Re: [PATCH] s390/io: Fix ioremap and iounmap undefinded issue for s390

[Niklas Schnelle <schnelle@linux.ibm.com>]

On Fri, 2021-09-03 at 17:40 +0800, Chunyan Zhang wrote:
> On Fri, 3 Sept 2021 at 16:24, Niklas Schnelle <schnelle@linux.ibm.com> wrote:
> > On Fri, 2021-09-03 at 16:03 +0800, Chunyan Zhang wrote:
> > > From: Chunyan Zhang <chunyan.zhang@unisoc.com>
> > > 
> > > There would not be ioremap and iounmap implementations if CONFIG_PCI is
> > > not set for s390, so add default declarations of these two functions
> > > for the case to avoid 'undefined reference' issue.
> > > 
> > > Fixes: 71ba41c9b1d9 ("s390/pci: provide support for MIO instructions")
> > > Reported-by: kernel test robot <lkp@intel.com>
> > > Signed-off-by: Chunyan Zhang <chunyan.zhang@unisoc.com>
> > > ---
> > > The issue was reported from https://lkml.org/lkml/2021/8/1/18
> > 
> > Thanks for the patch but I'm a little skeptical about adding
> > ioremap()/iounmap() stubs that don't do anything useful and on top ofn
> > that would do so silently.
> > 
> > In the above discussion you said that TIMER_OF should depend on
> > HAS_IOMEM. In arch/s390/Kconfig HAS_IOMEM is set if and only if
> > CONFIG_PCI is set so that sounds to me like it would prevent the
> > undefined reference without the risk of someone trying to use io*map()
> > without CONFIG_PCI.
> 
> Humm... you can ignore my reply on that time, I later found that's not
> correct :)
> 
> TIMER_OF would be selected by other configs and it seems not able to
> depends on HAS_IOMEM, unless all configs which select TIMER_OF depend
> on HAS_IOMEM, that would be a big change.

Hmm I already see a lot of the options in drivers/clocksource/Kconfig
depend on HAS_IOMEM and drivers/clocksource/timer-of.c definitely calls
iounmap() so wouldn't that be the logical change even if it is bigger? 

> 
> Actually HAS_IOMEM is set as default on other architectures, but not
> for s390 which redefined it.

Yes because most architectures always have IOMEM and io*map() functions
I believe. s390 is an exception here as the mainframe native
functionality all works without MMIO and you can run a fully functional
system including networking and block devices without any MMIO, PCI and
without ioremap()/iounmap().

> 
> > At the very least I think the functions should do a WARN_ONCE() but
> > then we have the same situation as discussed below with Linus making it
> > pretty clear that he prefers these cases to be compile time checked:
> 
> Ok, if I understand correctly, if io*map is not implemented for some
> case, there should be a *compile-time* error rather than adding a stub
> function to make this kind of errors disappeared.
> 
> Please correct me if I missed something.

Ideally not a compile time error but a compile time flag such as a
Kconfig option that would make sure that if HAS_IOMEM isn't set we
don't get drivers compiled which depend on working io*map(). After all
these drivers will surely not be functional.

> 
> Thanks,
> Chunyan
> 
> > https://lkml.org/lkml/2021/7/2/511
> > 
> > 
> > > ---
> > >  arch/s390/include/asm/io.h | 15 ++++++++++++---
> > >  1 file changed, 12 insertions(+), 3 deletions(-)
> > > 
> > > diff --git a/arch/s390/include/asm/io.h b/arch/s390/include/asm/io.h
> > > index e3882b012bfa..9438370c6445 100644
> > > --- a/arch/s390/include/asm/io.h
> > > +++ b/arch/s390/include/asm/io.h
> > > @@ -23,11 +23,8 @@ void unxlate_dev_mem_ptr(phys_addr_t phys, void *addr);
> > >  #define IO_SPACE_LIMIT 0
> > > 
> > > 
> > .. snip ..
> >