Re: ioctl FIBMAP for dax gone in v4.17-rc1

From: Dave Chinner <david@fromorbit.com>
To: "Darrick J. Wong" <darrick.wong@oracle.com>
Cc: Andreas Dilger <adilger@dilger.ca>,
	Dan Williams <dan.j.williams@intel.com>,
	Christoph Hellwig <hch@infradead.org>,
	Xiong Zhou <xzhou@redhat.com>,
	linux-fsdevel <linux-fsdevel@vger.kernel.org>,
	linux-xfs <linux-xfs@vger.kernel.org>,
	Eric Sandeen <sandeen@sandeen.net>,
	fstests <fstests@vger.kernel.org>,
	linux-nvdimm <linux-nvdimm@lists.01.org>
Subject: Re: ioctl FIBMAP for dax gone in v4.17-rc1
Date: Wed, 18 Apr 2018 09:36:18 +1000	[thread overview]
Message-ID: <20180417233618.GB27893@dastard> (raw)
In-Reply-To: <20180417175007.GB5210@magnolia>

On Tue, Apr 17, 2018 at 10:50:07AM -0700, Darrick J. Wong wrote:
> On Tue, Apr 17, 2018 at 11:40:00AM -0600, Andreas Dilger wrote:
> > On Apr 17, 2018, at 10:57 AM, Darrick J. Wong <darrick.wong@oracle.com> wrote:
> > > 
> > > On Tue, Apr 17, 2018 at 09:53:47AM -0700, Dan Williams wrote:
> > >> On Tue, Apr 17, 2018 at 9:10 AM, Christoph Hellwig <hch@infradead.org> wrote:
> > >>> On Tue, Apr 17, 2018 at 10:40:59PM +0800, Xiong Zhou wrote:
> > >>>> We got these in v4.17-rc1:
> > >>>> 6e2608d xfs, dax: introduce xfs_dax_aops
> > >>>> fb094c9 ext2, dax: introduce ext2_dax_aops
> > >>>> 5f0663b ext4, dax: introduce ext4_dax_aops
> > >>>> 
> > >>>> And we don't have ->bmap call in these aops, which may lead
> > >>>> to the ioctl call failure.
> > >>>> 
> > >>>> Do we have any plan of adding/supporting it ?
> > >>>> 
> > >>>> xfstests generic/223 covers this issue. If we are not going
> > >>>> to support this call for dax, we need to fix the testcase.
> > >>> 
> > >>> Not supporting ->bmap is a good thing as it is hightly dangerous.
> > >> 
> > >> I take this to mean "don't fix, it is another casualty of dax being
> > >> experimental and it won't be coming back". I can get on board with
> > >> that.
> > >> 
> > >> Otherwise, I was about to send a series adding bmap to {xfs,ext2,ext4}_dax_ops.
> > > 
> > > Frankly I'd rather see the swapfile code learn how to iomap and then we
> > > can get rid of bmap in xfs entirely.
> > 
> > Is anyone still using LILO to boot?

Yup. I do.

[ because grub has a history of corrupting filesystems on my
machines with it's obnoxious "probe for bootable OS images on update
by attempting to mount every block device in the system" behaviour. ]

> > It needed FIBMAP support to map the
> > kernel image for booting.

There's a difference between FIBMAP and ->bmap. One's a syscall
parameter under ABI constraints and the other is an internal
implementation.

We have to keep supporting FIBMAP forever, but we can re-implement
it internally using ->fiemap pretty damn easily - just replace the
bmap() implementation with a ->fiemap call and replace all the
direct ->bmap() calls with bmap(). Hence FIBMAP will continue to
work on all filesystems (including DAX!) and we get the ->bmap
crap out of the kernel.

Cheers,

Dave.
-- 
Dave Chinner
david@fromorbit.com