[dm-devel] [PATCH v5 2/8] drivers/pmem: Allow pmem_clear_poison() to accept arbitrary offset and len

[Vivek Goyal]

On Thu, Feb 27, 2020 at 02:11:43PM +1100, Dave Chinner wrote:
> On Wed, Feb 26, 2020 at 11:57:56AM -0500, Vivek Goyal wrote:
> > On Tue, Feb 25, 2020 at 02:49:30PM -0800, Dan Williams wrote:
> > [..]
> > > > > I'm ok with replacing blkdev_issue_zeroout() with a dax operation
> > > > > callback that deals with page aligned entries. That change at least
> > > > > makes the error boundary symmetric across copy_from_iter() and the
> > > > > zeroing path.
> > > >
> > > > IIUC, you are suggesting that modify dax_zero_page_range() to take page
> > > > aligned start and size and call this interface from
> > > > __dax_zero_page_range() and get rid of blkdev_issue_zeroout() in that
> > > > path?
> > > >
> > > > Something like.
> > > >
> > > > __dax_zero_page_range() {
> > > >   if(page_aligned_io)
> > > >         call_dax_page_zero_range()
> > > >   else
> > > >         use_direct_access_and_memcpy;
> > > > }
> > > >
> > > > And other callers of blkdev_issue_zeroout() in filesystems can migrate
> > > > to calling dax_zero_page_range() instead.
> > > >
> > > > If yes, I am not seeing what advantage do we get by this change.
> > > >
> > > > - __dax_zero_page_range() seems to be called by only partial block
> > > >   zeroing code. So dax_zero_page_range() call will remain unused.
> > > >
> > > >
> > > > - dax_zero_page_range() will be exact replacement of
> > > >   blkdev_issue_zeroout() so filesystems will not gain anything. Just that
> > > >   it will create a dax specific hook.
> > > >
> > > > In that case it might be simpler to just get rid of blkdev_issue_zeroout()
> > > > call from __dax_zero_page_range() and make sure there are no callers of
> > > > full block zeroing from this path.
> > > 
> > > I think you're right. The path I'm concerned about not regressing is
> > > the error clearing on new block allocation and we get that already via
> > > xfs_zero_extent() and sb_issue_zeroout().
> > 
> > Well I was wrong. I found atleast one user which uses __dax_zero_page_range()
> > to zero full PAGE_SIZE blocks.
> > 
> > xfs_io -c "allocsp 32K 0" foo.txt
> 
> That ioctl interface is deprecated and likely unused by any new
> application written since 1999. It predates unwritten extents (1998)
> and I don't think any native linux applications have ever used it. A
> newly written DAX aware application would almost certainly not use
> this interface.
> 
> IOWs, I wouldn't use it as justification for some special case
> behaviour; I'm more likely to say "rip that ancient ioctl out" than
> to jump through hoops because it's implementation behaviour.

Hi Dave,

Do you see any other path in xfs using iomap_zero_range() and zeroing
full block. iomap_zero_range() already skips IOMAP_HOLE and
IOMAP_UNWRITTEN. So it has to be a full block zeroing which is of not type
IOMAP_HOLE and IOMAP_UNWRITTEN.

My understanding is that ext4 and xfs both are initializing full blocks
using blkdev_issue_zeroout(). Only partial blocks are being zeroed using
this dax zeroing path.

If there are no callers of full block zeroing through
__dax_zero_page_range(), then I can simply get rid of
blkdev_issue_zerout() call from __dax_zero_page_range().

Thanks
Vivek