Re: [PATCH 3/3] bus: hisi_lpc: Expand build test coverage

From: John Garry
Date: Thu Oct 03 2019 - 05:27:09 EST


On 02/10/2019 19:22, Arnd Bergmann wrote:
On Wed, Oct 2, 2019 at 6:25 PM John Garry <john.garry@xxxxxxxxxx> wrote:
On 02/10/2019 16:43, Arnd Bergmann wrote:
On Tue, Oct 1, 2019 at 6:07 PM John Garry <john.garry@xxxxxxxxxx> wrote:

Currently the driver will only ever be built for ARM64 because it selects
CONFIG_INDIRECT_PIO, which itself depends on ARM64.

Expand build test coverage for the driver to other architectures by only
selecting CONFIG_INDIRECT_PIO for ARM64, when we really want it.

Signed-off-by: John Garry <john.garry@xxxxxxxxxx>


Hi Arnd,

Good idea, but doesn't this cause a link failure against
logic_pio_register_range() when INDIRECT_PIO is disabled?

No, it shouldn't do. Function
lib/logic_pio.c::logic_pio_register_range() is built always, outside any
INDIRECT_PIO checking.

Ok, I see.

A some stage, for completeness we should probably change
logic_pio_register_range() and friends to be under PCI_IOBASE. But then
we would need stubs for !PCI_IOBASE, due to this above change and also
references from the device tree code. I'd have to consider this a bit
more. Let me know what you think.

It's probably not to do this with the usual 'static inline' stubs in the
header files. There is no rush here, but it would be nice to not build
this code when it is not needed.

Agreed, I'll add it to my todo list.

So we want to build logic_pio.c for archs which define PCI_IOBASE, but there is no specific config option for that.


I wonder if this one-line change would take care of the !CONFIG_OF
case already (it probably doesn't):

--- a/lib/Makefile
+++ b/lib/Makefile
@@ -107,7 +107,7 @@ obj-$(CONFIG_HAS_IOMEM) += iomap_copy.o devres.o
obj-$(CONFIG_CHECK_SIGNATURE) += check_signature.o
obj-$(CONFIG_DEBUG_LOCKING_API_SELFTESTS) += locking-selftest.o

-obj-y += logic_pio.o
+lib-y += logic_pio.o

So that means that this obj will not be included in the vmlinux file for !CONFIG_OF (for archs with !PCI_IOBASE or !PCI).

It is an improvement, but I still would like to make the complete change and not to build at all for archs which don't define PCI_IOBASE.


obj-$(CONFIG_GENERIC_HWEIGHT) += hweight.o

On a related note: At some point we may want to add an indirect
method for readl()/writel(), to deal with some of the weirder 32-bit
ARM platforms. We'll have to see how well this can fit into the
infrastructure we already have for indirect PIO.

It should be possible. We would just need to manage 2 separate logical spaces instead of one: a. mmio b. pio

I'm guessing that all the code for logical <-> hw bus address translations would be the same.

Thanks,
John


Arnd

.