drivers/mtd/maps/physmap-bt1-rom.c:78:18: sparse: sparse: cast removes address space '__iomem' of expression

[Serge Semin]

On Thu, Nov 12, 2020 at 04:43:01PM +0100, Miquel Raynal wrote:
> Hi Serge,
> 
> Serge Semin <Sergey.Semin at baikalelectronics.ru> wrote on Thu, 12 Nov
> 2020 18:27:39 +0300:
> 
> > Hello Vignesh
> > 
> > On Thu, Nov 12, 2020 at 08:30:42PM +0530, Vignesh Raghavendra wrote:
> > > 
> > > 
> > > On 11/12/20 1:57 PM, Miquel Raynal wrote:  
> > > > Hi Sergey,
> > > > 
> > > > Serge Semin <Sergey.Semin at baikalelectronics.ru> wrote on Wed, 11 Nov
> > > > 2020 22:22:59 +0300:
> > > >   
> > > >> On Tue, Nov 10, 2020 at 04:35:56PM +0100, Miquel Raynal wrote:  
> > > >>> Hi Serge,
> > > >>>
> > > >>> Serge Semin <Sergey.Semin at baikalelectronics.ru> wrote on Tue, 10 Nov
> > > >>> 2020 14:38:27 +0300:
> > > >>>     
> > > >>>> Hello Miquel,
> > > >>>>
> > > >>>> A situation noted by the warning below won't cause any problem because
> > > >>>> the casting is done to a non-dereferenced variable. It is utilized
> > > >>>> as a pointer bias later in that function. Shall we just ignore the
> > > >>>> warning or still fix it somehow?    
> > > >>>     
> > > >>  
> > > >>> Do you think the cast to a !__iomem value is mandatory here?    
> > > >>
> > > >> It's not mandatory to have the casting with no __iomem, but wouldn't
> > > >> doing like this:
> > > >> + 	shift = (ssize_t __iomem)src & 0x3;
> > > >> be looking weird? Really, is there a good way to somehow extract the first
> > > >> two bits of a __iomem pointer without getting the sparse warning?  
> > > > 
> > > > I asked around me, what about trying uintptr_t?
> > > >   
> > >   
> > 
> > > One more way is to use __force to tell sparse that this casting is
> > > intentional:
> > > 
> > >        shift = (__force ssize_t)src & 0x3;  
> > 
> > Oh, great! That solution is actually much better than using some
> > currently unexplained sparse peculiarity! I was thinking about applying
> > some other attribute, but __force just didn't come to my mind. Thank
> > you very much for the suggestion. I'll post the fix with the solution
> > suggested by you.
> 

> Is the ssize_t cast the right one btw? I would definitely prefer an
> unsigned type here.

The reason of me deciding to use the ssize_t type here was to prevent
the types casting across the "shift", "chunk" and "len" variables
within this method. It seemed a bit better than having a standard type
like "unsigned int" here seeing the ssize_t type width won't exceed
the long type size anyway. Moreover since the "len" variable has got
the ssize_t type and I couldn't change it (the method is the map_info
callback), I've decided to stick with what is available and defined
"shift" and "chunk" as ssize_t-es. Another callback method
bt1_rom_map_read() in his module has been designed in the same way.

Do you think it's better to change it in favor of using a different
type like "unsigned int" here anyway? If so for unification I'd need to
change bt1_rom_map_read() (though the "shift" variable has been
defined as "unsigned long" there in the first place because the offs
argument has got that type).

What to do with the __force attribute here? It does seem appropriate
even if for some mystical reasons we haven't got the sparse warning
for the unsigned types.

-Sergey

> 
> Thanks,
> Miquèl