On Tue, Jan 07, 2020 at 12:33:54PM +0100, Uwe Kleine-König wrote:
> On Tue, Jan 07, 2020 at 12:03:59PM +0100, Krzysztof Kozlowski wrote:
> > On Tue, Jan 07, 2020 at 11:42:34AM +0100, Uwe Kleine-König wrote:
> > > > I guess other solution would be to add stubs for few clk functions...
> > > >
> > > > > Also HAS_IOMEM is a typical requirement, but I tested with an ARCH=um
> > > > > config (which does't have HAS_IOMEM) and they all compile fine.
> > > >
> > > > Because of !HAS_IOMEM, since some time ARCH=um does not support
> > > > COMPILE_TEST. Therefore HAS_IOMEM dependency is not needed for compile
> > > > testing (and for regular build it is selected by ARCH).
> > >
> > > Hehe, I didn't notice because for testing I just dropped the "depends on
> > > ..." lines in Kconfig instead of adding "|| COMPILE_TEST" :-) Still they
> > > compile fine on UML.
> > >
> > > Ah, since bc083a64b6c0 ("init/Kconfig: make COMPILE_TEST depend on
> > > !UML") == v4.8-rc1~52^2~83 COMPILE_TEST cannot be enabled on UML, but
> > > later 1bcbfbfdeb00 ("um: add dummy ioremap and iounmap functions")
> > > == v4.13-rc1~8^2~6 UM got a dummy implementation. So maybe we could
> > > revert bc083a64b6c0 today? (And if not, a comment about why near the
> > > "depends on !UML" in init/Kconfig would be great.)
> > >
> > > Orthogonal to that, I wonder if depending on HAS_IOMEM is right even
> > > though the compile testers won't notice it missing. Or should HAS_IOMEM
> > > be dropped?
> >
> > I think yes, it can be dropped, but this would require:
> > 1. Fixing any dependencies on HAS_IOMEM, e.g.:
> > WARNING: unmet direct dependencies detected for MFD_SYSCON
> > Depends on [n]: HAS_IOMEM [=n]
> > Selected by [y]:
> > - PHY_DA8XX_USB [=y] && (ARCH_DAVINCI_DA8XX || COMPILE_TEST [=y])
>
> I don't understand that warning. What did you modify to trigger that?
> Probably related to the big "if HAS_IOMEM" in drivers/mfd/Kconfig?!
OK, that's actually from my other patch to illustrate the problem:
https://lore.kernel.org/linux-arm-kernel/20200103164710.4829-2-krzk@xxxxxxxxxx/
After reverting of bc083a64b6c0, every driver that selects MFD_SYSCON
(or some other parts) has to depend on HAS_IOMEM.
>
> > 2. Checking if all of stubs are implemented (not only IOMEM but maybe
> > DMA as well). Also 1bcbfbfdeb00 brought only few stubs. Still we
> > need devm_ioremap_resource() and others.
>
> A problem is that it's unclear (to me at least) what the presence of
> HAS_IOMEM actually implies. I thought it's about ioremap + readl +
> writel (including their respective variants). Does it really include dma
> stuff, too?
>
> > Quick test shows mentioned "unmet direct dependencies" and:
> > phy-pxa-usb.c:(.text+0x2f5): undefined reference to `devm_ioremap_resource'
> > dma-iommu.c:(.text+0x179): undefined reference to `dma_pgprot'
>
> dma_pgprot seems to depend on HAS_DMA though, not HAS_IOMEM.
>
> (Oh, HAS_DMA is defined using "depends on !NO_DMA" + "default y".
> NO_DMA has three different definitions. Two of them (in
> drivers/crypto/ccree/cc_hw_queue_defs.h and arch/arm/include/asm/dma.h)
> unrelated.)
Yes, HAS_DMA is the second missing piece for UM.
Best regards,
Krzysztof
