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

[Toshi Kani]

On Sat, 2016-03-05 at 12:40 +0100, Ingo Molnar wrote:
> * Toshi Kani <toshi.kani@xxxxxxx> wrote:
> 
> > > So I'd say that since ioremap() in itself is fragile enough, we
> > > should work towards eliminating overlapping ranges.
> > > 
> > > The thing is, the whole vmap_area logic is based around non-
> > > overlapping ranges, sorted into the vmap_area_root rbtree.
> > > 
> > > Just check the logic in mm/vmalloc.c::alloc_vmap_area(): it's based
> > > on finding holes in the kernel-virtual allocations. 'Overlapping
> > > ranges' is very much not part of that logic, at least to my
> > > understanding.
> > > 
> > > How are overlapping ioremap()s even possible with that logic? The
> > > allocator searches for holes, not allowing for overlaps. What am I
> > > missing?
> > > 
> > > Could you outline a specific case where it's done intentionally - and
> > > the purpose behind that intention?
> > 
> > The term "overlapping" is a bit misleading. [...]
> 
> A bit? It was totally misleading ...
> 
> You meant virtual aliases for the same physical address, and those of
> course are allowed, as long the cache attributes are compatible, that is
> what the whole memtype infrastructure is about, as you yourself note:
> 
> > [...] ÂThis is "alias" mapping -- a physical address range is mapped to
> > multiple virtual address ranges. ÂThere is no overlapping in VMA.
> > 
> > Such alias mappings are used by multiple modules. ÂFor instance, a PMEM
> > range is mapped to the kernel and user spaces. Â/dev/mem is another
> > example that creates a user space mapping to a physical address where
> > other mappings may already exist.
> > 
> > Hence, alias mapping itself is a supported use-case. ÂHowever, alias
> > mapping with different cache types is not as it causes undefined
> > behavior. ÂTherefore, PAT module protects from this case by tracking
> > cache types used for mapping physical ranges. ÂWhen a different cache
> > type is requested, is_new_memtype_allowed() checks if the request needs
> > to be failed or can be changed to the existing type.
> 
> So where is the problem? The memtype implementation and hence most
> ioremap() users are supposed to be safe. set_memory_*() APIs are supposed
> to be safe as well, as they too go via the memtype API.

Let me try to summarize...

The original issue Luis brought up was that drivers written to work with
MTRR may create a single ioremap range covering multiple cache attributes
since MTRR can overwrite cache attribute of a certain range. ÂConverting
such drivers with PAT-based ioremap interfaces, i.e. ioremap_wc() and
ioremap_nocache(), requires a separate ioremap map for each cache
attribute, which can be challenging as it may result in overlapping ioremap
ranges (in his term) with different cache attributes.

So, Luis asked about 'sematics of overlapping ioremap()' calls. ÂHence, I
responded that aliasing mapping itself is supported, but alias with
different cache attribute is not. ÂWe have checks in place to detect such
condition. ÂOverlapping ioremap calls with a different cache attribute
either fails or gets redirected to the existing cache attribute on x86.

> > I agree that the current implementation is fragile, and some interfaces
> > skip such check at all, ex.Âvm_insert_pfn().
> 
> Most of those are really just low level interfaces forl cases that skip the memtype infrastructure.

Yes, and I'm just stating the fact that some pfn map use-cases, such as
mmap, are not tracked in memtype. ÂFor example, drm_gem_mmap()
->Âdrm_gem_mmap_obj() sets its VMA as WC attribute. Âi915_gem_fault() then
creates a WC map withÂvm_insert_pfn() at fault.

Thanks,
-Toshi