Hi,

On Wednesday 06 Jan 2021 at 16:13:53 (+0000), Al Viro wrote:
> On Wed, Jan 06, 2021 at 03:52:14PM +0000, Ionela Voinescu wrote:
> > > > > > vim +367 arch/arm64/kernel/topology.c
> > > > > > 
> > > > > >    362	
> > > > > >    363	int cpc_read_ffh(int cpu, struct cpc_reg *reg, u64 *val)
> > > > > >    364	{
> > > > > >    365		int ret = -EOPNOTSUPP;
> > > > > >    366	
> > > > > >  > 367		switch ((u64)reg->address) {
> > > > > 
> > > > > That's not a dereference but I guess sparse complains of dropping the
> > > > > __iomem. We could change the cast to (__force u64) to silence sparse.
> > > > > 
> > > > > Thanks for the report.
> > > > > 
> > > > 
> > > > Nothing I've tried seemed to silence sparse here, including casting to
> > > > (__force u64).
> > > 
> > > Would it work if we changed the case lines to (u64 __iomem)0x0?
> > > 
> > 
> > No, it does not. We still get the same warning on the switch line even
> > if there is no cast. Same if we directly check for:
> > 
> > if (reg->address == (u64 __iomem)0x0)
> 
> Folks, could you stop with the voodoo?  This u64 __iomem address thing is completely
> wrong.  What it says is "address of that field shall be an iomem pointer",
> which makes no sense whatsoever.
> 
> Just what had been intended?  __iomem is a qualifier of the same sort
> as const or volatile - this mess makes as much sense as
> struct cpc_reg {
>         u8 descriptor;
>         u16 length;
>         u8 space_id;   
>         u8 bit_width;   
>         u8 bit_offset;
>         u8 access_width;
>         u64 const address;
> } __packed;
> 
> Which would *NOT* be read as "reg->address is a numeric representation of
> address of something unmodifiable" - it would be "the value stored in
> reg->address can not be modified".
> 
> This annotation says "reg->address (somehow) lives in iomem", resulting in
> "so why the hell are you trying to read it by plain dereferencing of
> reg + field offset?" from sparse.
> 
> Get rid of this misannotation and don't breed force-cast to confuse
> everything hard enough to STFU.

Thanks, it makes sense, and removing the attribute solves the other
similar warnings in cppc_acpi. I'll double check and submit a patch for
that.

Thanks,
Ionela.