Re: [PATCH v7 2/9] dax: fix conversion of holes to PMDs

[Ross Zwisler]

On Thu, Jan 07, 2016 at 03:34:55PM -0700, Ross Zwisler wrote:
> On Wed, Jan 06, 2016 at 11:04:35AM -0800, Dan Williams wrote:
> > On Wed, Jan 6, 2016 at 10:00 AM, Ross Zwisler
> > <ross.zwisler@xxxxxxxxxxxxxxx> wrote:
> > > When we get a DAX PMD fault for a write it is possible that there could be
> > > some number of 4k zero pages already present for the same range that were
> > > inserted to service reads from a hole.  These 4k zero pages need to be
> > > unmapped from the VMAs and removed from the struct address_space radix tree
> > > before the real DAX PMD entry can be inserted.
> > >
> > > For PTE faults this same use case also exists and is handled by a
> > > combination of unmap_mapping_range() to unmap the VMAs and
> > > delete_from_page_cache() to remove the page from the address_space radix
> > > tree.
> > >
> > > For PMD faults we do have a call to unmap_mapping_range() (protected by a
> > > buffer_new() check), but nothing clears out the radix tree entry.  The
> > > buffer_new() check is also incorrect as the current ext4 and XFS filesystem
> > > code will never return a buffer_head with BH_New set, even when allocating
> > > new blocks over a hole.  Instead the filesystem will zero the blocks
> > > manually and return a buffer_head with only BH_Mapped set.
> > >
> > > Fix this situation by removing the buffer_new() check and adding a call to
> > > truncate_inode_pages_range() to clear out the radix tree entries before we
> > > insert the DAX PMD.
> > >
> > > Signed-off-by: Ross Zwisler <ross.zwisler@xxxxxxxxxxxxxxx>
> > 
> > Replaced the current contents of v6 in -mm from next-20160106 with
> > this v7 set and it looks good.
> > 
> > Reported-by: Dan Williams <dan.j.williams@xxxxxxxxx>
> > Tested-by: Dan Williams <dan.j.williams@xxxxxxxxx>
> > 
> > One question below...
> > 
> > > ---
> > >  fs/dax.c | 20 ++++++++++----------
> > >  1 file changed, 10 insertions(+), 10 deletions(-)
> > >
> > > diff --git a/fs/dax.c b/fs/dax.c
> > > index 03cc4a3..9dc0c97 100644
> > > --- a/fs/dax.c
> > > +++ b/fs/dax.c
> > > @@ -594,6 +594,7 @@ int __dax_pmd_fault(struct vm_area_struct *vma, unsigned long address,
> > >         bool write = flags & FAULT_FLAG_WRITE;
> > >         struct block_device *bdev;
> > >         pgoff_t size, pgoff;
> > > +       loff_t lstart, lend;
> > >         sector_t block;
> > >         int result = 0;
> > >
> > > @@ -647,15 +648,13 @@ int __dax_pmd_fault(struct vm_area_struct *vma, unsigned long address,
> > >                 goto fallback;
> > >         }
> > >
> > > -       /*
> > > -        * If we allocated new storage, make sure no process has any
> > > -        * zero pages covering this hole
> > > -        */
> > > -       if (buffer_new(&bh)) {
> > > -               i_mmap_unlock_read(mapping);
> > > -               unmap_mapping_range(mapping, pgoff << PAGE_SHIFT, PMD_SIZE, 0);
> > > -               i_mmap_lock_read(mapping);
> > > -       }
> > > +       /* make sure no process has any zero pages covering this hole */
> > > +       lstart = pgoff << PAGE_SHIFT;
> > > +       lend = lstart + PMD_SIZE - 1; /* inclusive */
> > > +       i_mmap_unlock_read(mapping);
> > > +       unmap_mapping_range(mapping, lstart, PMD_SIZE, 0);
> > > +       truncate_inode_pages_range(mapping, lstart, lend);
> > 
> > Do we need to do both unmap and truncate given that
> > truncate_inode_page() optionally does an unmap_mapping_range()
> > internally?
> 
> Ah, indeed it does.  Sure, having just the call to truncate_inode_page() seems
> cleaner.  I'll re-test and send this out in v8.

Actually, in testing it doesn't look like unmap_mapping_range() in
truncate_inode_page() gets called.  We fail the page_mapped(page) check for
our read-only zero pages.  I think we need to keep the unmap_mapping_range()
call in __dax_pmd_fault().