On Mon, 2015-11-23 at 12:53 -0800, Dan Williams wrote:
> On Mon, Nov 23, 2015 at 12:04 PM, Toshi Kani <toshi.kani@hpe.com> wrote:
> > The following oops was observed when mmap() with MAP_POPULATE
> > pre-faulted pmd mappings of a DAX file.  follow_trans_huge_pmd()
> > expects that a target address has a struct page.
> > 
> >   BUG: unable to handle kernel paging request at ffffea0012220000
> >   follow_trans_huge_pmd+0xba/0x390
> >   follow_page_mask+0x33d/0x420
> >   __get_user_pages+0xdc/0x800
> >   populate_vma_page_range+0xb5/0xe0
> >   __mm_populate+0xc5/0x150
> >   vm_mmap_pgoff+0xd5/0xe0
> >   SyS_mmap_pgoff+0x1c1/0x290
> >   SyS_mmap+0x1b/0x30
> > 
> > Fix it by making the PMD pre-fault handling consistent with PTE.
> > After pre-faulted in faultin_page(), follow_page_mask() calls
> > follow_trans_huge_pmd(), which is changed to call follow_pfn_pmd()
> > for VM_PFNMAP or VM_MIXEDMAP.  follow_pfn_pmd() handles FOLL_TOUCH
> > and returns with -EEXIST.
> 
> As of 4.4.-rc2 DAX pmd mappings are disabled.  So we have time to do
> something more comprehensive in 4.5.

Yes, I noticed during my testing that I could not use pmd...

> > Reported-by: Mauricio Porto <mauricio.porto@hpe.com>
> > Signed-off-by: Toshi Kani <toshi.kani@hpe.com>
> > Cc: Andrew Morton <akpm@linux-foundation.org>
> > Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
> > Cc: Matthew Wilcox <willy@linux.intel.com>
> > Cc: Dan Williams <dan.j.williams@intel.com>
> > Cc: Ross Zwisler <ross.zwisler@linux.intel.com>
> > ---
> >  mm/huge_memory.c |   34 ++++++++++++++++++++++++++++++++++
> >  1 file changed, 34 insertions(+)
> > 
> > diff --git a/mm/huge_memory.c b/mm/huge_memory.c
> > index d5b8920..f56e034 100644
> > --- a/mm/huge_memory.c
> > +++ b/mm/huge_memory.c
> [..]
> > @@ -1288,6 +1315,13 @@ struct page *follow_trans_huge_pmd(struct
> > vm_area_struct *vma,
> >         if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
> >                 goto out;
> > 
> > +       /* pfn map does not have a struct page */
> > +       if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP)) {
> > +               ret = follow_pfn_pmd(vma, addr, pmd, flags);
> > +               page = ERR_PTR(ret);
> > +               goto out;
> > +       }
> > +
> >         page = pmd_page(*pmd);
> >         VM_BUG_ON_PAGE(!PageHead(page), page);
> >         if (flags & FOLL_TOUCH) {
> 
> I think it is already problematic that dax pmd mappings are getting
> confused with transparent huge pages.  

We had the same issue with dax pte mapping [1], and this change extends the pfn
map handling to pmd.  So, this problem is not specific to pmd.

[1] https://lkml.org/lkml/2015/6/23/181

> They're more closely related to
> a hugetlbfs pmd mappings in that they are mapping an explicit
> allocation.  I have some pending patches to address this dax-pmd vs
> hugetlb-pmd vs thp-pmd classification that I will post shortly.

Not sure which way is better, but I am certainly interested in your changes.

> By the way, I'm collecting DAX pmd regression tests [1], is this just
> a simple crash upon using MAP_POPULATE?
> 
> [1]: https://github.com/pmem/ndctl/blob/master/lib/test-dax-pmd.c

Yes, this issue is easy to reproduce with MAP_POPULATE.  In case it helps,
attached are the test I used for testing the patches.  Sorry, the code is messy
since it was only intended for my internal use...

 - The test was originally written for the pte change [1] and comments in
test.sh (ex. mlock fail, ok) reflect the results without the pte change.
 - For the pmd test, I modified test-mmap.c to call posix_memalign() before
mmap().  By calling free(), the 2MB-aligned address from posix_memalign() can be
used for mmap().  This keeps the mmap'd address aligned on 2MB.
 - I created test file(s) with dd (i.e. all blocks written) in my test.
 - The other infinite loop issue (fixed by my other patch) was found by the test
case with option "-LMSr".

Thanks,
-Toshi