000-b003ffff : /axi1/pcie_bridge_1


pcie_bridge_0: pcie_bridge {
#address-cells = <1>;
#size-cells = <1>;
compatible = "xlnx,axi-pcie-1.00.a";
reg = < 0x30030000 0x10000 >;
interrupt-parent = <&xps_intc_0>;
interrupts = < 0 2 >;
qtec,apertures =<
0x80000000 0x20000000
0xa0000000 0x20000000
0xc0000000 0x20000000
0xe0000000 0x20000000
>;
};

pcie_bridge_1: pcie_bridge_1 {
#address-cells = <1>;
#size-cells = <1>;
compatible = "xlnx,axi-pcie-1.00.a";
reg = < 0x30030000 0x10000 >;
interrupt-parent = <&xps_intc_0>;
interrupts = < 31 2 >;
qtec,apertures =<
0x80000000 0x20000000
0xa0000000 0x20000000
0xc0000000 0x20000000
0xe0000000 0x20000000
>;
};

And for two devices requesting the same region via devm_ioremap_resource()



a0000000-dfffffff : PCI Bus 0000:00
b0000000-cfffffff : PCI Bus 0000:01
b0030000-b003ffff : /axi1/pcie_bridge
b0030000-b003ffff : /axi1/pcie_bridge
b0030000-b003ffff : /axi1/pcie_bridge

pcie_bridge_0: pcie_bridge {
#address-cells = <1>;
#size-cells = <1>;
compatible = "xlnx,axi-pcie-1.00.a";
reg = < 0x30030000 0x10000 >;
interrupt-parent = <&xps_intc_0>;
interrupts = < 0 2 >;
qtec,apertures =<
0x80000000 0x20000000
0xa0000000 0x20000000
0xc0000000 0x20000000
0xe0000000 0x20000000
>;
};

packer_0: packer_0{
#address-cells = <1>;
#size-cells = <1>;
compatible = "qtec,axi_matrix_packer-1.00.a";
reg = < 0x30060000 0x10000 >;
qtec,pcie_bridge= <&pcie_bridge_0>;
};


If you can think of any other corner case, please let me know, and I
will try it, now I have a setup for this.


>
>
> Start with the example I pointed out. When a driver does a
> request_mem_region(), which resource does it end up being a parent of if
> the regions overlap? Can you write a unittest that demonstrates the code
> has the correct behaviour? Will a driver end up getting the wrong
> device's resource structure as the parent? (hint: yes it will)
>
> > I can load and unload the device tree perfectly.
>
> Merely making it work for your use-case isn't the issue. It's whether or
> not making this change will break the core behavour of the resource
> code.
>
> g.




--
Ricardo Ribalda
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/