Re: Overlapping ioremap() calls, set_memory_*() semantics

[Maciej W. Rozycki]

On Wed, 9 Mar 2016, Ingo Molnar wrote:

> > > So to go back to the original suggestion from Luis, I've quoted it, but
> > > with a s/overlapping/aliased substitution:
> > > 
> > > > I had suggested long ago then that one possible resolution was for us
> > > > to add an API that *enables* aliased ioremap() calls, and only use it
> > > > on select locations in the kernel. This means we only have to convert a
> > > > few users to that call to white list such semantics, and by default
> > > > we'd disable aliased calls. To kick things off -- is this strategy
> > > > agreeable for all other architectures?
> > > 
> > > I'd say that since the overwhelming majority of ioremap() calls are not
> > > aliased, ever, thus making it 'harder' to accidentally alias is probably
> > > a good idea.
> > 
> > Did you mean 'aliased' or 'aliased with different cache attribute'?  The former 
> > check might be too strict.
> 
> I'd say even 'same attribute' aliasing is probably relatively rare.

 Please note that aliased cached mappings (any kinds of, not necessarily 
from `ioremap') cause a lot of headache (read: handling trouble) with 
architectures such as MIPS which support virtually indexed caches which 
suffer from cache aliasing.  There is a risk of data corruption if the 
same physical memory address space location is accessed through different 
virtual mappings as not all hardware catches duplicate cache entries 
created in such a case.

 We handle it in software for user mappings (although I keep having a 
feeling something always keeps escaping, due to the vast diversity of 
cache configurations possible), however I don't think we do for `ioremap', 
so disallowing aliased `ioremap' mappings by default sounds like a good 
idea to me.

 FWIW,

  Maciej