On 11/15/17 11:44 +0100, David Hildenbrand wrote:
> On 08.11.2017 08:56, Haozhong Zhang wrote:
> > It will be used by KVM to check whether a pfn should be
> > mapped to guest as UC.
> >
> > Signed-off-by: Haozhong Zhang <haozhong.zh...@intel.com>
> > ---
> > arch/x86/include/asm/pat.h | 2 ++
> > arch/x86/mm/pat.c | 16 ++++++++++++++++
> > 2 files changed, 18 insertions(+)
> >
> > diff --git a/arch/x86/include/asm/pat.h b/arch/x86/include/asm/pat.h
> > index fffb2794dd89..fabb0cf00e77 100644
> > --- a/arch/x86/include/asm/pat.h
> > +++ b/arch/x86/include/asm/pat.h
> > @@ -21,4 +21,6 @@ int io_reserve_memtype(resource_size_t start,
> > resource_size_t end,
> >
> > void io_free_memtype(resource_size_t start, resource_size_t end);
> >
> > +bool pat_pfn_is_uc_or_uc_minus(unsigned long pfn);
> > +
> > #endif /* _ASM_X86_PAT_H */
> > diff --git a/arch/x86/mm/pat.c b/arch/x86/mm/pat.c
> > index fe7d57a8fb60..e1282dd4eeb8 100644
> > --- a/arch/x86/mm/pat.c
> > +++ b/arch/x86/mm/pat.c
> > @@ -677,6 +677,22 @@ static enum page_cache_mode lookup_memtype(u64 paddr)
> > return rettype;
> > }
> >
> > +/**
> > + * Check with PAT whether the memory type of a pfn is UC or UC-.
> > + *
> > + * Only to be called when PAT is enabled.
> > + *
> > + * Returns true, if the memory type of @pfn is UC or UC-.
> > + * Otherwise, returns false.
> > + */
> > +bool pat_pfn_is_uc_or_uc_minus(unsigned long pfn)
> > +{
> > + enum page_cache_mode cm = lookup_memtype(PFN_PHYS(pfn));
> > +
> > + return cm == _PAGE_CACHE_MODE_UC || cm == _PAGE_CACHE_MODE_UC_MINUS;
> > +}
> > +EXPORT_SYMBOL_GPL(pat_pfn_is_uc_or_uc_minus);
> > +
> > /**
> > * io_reserve_memtype - Request a memory type mapping for a region of
> > memory
> > * @start: start (physical address) of the region
> >
>
> Wonder if we should check for pat internally. And if we should simply
> return the memtype via lookup_memtype() instead of creating such a
> strange named function (by providing e.g. a lookup_memtype() variant
> that can be called with !pat_enabled()).
>
> The caller can easily check against _PAGE_CACHE_MODE_UC ...
>
Yes, the better solution should work for both PAT enabled and disabled
cases, like what __vm_insert_mixed() does: use vma->vm_page_prot if
PAT is disabled, and refer to track_pfn_insert() in addition if PAT is
enabled.
The early RFC patch [1] got the cache mode in a similar way via a new
function kvm_vcpu_gfn_to_pgprot(). However, as explained in RFC, it
does not work, because the existing MMIO check (where kvm_vcpu_gfn_to_pgprot()
is called) in KVM is performed with a spinlock (vcpu->kvm->mmu_lock)
being taken, but kvm_vcpu_gfn_to_pgprot() has to touch a semaphore
(vcpu->kvm->mm->mmap_sem). Besides, KVM may prefetch and check MMIO of
other pfns within vcpu->kvm->mmu_lock, and the prefectched pfns cannot
be predicted in advance, which means we have to keep the MMIO check
within vcpu->kvm->mmu_lock.
Therefore, I only make a suboptimal fix in this patchset that only
fixes PAT enabled cases, which I suppose is the usual usage scenario
of NVDIMM.
[1] https://patchwork.kernel.org/patch/10016261/
Haozhong
