Re: [PATCH, RFC 2/2] Implement sharing/unsharing of PMDs for FS/DAX

[Dan Williams <dan.j.williams@intel.com>]

On Fri, May 24, 2019 at 9:07 AM Larry Bassel <larry.bassel@oracle.com> wrote:
> On 14 May 19 16:01, Kirill A. Shutemov wrote:
> > On Thu, May 09, 2019 at 09:05:33AM -0700, Larry Bassel wrote:
[..]
> > > diff --git a/mm/memory.c b/mm/memory.c
> > > index f7d962d..4c1814c 100644
> > > --- a/mm/memory.c
> > > +++ b/mm/memory.c
> > > @@ -3845,6 +3845,109 @@ static vm_fault_t handle_pte_fault(struct vm_fault *vmf)
> > >     return 0;
> > >  }
> > >
> > > +#ifdef CONFIG_MAY_SHARE_FSDAX_PMD
> > > +static pmd_t *huge_pmd_offset(struct mm_struct *mm,
> > > +                         unsigned long addr, unsigned long sz)
> >
> > Could you explain what this function suppose to do?
> >
> > As far as I can see vma_mmu_pagesize() is always PAGE_SIZE of DAX
> > filesystem. So we have 'sz' == PAGE_SIZE here.
>
> I thought so too, but in my testing I found that vma_mmu_pagesize() returns
> 4KiB, which differs from the DAX filesystem's 2MiB pagesize.

A given filesystem-dax vma is allowed to support both 4K and 2M
mappings, so the vma_mmu_pagesize() is not granular enough to describe
the capabilities of a filesystem-dax vma. In the device-dax case,
where there are mapping guarantees, the implementation does arrange
for vma_mmu_pagesize() to reflect the right page size.