Re: [PATCH v2 2/2] ext4: handle layout changes to pinned DAX mappings

[Ross Zwisler <ross.zwisler-VuQAYsv1563Yd54FQh9/CA@public.gmane.org>]

On Mon, Jul 02, 2018 at 09:59:48AM +0200, Lukas Czerner wrote:
> On Fri, Jun 29, 2018 at 09:13:00AM -0600, Ross Zwisler wrote:
> > On Fri, Jun 29, 2018 at 02:02:23PM +0200, Lukas Czerner wrote:
> > > On Wed, Jun 27, 2018 at 03:22:52PM -0600, Ross Zwisler wrote:
> > > > Follow the lead of xfs_break_dax_layouts() and add synchronization between
> > > > operations in ext4 which remove blocks from an inode (hole punch, truncate
> > > > down, etc.) and pages which are pinned due to DAX DMA operations.
> > > > 
> > > > Signed-off-by: Ross Zwisler <ross.zwisler-VuQAYsv1563Yd54FQh9/CA@public.gmane.org>
> > > > Reviewed-by: Jan Kara <jack-AlSwsSmVLrQ@public.gmane.org>
> > > > ---
> > > >  fs/ext4/ext4.h     |  1 +
> > > >  fs/ext4/extents.c  | 12 ++++++++++++
> > > >  fs/ext4/inode.c    | 46 ++++++++++++++++++++++++++++++++++++++++++++++
> > > >  fs/ext4/truncate.h |  4 ++++
> > > >  4 files changed, 63 insertions(+)
> > > > 
> > > > diff --git a/fs/ext4/ext4.h b/fs/ext4/ext4.h
> > > > index 0b127853c584..34bccd64d83d 100644
> > > > --- a/fs/ext4/ext4.h
> > > > +++ b/fs/ext4/ext4.h
> > > > @@ -2460,6 +2460,7 @@ extern int ext4_get_inode_loc(struct inode *, struct ext4_iloc *);
> > > >  extern int ext4_inode_attach_jinode(struct inode *inode);
> > > >  extern int ext4_can_truncate(struct inode *inode);
> > > >  extern int ext4_truncate(struct inode *);
> > > > +extern int ext4_break_layouts(struct inode *);
> > > >  extern int ext4_punch_hole(struct inode *inode, loff_t offset, loff_t length);
> > > >  extern int ext4_truncate_restart_trans(handle_t *, struct inode *, int nblocks);
> > > >  extern void ext4_set_inode_flags(struct inode *);
> > > > diff --git a/fs/ext4/extents.c b/fs/ext4/extents.c
> > > > index 0057fe3f248d..a6aef06f455b 100644
> > > > --- a/fs/ext4/extents.c
> > > > +++ b/fs/ext4/extents.c
> > > > @@ -4820,6 +4820,13 @@ static long ext4_zero_range(struct file *file, loff_t offset,
> > > >  		 * released from page cache.
> > > >  		 */
> > > >  		down_write(&EXT4_I(inode)->i_mmap_sem);
> > > > +
> > > > +		ret = ext4_break_layouts(inode);
> > > > +		if (ret) {
> > > > +			up_write(&EXT4_I(inode)->i_mmap_sem);
> > > > +			goto out_mutex;
> > > > +		}
> > > > +
> > > >  		ret = ext4_update_disksize_before_punch(inode, offset, len);
> > > >  		if (ret) {
> > > >  			up_write(&EXT4_I(inode)->i_mmap_sem);
> > > > @@ -5493,6 +5500,11 @@ int ext4_collapse_range(struct inode *inode, loff_t offset, loff_t len)
> > > >  	 * page cache.
> > > >  	 */
> > > >  	down_write(&EXT4_I(inode)->i_mmap_sem);
> > > > +
> > > > +	ret = ext4_break_layouts(inode);
> > > > +	if (ret)
> > > > +		goto out_mmap;
> > > 
> > > Hi,
> > > 
> > > don't we need to do the same for ext4_insert_range() since we're about
> > > to truncate_pagecache() as well ?
> > > 
> > > /thinking out loud/
> > > Xfs seems to do this before every fallocate operation, but in ext4
> > > it does not seem to be needed at least for simply allocating falocate...
> > 
> > I saw the case in ext4_insert_range(), and decided that we didn't need to
> > worry about synchronizing with DAX because no blocks were being removed from
> > the inode's extent map.  IIUC the truncate_pagecache() call is needed because
> > we are unmapping and removing any page cache mappings for the part of the file
> > after the insert because those blocks are now at a different offset in the
> > inode.  Because at the end of the operation we haven't removed any DAX pages
> > from the inode, we have nothing that we need to synchronize.
> > 
> > Hmm, unless this is a failure case we care about fixing?
> >  1) schedule I/O via O_DIRECT to page X
> >  2) fallocate(FALLOC_FL_INSERT_RANGE) to block < X, shifting X to a larger
> >     offset
> >  3) O_DIRECT I/O from 1) completes, but ends up writing into the *new* block
> >     that resides at X - the I/O from 1) completes
> > 
> > In this case the user is running I/O and issuing the fallocate at the same
> > time, and the sequencing could have worked out that #1 and #2 were reversed,
> > giving you the same behavior.  IMO this seems fine and that we shouldn't have
> > the DAX synchronization call in ext4_insert_range(), but I'm happy to add it
> > if I'm wrong.
> 
> Hi,
> 
> I think you're right, this case might mot matter much. I am just worried
> about unforeseen consequences of changing the layout with dax pages
> mapped. I guess we can also add this later fi we discover anything.
> 
> You can add
> 
> Reviewed-by: Lukas Czerner <lczerner-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
> 
> Thanks!
> -Lukas

Thank you for the review.  I'll add a comment to help explain my reasoning, as
Jan suggested.