Re: [PATCH 0/9 v3] ext4: Punch hole and DAX fixes

[Jan Kara <jack@suse.cz>]

On Tue 10-11-15 11:00:27, Dave Chinner wrote:
> On Mon, Nov 09, 2015 at 05:51:49PM +0100, Jan Kara wrote:
> > On Mon 09-11-15 17:22:56, Jan Kara wrote:
> > >   Hi,
> > > 
> > > On Fri 06-11-15 17:57:04, Boylston, Brian wrote:
> > > > I've written a test tool (included below) that exercises page faults on
> > > > hole-y portions of an mmapped file.  The file is created, sized using
> > > > various methods, mmapped, and then two threads race to write a marker to
> > > > different offsets within each mapped page.  Once the threads have
> > > > finished marking each page, the pages are checked for the presence of
> > > > the markers.
> > > > 
> > > > With vanilla 4.2 and 4.3 kernels, this test easily exposes corruption on
> > > > pmem-backed, DAX-mounted xfs and ext4 file systems.
> > > > 
> > > > With 4.3 and this ext4 patch set, the data corruption is still seen:
> > > > 
> > > > $ ./holetest -f /pmem1/brian/holetest 1000
> > > > holetest r207
> > > 
> > > Thanks for the test. I'll try to reproduce locally and have a look why
> > > my block zeroing patch didn't work as expected.
> > 
> > Ah, OK, I see what's going on. So ext4 with my patches still returns
> > buffer_new buffer even though it is zeroed out and thus generic DAX code
> > still tries to zero out the buffer again which indeed causes the corrution
> > (will test everything tomorrow with that code disabled).  Now I have
> > decided that block mapping function should return buffer_new buffer even
> > though it is zeroed out because e.g. if block zeroing was used for page
> > cache writes, we'd still need code in fs/buffer.c to do proper zeroing of
> > parts of the block that are not written. And that happens based on
> > buffer_new flag.
> 
> XFS special cases this for DAX in __xfs_get_blocks():
> 
>         if (IS_DAX(inode) && create) {
>                 ASSERT(!ISUNWRITTEN(&imap));
>                 /* zeroing is not needed at a higher layer */
>                 new = 0;
>         }
> 
> And so will not set the buffer_new() fo rhte DAX case as we've
> already directly zeroed the region the DAX code s about to write
> into...

OK, for now I did something similar in the ext4 mapping function for DAX
faults.

> > The zeroing code in __dax_fault() needs to go away anyway so whether we
> > return buffer_new buffer is not really substantial but I'd like to get some
> > agreement and consistency among filesystems in with which flags zeroed
> > blocks are returned. Thoughts?
> 
> There is no consistency to begin with, especially w.r.t. unwritten
> extent behaviour as the upper layers don't all understand that
> buffer_unwritten is a valid flag for getblock to return. Hence we
> have hacks in XFS setting buffer_new() in strange cases to get the
> upper level code to zero stuff that really needs zeroing...

In ext4 we set buffer as new in two cases:

1) When it was freshly allocated (regardless whether into unwritten or
   normal extent).
2) When it was converted from unwritten to written state.

This seems to do the job...

								Honza
-- 
Jan Kara <jack@suse.com>
SUSE Labs, CR